diff --git a/.flake8 b/.flake8
new file mode 100644
index 000000000..cf276d0ba
--- /dev/null
+++ b/.flake8
@@ -0,0 +1,35 @@
+[flake8]
+show-source=true
+statistics=true
+max-line-length = 88
+per-file-ignores =
+    # line too long
+    icefall/diagnostics.py: E501,
+    egs/*/ASR/*/conformer.py: E501,
+    egs/*/ASR/pruned_transducer_stateless*/*.py: E501,
+    egs/*/ASR/*/optim.py: E501,
+    egs/*/ASR/*/scaling.py: E501,
+    egs/librispeech/ASR/lstm_transducer_stateless*/*.py: E501, E203
+    egs/librispeech/ASR/conv_emformer_transducer_stateless*/*.py: E501, E203
+    egs/librispeech/ASR/conformer_ctc*/*py: E501,
+    egs/librispeech/ASR/zipformer_mmi/*.py: E501, E203
+    egs/librispeech/ASR/zipformer/*.py: E501, E203
+    egs/librispeech/ASR/RESULTS.md: E999,
+    egs/ljspeech/TTS/vits/*.py: E501, E203
+    # invalid escape sequence (cause by tex formular), W605
+    icefall/utils.py: E501, W605
+
+exclude =
+  .git,
+  **/data/**,
+  icefall/shared/make_kn_lm.py,
+  icefall/__init__.py
+  icefall/ctc/__init__.py
+
+ignore =
+  # E203 white space before ":"
+  E203,
+  # W503 line break before binary operator
+  W503,
+  # E226 missing whitespace around arithmetic operator
+  E226,
diff --git a/.git-blame-ignore-revs b/.git-blame-ignore-revs
new file mode 100644
index 000000000..5d65b98e9
--- /dev/null
+++ b/.git-blame-ignore-revs
@@ -0,0 +1,3 @@
+# Migrate to 88 characters per line (see: https://github.com/lhotse-speech/lhotse/issues/890)
+107df3b115a58f1b68a6458c3f94a130004be34c
+d31db010371a4128856480382876acdc0d1739ed
diff --git a/.github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh b/.github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
new file mode 100755
index 000000000..0bec8c0c4
--- /dev/null
+++ b/.github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+
+# This script computes fbank features for the test-clean and test-other datasets.
+# The computed features are saved to ~/tmp/fbank-libri and are
+# cached for later runs
+
+set -e
+
+export PYTHONPATH=$PWD:$PYTHONPATH
+echo $PYTHONPATH
+
+mkdir ~/tmp/fbank-libri
+cd egs/librispeech/ASR
+mkdir -p data
+cd data
+[ ! -e fbank ] && ln -s ~/tmp/fbank-libri fbank
+cd ..
+./local/compute_fbank_librispeech.py --dataset 'test-clean test-other'
+ls -lh data/fbank/
diff --git a/.github/scripts/download-gigaspeech-dev-test-dataset.sh b/.github/scripts/download-gigaspeech-dev-test-dataset.sh
new file mode 100755
index 000000000..f3564efc7
--- /dev/null
+++ b/.github/scripts/download-gigaspeech-dev-test-dataset.sh
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+
+# This script downloads the pre-computed fbank features for
+# dev and test datasets of GigaSpeech.
+#
+# You will find directories `~/tmp/giga-dev-dataset-fbank` after running
+# this script.
+
+set -e
+
+mkdir -p ~/tmp
+cd ~/tmp
+
+git lfs install
+git clone https://huggingface.co/csukuangfj/giga-dev-dataset-fbank
+
+ls -lh giga-dev-dataset-fbank/data/fbank
diff --git a/.github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh b/.github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
new file mode 100755
index 000000000..11704526c
--- /dev/null
+++ b/.github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+
+# This script downloads the test-clean and test-other datasets
+# of LibriSpeech and unzip them to the folder ~/tmp/download,
+# which is cached by GitHub actions for later runs.
+#
+# You will find directories ~/tmp/download/LibriSpeech after running
+# this script.
+
+set -e
+
+mkdir ~/tmp/download
+cd egs/librispeech/ASR
+ln -s ~/tmp/download .
+cd download
+wget -q --no-check-certificate https://www.openslr.org/resources/12/test-clean.tar.gz
+tar xf test-clean.tar.gz
+rm test-clean.tar.gz
+
+wget -q --no-check-certificate https://www.openslr.org/resources/12/test-other.tar.gz
+tar xf test-other.tar.gz
+rm test-other.tar.gz
+pwd
+ls -lh
+ls -lh LibriSpeech
diff --git a/.github/scripts/install-kaldifeat.sh b/.github/scripts/install-kaldifeat.sh
new file mode 100755
index 000000000..de30f7dfe
--- /dev/null
+++ b/.github/scripts/install-kaldifeat.sh
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+
+# This script installs kaldifeat into the directory ~/tmp/kaldifeat
+# which is cached by GitHub actions for later runs.
+
+set -e
+
+mkdir -p ~/tmp
+cd ~/tmp
+git clone https://github.com/csukuangfj/kaldifeat
+cd kaldifeat
+mkdir build
+cd build
+cmake -DCMAKE_BUILD_TYPE=Release ..
+make -j2 _kaldifeat
diff --git a/.github/scripts/multi-zh-hans.sh b/.github/scripts/multi-zh-hans.sh
new file mode 100755
index 000000000..427d8887b
--- /dev/null
+++ b/.github/scripts/multi-zh-hans.sh
@@ -0,0 +1,158 @@
+#!/usr/bin/env bash
+
+set -ex
+
+git config --global user.name "k2-fsa"
+git config --global user.email "csukuangfj@gmail.com"
+git config --global lfs.allowincompletepush true
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "pwd: $PWD"
+
+cd egs/multi_zh-hans/ASR
+
+repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-streaming-2023-11-05
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+cd exp/
+git lfs pull --include pretrained.pt
+rm -fv epoch-20.pt
+rm -fv *.onnx
+ln -s pretrained.pt epoch-20.pt
+cd ../data/lang_bpe_2000
+ls -lh
+git lfs pull --include L.pt L_disambig.pt Linv.pt bpe.model
+git lfs pull --include "*.model"
+ls -lh
+popd
+
+log "----------------------------------------"
+log "Export streaming ONNX CTC models "
+log "----------------------------------------"
+./zipformer/export-onnx-streaming-ctc.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  --causal 1 \
+  --avg 1 \
+  --epoch 20 \
+  --use-averaged-model 0 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --use-ctc 1
+
+ls -lh $repo/exp/
+
+log "------------------------------------------------------------"
+log "Test exported streaming ONNX CTC models (greedy search)     "
+log "------------------------------------------------------------"
+
+test_wavs=(
+DEV_T0000000000.wav
+DEV_T0000000001.wav
+DEV_T0000000002.wav
+TEST_MEETING_T0000000113.wav
+TEST_MEETING_T0000000219.wav
+TEST_MEETING_T0000000351.wav
+)
+
+for w in ${test_wavs[@]}; do
+  ./zipformer/onnx_pretrained-streaming-ctc.py \
+    --model-filename $repo/exp/ctc-epoch-20-avg-1-chunk-16-left-128.int8.onnx \
+    --tokens $repo/data/lang_bpe_2000/tokens.txt \
+    $repo/test_wavs/$w
+done
+
+log "Upload onnx CTC models to huggingface"
+url=https://huggingface.co/k2-fsa/sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $url
+dst=$(basename $url)
+cp -v $repo/exp/ctc*.onnx $dst
+cp -v $repo/data/lang_bpe_2000/tokens.txt $dst
+cp -v $repo/data/lang_bpe_2000/bpe.model $dst
+mkdir -p $dst/test_wavs
+cp -v $repo/test_wavs/*.wav $dst/test_wavs
+cd $dst
+git lfs track "*.onnx" "bpe.model"
+ls -lh
+file bpe.model
+git status
+git add .
+git commit -m "upload model" && git push https://k2-fsa:${HF_TOKEN}@huggingface.co/k2-fsa/$dst main || true
+
+log "Upload models to https://github.com/k2-fsa/sherpa-onnx"
+rm -rf .git
+rm -fv .gitattributes
+cd ..
+tar cjfv $dst.tar.bz2 $dst
+ls -lh *.tar.bz2
+mv -v $dst.tar.bz2 ../../../
+
+log "----------------------------------------"
+log "Export streaming ONNX transducer models "
+log "----------------------------------------"
+
+./zipformer/export-onnx-streaming.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  --causal 1 \
+  --avg 1 \
+  --epoch 20 \
+  --use-averaged-model 0 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --use-ctc 0
+
+ls -lh $repo/exp
+
+log "------------------------------------------------------------"
+log "Test exported streaming ONNX transducer models (Python code)"
+log "------------------------------------------------------------"
+
+log "test fp32"
+./zipformer/onnx_pretrained-streaming.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-20-avg-1-chunk-16-left-128.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-20-avg-1-chunk-16-left-128.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-20-avg-1-chunk-16-left-128.onnx \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav
+
+log "test int8"
+./zipformer/onnx_pretrained-streaming.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-20-avg-1-chunk-16-left-128.int8.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-20-avg-1-chunk-16-left-128.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-20-avg-1-chunk-16-left-128.int8.onnx \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav
+
+log "Upload onnx transducer models to huggingface"
+
+url=https://huggingface.co/k2-fsa/sherpa-onnx-streaming-zipformer-multi-zh-hans-2023-12-12
+GIT_LFS_SKIP_SMUDGE=1 git clone $url
+dst=$(basename $url)
+cp -v $repo/exp/encoder*.onnx $dst
+cp -v $repo/exp/decoder*.onnx $dst
+cp -v $repo/exp/joiner*.onnx $dst
+cp -v $repo/data/lang_bpe_2000/tokens.txt $dst
+cp -v $repo/data/lang_bpe_2000/bpe.model $dst
+mkdir -p $dst/test_wavs
+cp -v $repo/test_wavs/*.wav $dst/test_wavs
+cd $dst
+git lfs track "*.onnx" bpe.model
+git add .
+git commit -m "upload model" && git push https://k2-fsa:${HF_TOKEN}@huggingface.co/k2-fsa/$dst main || true
+
+log "Upload models to https://github.com/k2-fsa/sherpa-onnx"
+rm -rf .git
+rm -fv .gitattributes
+cd ..
+tar cjfv $dst.tar.bz2 $dst
+ls -lh *.tar.bz2
+mv -v $dst.tar.bz2 ../../../
diff --git a/.github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh b/.github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
new file mode 100755
index 000000000..1b48aae27
--- /dev/null
+++ b/.github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+
+# This script assumes that test-clean and test-other are downloaded
+# to egs/librispeech/ASR/download/LibriSpeech and generates manifest
+# files in egs/librispeech/ASR/data/manifests
+
+set -e
+
+cd egs/librispeech/ASR
+[ ! -e download ] && ln -s ~/tmp/download .
+mkdir -p data/manifests
+lhotse prepare librispeech -j 2 -p test-clean -p test-other ./download/LibriSpeech data/manifests
+ls -lh data/manifests
diff --git a/.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh b/.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
new file mode 100755
index 000000000..c3640cfde
--- /dev/null
+++ b/.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+git lfs install
+
+fbank_url=https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+log "Downloading pre-commputed fbank from $fbank_url"
+
+git clone https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+ln -s $PWD/aishell-test-dev-manifests/data .
+
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+log "Downloading pre-trained model from $repo_url"
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-epoch-29-avg-5-torch-1.10.0.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_char data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless3/exp
+
+  log "Decoding test and dev"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless3/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless3/exp
+  done
+
+  rm pruned_transducer_stateless3/exp/*.pt
+fi
diff --git a/.github/scripts/run-aishell-zipformer-2023-10-24.sh b/.github/scripts/run-aishell-zipformer-2023-10-24.sh
new file mode 100755
index 000000000..865e29799
--- /dev/null
+++ b/.github/scripts/run-aishell-zipformer-2023-10-24.sh
@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+git lfs install
+
+fbank_url=https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+log "Downloading pre-commputed fbank from $fbank_url"
+
+git clone https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+ln -s $PWD/aishell-test-dev-manifests/data .
+
+log "======================="
+log "CI testing large model"
+repo_url=https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-large-2023-10-24/
+log "Downloading pre-trained model from $repo_url"
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for method in modified_beam_search greedy_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --context-size 1 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_char/tokens.txt \
+    --num-encoder-layers 2,2,4,5,4,2 \
+    --feedforward-dim 512,768,1536,2048,1536,768 \
+    --encoder-dim 192,256,512,768,512,256 \
+    --encoder-unmasked-dim 192,192,256,320,256,192 \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+log "======================="
+log "CI testing medium model"
+repo_url=https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-2023-10-24/
+log "Downloading pre-trained model from $repo_url"
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+
+for method in modified_beam_search greedy_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --context-size 1 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_char/tokens.txt \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+
+log "======================="
+log "CI testing small model"
+repo_url=https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-small-2023-10-24/
+log "Downloading pre-trained model from $repo_url"
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+
+for method in modified_beam_search greedy_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --context-size 1 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_char/tokens.txt \
+    --num-encoder-layers 2,2,2,2,2,2 \
+    --feedforward-dim 512,768,768,768,768,768 \
+    --encoder-dim 192,256,256,256,256,256 \
+    --encoder-unmasked-dim 192,192,192,192,192,192 \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
diff --git a/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh b/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh
new file mode 100755
index 000000000..b61a9d7b6
--- /dev/null
+++ b/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh
@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/gigaspeech/ASR
+
+repo_url=https://huggingface.co/wgb14/icefall-asr-gigaspeech-pruned-transducer-stateless2
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless2/exp
+  ln -s $PWD/$repo/exp/pretrained-iter-3488000-avg-20.pt pruned_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh data/lang_bpe_500
+  ls -lh data/fbank
+  ls -lh pruned_transducer_stateless2/exp
+
+  ln -s data/fbank/cuts_DEV.jsonl.gz data/fbank/gigaspeech_cuts_DEV.jsonl.gz
+  ln -s data/fbank/cuts_TEST.jsonl.gz data/fbank/gigaspeech_cuts_TEST.jsonl.gz
+
+  log "Decoding dev and test"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  # Test only greedy_search to reduce CI running time
+  # for method in greedy_search fast_beam_search modified_beam_search; do
+  for method in greedy_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless2/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless2/exp
+  done
+
+  rm pruned_transducer_stateless2/exp/*.pt
+fi
diff --git a/.github/scripts/run-gigaspeech-zipformer-2023-10-17.sh b/.github/scripts/run-gigaspeech-zipformer-2023-10-17.sh
new file mode 100755
index 000000000..329896ef6
--- /dev/null
+++ b/.github/scripts/run-gigaspeech-zipformer-2023-10-17.sh
@@ -0,0 +1,158 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/gigaspeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "exp/jit_script.pt"
+git lfs pull --include "exp/pretrained.pt"
+rm epoch-30.pt
+ln -s pretrained.pt epoch-30.pt
+rm *.onnx
+ls -lh
+popd
+
+log "----------------------------------------"
+log "Export ONNX transducer models "
+log "----------------------------------------"
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 30 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+ls -lh $repo/exp
+
+log "------------------------------------------------------------"
+log "Test exported ONNX transducer models (Python code)          "
+log "------------------------------------------------------------"
+
+log "test fp32"
+./zipformer/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-30-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-30-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-30-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "test int8"
+./zipformer/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-30-avg-1.int8.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-30-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-30-avg-1.int8.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Upload models to huggingface"
+git config --global user.name "k2-fsa"
+git config --global user.email "xxx@gmail.com"
+
+url=https://huggingface.co/k2-fsa/sherpa-onnx-zipformer-gigaspeech-2023-12-12
+GIT_LFS_SKIP_SMUDGE=1 git clone $url
+dst=$(basename $url)
+cp -v $repo/exp/*.onnx $dst
+cp -v $repo/data/lang_bpe_500/tokens.txt $dst
+cp -v $repo/data/lang_bpe_500/bpe.model $dst
+mkdir -p $dst/test_wavs
+cp -v $repo/test_wavs/*.wav $dst/test_wavs
+cd $dst
+git lfs track "*.onnx"
+git add .
+git commit -m "upload model" && git push https://k2-fsa:${HF_TOKEN}@huggingface.co/k2-fsa/$dst main || true
+
+log "Upload models to https://github.com/k2-fsa/sherpa-onnx"
+rm -rf .git
+rm -fv .gitattributes
+cd ..
+tar cjfv $dst.tar.bz2 $dst
+ls -lh
+mv -v $dst.tar.bz2 ../../../
+
+log "Export to torchscript model"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer/jit_pretrained.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --nn-model-filename $repo/exp/jit_script.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for method in greedy_search modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer/exp/epoch-30.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./zipformer/decode.py \
+      --decoding-method $method \
+      --epoch 30 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  rm zipformer/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh b/.github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh
new file mode 100755
index 000000000..f6fe8c9b2
--- /dev/null
+++ b/.github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh
@@ -0,0 +1,122 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conformer-ctc3-2022-11-27
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lm/G_4_gram.pt"
+git lfs pull --include "exp/jit_trace.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Decode with models exported by torch.jit.trace()"
+
+for m in ctc-decoding 1best; do
+  ./conformer_ctc3/jit_pretrained.py \
+    --model-filename $repo/exp/jit_trace.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+log "Export to torchscript model"
+
+./conformer_ctc3/export.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --jit-trace 1 \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.trace()"
+
+for m in ctc-decoding 1best; do
+  ./conformer_ctc3/jit_pretrained.py \
+    --model-filename $repo/exp/jit_trace.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for m in ctc-decoding 1best; do
+  ./conformer_ctc3/pretrained.py \
+    --checkpoint $repo/exp/pretrained.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode" ]]; then
+  mkdir -p conformer_ctc3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt conformer_ctc3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh conformer_ctc3/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in ctc-decoding 1best; do
+    log "Decoding with $method"
+    ./conformer_ctc3/decode.py \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --exp-dir conformer_ctc3/exp/ \
+      --max-duration $max_duration \
+      --decoding-method $method \
+      --lm-dir data/lm
+  done
+
+  rm conformer_ctc3/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh b/.github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh
new file mode 100755
index 000000000..d547bdd45
--- /dev/null
+++ b/.github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh
@@ -0,0 +1,191 @@
+#!/usr/bin/env bash
+#
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+abs_repo=$(realpath $repo)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-iter-468000-avg-16.pt pretrained.pt
+ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+log "Test exporting with torch.jit.trace()"
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --jit-trace 1
+
+log "Decode with models exported by torch.jit.trace()"
+
+./lstm_transducer_stateless2/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo/exp/encoder_jit_trace.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_trace.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_trace.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./lstm_transducer_stateless2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./lstm_transducer_stateless2/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+
+if [[ x"${GITHUB_EVENT_LABEL_NAME}" == x"shallow-fusion" ]]; then
+  lm_repo_url=https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+  log "Download pre-trained RNN-LM model from ${lm_repo_url}"
+  GIT_LFS_SKIP_SMUDGE=1 git clone $lm_repo_url
+  lm_repo=$(basename $lm_repo_url)
+  pushd $lm_repo
+  git lfs pull --include "exp/pretrained.pt"
+  mv exp/pretrained.pt exp/epoch-88.pt
+  popd
+
+  mkdir -p lstm_transducer_stateless2/exp
+  ln -sf $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh lstm_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other with RNN LM"
+
+  ./lstm_transducer_stateless2/decode.py \
+    --use-averaged-model 0 \
+    --epoch 999 \
+    --avg 1 \
+    --exp-dir lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam 4 \
+    --use-shallow-fusion 1 \
+    --lm-type rnn \
+    --lm-exp-dir $lm_repo/exp \
+    --lm-epoch 88 \
+    --lm-avg 1 \
+    --lm-scale 0.3 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+fi
+
+if [[ x"${GITHUB_EVENT_LABEL_NAME}" == x"LODR" ]]; then
+  bigram_repo_url=https://huggingface.co/marcoyang/librispeech_bigram
+  log "Download bi-gram LM from ${bigram_repo_url}"
+  GIT_LFS_SKIP_SMUDGE=1 git clone $bigram_repo_url
+  bigramlm_repo=$(basename $bigram_repo_url)
+  pushd $bigramlm_repo
+  git lfs pull --include "2gram.fst.txt"
+  cp 2gram.fst.txt $abs_repo/data/lang_bpe_500/.
+  popd
+
+  lm_repo_url=https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+  log "Download pre-trained RNN-LM model from ${lm_repo_url}"
+  GIT_LFS_SKIP_SMUDGE=1 git clone $lm_repo_url
+  lm_repo=$(basename $lm_repo_url)
+  pushd $lm_repo
+  git lfs pull --include "exp/pretrained.pt"
+  mv exp/pretrained.pt exp/epoch-88.pt
+  popd
+
+  mkdir -p lstm_transducer_stateless2/exp
+  ln -sf $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh lstm_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other"
+
+  ./lstm_transducer_stateless2/decode.py \
+    --use-averaged-model 0 \
+    --epoch 999 \
+    --avg 1 \
+    --exp-dir lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_LODR \
+    --beam 4 \
+    --use-shallow-fusion 1 \
+    --lm-type rnn \
+    --lm-exp-dir $lm_repo/exp \
+    --lm-scale 0.4 \
+    --lm-epoch 88 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1 \
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16
+fi
+
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
+  mkdir -p lstm_transducer_stateless2/exp
+  ln -s $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh lstm_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./lstm_transducer_stateless2/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir lstm_transducer_stateless2/exp
+  done
+
+  rm lstm_transducer_stateless2/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh
new file mode 100755
index 000000000..412e3ad56
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh
@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in fast_beam_search modified_beam_search beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless/exp
+  done
+
+  rm pruned_transducer_stateless/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh
new file mode 100755
index 000000000..243b669ed
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh
@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless2-2022-04-29
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-38-avg-10.pt"
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-epoch-38-avg-10.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless2/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless2/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless2/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless2/exp
+  done
+
+  rm pruned_transducer_stateless2/exp/*.pt
+  rm -r data/lang_bpe_500
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh
new file mode 100755
index 000000000..2d0f80304
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh
@@ -0,0 +1,85 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-04-29
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-25-avg-6.pt"
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-epoch-25-avg-6.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless3/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless3/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless3/exp
+  done
+
+  rm pruned_transducer_stateless3/exp/*.pt
+  rm -r data/lang_bpe_500
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-05-13.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-05-13.sh
new file mode 100755
index 000000000..3d5814c48
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-05-13.sh
@@ -0,0 +1,123 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-iter-1224000-avg-14.pt pretrained.pt
+ln -s pretrained-iter-1224000-avg-14.pt epoch-99.pt
+popd
+
+
+log "Export to torchscript model"
+./pruned_transducer_stateless3/export.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit-trace 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.trace()"
+
+./pruned_transducer_stateless3/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo/exp/encoder_jit_trace.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_trace.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_trace.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless3/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo/exp/encoder_jit_script.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_script.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_script.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless3/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless3/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless3/exp
+  done
+
+  rm pruned_transducer_stateless3/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh
new file mode 100755
index 000000000..3d2442d54
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh
@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-epoch-39-avg-7.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless5/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --num-encoder-layers 18 \
+    --dim-feedforward 2048 \
+    --nhead 8 \
+    --encoder-dim 512 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless5/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav \
+    --num-encoder-layers 18 \
+    --dim-feedforward 2048 \
+    --nhead 8 \
+    --encoder-dim 512 \
+    --decoder-dim 512 \
+    --joiner-dim 512
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless5/exp
+  ln -s $PWD/$repo/exp/pretrained-epoch-39-avg-7.pt pruned_transducer_stateless5/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless5/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless5/decode.py \
+      --decoding-method $method \
+      --use-averaged-model 0 \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless5/exp \
+      --num-encoder-layers 18 \
+      --dim-feedforward 2048 \
+      --nhead 8 \
+      --encoder-dim 512 \
+      --decoder-dim 512 \
+      --joiner-dim 512
+  done
+
+  rm pruned_transducer_stateless5/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh
new file mode 100755
index 000000000..961dde4f4
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh
@@ -0,0 +1,106 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless7/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless7/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless7/exp
+  done
+
+  rm pruned_transducer_stateless7/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh
new file mode 100755
index 000000000..ba7139efb
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh
@@ -0,0 +1,150 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lm/G_4_gram.pt"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7_ctc/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
+    --model-filename $repo/exp/cpu_jit.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7_ctc/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7_ctc/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc/pretrained_ctc.py \
+    --checkpoint $repo/exp/pretrained.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless7_ctc/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_ctc/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7_ctc/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless7_ctc/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless7_ctc/exp
+  done
+
+  for m in ctc-decoding 1best; do
+    ./pruned_transducer_stateless7_ctc/ctc_decode.py \
+        --epoch 999 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+        --max-duration $max_duration \
+        --use-averaged-model 0 \
+        --decoding-method $m \
+        --hlg-scale 0.6 \
+        --lm-dir data/lm
+  done
+
+  rm pruned_transducer_stateless7_ctc/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh
new file mode 100755
index 000000000..1ecbc4798
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh
@@ -0,0 +1,147 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+    --model-filename $repo/exp/cpu_jit.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py \
+    --checkpoint $repo/exp/pretrained.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless7_ctc_bs/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_ctc_bs/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7_ctc_bs/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless7_ctc_bs/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless7_ctc_bs/exp
+  done
+
+  for m in ctc-decoding 1best; do
+    ./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+        --epoch 999 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+        --max-duration $max_duration \
+        --use-averaged-model 0 \
+        --decoding-method $m \
+        --hlg-scale 0.6
+  done
+  
+  rm pruned_transducer_stateless7_ctc_bs/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless7-streaming-2022-12-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless7-streaming-2022-12-29.sh
new file mode 100755
index 000000000..37b192a57
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless7-streaming-2022-12-29.sh
@@ -0,0 +1,148 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "exp/encoder_jit_trace.pt"
+git lfs pull --include "exp/decoder_jit_trace.pt"
+git lfs pull --include "exp/joiner_jit_trace.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --decode-chunk-len 32 \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7_streaming/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  --decode-chunk-len 32 \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Export to torchscript model by torch.jit.trace()"
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --decode-chunk-len 32 \
+  --epoch 99 \
+  --avg 1
+
+log "Decode with models exported by torch.jit.trace()"
+
+./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo/exp/encoder_jit_trace.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_trace.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_trace.pt \
+  --decode-chunk-len 32 \
+  $repo/test_wavs/1089-134686-0001.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7_streaming/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --decode-chunk-len 32 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7_streaming/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --decode-chunk-len 32 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless7_streaming/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_streaming/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7_streaming/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+  num_decode_stream=200
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "decoding with $method"
+
+    ./pruned_transducer_stateless7_streaming/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --decode-chunk-len 32 \
+      --exp-dir pruned_transducer_stateless7_streaming/exp
+  done
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless7_streaming/streaming_decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --decode-chunk-len 32 \
+      --num-decode-streams $num_decode_stream
+      --exp-dir pruned_transducer_stateless7_streaming/exp
+  done
+
+  rm pruned_transducer_stateless7_streaming/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh
new file mode 100755
index 000000000..4f2bfac24
--- /dev/null
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh
@@ -0,0 +1,115 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Export to torchscript model"
+./pruned_transducer_stateless8/export.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model false \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless8/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless8/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless8/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless8/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless8/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless8/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless8/exp
+  done
+
+  rm pruned_transducer_stateless8/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-streaming-pruned-transducer-stateless2-2022-06-26.sh b/.github/scripts/run-librispeech-streaming-pruned-transducer-stateless2-2022-06-26.sh
new file mode 100755
index 000000000..5cbdad16d
--- /dev/null
+++ b/.github/scripts/run-librispeech-streaming-pruned-transducer-stateless2-2022-06-26.sh
@@ -0,0 +1,101 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless2_20220625
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-epoch-24-avg-10.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless2/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless2/exp
+  ln -s $PWD/$repo/exp/pretrained-epoch-24-avg-10.pt pruned_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Simulate streaming decoding with $method"
+
+    ./pruned_transducer_stateless2/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless2/exp \
+      --simulate-streaming 1 \
+      --causal-convolution 1
+  done
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Real streaming decoding with $method"
+
+    ./pruned_transducer_stateless2/streaming_decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --num-decode-streams 100 \
+      --exp-dir pruned_transducer_stateless2/exp \
+      --left-context 32 \
+      --decode-chunk-size 8 \
+      --right-context 0
+  done
+
+  rm pruned_transducer_stateless2/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh b/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh
new file mode 100755
index 000000000..f4e2124b1
--- /dev/null
+++ b/.github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh
@@ -0,0 +1,116 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "exp/jit_script_chunk_16_left_128.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer/jit_pretrained_streaming.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --nn-model-filename $repo/exp/jit_script_chunk_16_left_128.pt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+for method in greedy_search modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --causal 1 \
+    --chunk-size 16 \
+    --left-context-frames 128 \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Simulated streaming decoding with $method"
+
+    ./zipformer/decode.py \
+      --causal 1 \
+      --chunk-size 16 \
+      --left-context-frames 128 \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Chunk-wise streaming decoding with $method"
+
+    ./zipformer/streaming_decode.py \
+      --causal 1 \
+      --chunk-size 16 \
+      --left-context-frames 128 \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  rm zipformer/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh b/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh
new file mode 100755
index 000000000..ff77855a2
--- /dev/null
+++ b/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh
@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless2-torchaudio-2022-04-19
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in fast_beam_search modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless2/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless2/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless2/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless2/exp
+  done
+
+  rm transducer_stateless2/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-zipformer-2023-05-18.sh b/.github/scripts/run-librispeech-zipformer-2023-05-18.sh
new file mode 100755
index 000000000..fb1a0149d
--- /dev/null
+++ b/.github/scripts/run-librispeech-zipformer-2023-05-18.sh
@@ -0,0 +1,94 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "exp/jit_script.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer/jit_pretrained.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --nn-model-filename $repo/exp/jit_script.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for method in greedy_search modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./zipformer/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  rm zipformer/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh b/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh
new file mode 100755
index 000000000..0026d2109
--- /dev/null
+++ b/.github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh
@@ -0,0 +1,117 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lm/G_4_gram.pt"
+git lfs pull --include "exp/jit_script.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --use-transducer 1 \
+  --use-ctc 1 \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+for method in ctc-decoding 1best; do
+  ./zipformer/jit_pretrained_ctc.py \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --model-filename $repo/exp/jit_script.pt \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $method \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in ctc-decoding 1best; do
+  log "$method"
+
+  ./zipformer/pretrained_ctc.py \
+    --use-transducer 1 \
+    --use-ctc 1 \
+    --method $method \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --G $repo/data/lm/G_4_gram.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in ctc-decoding 1best; do
+    log "Decoding with $method"
+
+    ./zipformer/ctc_decode.py \
+      --use-transducer 1 \
+      --use-ctc 1 \
+      --decoding-method $method \
+      --nbest-scale 1.0 \
+      --hlg-scale 0.6 \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  rm zipformer/exp/*.pt
+fi
diff --git a/.github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh b/.github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh
new file mode 100755
index 000000000..c59921055
--- /dev/null
+++ b/.github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh
@@ -0,0 +1,102 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/3gram.pt"
+git lfs pull --include "data/lang_bpe_500/4gram.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer_mmi/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer_mmi/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  --lang-dir $repo/data/lang_bpe_500 \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for method in 1best nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+  log "$method"
+
+  ./zipformer_mmi/pretrained.py \
+    --method $method \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_bpe_500 \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer_mmi/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer_mmi/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer_mmi/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in 1best nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+    log "Decoding with $method"
+
+    ./zipformer_mmi/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --nbest-scale 1.2 \
+      --hp-scale 1.0 \
+      --max-duration $max_duration \
+      --lang-dir $repo/data/lang_bpe_500 \
+      --exp-dir zipformer_mmi/exp
+  done
+
+  rm zipformer_mmi/exp/*.pt
+fi
diff --git a/.github/scripts/run-multi-corpora-zipformer.sh b/.github/scripts/run-multi-corpora-zipformer.sh
new file mode 100755
index 000000000..90f859f43
--- /dev/null
+++ b/.github/scripts/run-multi-corpora-zipformer.sh
@@ -0,0 +1,135 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/multi_zh-hans/ASR
+
+log "==== Test icefall-asr-multi-zh-hans-zipformer-2023-9-2 ===="
+repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s epoch-20.pt epoch-99.pt
+popd
+
+ls -lh $repo/exp/*.pt
+
+
+./zipformer/pretrained.py \
+  --checkpoint $repo/exp/epoch-99.pt \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  --method greedy_search \
+$repo/test_wavs/DEV_T0000000000.wav \
+$repo/test_wavs/DEV_T0000000001.wav \
+$repo/test_wavs/DEV_T0000000002.wav
+
+for method in modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/epoch-99.pt \
+    --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav
+done
+
+rm -rf $repo
+
+log "==== Test icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 ===="
+repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24/
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s epoch-20.pt epoch-99.pt
+popd
+
+ls -lh $repo/exp/*.pt
+
+
+./zipformer/pretrained.py \
+  --checkpoint $repo/exp/epoch-99.pt \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  --use-ctc 1 \
+  --method greedy_search \
+$repo/test_wavs/DEV_T0000000000.wav \
+$repo/test_wavs/DEV_T0000000001.wav \
+$repo/test_wavs/DEV_T0000000002.wav
+
+for method in modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --use-ctc 1 \
+    --checkpoint $repo/exp/epoch-99.pt \
+    --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav
+done
+
+rm -rf $repo
+
+cd ../../../egs/multi_zh_en/ASR
+log "==== Test icefall-asr-zipformer-multi-zh-en-2023-11-22 ===="
+repo_url=https://huggingface.co/zrjin/icefall-asr-zipformer-multi-zh-en-2023-11-22/
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+./zipformer/pretrained.py \
+  --checkpoint $repo/exp/pretrained.pt \
+  --bpe-model $repo/data/lang_bbpe_2000/bbpe.model \
+  --method greedy_search \
+$repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_29.wav \
+$repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_55.wav \
+$repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_75.wav
+
+for method in modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bbpe_2000/bbpe.model \
+  $repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_29.wav \
+  $repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_55.wav \
+  $repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_75.wav
+done
+
+rm -rf $repo
diff --git a/.github/scripts/run-pre-trained-ctc.sh b/.github/scripts/run-pre-trained-ctc.sh
new file mode 100755
index 000000000..7d6449c9a
--- /dev/null
+++ b/.github/scripts/run-pre-trained-ctc.sh
@@ -0,0 +1,240 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+pushd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC greedy search"
+
+./zipformer/onnx_pretrained_ctc.py \
+  --nn-model $repo/model.onnx \
+  --tokens $repo/tokens.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "CTC H decoding"
+
+./zipformer/onnx_pretrained_ctc_H.py \
+  --nn-model $repo/model.onnx \
+  --tokens $repo/tokens.txt \
+  --H $repo/H.fst \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "CTC HL decoding"
+
+./zipformer/onnx_pretrained_ctc_HL.py \
+  --nn-model $repo/model.onnx \
+  --words $repo/words.txt \
+  --HL $repo/HL.fst \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "CTC HLG decoding"
+
+./zipformer/onnx_pretrained_ctc_HLG.py \
+  --nn-model $repo/model.onnx \
+  --words $repo/words.txt \
+  --HLG $repo/HLG.fst \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+rm -rf $repo
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/L_disambig.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/lexicon.txt"
+git lfs pull --include "data/lang_bpe_500/lexicon_disambig.txt"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "data/lang_bpe_500/words.txt"
+git lfs pull --include "data/lm/G_3_gram.fst.txt"
+
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC decoding"
+
+./conformer_ctc/pretrained.py \
+  --method ctc-decoding \
+  --num-classes 500 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "HLG decoding"
+
+./conformer_ctc/pretrained.py \
+  --method 1best \
+  --num-classes 500 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --words-file $repo/data/lang_bpe_500/words.txt \
+  --HLG $repo/data/lang_bpe_500/HLG.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "CTC decoding on CPU with kaldi decoders using OpenFst"
+
+log "Exporting model with torchscript"
+
+pushd $repo/exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+./conformer_ctc/export.py \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --jit 1
+
+ls -lh $repo/exp
+
+
+log "Generating H.fst, HL.fst"
+
+./local/prepare_lang_fst.py  --lang-dir $repo/data/lang_bpe_500 --ngram-G $repo/data/lm/G_3_gram.fst.txt
+
+ls -lh $repo/data/lang_bpe_500
+
+log "Decoding with H on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_H.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --H $repo/data/lang_bpe_500/H.fst \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Decoding with HL on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HL.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HL $repo/data/lang_bpe_500/HL.fst \
+  --words $repo/data/lang_bpe_500/words.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Decoding with HLG on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HLG $repo/data/lang_bpe_500/HLG.fst \
+  --words $repo/data/lang_bpe_500/words.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+
+popd
+
+log "Test aishell"
+
+pushd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall_asr_aishell_conformer_ctc
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "data/lang_char/H.fst"
+git lfs pull --include "data/lang_char/HL.fst"
+git lfs pull --include "data/lang_char/HLG.fst"
+
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC decoding"
+
+log "Exporting model with torchscript"
+
+pushd $repo/exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+./conformer_ctc/export.py \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --jit 1
+
+ls -lh $repo/exp
+
+ls -lh $repo/data/lang_char
+
+log "Decoding with H on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_H.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --H $repo/data/lang_char/H.fst \
+  --tokens $repo/data/lang_char/tokens.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "Decoding with HL on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HL.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HL $repo/data/lang_char/HL.fst \
+  --words $repo/data/lang_char/words.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "Decoding with HLG on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HLG $repo/data/lang_char/HLG.fst \
+  --words $repo/data/lang_char/words.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+rm -rf $repo
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh
new file mode 100755
index 000000000..7b686328d
--- /dev/null
+++ b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh
@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-100h-transducer-stateless-multi-datasets-bpe-500-2022-02-21
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_multi_datasets/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./transducer_stateless_multi_datasets/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless_multi_datasets/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless_multi_datasets/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh transducer_stateless_multi_datasets/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless_multi_datasets/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless_multi_datasets/exp
+  done
+
+  rm transducer_stateless_multi_datasets/exp/*.pt
+fi
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh
new file mode 100755
index 000000000..a8eeeb514
--- /dev/null
+++ b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh
@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-multi-datasets-bpe-500-2022-03-01
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_multi_datasets/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./transducer_stateless_multi_datasets/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless_multi_datasets/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless_multi_datasets/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh transducer_stateless_multi_datasets/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless_multi_datasets/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless_multi_datasets/exp
+  done
+
+  rm transducer_stateless_multi_datasets/exp/*.pt
+fi
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-modified-2-aishell.sh b/.github/scripts/run-pre-trained-transducer-stateless-modified-2-aishell.sh
new file mode 100755
index 000000000..0644d9be0
--- /dev/null
+++ b/.github/scripts/run-pre-trained-transducer-stateless-modified-2-aishell.sh
@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2-2022-03-01
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_modified-2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless_modified-2/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-modified-aishell.sh b/.github/scripts/run-pre-trained-transducer-stateless-modified-aishell.sh
new file mode 100755
index 000000000..79fb64311
--- /dev/null
+++ b/.github/scripts/run-pre-trained-transducer-stateless-modified-aishell.sh
@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2022-03-01
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_modified/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless_modified/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
diff --git a/.github/scripts/run-pre-trained-transducer-stateless.sh b/.github/scripts/run-pre-trained-transducer-stateless.sh
new file mode 100755
index 000000000..2e2360435
--- /dev/null
+++ b/.github/scripts/run-pre-trained-transducer-stateless.sh
@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in fast_beam_search modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh transducer_stateless/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless/exp
+  done
+
+  rm transducer_stateless/exp/*.pt
+fi
diff --git a/.github/scripts/run-pre-trained-transducer.sh b/.github/scripts/run-pre-trained-transducer.sh
new file mode 100755
index 000000000..b865f8d13
--- /dev/null
+++ b/.github/scripts/run-pre-trained-transducer.sh
@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-bpe-500-2021-12-23
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "Beam search decoding"
+
+./transducer/pretrained.py \
+  --method beam_search \
+  --beam-size 4 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
diff --git a/.github/scripts/run-swbd-conformer-ctc-2023-08-26.sh b/.github/scripts/run-swbd-conformer-ctc-2023-08-26.sh
new file mode 100755
index 000000000..d8cc020e1
--- /dev/null
+++ b/.github/scripts/run-swbd-conformer-ctc-2023-08-26.sh
@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/swbd/ASR
+
+repo_url=https://huggingface.co/zrjin/icefall-asr-swbd-conformer-ctc-2023-8-26
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s epoch-98.pt epoch-99.pt
+popd
+
+ls -lh $repo/exp/*.pt
+
+for method in ctc-decoding 1best; do
+  log "$method"
+
+  ./conformer_ctc/pretrained.py \
+    --method $method \
+    --checkpoint $repo/exp/epoch-99.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --words-file $repo/data/lang_bpe_500/words.txt \
+    --HLG  $repo/data/lang_bpe_500/HLG.pt \
+    --G $repo/data/lm/G_4_gram.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+done
diff --git a/.github/scripts/run-wenetspeech-pruned-transducer-stateless2.sh b/.github/scripts/run-wenetspeech-pruned-transducer-stateless2.sh
new file mode 100755
index 000000000..a3a2d3080
--- /dev/null
+++ b/.github/scripts/run-wenetspeech-pruned-transducer-stateless2.sh
@@ -0,0 +1,119 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained_epoch_10_avg_2.pt pretrained.pt
+ln -s pretrained_epoch_10_avg_2.pt epoch-99.pt
+popd
+
+log "Test exporting to ONNX format"
+
+./pruned_transducer_stateless2/export-onnx.py \
+  --exp-dir $repo/exp \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1
+
+log "Export to torchscript model"
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir $repo/exp \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir $repo/exp \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --jit-trace 1
+
+ls -lh $repo/exp/*.onnx
+ls -lh $repo/exp/*.pt
+
+log "Decode with ONNX models"
+
+./pruned_transducer_stateless2/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-10-avg-2.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-10-avg-2.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-10-avg-2.onnx \
+  --onnx-joiner-encoder-proj-filename $repo/exp/joiner_encoder_proj-epoch-10-avg-2.onnx \
+  --onnx-joiner-decoder-proj-filename $repo/exp/joiner_decoder_proj-epoch-10-avg-2.onnx
+
+./pruned_transducer_stateless2/onnx_pretrained.py \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav
+
+log "Decode with models exported by torch.jit.trace()"
+
+./pruned_transducer_stateless2/jit_pretrained.py \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --encoder-model-filename $repo/exp/encoder_jit_trace.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_trace.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_trace.pt \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav
+
+./pruned_transducer_stateless2/jit_pretrained.py \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --encoder-model-filename $repo/exp/encoder_jit_script.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_script.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_script.pt \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint $repo/exp/epoch-99.pt \
+    --lang-dir $repo/data/lang_char \
+    --decoding-method greedy_search \
+    --max-sym-per-frame $sym \
+    $repo/test_wavs/DEV_T0000000000.wav \
+    $repo/test_wavs/DEV_T0000000001.wav \
+    $repo/test_wavs/DEV_T0000000002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless2/pretrained.py \
+    --decoding-method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/epoch-99.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/DEV_T0000000000.wav \
+    $repo/test_wavs/DEV_T0000000001.wav \
+    $repo/test_wavs/DEV_T0000000002.wav
+done
diff --git a/.github/scripts/test-ncnn-export.sh b/.github/scripts/test-ncnn-export.sh
new file mode 100755
index 000000000..4073c594a
--- /dev/null
+++ b/.github/scripts/test-ncnn-export.sh
@@ -0,0 +1,230 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+pushd egs/librispeech/ASR
+
+log  "Install ncnn and pnnx"
+
+# We are using a modified ncnn here. Will try to merge it to the official repo
+# of ncnn
+git clone https://github.com/csukuangfj/ncnn
+pushd ncnn
+git submodule init
+git submodule update python/pybind11
+python3 setup.py bdist_wheel
+ls -lh dist/
+pip install dist/*.whl
+cd tools/pnnx
+mkdir build
+cd build
+
+echo "which python3"
+
+which python3
+#/opt/hostedtoolcache/Python/3.8.16/x64/bin/python3
+
+cmake -D Python3_EXECUTABLE=$(which python3) ..
+make -j4 pnnx
+
+./src/pnnx || echo "pass"
+
+popd
+
+export PATH=$PWD/ncnn/tools/pnnx/build/src:$PATH
+
+log "=========================================================================="
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.trace()"
+
+./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir $repo/exp \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+pnnx $repo/exp/encoder_jit_trace-pnnx.pt
+pnnx $repo/exp/decoder_jit_trace-pnnx.pt
+pnnx $repo/exp/joiner_jit_trace-pnnx.pt
+
+python3 ./conv_emformer_transducer_stateless2/streaming-ncnn-decode.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+  --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+  --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+  --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+  --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+  --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+  $repo/test_wavs/1089-134686-0001.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+
+cd exp
+ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.trace()"
+
+./lstm_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0
+
+pnnx $repo/exp/encoder_jit_trace-pnnx.pt
+pnnx $repo/exp/decoder_jit_trace-pnnx.pt
+pnnx $repo/exp/joiner_jit_trace-pnnx.pt
+
+python3 ./lstm_transducer_stateless2/streaming-ncnn-decode.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+  --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+  --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+  --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+  --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+  --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+  $repo/test_wavs/1089-134686-0001.wav
+
+python3 ./lstm_transducer_stateless2/ncnn-decode.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+  --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+  --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+  --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+  --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+  --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+  $repo/test_wavs/1089-134686-0001.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+./pruned_transducer_stateless7_streaming/export-for-ncnn.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --exp-dir $repo/exp \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  \
+  --decode-chunk-len 32 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims "1024,1024,2048,2048,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+pnnx $repo/exp/encoder_jit_trace-pnnx.pt
+pnnx $repo/exp/decoder_jit_trace-pnnx.pt
+pnnx $repo/exp/joiner_jit_trace-pnnx.pt
+
+python3 ./pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+  --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+  --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+  --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+  --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+  --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+  $repo/test_wavs/1089-134686-0001.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/pfluo/k2fsa-zipformer-chinese-english-mixed
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char_bpe/L.pt"
+git lfs pull --include "data/lang_char_bpe/L_disambig.pt"
+git lfs pull --include "data/lang_char_bpe/Linv.pt"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+./pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py \
+  --tokens $repo/data/lang_char_bpe/tokens.txt \
+  --exp-dir $repo/exp \
+  --use-averaged-model 0 \
+  --epoch 9999 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims "1024,1024,1536,1536,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+pnnx $repo/exp/encoder_jit_trace-pnnx.pt
+pnnx $repo/exp/decoder_jit_trace-pnnx.pt
+pnnx $repo/exp/joiner_jit_trace-pnnx.pt
+
+python3 ./pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py \
+  --tokens $repo/data/lang_char_bpe/tokens.txt \
+  --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+  --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+  --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+  --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+  --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+  --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+  $repo/test_wavs/0.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
diff --git a/.github/scripts/test-onnx-export.sh b/.github/scripts/test-onnx-export.sh
new file mode 100755
index 000000000..fcfc11fa6
--- /dev/null
+++ b/.github/scripts/test-onnx-export.sh
@@ -0,0 +1,465 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+log "=========================================================================="
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.script()"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+log "Test export to ONNX format"
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal False \
+  --chunk-size "16,32,64,-1" \
+  --left-context-frames "64,128,256,-1"
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./zipformer/onnx_check.py \
+  --jit-filename $repo/exp/jit_script.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./zipformer/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+rm -rf $repo
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+log "Test export streaming model to ONNX format"
+./zipformer/export-onnx-streaming.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal True \
+  --chunk-size 16 \
+  --left-context-frames 64
+
+ls -lh $repo/exp
+
+log "Run onnx_pretrained-streaming.py"
+
+./zipformer/onnx_pretrained-streaming.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1-chunk-16-left-64.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1-chunk-16-left-64.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1-chunk-16-left-64.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+rm -rf $repo
+
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.trace()"
+
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --exp-dir $repo/exp/
+
+log "Test exporting to ONNX format"
+
+./pruned_transducer_stateless7_streaming/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --exp-dir $repo/exp/
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./pruned_transducer_stateless7_streaming/onnx_check.py \
+  --jit-encoder-filename $repo/exp/encoder_jit_trace.pt \
+  --jit-decoder-filename $repo/exp/decoder_jit_trace.pt \
+  --jit-joiner-filename $repo/exp/joiner_jit_trace.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./pruned_transducer_stateless7_streaming/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-1224000-avg-14.pt"
+
+cd exp
+ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+popd
+
+log "Export via torch.jit.script()"
+
+./pruned_transducer_stateless3/export.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --jit 1
+
+log "Test exporting to ONNX format"
+
+./pruned_transducer_stateless3/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./pruned_transducer_stateless3/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-9999-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./pruned_transducer_stateless3/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-39-avg-7.pt"
+
+cd exp
+ln -s pretrained-epoch-39-avg-7.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.script()"
+
+./pruned_transducer_stateless5/export.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --jit 1
+
+log "Test exporting to ONNX format"
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./pruned_transducer_stateless5/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./pruned_transducer_stateless5/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.script()"
+
+./pruned_transducer_stateless7/export.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024" \
+  --jit 1
+
+log "Test exporting to ONNX format"
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024"
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./pruned_transducer_stateless7/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./pruned_transducer_stateless7/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "=========================================================================="
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-99.pt
+popd
+
+log "Test exporting to ONNX format"
+
+./conv_emformer_transducer_stateless2/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+log "Run onnx_pretrained.py"
+
+./conv_emformer_transducer_stateless2/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1221-135766-0001.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+log "=========================================================================="
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+
+cd exp
+ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.trace()"
+
+./lstm_transducer_stateless2/export.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --jit-trace 1
+
+log "Test exporting to ONNX format"
+
+./lstm_transducer_stateless2/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./lstm_transducer_stateless2/onnx_check.py \
+  --jit-encoder-filename $repo/exp/encoder_jit_trace.pt \
+  --jit-decoder-filename $repo/exp/decoder_jit_trace.pt \
+  --jit-joiner-filename $repo/exp/joiner_jit_trace.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./lstm_transducer_stateless2/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1221-135766-0001.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
diff --git a/.github/workflows/build-doc.yml b/.github/workflows/build-doc.yml
new file mode 100644
index 000000000..d7fe2c964
--- /dev/null
+++ b/.github/workflows/build-doc.yml
@@ -0,0 +1,69 @@
+# Copyright      2022  Xiaomi Corp.       (author: Fangjun Kuang)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# refer to https://github.com/actions/starter-workflows/pull/47/files
+
+# You can access it at https://k2-fsa.github.io/icefall/
+name: Generate doc
+on:
+  push:
+    branches:
+    - master
+    - doc
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: build_doc-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build-doc:
+    if: github.event.label.name == 'doc' || github.event_name == 'push'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.8"]
+    steps:
+      # refer to https://github.com/actions/checkout
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Display Python version
+        run: python -c "import sys; print(sys.version)"
+
+      - name: Build doc
+        shell: bash
+        run: |
+          cd docs
+          python3 -m pip install -r ./requirements.txt
+          make html
+          touch build/html/.nojekyll
+
+      - name: Deploy
+        uses: peaceiris/actions-gh-pages@v3
+        with:
+          github_token: ${{ secrets.GITHUB_TOKEN }}
+          publish_dir: ./docs/build/html
+          publish_branch: gh-pages
diff --git a/.github/workflows/build-docker-image.yml b/.github/workflows/build-docker-image.yml
new file mode 100644
index 000000000..e5d96dcdf
--- /dev/null
+++ b/.github/workflows/build-docker-image.yml
@@ -0,0 +1,52 @@
+# see also
+# https://docs.github.com/en/actions/publishing-packages/publishing-docker-images#publishing-images-to-github-packages
+name: Build docker image
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: build_docker-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build-docker-image:
+    name: ${{ matrix.image }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        image: ["torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"]
+
+    steps:
+      # refer to https://github.com/actions/checkout
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Rename
+        shell: bash
+        run: |
+          image=${{ matrix.image }}
+          mv -v ./docker/$image.dockerfile ./Dockerfile
+
+      - name: Free space
+        shell: bash
+        run: |
+          df -h
+          rm -rf /opt/hostedtoolcache
+          df -h
+
+      - name: Log in to Docker Hub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKER_USERNAME }}
+          password: ${{ secrets.DOCKER_PASSWORD }}
+
+      - name: Build and push
+        uses: docker/build-push-action@v4
+        with:
+          context: .
+          file: ./Dockerfile
+          push: true
+          tags: k2fsa/icefall:${{ matrix.image }}
diff --git a/.github/workflows/multi-zh-hans.yml b/.github/workflows/multi-zh-hans.yml
new file mode 100644
index 000000000..9081047de
--- /dev/null
+++ b/.github/workflows/multi-zh-hans.yml
@@ -0,0 +1,79 @@
+name: run-multi-zh-hans
+
+on:
+  push:
+    branches:
+      - master
+
+  workflow_dispatch:
+
+concurrency:
+  group: run-multi-zh-hans-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: write
+
+jobs:
+  multi-zh-hans:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: export-model
+        shell: bash
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/multi-zh-hans.sh
+          ls -lh
+
+      - name: upload model to https://github.com/k2-fsa/sherpa-onnx
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.tar.bz2
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: asr-models
diff --git a/.github/workflows/run-aishell-2022-06-20.yml b/.github/workflows/run-aishell-2022-06-20.yml
new file mode 100644
index 000000000..53fcb2c03
--- /dev/null
+++ b/.github/workflows/run-aishell-2022-06-20.yml
@@ -0,0 +1,123 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-aishell-2022-06-20
+# pruned RNN-T + reworked model with random combiner
+# https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_aishell_2022_06_20-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_aishell_2022_06_20:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
+
+      - name: Display decoding results for aishell pruned_transducer_stateless3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/aishell/ASR/
+          tree ./pruned_transducer_stateless3/exp
+
+          cd pruned_transducer_stateless3
+          echo "results for pruned_transducer_stateless3"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+      - name: Upload decoding results for aishell pruned_transducer_stateless3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: aishell-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless3-2022-06-20
+          path: egs/aishell/ASR/pruned_transducer_stateless3/exp/
diff --git a/.github/workflows/run-aishell-zipformer-2023-10-24.yml b/.github/workflows/run-aishell-zipformer-2023-10-24.yml
new file mode 100644
index 000000000..f2fb44a5f
--- /dev/null
+++ b/.github/workflows/run-aishell-zipformer-2023-10-24.yml
@@ -0,0 +1,95 @@
+# Copyright      2023  Zengrui Jin (Xiaomi Corp.)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-aishell-zipformer-2023-10-24
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_aishell_zipformer_2023_10_24-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_aishell_zipformer_2023_10_24:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'zipformer' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-aishell-zipformer-2023-10-24.sh
+
+      
\ No newline at end of file
diff --git a/.github/workflows/run-docker-image.yml b/.github/workflows/run-docker-image.yml
new file mode 100644
index 000000000..d048923b6
--- /dev/null
+++ b/.github/workflows/run-docker-image.yml
@@ -0,0 +1,105 @@
+name: Run docker image
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: run_docker_image-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run-docker-image:
+    name: ${{ matrix.image }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        image: ["torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"]
+    steps:
+      # refer to https://github.com/actions/checkout
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Run the build process with Docker
+        uses: addnab/docker-run-action@v3
+        with:
+            image: k2fsa/icefall:${{ matrix.image }}
+            shell: bash
+            run: |
+              uname -a
+              cat /etc/*release
+
+              find / -name libcuda* 2>/dev/null
+
+              ls -lh /usr/local/
+              ls -lh /usr/local/cuda*
+
+              nvcc --version
+
+              ls -lh /usr/local/cuda-*/compat/*
+
+              # For torch1.9.0-cuda10.2
+              export LD_LIBRARY_PATH=/usr/local/cuda-10.2/compat:$LD_LIBRARY_PATH
+
+              # For torch1.12.1-cuda11.3
+              export LD_LIBRARY_PATH=/usr/local/cuda-11.3/compat:$LD_LIBRARY_PATH
+
+              # For torch2.0.0-cuda11.7
+              export LD_LIBRARY_PATH=/usr/local/cuda-11.7/compat:$LD_LIBRARY_PATH
+
+              # For torch2.1.0-cuda11.8
+              export LD_LIBRARY_PATH=/usr/local/cuda-11.8/compat:$LD_LIBRARY_PATH
+
+              # For torch2.1.0-cuda12.1
+              export LD_LIBRARY_PATH=/usr/local/cuda-12.1/compat:$LD_LIBRARY_PATH
+
+
+              which nvcc
+              cuda_dir=$(dirname $(which nvcc))
+              echo "cuda_dir: $cuda_dir"
+
+              find $cuda_dir -name libcuda.so*
+              echo "--------------------"
+
+              find / -name libcuda.so* 2>/dev/null
+
+              # for torch1.13.0-cuda11.6
+              if [ -e /opt/conda/lib/stubs/libcuda.so ]; then
+                cd /opt/conda/lib/stubs && ln -s libcuda.so libcuda.so.1 && cd -
+                export LD_LIBRARY_PATH=/opt/conda/lib/stubs:$LD_LIBRARY_PATH
+              fi
+
+              find / -name libcuda.so* 2>/dev/null
+              echo "LD_LIBRARY_PATH: $LD_LIBRARY_PATH"
+
+              python3 --version
+              which python3
+
+              python3 -m pip list
+
+              echo "----------torch----------"
+              python3 -m torch.utils.collect_env
+
+              echo "----------k2----------"
+              python3 -c "import k2; print(k2.__file__)"
+              python3 -c "import k2; print(k2.__dev_version__)"
+              python3 -m k2.version
+
+              echo "----------lhotse----------"
+              python3 -c "import lhotse; print(lhotse.__file__)"
+              python3 -c "import lhotse; print(lhotse.__version__)"
+
+              echo "----------kaldifeat----------"
+              python3 -c "import kaldifeat; print(kaldifeat.__file__)"
+              python3 -c "import kaldifeat; print(kaldifeat.__version__)"
+
+              echo "Test yesno recipe"
+
+              cd egs/yesno/ASR
+
+              ./prepare.sh
+
+              ./tdnn/train.py
+
+              ./tdnn/decode.py
diff --git a/.github/workflows/run-gigaspeech-2022-05-13.yml b/.github/workflows/run-gigaspeech-2022-05-13.yml
new file mode 100644
index 000000000..3121520c1
--- /dev/null
+++ b/.github/workflows/run-gigaspeech-2022-05-13.yml
@@ -0,0 +1,126 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-gigaspeech-2022-05-13
+# stateless transducer + k2 pruned rnnt-loss + reworked conformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_gigaspeech_2022_05_13-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_gigaspeech_2022_05_13:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Download GigaSpeech dev/test dataset
+        shell: bash
+        run: |
+          sudo apt-get install -y -q git-lfs
+
+          .github/scripts/download-gigaspeech-dev-test-dataset.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          ln -s ~/tmp/giga-dev-dataset-fbank/data egs/gigaspeech/ASR/
+
+          ls -lh egs/gigaspeech/ASR/data/fbank
+
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh
+
+      - name: Display decoding results for gigaspeech pruned_transducer_stateless2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/gigaspeech/ASR/
+          tree ./pruned_transducer_stateless2/exp
+
+          sudo apt-get -qq install tree
+
+          cd pruned_transducer_stateless2
+          echo "results for pruned_transducer_stateless2"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+
+      - name: Upload decoding results for gigaspeech pruned_transducer_stateless2
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-gigaspeech-pruned_transducer_stateless2-2022-05-12
+          path: egs/gigaspeech/ASR/pruned_transducer_stateless2/exp/
diff --git a/.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml b/.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml
new file mode 100644
index 000000000..87090e310
--- /dev/null
+++ b/.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml
@@ -0,0 +1,140 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-gigaspeech-zipformer-2023-10-17
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+  workflow_dispatch:
+
+concurrency:
+  group: run_gigaspeech_2023_10_17_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_gigaspeech_2023_10_17_zipformer:
+    if: github.event.label.name == 'zipformer' ||github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p egs/gigaspeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/gigaspeech/ASR/data/fbank
+          ls -lh egs/gigaspeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-gigaspeech-zipformer-2023-10-17.sh
+
+      - name: upload model to https://github.com/k2-fsa/sherpa-onnx
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.tar.bz2
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: asr-models
+
+      - name: Display decoding results for gigaspeech zipformer
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/gigaspeech/ASR/
+          tree ./zipformer/exp
+
+          cd zipformer
+          echo "results for zipformer"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for gigaspeech zipformer
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11
+          path: egs/gigaspeech/ASR/zipformer/exp/
diff --git a/.github/workflows/run-librispeech-2022-03-12.yml b/.github/workflows/run-librispeech-2022-03-12.yml
new file mode 100644
index 000000000..f092e3c80
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-03-12.yml
@@ -0,0 +1,159 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-03-12
+# stateless transducer + k2 pruned rnnt-loss
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_03_12-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_03_12:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh
+
+      - name: Display decoding results for pruned_transducer_stateless
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless/exp
+
+          cd pruned_transducer_stateless
+          echo "results for pruned_transducer_stateless"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for pruned_transducer_stateless
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless-2022-03-12
+          path: egs/librispeech/ASR/pruned_transducer_stateless/exp/
diff --git a/.github/workflows/run-librispeech-2022-04-29.yml b/.github/workflows/run-librispeech-2022-04-29.yml
new file mode 100644
index 000000000..f8f4d9977
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-04-29.yml
@@ -0,0 +1,185 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-04-29
+# stateless pruned transducer (reworked model) + giga speech
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_04_29-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_04_29:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh
+
+      - name: Display decoding results for pruned_transducer_stateless2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree pruned_transducer_stateless2/exp
+          cd pruned_transducer_stateless2/exp
+          echo "===greedy search==="
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Display decoding results for pruned_transducer_stateless3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree pruned_transducer_stateless3/exp
+          cd pruned_transducer_stateless3/exp
+          echo "===greedy search==="
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for pruned_transducer_stateless2
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless2-2022-04-29
+          path: egs/librispeech/ASR/pruned_transducer_stateless2/exp/
+
+      - name: Upload decoding results for pruned_transducer_stateless3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless3-2022-04-29
+          path: egs/librispeech/ASR/pruned_transducer_stateless3/exp/
diff --git a/.github/workflows/run-librispeech-2022-05-13.yml b/.github/workflows/run-librispeech-2022-05-13.yml
new file mode 100644
index 000000000..dc20185da
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-05-13.yml
@@ -0,0 +1,159 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-05-13
+# stateless transducer + k2 pruned rnnt-loss + deeper model
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_05_13-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_05_13:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless5
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless5/exp
+
+          cd pruned_transducer_stateless5
+          echo "results for pruned_transducer_stateless5"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless5
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless5-2022-05-13
+          path: egs/librispeech/ASR/pruned_transducer_stateless5/exp/
diff --git a/.github/workflows/run-librispeech-2022-11-11-stateless7.yml b/.github/workflows/run-librispeech-2022-11-11-stateless7.yml
new file mode 100644
index 000000000..7e378c9a1
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-11-11-stateless7.yml
@@ -0,0 +1,159 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-11-11-stateless7
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_11_11_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_11_11_zipformer:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7/exp
+
+          cd pruned_transducer_stateless7
+          echo "results for pruned_transducer_stateless7"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-2022-11-11
+          path: egs/librispeech/ASR/pruned_transducer_stateless7/exp/
diff --git a/.github/workflows/run-librispeech-2022-11-14-stateless8.yml b/.github/workflows/run-librispeech-2022-11-14-stateless8.yml
new file mode 100644
index 000000000..a2c1a0ad6
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-11-14-stateless8.yml
@@ -0,0 +1,159 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-11-14-stateless8
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_11_14_zipformer_stateless8-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_11_14_zipformer_stateless8:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless8
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless8/exp
+
+          cd pruned_transducer_stateless8
+          echo "results for pruned_transducer_stateless8"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless8
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless8-2022-11-14
+          path: egs/librispeech/ASR/pruned_transducer_stateless8/exp/
diff --git a/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml b/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml
new file mode 100644
index 000000000..500ab1736
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml
@@ -0,0 +1,163 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-12-01-stateless7-ctc
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+jobs:
+  run_librispeech_2022_11_11_zipformer:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7_ctc
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7_ctc/exp
+
+          cd pruned_transducer_stateless7_ctc
+          echo "results for pruned_transducer_stateless7_ctc"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===ctc decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7_ctc
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-ctc-2022-12-01
+          path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc/exp/
diff --git a/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml b/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml
new file mode 100644
index 000000000..1a7f9f594
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml
@@ -0,0 +1,167 @@
+# Copyright      2022  Zengwei Yao
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-12-08-zipformer-mmi
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_12_08_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_12_08_zipformer:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh
+
+      - name: Display decoding results for librispeech zipformer-mmi
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer-mmi/exp
+
+          cd zipformer-mmi
+          echo "results for zipformer-mmi"
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest==="
+          find exp/nbest -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-LG==="
+          find exp/nbest-rescoring-LG -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-LG -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-3-gram==="
+          find exp/nbest-rescoring-3-gram -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-3-gram -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-4-gram==="
+          find exp/nbest-rescoring-4-gram -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-4-gram -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech zipformer-mmi
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer_mmi-2022-12-08
+          path: egs/librispeech/ASR/zipformer_mmi/exp/
diff --git a/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml b/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml
new file mode 100644
index 000000000..68014e20c
--- /dev/null
+++ b/.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml
@@ -0,0 +1,172 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-12-29-stateless7-streaming
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_12_29_zipformer_streaming-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_12_29_zipformer_streaming:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event.label.name == 'streaming-zipformer' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-streaming-2022-12-29.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7_streaming
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7_streaming/exp
+
+          cd pruned_transducer_stateless7_streaming
+          echo "results for pruned_transducer_stateless7_streaming"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===streaming greedy search==="
+          find exp/streaming/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===streaming fast_beam_search==="
+          find exp/streaming/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===streaming modified beam search==="
+          find exp/streaming/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7_streaming
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-streaming-2022-12-29
+          path: egs/librispeech/ASR/pruned_transducer_stateless7_streaming/exp/
diff --git a/.github/workflows/run-librispeech-2023-01-29-stateless7-ctc-bs.yml b/.github/workflows/run-librispeech-2023-01-29-stateless7-ctc-bs.yml
new file mode 100644
index 000000000..821abc25d
--- /dev/null
+++ b/.github/workflows/run-librispeech-2023-01-29-stateless7-ctc-bs.yml
@@ -0,0 +1,163 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2023-01-29-stateless7-ctc-bs
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+jobs:
+  run_librispeech_2023_01_29_zipformer_ctc_bs:
+    if: github.event.label.name == 'run-decode' || github.event.label.name == 'blank-skip' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2023-01-29.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7_ctc_bs
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7_ctc_bs/exp
+
+          cd pruned_transducer_stateless7_ctc_bs
+          echo "results for pruned_transducer_stateless7_ctc_bs"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===ctc decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7_ctc_bs
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless7-ctc-bs-2023-01-29
+          path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/exp/
diff --git a/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml b/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml
new file mode 100644
index 000000000..905515dc4
--- /dev/null
+++ b/.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml
@@ -0,0 +1,155 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-conformer-ctc3-2022-11-28
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_11_28_conformer_ctc3-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_11_28_conformer_ctc3:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh
+
+      - name: Display decoding results for librispeech conformer_ctc3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./conformer_ctc3/exp
+
+          cd conformer_ctc3
+          echo "results for conformer_ctc3"
+          echo "===ctc-decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech conformer_ctc3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-conformer_ctc3-2022-11-28
+          path: egs/librispeech/ASR/conformer_ctc3/exp/
diff --git a/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml b/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml
new file mode 100644
index 000000000..501fae38c
--- /dev/null
+++ b/.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml
@@ -0,0 +1,163 @@
+name: run-librispeech-lstm-transducer2-2022-09-03
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_lstm_transducer_stateless2_2022_09_03-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_lstm_transducer_stateless2_2022_09_03:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'LODR' || github.event.label.name == 'shallow-fusion' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh
+
+      - name: Display decoding results for lstm_transducer_stateless2
+        if: github.event_name == 'schedule'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree lstm_transducer_stateless2/exp
+          cd lstm_transducer_stateless2/exp
+          echo "===greedy search==="
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Display decoding results for lstm_transducer_stateless2
+        if: github.event.label.name == 'shallow-fusion'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree lstm_transducer_stateless2/exp
+          cd lstm_transducer_stateless2/exp
+          echo "===modified_beam_search_lm_shallow_fusion==="
+          echo "===Using RNNLM==="
+          find modified_beam_search_lm_shallow_fusion  -name "log-*rnn*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search_lm_shallow_fusion  -name "log-*rnn*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Display decoding results for lstm_transducer_stateless2
+        if: github.event.label.name == 'LODR'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree lstm_transducer_stateless2/exp
+          cd lstm_transducer_stateless2/exp
+          echo "===modified_beam_search_rnnlm_LODR==="
+          find modified_beam_search_LODR  -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search_LODR  -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for lstm_transducer_stateless2
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'shallow-fusion' || github.event.label.name == 'LODR'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-lstm_transducer_stateless2-2022-09-03
+          path: egs/librispeech/ASR/lstm_transducer_stateless2/exp/
diff --git a/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml b/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml
new file mode 100644
index 000000000..3fb0920bc
--- /dev/null
+++ b/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml
@@ -0,0 +1,157 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-pruned-transducer-stateless3-2022-05-13
+# stateless pruned transducer (reworked model) + giga speech
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_pruned_transducer_stateless3_2022_05_13-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_pruned_transducer_stateless3_2022_05_13:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless3-2022-05-13.sh
+
+      - name: Display decoding results for pruned_transducer_stateless3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree pruned_transducer_stateless3/exp
+          cd pruned_transducer_stateless3/exp
+          echo "===greedy search==="
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for pruned_transducer_stateless3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless3-2022-04-29
+          path: egs/librispeech/ASR/pruned_transducer_stateless3/exp/
diff --git a/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml b/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml
new file mode 100644
index 000000000..67a6f6fc4
--- /dev/null
+++ b/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml
@@ -0,0 +1,159 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-streaming-2022-06-26
+# streaming conformer stateless transducer2
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_streaming_2022_06_26-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_streaming_2022_06_26:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-streaming-pruned-transducer-stateless2-2022-06-26.sh
+
+      - name: Display decoding results
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless2/exp
+
+          cd pruned_transducer_stateless2
+          echo "results for pruned_transducer_stateless2"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified_beam_search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for pruned_transducer_stateless2
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-pruned_transducer_stateless2-2022-06-26
+          path: egs/librispeech/ASR/pruned_transducer_stateless2/exp/
diff --git a/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml b/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml
new file mode 100644
index 000000000..5145fb43c
--- /dev/null
+++ b/.github/workflows/run-librispeech-streaming-zipformer-2023-05-18.yml
@@ -0,0 +1,174 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-streaming-zipformer-2023-05-18
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2023_05_18_streaming_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2023_05_18_streaming_zipformer:
+    if: github.event.label.name == 'zipformer' ||github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-streaming-zipformer-2023-05-18.sh
+
+      - name: Display decoding results for librispeech zipformer
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer/exp
+
+          cd zipformer
+
+          echo "results for zipformer, simulated streaming decoding"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "results for zipformer, chunk-wise streaming decoding"
+          echo "===greedy search==="
+          find exp/streaming/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/streaming/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/streaming/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+
+      - name: Upload decoding results for librispeech zipformer
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11
+          path: egs/librispeech/ASR/zipformer/exp/
diff --git a/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml b/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml
new file mode 100644
index 000000000..35ca08a31
--- /dev/null
+++ b/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml
@@ -0,0 +1,159 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-2022-04-19
+# stateless transducer + torchaudio rnn-t loss
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_04_19-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_04_19:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh
+
+      - name: Display decoding results
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./transducer_stateless2/exp
+
+          cd transducer_stateless2
+          echo "results for transducer_stateless2"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified_beam_search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for transducer_stateless2
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless2-2022-04-19
+          path: egs/librispeech/ASR/transducer_stateless2/exp/
diff --git a/.github/workflows/run-librispeech-zipformer-2023-05-18.yml b/.github/workflows/run-librispeech-zipformer-2023-05-18.yml
new file mode 100644
index 000000000..e9d235ad1
--- /dev/null
+++ b/.github/workflows/run-librispeech-zipformer-2023-05-18.yml
@@ -0,0 +1,159 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-zipformer-2023-05-18
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2023_05_18_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2023_05_18_zipformer:
+    if: github.event.label.name == 'zipformer' ||github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-zipformer-2023-05-18.sh
+
+      - name: Display decoding results for librispeech zipformer
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer/exp
+
+          cd zipformer
+          echo "results for zipformer"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech zipformer
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11
+          path: egs/librispeech/ASR/zipformer/exp/
diff --git a/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml b/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml
new file mode 100644
index 000000000..48f0b1532
--- /dev/null
+++ b/.github/workflows/run-librispeech-zipformer-ctc-2023-06-14.yml
@@ -0,0 +1,155 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-librispeech-zipformer-ctc-2023-06-14
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2023_06_14_zipformer-ctc-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2023_06_14_zipformer_ctc:
+    if: github.event.label.name == 'zipformer' ||github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-zipformer-ctc-2023-06-14.sh
+
+      - name: Display decoding results for librispeech zipformer
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer/exp
+
+          cd zipformer
+          echo "results for zipformer"
+          echo "===ctc-decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech zipformer
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11
+          path: egs/librispeech/ASR/zipformer/exp/
diff --git a/.github/workflows/run-multi-corpora-zipformer.yml b/.github/workflows/run-multi-corpora-zipformer.yml
new file mode 100644
index 000000000..38f7eb908
--- /dev/null
+++ b/.github/workflows/run-multi-corpora-zipformer.yml
@@ -0,0 +1,84 @@
+# Copyright      2023   Xiaomi Corp.    (author: Zengrui Jin)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-multi-corpora-zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run_multi-corpora_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_multi-corpora_zipformer:
+    if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'multi-zh_hans' || github.event.label.name == 'zipformer' || github.event.label.name == 'multi-corpora'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-multi-corpora-zipformer.sh
diff --git a/.github/workflows/run-pretrained-ctc.yml b/.github/workflows/run-pretrained-ctc.yml
new file mode 100644
index 000000000..074a63dfc
--- /dev/null
+++ b/.github/workflows/run-pretrained-ctc.yml
@@ -0,0 +1,87 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-ctc
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  workflow_dispatch:
+    inputs:
+      test-run:
+        description: 'Test (y/n)?'
+        required: true
+        default: 'y'
+
+concurrency:
+  group: run_pre_trained_ctc-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_ctc:
+    if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event.inputs.test-run == 'y' || github.event.label.name == 'ctc'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+          .github/scripts/run-pre-trained-ctc.sh
diff --git a/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml b/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml
new file mode 100644
index 000000000..f8caee8e5
--- /dev/null
+++ b/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml
@@ -0,0 +1,158 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-trandsucer-stateless-multi-datasets-librispeech-100h
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_pre_trained_transducer_stateless_multi_datasets_librispeech_100h-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_transducer_stateless_multi_datasets_librispeech_100h:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh
+
+      - name: Display decoding results for transducer_stateless_multi_datasets
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./transducer_stateless_multi_datasets/exp
+
+          cd transducer_stateless_multi_datasets
+          echo "results for transducer_stateless_multi_datasets"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for transducer_stateless_multi_datasets
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless_multi_datasets-100h-2022-02-21
+          path: egs/librispeech/ASR/transducer_stateless_multi_datasets/exp/
diff --git a/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml b/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml
new file mode 100644
index 000000000..7c3910eb8
--- /dev/null
+++ b/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml
@@ -0,0 +1,158 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-trandsucer-stateless-multi-datasets-librispeech-960h
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_pre_trained_transducer_stateless_multi_datasets_librispeech_960h-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_transducer_stateless_multi_datasets_librispeech_960h:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh
+
+      - name: Display decoding results for transducer_stateless_multi_datasets
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./transducer_stateless_multi_datasets/exp
+
+          cd transducer_stateless_multi_datasets
+          echo "results for transducer_stateless_multi_datasets"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for transducer_stateless_multi_datasets
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless_multi_datasets-100h-2022-03-01
+          path: egs/librispeech/ASR/transducer_stateless_multi_datasets/exp/
diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml
new file mode 100644
index 000000000..ce6d6f92d
--- /dev/null
+++ b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml
@@ -0,0 +1,80 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-trandsucer-stateless-modified-2-aishell
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run_pre_trained_transducer_stateless_modified_2_aishell-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_transducer_stateless_modified_2_aishell:
+    if: github.event.label.name == 'ready' || github.event_name == 'push'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+          .github/scripts/run-pre-trained-transducer-stateless-modified-2-aishell.sh
diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml
new file mode 100644
index 000000000..f0cebd94a
--- /dev/null
+++ b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml
@@ -0,0 +1,80 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-trandsucer-stateless-modified-aishell
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run_pre_trained_transducer_stateless_modified_aishell-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_transducer_stateless_modified_aishell:
+    if: github.event.label.name == 'ready' || github.event_name == 'push'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+          .github/scripts/run-pre-trained-transducer-stateless-modified-aishell.sh
diff --git a/.github/workflows/run-pretrained-transducer-stateless.yml b/.github/workflows/run-pretrained-transducer-stateless.yml
new file mode 100644
index 000000000..1b69b97bf
--- /dev/null
+++ b/.github/workflows/run-pretrained-transducer-stateless.yml
@@ -0,0 +1,158 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-transducer-stateless
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_pre_trained_transducer_stateless-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_transducer_stateless:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-pre-trained-transducer-stateless.sh
+
+      - name: Display decoding results for transducer_stateless
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./transducer_stateless/exp
+
+          cd transducer_stateless
+          echo "results for transducer_stateless"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for transducer_stateless
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-transducer_stateless-2022-02-07
+          path: egs/librispeech/ASR/transducer_stateless/exp/
diff --git a/.github/workflows/run-pretrained-transducer.yml b/.github/workflows/run-pretrained-transducer.yml
new file mode 100644
index 000000000..91d87f1c9
--- /dev/null
+++ b/.github/workflows/run-pretrained-transducer.yml
@@ -0,0 +1,80 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-pre-trained-transducer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run_pre_trained_transducer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_pre_trained_transducer:
+    if: github.event.label.name == 'ready' || github.event_name == 'push'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          make -j2 _kaldifeat
+
+      - name: Inference with pre-trained model
+        shell: bash
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+          .github/scripts/run-pre-trained-transducer.sh
diff --git a/.github/workflows/run-ptb-rnn-lm.yml b/.github/workflows/run-ptb-rnn-lm.yml
new file mode 100644
index 000000000..f8d9c02c5
--- /dev/null
+++ b/.github/workflows/run-ptb-rnn-lm.yml
@@ -0,0 +1,71 @@
+name: run-ptb-rnn-lm-training
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_ptb_rnn_lm_training-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_ptb_rnn_lm_training:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'rnnlm' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.8"]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Prepare data
+        shell: bash
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          cd egs/ptb/LM
+          ./prepare.sh
+
+      - name: Run training
+        shell: bash
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          cd egs/ptb/LM
+          ./train-rnn-lm.sh --world-size 1 --num-epochs 5 --use-epoch 4 --use-avg 2
+
+      - name: Upload pretrained models
+        uses: actions/upload-artifact@v2
+        if: github.event.label.name == 'ready' || github.event.label.name == 'rnnlm' || github.event_name == 'push' || github.event_name == 'schedule'
+        with:
+          name: python-${{ matrix.python-version }}-ubuntu-rnn-lm-ptb
+          path: egs/ptb/LM/my-rnnlm-exp/
diff --git a/.github/workflows/run-swbd-conformer-ctc.yml b/.github/workflows/run-swbd-conformer-ctc.yml
new file mode 100644
index 000000000..842691d38
--- /dev/null
+++ b/.github/workflows/run-swbd-conformer-ctc.yml
@@ -0,0 +1,84 @@
+# Copyright      2023   Xiaomi Corp.    (author: Zengrui Jin)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-swbd-conformer_ctc
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run-swbd-conformer_ctc-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run-swbd-conformer_ctc:
+    if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'swbd'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-swbd-conformer-ctc-2023-08-26.sh
diff --git a/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml b/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml
new file mode 100644
index 000000000..319a5558a
--- /dev/null
+++ b/.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml
@@ -0,0 +1,84 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-wenetspeech-pruned-transducer-stateless2
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run_wenetspeech_pruned_transducer_stateless2-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_wenetspeech_pruned_transducer_stateless2:
+    if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'wenetspeech'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-wenetspeech-pruned-transducer-stateless2.sh
diff --git a/.github/workflows/run-yesno-recipe.yml b/.github/workflows/run-yesno-recipe.yml
new file mode 100644
index 000000000..9ac848535
--- /dev/null
+++ b/.github/workflows/run-yesno-recipe.yml
@@ -0,0 +1,185 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-yesno-recipe
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
+
+concurrency:
+  group: run-yesno-recipe-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run-yesno-recipe:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        # os: [ubuntu-latest, macos-10.15]
+        # TODO: enable macOS for CPU testing
+        os: [ubuntu-latest]
+        python-version: [3.8]
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install libnsdfile and libsox
+        if: startsWith(matrix.os, 'ubuntu')
+        run: |
+          sudo apt update
+          sudo apt install -q -y libsndfile1-dev libsndfile1 ffmpeg
+          sudo apt install -q -y --fix-missing sox libsox-dev libsox-fmt-all
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+          pip install --no-deps --force-reinstall k2==1.24.4.dev20231021+cpu.torch1.13.1 -f https://k2-fsa.github.io/k2/cpu.html
+          pip install kaldifeat==1.25.1.dev20231022+cpu.torch1.13.1 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+
+      - name: Run yesno recipe
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          ./prepare.sh
+          python3 ./tdnn/train.py
+          python3 ./tdnn/decode.py
+
+      - name: Test exporting to pretrained.pt
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export.py --epoch 14 --avg 2
+
+          python3 ./tdnn/pretrained.py \
+            --checkpoint ./tdnn/exp/pretrained.pt \
+            --HLG ./data/lang_phone/HLG.pt \
+            --words-file ./data/lang_phone/words.txt \
+            download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+            download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+      - name: Test exporting to torchscript
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export.py --epoch 14 --avg 2 --jit 1
+
+          python3 ./tdnn/jit_pretrained.py \
+            --nn-model ./tdnn/exp/cpu_jit.pt \
+            --HLG ./data/lang_phone/HLG.pt \
+            --words-file ./data/lang_phone/words.txt \
+            download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+            download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+      - name: Test exporting to onnx
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export_onnx.py --epoch 14 --avg 2
+
+          echo "Test float32 model"
+          python3 ./tdnn/onnx_pretrained.py \
+            --nn-model ./tdnn/exp/model-epoch-14-avg-2.onnx \
+            --HLG ./data/lang_phone/HLG.pt \
+            --words-file ./data/lang_phone/words.txt \
+            download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+            download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+
+          echo "Test int8 model"
+          python3 ./tdnn/onnx_pretrained.py \
+            --nn-model ./tdnn/exp/model-epoch-14-avg-2.int8.onnx \
+            --HLG ./data/lang_phone/HLG.pt \
+            --words-file ./data/lang_phone/words.txt \
+            download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+            download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+      - name: Test decoding with H
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export.py --epoch 14 --avg 2 --jit 1
+
+          python3 ./tdnn/jit_pretrained_decode_with_H.py \
+              --nn-model ./tdnn/exp/cpu_jit.pt \
+              --H ./data/lang_phone/H.fst \
+              --tokens ./data/lang_phone/tokens.txt \
+              ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+              ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+              ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+      - name: Test decoding with HL
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export.py --epoch 14 --avg 2 --jit 1
+
+          python3 ./tdnn/jit_pretrained_decode_with_HL.py \
+              --nn-model ./tdnn/exp/cpu_jit.pt \
+              --HL ./data/lang_phone/HL.fst \
+              --words ./data/lang_phone/words.txt \
+              ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+              ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+              ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+      - name: Show generated files
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          cd egs/yesno/ASR
+          ls -lh tdnn/exp
+          ls -lh data/lang_phone
diff --git a/.github/workflows/style_check.yml b/.github/workflows/style_check.yml
new file mode 100644
index 000000000..fc1dcbfd4
--- /dev/null
+++ b/.github/workflows/style_check.yml
@@ -0,0 +1,69 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: style_check
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
+
+concurrency:
+  group: style_check-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  style_check:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v1
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m pip install --upgrade pip black==22.3.0 flake8==5.0.4 click==8.1.0
+          # Click issue fixed in https://github.com/psf/black/pull/2966
+
+      - name: Run flake8
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          # stop the build if there are Python syntax errors or undefined names
+          flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
+          # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+          flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 \
+            --statistics --extend-ignore=E203,E266,E501,F401,E402,F403,F841,W503
+
+      - name: Run black
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          black --check --diff .
diff --git a/.github/workflows/test-ncnn-export.yml b/.github/workflows/test-ncnn-export.yml
new file mode 100644
index 000000000..5709f8ebb
--- /dev/null
+++ b/.github/workflows/test-ncnn-export.yml
@@ -0,0 +1,75 @@
+name: test-ncnn-export
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: test_ncnn_export-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test_ncnn_export:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'ncnn' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Test ncnn export
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/test-ncnn-export.sh
diff --git a/.github/workflows/test-onnx-export.yml b/.github/workflows/test-onnx-export.yml
new file mode 100644
index 000000000..c05cde3ba
--- /dev/null
+++ b/.github/workflows/test-onnx-export.yml
@@ -0,0 +1,75 @@
+name: test-onnx-export
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: test_onnx_export-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test_onnx_export:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'onnx' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Test ONNX export
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/test-onnx-export.sh
diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
new file mode 100644
index 000000000..363556bb7
--- /dev/null
+++ b/.github/workflows/test.yml
@@ -0,0 +1,131 @@
+# Copyright      2021  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: test
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
+
+concurrency:
+  group: test-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.8"]
+        torch: ["1.13.0"]
+        torchaudio: ["0.13.0"]
+        k2-version: ["1.24.3.dev20230719"]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v1
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install libnsdfile and libsox
+        if: startsWith(matrix.os, 'ubuntu')
+        run: |
+          sudo apt update
+          sudo apt install -q -y libsndfile1-dev libsndfile1 ffmpeg
+          sudo apt install -q -y --fix-missing libsox-dev libsox-fmt-all
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m pip install --upgrade pip pytest
+          # numpy 1.20.x does not support python 3.6
+          pip install numpy==1.19
+          pip install torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
+          pip install torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
+
+          pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.github.io/k2/cpu.html
+          pip install git+https://github.com/lhotse-speech/lhotse
+          # icefall requirements
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+          pip install kaldifst
+          pip install onnxruntime matplotlib
+          pip install -r requirements.txt
+
+      - name: Install graphviz
+        if: startsWith(matrix.os, 'ubuntu')
+        shell: bash
+        run: |
+          python3 -m pip install -qq graphviz
+          sudo apt-get -qq install graphviz
+
+      - name: Run tests
+        if: startsWith(matrix.os, 'ubuntu')
+        run: |
+          ls -lh
+          export PYTHONPATH=$PWD:$PWD/lhotse:$PYTHONPATH
+          echo $PYTHONPATH
+          pytest -v -s ./test
+          # runt tests for conformer ctc
+          cd egs/librispeech/ASR/conformer_ctc
+          pytest -v -s
+
+          cd ../pruned_transducer_stateless
+          pytest -v -s
+
+          cd ../pruned_transducer_stateless2
+          pytest -v -s
+
+          cd ../pruned_transducer_stateless3
+          pytest -v -s
+
+          cd ../pruned_transducer_stateless4
+          pytest -v -s
+
+          echo $PYTHONPATH
+          cd ../pruned_transducer_stateless7
+          pytest -v -s
+
+          cd ../transducer_stateless
+          pytest -v -s
+
+          # cd ../transducer
+          # pytest -v -s
+
+          cd ../transducer_stateless2
+          pytest -v -s
+
+          cd ../transducer_lstm
+          pytest -v -s
+
+          cd ../zipformer
+          pytest -v -s
+
+      - uses: actions/upload-artifact@v2
+        with:
+          path: egs/librispeech/ASR/zipformer/swoosh.pdf
+          name: swoosh.pdf
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 000000000..fa18ca83c
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,38 @@
+icefall.egg-info/
+data
+__pycache__
+path.sh
+exp
+exp*/
+*.pt
+download
+dask-worker-space
+log
+*.bak
+*-bak
+*bak.py
+
+# Ignore Mac system files
+.DS_store
+
+# Ignore node_modules folder
+node_modules
+
+# ignore .nfs
+
+.nfs*
+
+# Ignore all text files
+*.txt
+
+# Ignore files related to API keys
+.env
+
+# Ignore SASS config files
+.sass-cache
+
+*.param
+*.bin
+.DS_Store
+*.fst
+*.arpa
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
new file mode 100644
index 000000000..1bb38f6ba
--- /dev/null
+++ b/.pre-commit-config.yaml
@@ -0,0 +1,40 @@
+repos:
+  - repo: https://github.com/psf/black
+    rev: 22.3.0
+    hooks:
+      - id: black
+        args: ["--line-length=88"]
+        additional_dependencies: ['click==8.1.0']
+        exclude: icefall\/__init__\.py
+
+  - repo: https://github.com/PyCQA/flake8
+    rev: 5.0.4
+    hooks:
+      - id: flake8
+        args: ["--max-line-length=88", "--extend-ignore=E203,E266,E501,F401,E402,F403,F841,W503"]
+
+      # What are we ignoring here?
+      # E203: whitespace before ':'
+      # E266: too many leading '#' for block comment
+      # E501: line too long
+      # F401: module imported but unused
+      # E402: module level import not at top of file
+      # F403: 'from module import *' used; unable to detect undefined names
+      # F841: local variable is assigned to but never used
+      # W503: line break before binary operator
+      # In addition, the default ignore list is:
+      # E121,E123,E126,E226,E24,E704,W503,W504
+
+  - repo: https://github.com/pycqa/isort
+    rev: 5.11.5
+    hooks:
+      - id: isort
+        args: ["--profile=black"]
+
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.2.0
+    hooks:
+      - id: check-executables-have-shebangs
+      - id: end-of-file-fixer
+      - id: mixed-line-ending
+      - id: trailing-whitespace
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 000000000..d64569567
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/README.md b/README.md
new file mode 100644
index 000000000..15e9e17e6
--- /dev/null
+++ b/README.md
@@ -0,0 +1,413 @@
+<div align="center">
+<img src="https://raw.githubusercontent.com/k2-fsa/icefall/master/docs/source/_static/logo.png" width=168>
+</div>
+
+## Introduction
+
+icefall contains ASR recipes for various datasets
+using <https://github.com/k2-fsa/k2>.
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy models
+trained with icefall.
+
+You can try pre-trained models from within your browser without the need
+to download or install anything by visiting <https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>
+See <https://k2-fsa.github.io/icefall/huggingface/spaces.html> for more details.
+
+## Installation
+
+Please refer to <https://icefall.readthedocs.io/en/latest/installation/index.html>
+for installation.
+
+## Recipes
+
+Please refer to <https://icefall.readthedocs.io/en/latest/recipes/index.html>
+for more information.
+
+We provide the following recipes:
+
+  - [yesno][yesno]
+  - [LibriSpeech][librispeech]
+  - [GigaSpeech][gigaspeech]
+  - [AMI][ami]
+  - [Aishell][aishell]
+  - [Aishell2][aishell2]
+  - [Aishell4][aishell4]
+  - [TIMIT][timit]
+  - [TED-LIUM3][tedlium3]
+  - [Aidatatang_200zh][aidatatang_200zh]
+  - [WenetSpeech][wenetspeech]
+  - [Alimeeting][alimeeting]
+  - [Switchboard][swbd]
+  - [TAL_CSASR][tal_csasr]
+
+### yesno
+
+This is the simplest ASR recipe in `icefall` and can be run on CPU.
+Training takes less than 30 seconds and gives you the following WER:
+
+```
+[test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
+```
+We provide a Colab notebook for this recipe: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing)
+
+
+### LibriSpeech
+
+Please see <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>
+for the **latest** results.
+
+We provide 5 models for this recipe:
+
+- [conformer CTC model][LibriSpeech_conformer_ctc]
+- [TDNN LSTM CTC model][LibriSpeech_tdnn_lstm_ctc]
+- [Transducer: Conformer encoder + LSTM decoder][LibriSpeech_transducer]
+- [Transducer: Conformer encoder + Embedding decoder][LibriSpeech_transducer_stateless]
+- [Transducer: Zipformer encoder + Embedding decoder][LibriSpeech_zipformer]
+
+#### Conformer CTC Model
+
+The best WER we currently have is:
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 2.42       | 5.73       |
+
+
+We provide a Colab notebook to run a pre-trained conformer CTC model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1huyupXAcHsUrKaWfI83iMEJ6J0Nh0213?usp=sharing)
+
+#### TDNN LSTM CTC Model
+
+The WER for this model is:
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 6.59       | 17.69      |
+
+We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1-iSfQMp2So-We_Uu49N4AAcMInB72u9z?usp=sharing)
+
+
+#### Transducer: Conformer encoder + LSTM decoder
+
+Using Conformer as encoder and LSTM as decoder.
+
+The best WER with greedy search is:
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 3.07       | 7.51       |
+
+We provide a Colab notebook to run a pre-trained RNN-T conformer model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1_u6yK9jDkPwG_NLrZMN2XK7Aeq4suMO2?usp=sharing)
+
+#### Transducer: Conformer encoder + Embedding decoder
+
+Using Conformer as encoder. The decoder consists of 1 embedding layer
+and 1 convolutional layer.
+
+The best WER using modified beam search with beam size 4 is:
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 2.56       | 6.27       |
+
+Note: No auxiliary losses are used in the training and no LMs are used
+in the decoding.
+
+We provide a Colab notebook to run a pre-trained transducer conformer + stateless decoder model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CO1bXJ-2khDckZIW8zjOPHGSKLHpTDlp?usp=sharing)
+
+
+#### k2 pruned RNN-T
+
+| Encoder         | Params | test-clean | test-other | epochs  | devices    |
+|-----------------|--------|------------|------------|---------|------------|
+| zipformer       | 65.5M  | 2.21       | 4.79       | 50      | 4 32G-V100 |
+| zipformer-small | 23.2M  | 2.42       | 5.73       | 50      | 2 32G-V100 |
+| zipformer-large | 148.4M | 2.06       | 4.63       | 50      | 4 32G-V100 |
+| zipformer-large | 148.4M | 2.00       | 4.38       | 174     | 8 80G-A100 |
+
+Note: No auxiliary losses are used in the training and no LMs are used
+in the decoding.
+
+#### k2 pruned RNN-T + GigaSpeech
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 1.78       | 4.08       |
+
+Note: No auxiliary losses are used in the training and no LMs are used
+in the decoding.
+
+#### k2 pruned RNN-T + GigaSpeech + CommonVoice
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 1.90       | 3.98       |
+
+Note: No auxiliary losses are used in the training and no LMs are used
+in the decoding.
+
+
+### GigaSpeech
+
+We provide three models for this recipe:
+
+- [Conformer CTC model][GigaSpeech_conformer_ctc]
+- [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][GigaSpeech_pruned_transducer_stateless2].
+- [Transducer: Zipformer encoder + Embedding decoder][GigaSpeech_zipformer]
+
+#### Conformer CTC
+
+|     |  Dev  | Test  |
+|-----|-------|-------|
+| WER | 10.47 | 10.58 |
+
+#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss
+
+|                      |  Dev  | Test  |
+|----------------------|-------|-------|
+|    greedy search     | 10.51 | 10.73 |
+|   fast beam search   | 10.50 | 10.69 |
+| modified beam search | 10.40 | 10.51 |
+
+#### Transducer: Zipformer encoder + Embedding decoder
+
+|                      |  Dev  | Test  |
+|----------------------|-------|-------|
+|    greedy search     | 10.31 | 10.50 |
+|   fast beam search   | 10.26 | 10.48 |
+| modified beam search | 10.25 | 10.38 |
+
+
+### Aishell
+
+We provide three models for this recipe: [conformer CTC model][Aishell_conformer_ctc],
+[TDNN LSTM CTC model][Aishell_tdnn_lstm_ctc], and [Transducer Stateless Model][Aishell_pruned_transducer_stateless7],
+
+#### Conformer CTC Model
+
+The best CER we currently have is:
+
+|     | test |
+|-----|------|
+| CER | 4.26 |
+
+#### TDNN LSTM CTC Model
+
+The CER for this model is:
+
+|     | test  |
+|-----|-------|
+| CER | 10.16 |
+
+We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model:  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1jbyzYq3ytm6j2nlEt-diQm-6QVWyDDEa?usp=sharing)
+
+#### Transducer Stateless Model
+
+The best CER we currently have is:
+
+|     | test |
+|-----|------|
+| CER | 4.38 |
+
+We provide a Colab notebook to run a pre-trained TransducerStateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/14XaT2MhnBkK-3_RqqWq3K90Xlbin-GZC?usp=sharing)
+
+
+### Aishell2
+
+We provide one model for this recipe: [Transducer Stateless Model][Aishell2_pruned_transducer_stateless5].
+
+#### Transducer Stateless Model
+
+The best WER we currently have is:
+
+|     |   dev-ios  |  test-ios  |
+|-----|------------|------------|
+| WER |    5.32    |    5.56    |
+
+
+### Aishell4
+
+We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][Aishell4_pruned_transducer_stateless5].
+
+#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss (trained with all subsets)
+
+The best CER we currently have is:
+
+|     |   test     |
+|-----|------------|
+| CER |   29.08    |
+
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1z3lkURVv9M7uTiIgf3Np9IntMHEknaks?usp=sharing)
+
+
+### TIMIT
+
+We provide two models for this recipe: [TDNN LSTM CTC model][TIMIT_tdnn_lstm_ctc]
+and [TDNN LiGRU CTC model][TIMIT_tdnn_ligru_ctc].
+
+#### TDNN LSTM CTC Model
+
+The best PER we currently have is:
+
+||TEST|
+|--|--|
+|PER| 19.71% |
+
+We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1Hs9DA4V96uapw_30uNp32OMJgkuR5VVd?usp=sharing)
+
+#### TDNN LiGRU CTC Model
+
+The PER for this model is:
+
+||TEST|
+|--|--|
+|PER| 17.66% |
+
+We provide a Colab notebook to run a pre-trained TDNN LiGRU CTC model:  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1z3lkURVv9M7uTiIgf3Np9IntMHEknaks?usp=sharing)
+
+
+### TED-LIUM3
+
+We provide two models for this recipe: [Transducer Stateless: Conformer encoder + Embedding decoder][TED-LIUM3_transducer_stateless] and [Pruned Transducer Stateless: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][TED-LIUM3_pruned_transducer_stateless].
+
+#### Transducer Stateless:  Conformer encoder + Embedding decoder
+
+The best WER using modified beam search with beam size 4 is:
+
+|     |  dev  |  test  |
+|-----|-------|--------|
+| WER |  6.91 |  6.33  |
+
+Note: No auxiliary losses are used in the training and no LMs are used in the decoding.
+
+We provide a Colab notebook to run a pre-trained Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1MmY5bBxwvKLNT4A2DJnwiqRXhdchUqPN?usp=sharing)
+
+#### Pruned Transducer Stateless: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss
+
+The best WER using modified beam search with beam size 4 is:
+
+|     |  dev  |  test  |
+|-----|-------|--------|
+| WER |  6.77 |  6.14  |
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1je_1zGrOkGVVd4WLzgkXRHxl-I27yWtz?usp=sharing)
+
+
+### Aidatatang_200zh
+
+We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][Aidatatang_200zh_pruned_transducer_stateless2].
+
+#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss
+
+|                      |  Dev  | Test  |
+|----------------------|-------|-------|
+|    greedy search     | 5.53  | 6.59  |
+|   fast beam search   | 5.30  | 6.34  |
+| modified beam search | 5.27  | 6.33  |
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1wNSnSj3T5oOctbh5IGCa393gKOoQw2GH?usp=sharing)
+
+
+### WenetSpeech
+
+We provide some models for this recipe: [Pruned stateless RNN-T_2: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][WenetSpeech_pruned_transducer_stateless2] and [Pruned stateless RNN-T_5: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][WenetSpeech_pruned_transducer_stateless5].
+
+#### Pruned stateless RNN-T_2: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss (trained with L subset, offline ASR)
+
+|                      |  Dev  | Test-Net | Test-Meeting |
+|----------------------|-------|----------|--------------|
+|    greedy search     | 7.80  |  8.75    |  13.49       |
+| modified beam search| 7.76  |  8.71    |  13.41       |
+|   fast beam search   | 7.94  |  8.74    |  13.80       |
+
+#### Pruned stateless RNN-T_5: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss (trained with L subset)
+**Streaming**:
+|                      |  Dev  | Test-Net | Test-Meeting |
+|----------------------|-------|----------|--------------|
+| greedy_search | 8.78 | 10.12 | 16.16 |
+| modified_beam_search | 8.53| 9.95 | 15.81 |
+| fast_beam_search| 9.01 | 10.47 | 16.28 |
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless2 model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1EV4e1CHa1GZgEF-bZgizqI9RyFFehIiN?usp=sharing)
+
+### Alimeeting
+
+We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][Alimeeting_pruned_transducer_stateless2].
+
+#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss (trained with far subset)
+
+|                      |  Eval  | Test-Net |
+|----------------------|--------|----------|
+|    greedy search     | 31.77  |  34.66   |
+|   fast beam search   | 31.39  |  33.02   |
+| modified beam search | 30.38  |  34.25   |
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tKr3f0mL17uO_ljdHGKtR7HOmthYHwJG?usp=sharing)
+
+
+### TAL_CSASR
+
+We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][TAL_CSASR_pruned_transducer_stateless5].
+
+#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss
+
+The best results for Chinese CER(%) and English WER(%) respectively (zh: Chinese, en: English):
+|decoding-method | dev | dev_zh | dev_en | test | test_zh | test_en |
+|--|--|--|--|--|--|--|
+|greedy_search| 7.30 | 6.48 | 19.19 |7.39| 6.66 | 19.13|
+|modified_beam_search| 7.15 | 6.35 | 18.95 | 7.22| 6.50 | 18.70 |
+|fast_beam_search| 7.18 | 6.39| 18.90 |  7.27| 6.55 | 18.77|
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1DmIx-NloI1CMU5GdZrlse7TRu4y3Dpf8?usp=sharing)
+
+## Deployment with C++
+
+Once you have trained a model in icefall, you may want to deploy it with C++,
+without Python dependencies.
+
+Please refer to the documentation
+<https://icefall.readthedocs.io/en/latest/recipes/Non-streaming-ASR/librispeech/conformer_ctc.html#deployment-with-c>
+for how to do this.
+
+We also provide a Colab notebook, showing you how to run a torch scripted model in [k2][k2] with C++.
+Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1BIGLWzS36isskMXHKcqC9ysN6pspYXs_?usp=sharing)
+
+
+[LibriSpeech_tdnn_lstm_ctc]: egs/librispeech/ASR/tdnn_lstm_ctc
+[LibriSpeech_conformer_ctc]: egs/librispeech/ASR/conformer_ctc
+[LibriSpeech_transducer]: egs/librispeech/ASR/transducer
+[LibriSpeech_transducer_stateless]: egs/librispeech/ASR/transducer_stateless
+[LibriSpeech_zipformer]: egs/librispeech/ASR/zipformer
+[Aishell_tdnn_lstm_ctc]: egs/aishell/ASR/tdnn_lstm_ctc
+[Aishell_conformer_ctc]: egs/aishell/ASR/conformer_ctc
+[Aishell_pruned_transducer_stateless7]: egs/aishell/ASR/pruned_transducer_stateless7_bbpe
+[Aishell2_pruned_transducer_stateless5]: egs/aishell2/ASR/pruned_transducer_stateless5
+[Aishell4_pruned_transducer_stateless5]: egs/aishell4/ASR/pruned_transducer_stateless5
+[TIMIT_tdnn_lstm_ctc]: egs/timit/ASR/tdnn_lstm_ctc
+[TIMIT_tdnn_ligru_ctc]: egs/timit/ASR/tdnn_ligru_ctc
+[TED-LIUM3_transducer_stateless]: egs/tedlium3/ASR/transducer_stateless
+[TED-LIUM3_pruned_transducer_stateless]: egs/tedlium3/ASR/pruned_transducer_stateless
+[GigaSpeech_conformer_ctc]: egs/gigaspeech/ASR/conformer_ctc
+[GigaSpeech_pruned_transducer_stateless2]: egs/gigaspeech/ASR/pruned_transducer_stateless2
+[GigaSpeech_zipformer]: egs/gigaspeech/ASR/zipformer
+[Aidatatang_200zh_pruned_transducer_stateless2]: egs/aidatatang_200zh/ASR/pruned_transducer_stateless2
+[WenetSpeech_pruned_transducer_stateless2]: egs/wenetspeech/ASR/pruned_transducer_stateless2
+[WenetSpeech_pruned_transducer_stateless5]: egs/wenetspeech/ASR/pruned_transducer_stateless5
+[Alimeeting_pruned_transducer_stateless2]: egs/alimeeting/ASR/pruned_transducer_stateless2
+[TAL_CSASR_pruned_transducer_stateless5]: egs/tal_csasr/ASR/pruned_transducer_stateless5
+[yesno]: egs/yesno/ASR
+[librispeech]: egs/librispeech/ASR
+[aishell]: egs/aishell/ASR
+[aishell2]: egs/aishell2/ASR
+[aishell4]: egs/aishell4/ASR
+[timit]: egs/timit/ASR
+[tedlium3]: egs/tedlium3/ASR
+[gigaspeech]: egs/gigaspeech/ASR
+[aidatatang_200zh]: egs/aidatatang_200zh/ASR
+[wenetspeech]: egs/wenetspeech/ASR
+[alimeeting]: egs/alimeeting/ASR
+[tal_csasr]: egs/tal_csasr/ASR
+[ami]: egs/ami
+[swbd]: egs/swbd/ASR
+[k2]: https://github.com/k2-fsa/k2
diff --git a/activate-icefall.sh b/activate-icefall.sh
new file mode 100644
index 000000000..6116ca47a
--- /dev/null
+++ b/activate-icefall.sh
@@ -0,0 +1 @@
+export PYTHONPATH=/var/data/share20/qc/k2/Github/icefall:$PYTHONPATH
diff --git a/contributing.md b/contributing.md
new file mode 100644
index 000000000..0a1f9936e
--- /dev/null
+++ b/contributing.md
@@ -0,0 +1,37 @@
+# Contributing to Our Project
+
+Thank you for your interest in contributing to our project! We use Git pre-commit hooks to ensure code quality and consistency. Before contributing, please follow these guidelines to enable and use the pre-commit hooks.
+
+## Pre-Commit Hooks
+
+We have set up pre-commit hooks to check that the files you're committing meet our coding and formatting standards. These checks include:
+
+- Ensuring there are no trailing spaces.
+- Formatting code with [black](https://github.com/psf/black).
+- Checking compliance with PEP8 using [flake8](https://flake8.pycqa.org/).
+- Verifying that files end with a newline character (and only a newline).
+- Sorting imports using [isort](https://pycqa.github.io/isort/).
+
+Please note that these hooks are disabled by default. To enable them, follow these steps:
+
+### Installation (Run only once)
+
+1. Install the `pre-commit` package using pip:
+   ```bash
+   pip install pre-commit
+   ```
+1. Install the Git hooks using:
+   ```bash
+   pre-commit install
+   ```
+### Making a Commit
+Once you have enabled the pre-commit hooks, follow these steps when making a commit:
+1. Make your changes to the codebase.
+2. Stage your changes by using git add for the files you modified.
+3. Commit your changes using git commit. The pre-commit hooks will run automatically at this point.
+4. If all hooks run successfully, you can write your commit message, and your changes will be successfully committed.
+5. If any hook fails, your commit will not be successful. Please read and follow the error messages provided, make the necessary changes, and then re-run git add and git commit.
+
+### Your Contribution
+Your contributions are valuable to us, and by following these guidelines, you help maintain code consistency and quality in our project. We appreciate your dedication to ensuring high-quality code. If you have questions or need assistance, feel free to reach out to us. Thank you for being part of our open-source community!
+
diff --git a/docker/README.md b/docker/README.md
new file mode 100644
index 000000000..19959bfe6
--- /dev/null
+++ b/docker/README.md
@@ -0,0 +1,129 @@
+# icefall dockerfile
+
+## Download from dockerhub
+
+You can find pre-built docker image for icefall at the following address:
+
+  <https://hub.docker.com/r/k2fsa/icefall/tags>
+
+Example usage:
+
+```bash
+docker run --gpus all --rm -it  k2fsa/icefall:torch1.13.0-cuda11.6 /bin/bash
+```
+
+
+## Build from dockerfile
+
+2 sets of configuration are provided - (a) Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8, and (b) Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8.
+
+If your NVIDIA driver supports CUDA Version: 11.3, please go for case (a) Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8.
+
+Otherwise, since the older PyTorch images are not updated with the [apt-key rotation by NVIDIA](https://developer.nvidia.com/blog/updating-the-cuda-linux-gpg-repository-key), you have to go for case (b) Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8. Ensure that your NVDIA driver supports at least CUDA 11.0.
+
+You can check the highest CUDA version within your NVIDIA driver's support with the `nvidia-smi` command below. In this example, the highest CUDA version is 11.0, i.e. case (b) Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8.
+
+```bash
+$ nvidia-smi
+Tue Sep 20 00:26:13 2022
++-----------------------------------------------------------------------------+
+| NVIDIA-SMI 450.119.03   Driver Version: 450.119.03   CUDA Version: 11.0     |
+|-------------------------------+----------------------+----------------------+
+| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
+| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
+|                               |                      |               MIG M. |
+|===============================+======================+======================|
+|   0  TITAN RTX           On   | 00000000:03:00.0 Off |                  N/A |
+| 41%   31C    P8     4W / 280W |     16MiB / 24219MiB |      0%      Default |
+|                               |                      |                  N/A |
++-------------------------------+----------------------+----------------------+
+|   1  TITAN RTX           On   | 00000000:04:00.0 Off |                  N/A |
+| 41%   30C    P8    11W / 280W |      6MiB / 24220MiB |      0%      Default |
+|                               |                      |                  N/A |
++-------------------------------+----------------------+----------------------+
+
++-----------------------------------------------------------------------------+
+| Processes:                                                                  |
+|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
+|        ID   ID                                                   Usage      |
+|=============================================================================|
+|    0   N/A  N/A      2085      G   /usr/lib/xorg/Xorg                  9MiB |
+|    0   N/A  N/A      2240      G   /usr/bin/gnome-shell                4MiB |
+|    1   N/A  N/A      2085      G   /usr/lib/xorg/Xorg                  4MiB |
++-----------------------------------------------------------------------------+
+
+```
+
+## Building images locally
+If your environment requires a proxy to access the Internet, remember to add those information into the Dockerfile directly.
+For most cases, you can uncomment these lines in the Dockerfile and add in your proxy details.
+
+```dockerfile
+ENV http_proxy=http://aaa.bb.cc.net:8080 \
+    https_proxy=http://aaa.bb.cc.net:8080
+```
+
+Then, proceed with these commands.
+
+### If you are case (a), i.e. your NVIDIA driver supports CUDA version >= 11.3:
+
+```bash
+cd docker/Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8
+docker build -t icefall/pytorch1.12.1 .
+```
+
+### If you are case (b), i.e. your NVIDIA driver can only support CUDA versions 11.0 <= x < 11.3:
+```bash
+cd docker/Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8
+docker build -t icefall/pytorch1.7.1 .
+```
+
+## Running your built local image
+Sample usage of the GPU based images. These commands are written with case (a) in mind, so please make the necessary changes to your image name if you are case (b).
+Note: use [nvidia-docker](https://github.com/NVIDIA/nvidia-docker) to run the GPU images.
+
+```bash
+docker run -it --runtime=nvidia --shm-size=2gb --name=icefall --gpus all icefall/pytorch1.12.1
+```
+
+### Tips:
+1. Since your data and models most probably won't be in the docker, you must use the -v flag to access the host machine. Do this by specifying `-v {/path/in/host/machine}:{/path/in/docker}`.
+
+2. Also, if your environment requires a proxy, this would be a good time to add it in too: `-e http_proxy=http://aaa.bb.cc.net:8080 -e https_proxy=http://aaa.bb.cc.net:8080`.
+
+Overall, your docker run command should look like this.
+
+```bash
+docker run -it --runtime=nvidia --shm-size=2gb --name=icefall --gpus all -v {/path/in/host/machine}:{/path/in/docker} -e http_proxy=http://aaa.bb.cc.net:8080 -e https_proxy=http://aaa.bb.cc.net:8080 icefall/pytorch1.12.1
+```
+
+You can explore more docker run options [here](https://docs.docker.com/engine/reference/commandline/run/) to suit your environment.
+
+### Linking to icefall in your host machine
+
+If you already have icefall downloaded onto your host machine, you can use that repository instead so that changes in your code are visible inside and outside of the container.
+
+Note: Remember to set the -v flag above during the first run of the container, as that is the only way for your container to access your host machine.
+Warning: Check that the icefall in your host machine is visible from within your container before proceeding to the commands below.
+
+Use these commands once you are inside the container.
+
+```bash
+rm -r /workspace/icefall
+ln -s {/path/in/docker/to/icefall} /workspace/icefall
+```
+
+## Starting another session in the same running container.
+```bash
+docker exec -it icefall /bin/bash
+```
+
+## Restarting a killed container that has been run before.
+```bash
+docker start -ai icefall
+```
+
+## Sample usage of the CPU based images:
+```bash
+docker run -it icefall /bin/bash
+```
diff --git a/docker/Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8/Dockerfile b/docker/Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8/Dockerfile
new file mode 100644
index 000000000..ff9e40604
--- /dev/null
+++ b/docker/Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8/Dockerfile
@@ -0,0 +1,74 @@
+FROM pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel
+
+# ENV http_proxy=http://aaa.bbb.cc.net:8080 \
+#	https_proxy=http://aaa.bbb.cc.net:8080
+
+# install normal source
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        g++ \
+        make \
+        automake \
+        autoconf \
+        bzip2 \
+        unzip \
+        wget \
+        sox \
+        libtool \
+        git \
+        subversion \
+        zlib1g-dev \
+        gfortran \
+        ca-certificates \
+        patch \
+        ffmpeg \
+        valgrind \
+		libssl-dev \
+	    vim \
+		curl
+
+# cmake
+RUN wget -P /opt https://cmake.org/files/v3.18/cmake-3.18.0.tar.gz && \
+    cd /opt && \
+    tar -zxvf cmake-3.18.0.tar.gz && \
+    cd cmake-3.18.0 && \
+    ./bootstrap && \
+    make && \
+    make install && \
+    rm -rf cmake-3.18.0.tar.gz && \
+    find /opt/cmake-3.18.0 -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
+    cd -
+
+# flac
+RUN wget -P /opt https://downloads.xiph.org/releases/flac/flac-1.3.2.tar.xz  && \
+    cd /opt && \
+    xz -d flac-1.3.2.tar.xz && \
+    tar -xvf flac-1.3.2.tar && \
+    cd flac-1.3.2 && \
+    ./configure && \
+    make && make install && \
+    rm -rf flac-1.3.2.tar && \
+    find /opt/flac-1.3.2  -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
+    cd -
+
+RUN conda install -y -c pytorch torchaudio=0.12 && \
+    pip install graphviz
+
+
+#install k2 from source
+RUN git clone https://github.com/k2-fsa/k2.git /opt/k2 && \
+    cd /opt/k2 && \
+    python3 setup.py install && \
+    cd -
+
+# install  lhotse
+RUN pip install git+https://github.com/lhotse-speech/lhotse
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+	cd /workspace/icefall && \
+	pip install -r requirements.txt
+
+RUN pip install kaldifeat
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8/Dockerfile b/docker/Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8/Dockerfile
new file mode 100644
index 000000000..5c7423fa5
--- /dev/null
+++ b/docker/Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8/Dockerfile
@@ -0,0 +1,90 @@
+FROM pytorch/pytorch:1.7.1-cuda11.0-cudnn8-devel
+
+# ENV http_proxy=http://aaa.bbb.cc.net:8080 \
+#	https_proxy=http://aaa.bbb.cc.net:8080
+
+RUN rm /etc/apt/sources.list.d/cuda.list && \
+	rm /etc/apt/sources.list.d/nvidia-ml.list && \
+	apt-key del 7fa2af80
+
+# install normal source
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        g++ \
+        make \
+        automake \
+        autoconf \
+        bzip2 \
+        unzip \
+        wget \
+        sox \
+        libtool \
+        git \
+        subversion \
+        zlib1g-dev \
+        gfortran \
+        ca-certificates \
+        patch \
+        ffmpeg \
+        valgrind \
+		libssl-dev \
+	    vim \
+		curl
+
+# Add new keys and reupdate
+RUN curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/3bf863cc.pub | apt-key add - && \
+	curl -fsSL https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/7fa2af80.pub | apt-key add - && \
+	echo "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64 /" > /etc/apt/sources.list.d/cuda.list && \
+	echo "deb https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64 /" > /etc/apt/sources.list.d/nvidia-ml.list && \
+	rm -rf /var/lib/apt/lists/* && \
+	mv /opt/conda/lib/libcufft.so.10 /opt/libcufft.so.10.bak && \
+    mv /opt/conda/lib/libcurand.so.10 /opt/libcurand.so.10.bak && \
+    mv /opt/conda/lib/libcublas.so.11 /opt/libcublas.so.11.bak && \
+    mv /opt/conda/lib/libnvrtc.so.11.0 /opt/libnvrtc.so.11.1.bak && \
+    # mv /opt/conda/lib/libnvToolsExt.so.1 /opt/libnvToolsExt.so.1.bak && \
+    mv /opt/conda/lib/libcudart.so.11.0 /opt/libcudart.so.11.0.bak && \
+	apt-get update && apt-get -y upgrade
+
+# cmake
+RUN wget -P /opt https://cmake.org/files/v3.18/cmake-3.18.0.tar.gz && \
+    cd /opt && \
+    tar -zxvf cmake-3.18.0.tar.gz && \
+    cd cmake-3.18.0 && \
+    ./bootstrap && \
+    make && \
+    make install && \
+    rm -rf cmake-3.18.0.tar.gz && \
+    find /opt/cmake-3.18.0 -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
+    cd -
+
+# flac
+RUN wget -P /opt https://downloads.xiph.org/releases/flac/flac-1.3.2.tar.xz  && \
+    cd /opt && \
+    xz -d flac-1.3.2.tar.xz && \
+    tar -xvf flac-1.3.2.tar && \
+    cd flac-1.3.2 && \
+    ./configure && \
+    make && make install && \
+    rm -rf flac-1.3.2.tar && \
+    find /opt/flac-1.3.2  -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
+    cd -
+
+RUN conda install -y -c pytorch torchaudio=0.7.1 && \
+    pip install graphviz
+
+#install k2 from source
+RUN git clone https://github.com/k2-fsa/k2.git /opt/k2 && \
+    cd /opt/k2 && \
+    python3 setup.py install && \
+    cd -
+
+# install  lhotse
+RUN pip install git+https://github.com/lhotse-speech/lhotse
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+	cd /workspace/icefall && \
+	pip install -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch1.12.1-cuda11.3.dockerfile b/docker/torch1.12.1-cuda11.3.dockerfile
new file mode 100644
index 000000000..deb5715cc
--- /dev/null
+++ b/docker/torch1.12.1-cuda11.3.dockerfile
@@ -0,0 +1,70 @@
+FROM pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.7
+ARG K2_VERSION="1.24.4.dev20230725+cuda11.3.torch1.12.1"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.3.torch1.12.1"
+ARG TORCHAUDIO_VERSION="0.12.1+cu113"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    	libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch1.13.0-cuda11.6.dockerfile b/docker/torch1.13.0-cuda11.6.dockerfile
new file mode 100644
index 000000000..afc6c1b84
--- /dev/null
+++ b/docker/torch1.13.0-cuda11.6.dockerfile
@@ -0,0 +1,72 @@
+FROM pytorch/pytorch:1.13.0-cuda11.6-cudnn8-runtime
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.9
+ARG K2_VERSION="1.24.4.dev20231021+cuda11.6.torch1.13.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.6.torch1.13.0"
+ARG TORCHAUDIO_VERSION="0.13.0+cu116"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    	libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+ENV LD_LIBRARY_PATH /opt/conda/lib/stubs:$LD_LIBRARY_PATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch1.9.0-cuda10.2.dockerfile b/docker/torch1.9.0-cuda10.2.dockerfile
new file mode 100644
index 000000000..9ff225b54
--- /dev/null
+++ b/docker/torch1.9.0-cuda10.2.dockerfile
@@ -0,0 +1,86 @@
+FROM pytorch/pytorch:1.9.0-cuda10.2-cudnn7-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.7
+ARG K2_VERSION="1.24.3.dev20230726+cuda10.2.torch1.9.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda10.2.torch1.9.0"
+ARG TORCHAUDIO_VERSION="0.9.0"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+# see https://developer.nvidia.com/blog/updating-the-cuda-linux-gpg-repository-key/
+
+RUN rm /etc/apt/sources.list.d/cuda.list && \
+	rm /etc/apt/sources.list.d/nvidia-ml.list && \
+	apt-key del 7fa2af80
+
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    	libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-keyring_1.0-1_all.deb && \
+    dpkg -i cuda-keyring_1.0-1_all.deb && \
+    rm -v cuda-keyring_1.0-1_all.deb && \
+    apt-get update && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip uninstall -y tqdm && \
+    pip install -U --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz \
+      tqdm>=4.63.0
+
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch2.0.0-cuda11.7.dockerfile b/docker/torch2.0.0-cuda11.7.dockerfile
new file mode 100644
index 000000000..db8076560
--- /dev/null
+++ b/docker/torch2.0.0-cuda11.7.dockerfile
@@ -0,0 +1,70 @@
+FROM pytorch/pytorch:2.0.0-cuda11.7-cudnn8-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20231021+cuda11.7.torch2.0.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.7.torch2.0.0"
+ARG TORCHAUDIO_VERSION="2.0.0+cu117"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    	libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch2.1.0-cuda11.8.dockerfile b/docker/torch2.1.0-cuda11.8.dockerfile
new file mode 100644
index 000000000..b006b0d96
--- /dev/null
+++ b/docker/torch2.1.0-cuda11.8.dockerfile
@@ -0,0 +1,70 @@
+FROM pytorch/pytorch:2.1.0-cuda11.8-cudnn8-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20231021+cuda11.8.torch2.1.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.8.torch2.1.0"
+ARG TORCHAUDIO_VERSION="2.1.0+cu118"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    	libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch2.1.0-cuda12.1.dockerfile b/docker/torch2.1.0-cuda12.1.dockerfile
new file mode 100644
index 000000000..1b078dc22
--- /dev/null
+++ b/docker/torch2.1.0-cuda12.1.dockerfile
@@ -0,0 +1,70 @@
+FROM pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20231021+cuda12.1.torch2.1.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda12.1.torch2.1.0"
+ARG TORCHAUDIO_VERSION="2.1.0+cu121"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    	libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docs/.gitignore b/docs/.gitignore
new file mode 100644
index 000000000..567609b12
--- /dev/null
+++ b/docs/.gitignore
@@ -0,0 +1 @@
+build/
diff --git a/docs/Makefile b/docs/Makefile
new file mode 100644
index 000000000..d0c3cbf10
--- /dev/null
+++ b/docs/Makefile
@@ -0,0 +1,20 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line, and also
+# from the environment for the first two.
+SPHINXOPTS    ?=
+SPHINXBUILD   ?= sphinx-build
+SOURCEDIR     = source
+BUILDDIR      = build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
diff --git a/docs/README.md b/docs/README.md
new file mode 100644
index 000000000..3abb38f8b
--- /dev/null
+++ b/docs/README.md
@@ -0,0 +1,24 @@
+
+## Usage
+
+```bash
+cd /path/to/icefall/docs
+pip install -r requirements.txt
+make clean
+make html
+cd build/html
+python3 -m http.server 8000
+```
+
+It prints:
+
+```
+Serving HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...
+```
+
+Open your browser and go to <http://0.0.0.0:8000/> to view the generated
+documentation.
+
+Done!
+
+**Hint**: You can change the port number when starting the server.
diff --git a/docs/make.bat b/docs/make.bat
new file mode 100644
index 000000000..6247f7e23
--- /dev/null
+++ b/docs/make.bat
@@ -0,0 +1,35 @@
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+	set SPHINXBUILD=sphinx-build
+)
+set SOURCEDIR=source
+set BUILDDIR=build
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+	echo.
+	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+	echo.installed, then set the SPHINXBUILD environment variable to point
+	echo.to the full path of the 'sphinx-build' executable. Alternatively you
+	echo.may add the Sphinx directory to PATH.
+	echo.
+	echo.If you don't have Sphinx installed, grab it from
+	echo.http://sphinx-doc.org/
+	exit /b 1
+)
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+
+:end
+popd
diff --git a/docs/source/_static/logo.png b/docs/source/_static/logo.png
new file mode 100644
index 000000000..84d42568c
Binary files /dev/null and b/docs/source/_static/logo.png differ
diff --git a/docs/source/conf.py b/docs/source/conf.py
new file mode 100644
index 000000000..5a534e126
--- /dev/null
+++ b/docs/source/conf.py
@@ -0,0 +1,101 @@
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+#
+# import os
+# import sys
+# sys.path.insert(0, os.path.abspath('.'))
+
+import sphinx_rtd_theme
+
+# -- Project information -----------------------------------------------------
+
+project = "icefall"
+copyright = "2021, icefall development team"
+author = "icefall development team"
+
+# The full version, including alpha/beta/rc tags
+release = "0.1"
+
+
+# -- General configuration ---------------------------------------------------
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    "sphinx.ext.todo",
+    "sphinx_rtd_theme",
+    "sphinxcontrib.youtube",
+]
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ["_templates"]
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This pattern also affects html_static_path and html_extra_path.
+exclude_patterns = []
+
+source_suffix = {
+    ".rst": "restructuredtext",
+}
+master_doc = "index"
+
+
+# -- Options for HTML output -------------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_theme = "sphinx_rtd_theme"
+html_theme_path = [sphinx_rtd_theme.get_html_theme_path()]
+html_show_sourcelink = True
+
+# Add any paths that contain custom static files (such as style sheets) here,
+# relative to this directory. They are copied after the builtin static files,
+# so a file named "default.css" will overwrite the builtin "default.css".
+html_static_path = ["_static", "installation/images"]
+
+pygments_style = "sphinx"
+
+numfig = True
+
+html_context = {
+    "display_github": True,
+    "github_user": "k2-fsa",
+    "github_repo": "icefall",
+    "github_version": "master",
+    "conf_py_path": "/docs/source/",
+}
+
+todo_include_todos = True
+
+rst_epilog = """
+.. _sherpa-ncnn: https://github.com/k2-fsa/sherpa-ncnn
+.. _sherpa-onnx: https://github.com/k2-fsa/sherpa-onnx
+.. _icefall: https://github.com/k2-fsa/icefall
+.. _git-lfs: https://git-lfs.com/
+.. _ncnn: https://github.com/tencent/ncnn
+.. _LibriSpeech: https://www.openslr.org/12
+.. _Gigaspeech: https://github.com/SpeechColab/GigaSpeech
+.. _musan: http://www.openslr.org/17/
+.. _ONNX: https://github.com/onnx/onnx
+.. _onnxruntime: https://github.com/microsoft/onnxruntime
+.. _torch: https://github.com/pytorch/pytorch
+.. _torchaudio: https://github.com/pytorch/audio
+.. _k2: https://github.com/k2-fsa/k2
+.. _lhotse: https://github.com/lhotse-speech/lhotse
+.. _yesno: https://www.openslr.org/1/
+.. _Next-gen Kaldi: https://github.com/k2-fsa
+.. _Kaldi: https://github.com/kaldi-asr/kaldi
+.. _lilcom: https://github.com/danpovey/lilcom
+"""
diff --git a/docs/source/contributing/code-style.rst b/docs/source/contributing/code-style.rst
new file mode 100644
index 000000000..cb08229c3
--- /dev/null
+++ b/docs/source/contributing/code-style.rst
@@ -0,0 +1,74 @@
+.. _follow the code style:
+
+Follow the code style
+=====================
+
+We use the following tools to make the code style to be as consistent as possible:
+
+  - `black <https://github.com/psf/black>`_, to format the code
+  - `flake8 <https://github.com/PyCQA/flake8>`_, to check the style and quality of the code
+  - `isort <https://github.com/PyCQA/isort>`_, to sort ``imports``
+
+The following versions of the above tools are used:
+
+  - ``black == 22.3.0``
+  - ``flake8 == 5.0.4``
+  - ``isort == 5.10.1``
+
+After running the following commands:
+
+  .. code-block::
+
+    $ git clone https://github.com/k2-fsa/icefall
+    $ cd icefall
+    $ pip install pre-commit
+    $ pre-commit install
+
+it will run the following checks whenever you run ``git commit``, **automatically**:
+
+    .. figure:: images/pre-commit-check.png
+       :width: 600
+       :align: center
+
+       pre-commit hooks invoked by ``git commit`` (Failed).
+
+If any of the above checks failed, your ``git commit`` was not successful.
+Please fix any issues reported by the check tools.
+
+.. HINT::
+
+  Some of the check tools, i.e., ``black`` and ``isort`` will modify
+  the files to be committed **in-place**. So please run ``git status``
+  after failure to see which file has been modified by the tools
+  before you make any further changes.
+
+After fixing all the failures, run ``git commit`` again and
+it should succeed this time:
+
+    .. figure:: images/pre-commit-check-success.png
+       :width: 600
+       :align: center
+
+       pre-commit hooks invoked by ``git commit`` (Succeeded).
+
+If you want to check the style of your code before ``git commit``, you
+can do the following:
+
+  .. code-block:: bash
+
+    $ pre-commit install
+    $ pre-commit run
+
+Or without installing the pre-commit hooks:
+
+  .. code-block:: bash
+
+    $ cd icefall
+    $ pip install black==22.3.0 flake8==5.0.4 isort==5.10.1
+    $ black --check your_changed_file.py
+    $ black your_changed_file.py  # modify it in-place
+    $
+    $ flake8 your_changed_file.py
+    $
+    $ isort --check your_changed_file.py  # modify it in-place
+    $ isort your_changed_file.py
diff --git a/docs/source/contributing/doc.rst b/docs/source/contributing/doc.rst
new file mode 100644
index 000000000..893d8a15e
--- /dev/null
+++ b/docs/source/contributing/doc.rst
@@ -0,0 +1,45 @@
+Contributing to Documentation
+=============================
+
+We use `sphinx <https://www.sphinx-doc.org/en/master/>`_
+for documentation.
+
+Before writing documentation, you have to prepare the environment:
+
+  .. code-block:: bash
+
+    $ cd docs
+    $ pip install -r requirements.txt
+
+After setting up the environment, you are ready to write documentation.
+Please refer to `reStructuredText Primer <https://www.sphinx-doc.org/en/master/usage/restructuredtext/basics.html>`_
+if you are not familiar with ``reStructuredText``.
+
+After writing some documentation, you can build the documentation **locally**
+to preview what it looks like if it is published:
+
+  .. code-block:: bash
+
+    $ cd docs
+    $ make html
+
+The generated documentation is in ``docs/build/html`` and can be viewed
+with the following commands:
+
+  .. code-block:: bash
+
+    $ cd docs/build/html
+    $ python3 -m http.server
+
+It will print::
+
+  Serving HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...
+
+Open your browser, go to `<http://0.0.0.0:8000/>`_, and you will see
+the following:
+
+    .. figure:: images/doc-contrib.png
+       :width: 600
+       :align: center
+
+       View generated documentation locally with ``python3 -m http.server``.
diff --git a/docs/source/contributing/how-to-create-a-recipe.rst b/docs/source/contributing/how-to-create-a-recipe.rst
new file mode 100644
index 000000000..168a856c3
--- /dev/null
+++ b/docs/source/contributing/how-to-create-a-recipe.rst
@@ -0,0 +1,156 @@
+How to create a recipe
+======================
+
+.. HINT::
+
+  Please read :ref:`follow the code style` to adjust your code style.
+
+.. CAUTION::
+
+  ``icefall`` is designed to be as Pythonic as possible. Please use
+  Python in your recipe if possible.
+
+Data Preparation
+----------------
+
+We recommend you to prepare your training/test/validate dataset
+with `lhotse <https://github.com/lhotse-speech/lhotse>`_.
+
+Please refer to `<https://lhotse.readthedocs.io/en/latest/index.html>`_
+for how to create a recipe in ``lhotse``.
+
+.. HINT::
+
+  The ``yesno`` recipe in ``lhotse`` is a very good example.
+
+  Please refer to `<https://github.com/lhotse-speech/lhotse/pull/380>`_,
+  which shows how to add a new recipe to ``lhotse``.
+
+Suppose you would like to add a recipe for a dataset named ``foo``.
+You can do the following:
+
+.. code-block::
+
+  $ cd egs
+  $ mkdir -p foo/ASR
+  $ cd foo/ASR
+  $ touch prepare.sh
+  $ chmod +x prepare.sh
+
+If your dataset is very simple, please follow
+`egs/yesno/ASR/prepare.sh <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/prepare.sh>`_
+to write your own ``prepare.sh``.
+Otherwise, please refer to
+`egs/librispeech/ASR/prepare.sh <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/prepare.sh>`_
+to prepare your data.
+
+
+Training
+--------
+
+Assume you have a fancy model, called ``bar`` for the ``foo`` recipe, you can
+organize your files in the following way:
+
+.. code-block::
+
+  $ cd egs/foo/ASR
+  $ mkdir bar
+  $ cd bar
+  $ touch README.md model.py train.py decode.py asr_datamodule.py pretrained.py
+
+For instance , the ``yesno`` recipe has a ``tdnn`` model and its directory structure
+looks like the following:
+
+.. code-block:: bash
+
+  egs/yesno/ASR/tdnn/
+  |-- README.md
+  |-- asr_datamodule.py
+  |-- decode.py
+  |-- model.py
+  |-- pretrained.py
+  `-- train.py
+
+**File description**:
+
+  - ``README.md``
+
+    It contains information of this recipe, e.g., how to run it, what the WER is, etc.
+
+  - ``asr_datamodule.py``
+
+    It provides code to create PyTorch dataloaders with train/test/validation dataset.
+
+  - ``decode.py``
+
+    It takes as inputs the checkpoints saved during the training stage to decode the test
+    dataset(s).
+
+  - ``model.py``
+
+    It contains the definition of your fancy neural network model.
+
+  - ``pretrained.py``
+
+    We can use this script to do inference with a pre-trained model.
+
+  - ``train.py``
+
+    It contains training code.
+
+
+.. HINT::
+
+  Please take a look at
+
+    - `egs/yesno/tdnn <https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR/tdnn>`_
+    - `egs/librispeech/tdnn_lstm_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/tdnn_lstm_ctc>`_
+    - `egs/librispeech/conformer_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conformer_ctc>`_
+
+  to get a feel what the resulting files look like.
+
+.. NOTE::
+
+  Every model in a recipe is kept to be as self-contained as possible.
+  We tolerate duplicate code among different recipes.
+
+
+The training stage should be invocable by:
+
+  .. code-block::
+
+    $ cd egs/foo/ASR
+    $ ./bar/train.py
+    $ ./bar/train.py --help
+
+
+Decoding
+--------
+
+Please refer to
+
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/conformer_ctc/decode.py>`_
+
+    If your model is transformer/conformer based.
+
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/tdnn_lstm_ctc/decode.py>`_
+
+    If your model is TDNN/LSTM based, i.e., there is no attention decoder.
+
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/tdnn/decode.py>`_
+
+    If there is no LM rescoring.
+
+The decoding stage should be invocable by:
+
+  .. code-block::
+
+    $ cd egs/foo/ASR
+    $ ./bar/decode.py
+    $ ./bar/decode.py --help
+
+Pre-trained model
+-----------------
+
+Please demonstrate how to use your model for inference in ``egs/foo/ASR/bar/pretrained.py``.
+If possible, please consider creating a Colab notebook to show that.
diff --git a/docs/source/contributing/images/doc-contrib.png b/docs/source/contributing/images/doc-contrib.png
new file mode 100644
index 000000000..00906ab83
Binary files /dev/null and b/docs/source/contributing/images/doc-contrib.png differ
diff --git a/docs/source/contributing/images/pre-commit-check-success.png b/docs/source/contributing/images/pre-commit-check-success.png
new file mode 100644
index 000000000..3c6ee9b1c
Binary files /dev/null and b/docs/source/contributing/images/pre-commit-check-success.png differ
diff --git a/docs/source/contributing/images/pre-commit-check.png b/docs/source/contributing/images/pre-commit-check.png
new file mode 100644
index 000000000..80784eced
Binary files /dev/null and b/docs/source/contributing/images/pre-commit-check.png differ
diff --git a/docs/source/contributing/index.rst b/docs/source/contributing/index.rst
new file mode 100644
index 000000000..21c747d33
--- /dev/null
+++ b/docs/source/contributing/index.rst
@@ -0,0 +1,22 @@
+Contributing
+============
+
+Contributions to ``icefall`` are very welcomed.
+There are many possible ways to make contributions and
+two of them are:
+
+  - To write documentation
+  - To write code
+
+    - (1) To follow the code style in the repository
+    - (2) To write a new recipe
+
+In this page, we describe how to contribute documentation
+and code to ``icefall``.
+
+.. toctree::
+   :maxdepth: 2
+
+   doc
+   code-style
+   how-to-create-a-recipe
diff --git a/docs/source/decoding-with-langugage-models/LODR.rst b/docs/source/decoding-with-langugage-models/LODR.rst
new file mode 100644
index 000000000..b6b6e8cbb
--- /dev/null
+++ b/docs/source/decoding-with-langugage-models/LODR.rst
@@ -0,0 +1,187 @@
+.. _LODR:
+
+LODR for RNN Transducer
+=======================
+
+
+As a type of E2E model, neural transducers are usually considered as having an internal
+language model, which learns the language level information on the training corpus.
+In real-life scenario, there is often a mismatch between the training corpus and the target corpus space.
+This mismatch can be a problem when decoding for neural transducer models with language models as its internal
+language can act "against" the external LM. In this tutorial, we show how to use
+`Low-order Density Ratio <https://arxiv.org/abs/2203.16776>`_ to alleviate this effect to further improve the performance
+of langugae model integration.
+
+.. note::
+
+    This tutorial is based on the recipe
+    `pruned_transducer_stateless7_streaming <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming>`_,
+    which is a streaming transducer model trained on `LibriSpeech`_.
+    However, you can easily apply LODR to other recipes.
+    If you encounter any problems, please open an issue here `icefall <https://github.com/k2-fsa/icefall/issues>`__.
+
+
+.. note::
+
+    For simplicity, the training and testing corpus in this tutorial are the same (`LibriSpeech`_). However,
+    you can change the testing set to any other domains (e.g `GigaSpeech`_) and prepare the language models
+    using that corpus.
+
+First, let's have a look at some background information. As the predecessor of LODR, Density Ratio (DR) is first proposed `here <https://arxiv.org/abs/2002.11268>`_
+to address the language information mismatch between the training
+corpus (source domain) and the testing corpus (target domain). Assuming that the source domain and the test domain
+are acoustically similar, DR derives the following formular for decoding with Bayes' theorem:
+
+.. math::
+
+    \text{score}\left(y_u|\mathit{x},y\right) =
+    \log p\left(y_u|\mathit{x},y_{1:u-1}\right) +
+    \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) -
+    \lambda_2 \log p_{\text{Source LM}}\left(y_u|\mathit{x},y_{1:u-1}\right)
+
+
+where :math:`\lambda_1` and :math:`\lambda_2` are the weights of LM scores for target domain and source domain respectively.
+Here, the source domain LM is trained on the training corpus. The only difference in the above formular compared to
+shallow fusion is the subtraction of the source domain LM.
+
+Some works treat the predictor and the joiner of the neural transducer as its internal LM. However, the LM is
+considered to be weak and can only capture low-level language information. Therefore, `LODR <https://arxiv.org/abs/2203.16776>`__ proposed to use
+a low-order n-gram LM as an approximation of the ILM of the neural transducer. This leads to the following formula
+during decoding for transducer model:
+
+.. math::
+
+    \text{score}\left(y_u|\mathit{x},y\right) =
+    \log p_{rnnt}\left(y_u|\mathit{x},y_{1:u-1}\right) +
+    \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) -
+    \lambda_2 \log p_{\text{bi-gram}}\left(y_u|\mathit{x},y_{1:u-1}\right)
+
+In LODR, an additional bi-gram LM estimated on the source domain (e.g training corpus) is required. Compared to DR,
+the only difference lies in the choice of source domain LM. According to the original `paper <https://arxiv.org/abs/2203.16776>`_,
+LODR achieves similar performance compared DR in both intra-domain and cross-domain settings.
+As a bi-gram is much faster to evaluate, LODR is usually much faster.
+
+Now, we will show you how to use LODR in ``icefall``.
+For illustration purpose, we will use a pre-trained ASR model from this `link <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>`_.
+If you want to train your model from scratch, please have a look at :ref:`non_streaming_librispeech_pruned_transducer_stateless`.
+The testing scenario here is intra-domain (we decode the model trained on `LibriSpeech`_ on `LibriSpeech`_ testing sets).
+
+As the initial step, let's download the pre-trained model.
+
+.. code-block:: bash
+
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+    $ cd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ git lfs pull --include "pretrained.pt"
+    $ ln -s pretrained.pt epoch-99.pt # create a symbolic link so that the checkpoint can be loaded
+    $ cd ../data/lang_bpe_500
+    $ git lfs pull --include bpe.model
+    $ cd ../../..
+
+To test the model, let's have a look at the decoding results **without** using LM. This can be done via the following command:
+
+.. code-block:: bash
+
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --exp-dir $exp_dir \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search
+
+The following WERs are achieved on test-clean and test-other:
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	3.11	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	7.93	best for test-other
+
+Then, we download the external language model and bi-gram LM that are necessary for LODR.
+Note that the bi-gram is estimated on the LibriSpeech 960 hours' text.
+
+.. code-block:: bash
+
+    $ # download the external LM
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+    $ # create a symbolic link so that the checkpoint can be loaded
+    $ pushd icefall-librispeech-rnn-lm/exp
+    $ git lfs pull --include "pretrained.pt"
+    $ ln -s pretrained.pt epoch-99.pt
+    $ popd
+    $
+    $ # download the bi-gram
+    $ git lfs install
+    $ git clone https://huggingface.co/marcoyang/librispeech_bigram
+    $ pushd data/lang_bpe_500
+    $ ln -s ../../librispeech_bigram/2gram.fst.txt .
+    $ popd
+
+Then, we perform LODR decoding by setting ``--decoding-method`` to ``modified_beam_search_lm_LODR``:
+
+.. code-block:: bash
+
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ lm_dir=./icefall-librispeech-rnn-lm/exp
+    $ lm_scale=0.42
+    $ LODR_scale=-0.24
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --beam-size 4 \
+        --exp-dir $exp_dir \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search_LODR \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --use-shallow-fusion 1 \
+        --lm-type rnn \
+        --lm-exp-dir $lm_dir \
+        --lm-epoch 99 \
+        --lm-scale $lm_scale \
+        --lm-avg 1 \
+        --rnn-lm-embedding-dim 2048 \
+        --rnn-lm-hidden-dim 2048 \
+        --rnn-lm-num-layers 3 \
+        --lm-vocab-size 500 \
+        --tokens-ngram 2 \
+        --ngram-lm-scale $LODR_scale
+
+There are two extra arguments that need to be given when doing LODR. ``--tokens-ngram`` specifies the order of n-gram. As we
+are using a bi-gram, we set it to 2. ``--ngram-lm-scale`` is the scale of the bi-gram, it should be a negative number
+as we are subtracting the bi-gram's score during decoding.
+
+The decoding results obtained with the above command are shown below:
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	2.61	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	6.74	best for test-other
+
+Recall that the lowest WER we obtained in :ref:`shallow_fusion` with beam size of 4 is ``2.77/7.08``, LODR
+indeed **further improves** the WER. We can do even better if we increase ``--beam-size``:
+
+.. list-table:: WER of LODR with different beam sizes
+   :widths: 25 25 50
+   :header-rows: 1
+
+   * - Beam size
+     - test-clean
+     - test-other
+   * - 4
+     - 2.61
+     - 6.74
+   * - 8
+     - 2.45
+     - 6.38
+   * - 12
+     - 2.4
+     - 6.23
diff --git a/docs/source/decoding-with-langugage-models/index.rst b/docs/source/decoding-with-langugage-models/index.rst
new file mode 100644
index 000000000..c49da9a4e
--- /dev/null
+++ b/docs/source/decoding-with-langugage-models/index.rst
@@ -0,0 +1,34 @@
+Decoding with language models
+=============================
+
+This section describes how to use external langugage models 
+during decoding to improve the WER of transducer models. To train an external language model,
+please refer to this tutorial: :ref:`train_nnlm`.
+
+The following decoding methods with external langugage models are available:
+
+
+.. list-table:: 
+   :widths: 25 50
+   :header-rows: 1
+
+   * - Decoding method
+     - beam=4
+   * - ``modified_beam_search``
+     - Beam search (i.e. really n-best decoding, the "beam" is the value of n), similar to the original RNN-T paper. Note, this method does not use language model. 
+   * - ``modified_beam_search_lm_shallow_fusion``
+     - As ``modified_beam_search``, but interpolate RNN-T scores with language model scores, also known as shallow fusion
+   * - ``modified_beam_search_LODR``
+     - As ``modified_beam_search_lm_shallow_fusion``, but subtract score of a (BPE-symbol-level) bigram backoff language model used as an approximation to the internal language model of RNN-T.
+   * - ``modified_beam_search_lm_rescore``
+     - As ``modified_beam_search``, but rescore the n-best hypotheses with external language model (e.g. RNNLM) and re-rank them.
+   * - ``modified_beam_search_lm_rescore_LODR``
+     - As ``modified_beam_search_lm_rescore``, but also subtract the score of a (BPE-symbol-level) bigram backoff language model during re-ranking.
+
+
+.. toctree::
+   :maxdepth: 2
+
+   shallow-fusion
+   LODR
+   rescoring
diff --git a/docs/source/decoding-with-langugage-models/rescoring.rst b/docs/source/decoding-with-langugage-models/rescoring.rst
new file mode 100644
index 000000000..4cabaa432
--- /dev/null
+++ b/docs/source/decoding-with-langugage-models/rescoring.rst
@@ -0,0 +1,255 @@
+.. _rescoring:
+
+LM rescoring for Transducer
+=================================
+
+LM rescoring is a commonly used approach to incorporate external LM information. Unlike shallow-fusion-based
+methods (see :ref:`shallow_fusion`, :ref:`LODR`), rescoring is usually performed to re-rank the n-best hypotheses after beam search.
+Rescoring is usually more efficient than shallow fusion since less computation is performed on the external LM.
+In this tutorial, we will show you how to use external LM to rescore the n-best hypotheses decoded from neural transducer models in
+`icefall <https://github.com/k2-fsa/icefall>`__.
+
+.. note::
+
+    This tutorial is based on the recipe 
+    `pruned_transducer_stateless7_streaming <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming>`_,
+    which is a streaming transducer model trained on `LibriSpeech`_. 
+    However, you can easily apply shallow fusion to other recipes.
+    If you encounter any problems, please open an issue `here <https://github.com/k2-fsa/icefall/issues>`_.
+
+.. note::
+
+    For simplicity, the training and testing corpus in this tutorial is the same (`LibriSpeech`_). However, you can change the testing set
+    to any other domains (e.g `GigaSpeech`_) and use an external LM trained on that domain.
+
+.. HINT::
+
+  We recommend you to use a GPU for decoding.
+
+For illustration purpose, we will use a pre-trained ASR model from this `link <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>`__.
+If you want to train your model from scratch, please have a look at :ref:`non_streaming_librispeech_pruned_transducer_stateless`.
+
+As the initial step, let's download the pre-trained model.
+
+.. code-block:: bash
+
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+    $ cd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ git lfs pull --include "pretrained.pt"
+    $ ln -s pretrained.pt epoch-99.pt # create a symbolic link so that the checkpoint can be loaded
+    $ cd ../data/lang_bpe_500
+    $ git lfs pull --include bpe.model
+    $ cd ../../..
+
+As usual, we first test the model's performance without external LM. This can be done via the following command:
+
+.. code-block:: bash
+
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --exp-dir $exp_dir \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search
+
+The following WERs are achieved on test-clean and test-other:
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	3.11	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	7.93	best for test-other
+
+Now, we will try to improve the above WER numbers via external LM rescoring. We will download 
+a pre-trained LM from this `link <https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm>`__.
+
+.. note::
+
+    This is an RNN LM trained on the LibriSpeech text corpus. So it might not be ideal for other corpus.
+    You may also train a RNN LM from scratch. Please refer to this `script <https://github.com/k2-fsa/icefall/blob/master/icefall/rnn_lm/train.py>`__
+    for training a RNN LM and this `script <https://github.com/k2-fsa/icefall/blob/master/icefall/transformer_lm/train.py>`__ to train a transformer LM.
+
+.. code-block:: bash
+
+    $ # download the external LM
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm 
+    $ # create a symbolic link so that the checkpoint can be loaded
+    $ pushd icefall-librispeech-rnn-lm/exp
+    $ git lfs pull --include "pretrained.pt"
+    $ ln -s pretrained.pt epoch-99.pt 
+    $ popd
+
+
+With the RNNLM available, we can rescore the n-best hypotheses generated from `modified_beam_search`. Here,
+`n` should be the number of beams, i.e ``--beam-size``. The command for LM rescoring is
+as follows. Note that the ``--decoding-method`` is set to `modified_beam_search_lm_rescore` and ``--use-shallow-fusion``
+is set to `False`.
+
+.. code-block:: bash
+    
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ lm_dir=./icefall-librispeech-rnn-lm/exp
+    $ lm_scale=0.43
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --beam-size 4 \
+        --exp-dir $exp_dir \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search_lm_rescore \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --use-shallow-fusion 0 \
+        --lm-type rnn \
+        --lm-exp-dir $lm_dir \
+        --lm-epoch 99 \
+        --lm-scale $lm_scale \
+        --lm-avg 1 \
+        --rnn-lm-embedding-dim 2048 \
+        --rnn-lm-hidden-dim 2048 \
+        --rnn-lm-num-layers 3 \
+        --lm-vocab-size 500
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	2.93	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	7.6	best for test-other
+
+Great! We made some improvements! Increasing the size of the n-best hypotheses will further boost the performance,
+see the following table:
+
+.. list-table:: WERs of LM rescoring with different beam sizes
+   :widths: 25 25 25
+   :header-rows: 1
+
+   * - Beam size
+     - test-clean
+     - test-other
+   * - 4
+     - 2.93
+     - 7.6
+   * - 8
+     - 2.67
+     - 7.11
+   * - 12
+     - 2.59
+     - 6.86
+
+In fact, we can also apply LODR (see :ref:`LODR`) when doing LM rescoring. To do so, we need to 
+download the bi-gram required by LODR:
+
+.. code-block:: bash
+
+    $ # download the bi-gram
+    $ git lfs install
+    $ git clone https://huggingface.co/marcoyang/librispeech_bigram
+    $ pushd data/lang_bpe_500
+    $ ln -s ../../librispeech_bigram/2gram.arpa .
+    $ popd
+
+Then we can performn LM rescoring + LODR by changing the decoding method to `modified_beam_search_lm_rescore_LODR`. 
+
+.. note:: 
+
+    This decoding method requires the dependency of `kenlm <https://github.com/kpu/kenlm>`_. You can install it
+    via this command: `pip install https://github.com/kpu/kenlm/archive/master.zip`. 
+
+.. code-block:: bash
+    
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ lm_dir=./icefall-librispeech-rnn-lm/exp
+    $ lm_scale=0.43
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --beam-size 4 \
+        --exp-dir $exp_dir \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search_lm_rescore_LODR \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --use-shallow-fusion 0 \
+        --lm-type rnn \
+        --lm-exp-dir $lm_dir \
+        --lm-epoch 99 \
+        --lm-scale $lm_scale \
+        --lm-avg 1 \
+        --rnn-lm-embedding-dim 2048 \
+        --rnn-lm-hidden-dim 2048 \
+        --rnn-lm-num-layers 3 \
+        --lm-vocab-size 500
+
+You should see the following WERs after executing the commands above:
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	2.9	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	7.57	best for test-other
+
+It's slightly better than LM rescoring. If we further increase the beam size, we will see
+further improvements from LM rescoring + LODR:
+
+.. list-table:: WERs of LM rescoring + LODR with different beam sizes
+   :widths: 25 25 25
+   :header-rows: 1
+
+   * - Beam size
+     - test-clean
+     - test-other
+   * - 4
+     - 2.9
+     - 7.57
+   * - 8
+     - 2.63
+     - 7.04
+   * - 12
+     - 2.52
+     - 6.73
+
+As mentioned earlier, LM rescoring is usually faster than shallow-fusion based methods.
+Here, we benchmark the WERs and decoding speed of them:
+
+.. list-table:: LM-rescoring-based methods vs shallow-fusion-based methods (The numbers in each field is WER on test-clean, WER on test-other and decoding time on test-clean)
+   :widths: 25 25 25 25
+   :header-rows: 1
+
+   * - Decoding method
+     - beam=4
+     - beam=8
+     - beam=12
+   * - ``modified_beam_search``
+     - 3.11/7.93; 132s
+     - 3.1/7.95; 177s
+     - 3.1/7.96; 210s
+   * - ``modified_beam_search_lm_shallow_fusion``
+     - 2.77/7.08; 262s
+     - 2.62/6.65; 352s
+     - 2.58/6.65; 488s
+   * - ``modified_beam_search_LODR``
+     - 2.61/6.74; 400s
+     - 2.45/6.38; 610s
+     - 2.4/6.23; 870s
+   * - ``modified_beam_search_lm_rescore``
+     - 2.93/7.6; 156s
+     - 2.67/7.11; 203s
+     - 2.59/6.86; 255s
+   * - ``modified_beam_search_lm_rescore_LODR``
+     - 2.9/7.57; 160s
+     - 2.63/7.04; 203s
+     - 2.52/6.73; 263s
+
+.. note::
+
+    Decoding is performed with a single 32G V100, we set ``--max-duration`` to 600. 
+    Decoding time here is only for reference and it may vary.
\ No newline at end of file
diff --git a/docs/source/decoding-with-langugage-models/shallow-fusion.rst b/docs/source/decoding-with-langugage-models/shallow-fusion.rst
new file mode 100644
index 000000000..684fefeb4
--- /dev/null
+++ b/docs/source/decoding-with-langugage-models/shallow-fusion.rst
@@ -0,0 +1,179 @@
+.. _shallow_fusion:
+
+Shallow fusion for Transducer
+=================================
+
+External language models (LM) are commonly used to improve WERs for E2E ASR models.
+This tutorial shows you how to perform ``shallow fusion`` with an external LM
+to improve the word-error-rate of a transducer model.
+
+.. note::
+
+    This tutorial is based on the recipe 
+    `pruned_transducer_stateless7_streaming <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming>`_,
+    which is a streaming transducer model trained on `LibriSpeech`_. 
+    However, you can easily apply shallow fusion to other recipes.
+    If you encounter any problems, please open an issue here `icefall <https://github.com/k2-fsa/icefall/issues>`_.
+
+.. note::
+
+    For simplicity, the training and testing corpus in this tutorial is the same (`LibriSpeech`_). However, you can change the testing set
+    to any other domains (e.g `GigaSpeech`_) and use an external LM trained on that domain.
+
+.. HINT::
+
+  We recommend you to use a GPU for decoding.
+
+For illustration purpose, we will use a pre-trained ASR model from this `link <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>`__.
+If you want to train your model from scratch, please have a look at :ref:`non_streaming_librispeech_pruned_transducer_stateless`.
+
+As the initial step, let's download the pre-trained model.
+
+.. code-block:: bash
+
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+    $ cd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ git lfs pull --include "pretrained.pt"
+    $ ln -s pretrained.pt epoch-99.pt # create a symbolic link so that the checkpoint can be loaded
+    $ cd ../data/lang_bpe_500
+    $ git lfs pull --include bpe.model
+    $ cd ../../..
+
+To test the model, let's have a look at the decoding results without using LM. This can be done via the following command:
+
+.. code-block:: bash
+
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --exp-dir $exp_dir \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search
+
+The following WERs are achieved on test-clean and test-other:
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	3.11	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	7.93	best for test-other
+
+These are already good numbers! But we can further improve it by using shallow fusion with external LM.
+Training a language model usually takes a long time, we can download a pre-trained LM from this `link <https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm>`__.
+
+.. code-block:: bash
+
+    $ # download the external LM
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm 
+    $ # create a symbolic link so that the checkpoint can be loaded
+    $ pushd icefall-librispeech-rnn-lm/exp
+    $ git lfs pull --include "pretrained.pt"
+    $ ln -s pretrained.pt epoch-99.pt 
+    $ popd
+
+.. note::
+
+    This is an RNN LM trained on the LibriSpeech text corpus. So it might not be ideal for other corpus.
+    You may also train a RNN LM from scratch. Please refer to this `script <https://github.com/k2-fsa/icefall/blob/master/icefall/rnn_lm/train.py>`__
+    for training a RNN LM and this `script <https://github.com/k2-fsa/icefall/blob/master/icefall/transformer_lm/train.py>`__ to train a transformer LM.
+
+To use shallow fusion for decoding, we can execute the following command:
+
+.. code-block:: bash
+    
+    $ exp_dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+    $ lm_dir=./icefall-librispeech-rnn-lm/exp
+    $ lm_scale=0.29
+    $ ./pruned_transducer_stateless7_streaming/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --use-averaged-model False \
+        --beam-size 4 \
+        --exp-dir $exp_dir \
+        --max-duration 600 \
+        --decode-chunk-len 32 \
+        --decoding-method modified_beam_search_lm_shallow_fusion \
+        --bpe-model ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/bpe.model \
+        --use-shallow-fusion 1 \
+        --lm-type rnn \
+        --lm-exp-dir $lm_dir \
+        --lm-epoch 99 \
+        --lm-scale $lm_scale \
+        --lm-avg 1 \
+        --rnn-lm-embedding-dim 2048 \
+        --rnn-lm-hidden-dim 2048 \
+        --rnn-lm-num-layers 3 \
+        --lm-vocab-size 500
+
+Note that we set ``--decoding-method modified_beam_search_lm_shallow_fusion`` and ``--use-shallow-fusion True``
+to use shallow fusion. ``--lm-type`` specifies the type of neural LM we are going to use, you can either choose
+between ``rnn`` or ``transformer``. The following three arguments are associated with the rnn:
+
+- ``--rnn-lm-embedding-dim``
+    The embedding dimension of the RNN LM
+
+- ``--rnn-lm-hidden-dim``
+    The hidden dimension of the RNN LM
+
+- ``--rnn-lm-num-layers``
+    The number of RNN layers in the RNN LM.
+
+
+The decoding result obtained with the above command are shown below.
+
+.. code-block:: text
+
+    $ For test-clean, WER of different settings are:
+    $ beam_size_4	2.77	best for test-clean
+    $ For test-other, WER of different settings are:
+    $ beam_size_4	7.08	best for test-other
+
+The improvement of shallow fusion is very obvious! The relative WER reduction on test-other is around 10.5%. 
+A few parameters can be tuned to further boost the performance of shallow fusion:
+
+- ``--lm-scale`` 
+
+    Controls the scale of the LM. If too small, the external language model may not be fully utilized; if too large, 
+    the LM score may dominant during decoding, leading to bad WER. A typical value of this is around 0.3.
+
+- ``--beam-size`` 
+    
+    The number of active paths in the search beam. It controls the trade-off between decoding efficiency and accuracy.
+
+Here, we also show how `--beam-size` effect the WER and decoding time:
+
+.. list-table:: WERs and decoding time (on test-clean) of shallow fusion with different beam sizes
+   :widths: 25 25 25 25
+   :header-rows: 1
+
+   * - Beam size
+     - test-clean
+     - test-other
+     - Decoding time on test-clean (s)
+   * - 4
+     - 2.77
+     - 7.08
+     - 262
+   * - 8
+     - 2.62
+     - 6.65
+     - 352
+   * - 12
+     - 2.58
+     - 6.65
+     - 488
+
+As we see, a larger beam size during shallow fusion improves the WER, but is also slower.
+
+
+
+
+
+
+
+ 
diff --git a/docs/source/docker/img/docker-hub.png b/docs/source/docker/img/docker-hub.png
new file mode 100644
index 000000000..a9e7715b0
Binary files /dev/null and b/docs/source/docker/img/docker-hub.png differ
diff --git a/docs/source/docker/index.rst b/docs/source/docker/index.rst
new file mode 100644
index 000000000..2c92a4cbc
--- /dev/null
+++ b/docs/source/docker/index.rst
@@ -0,0 +1,17 @@
+.. _icefall_docker:
+
+Docker
+======
+
+This section describes how to use pre-built docker images to run `icefall`_.
+
+.. hint::
+
+   If you only have CPUs available, you can still use the pre-built docker
+   images.
+
+.. toctree::
+   :maxdepth: 2
+
+   ./intro.rst
+
diff --git a/docs/source/docker/intro.rst b/docs/source/docker/intro.rst
new file mode 100644
index 000000000..9ead0df00
--- /dev/null
+++ b/docs/source/docker/intro.rst
@@ -0,0 +1,173 @@
+Introduction
+=============
+
+We have pre-built docker images hosted at the following address:
+
+  `<https://hub.docker.com/repository/docker/k2fsa/icefall/general>`_
+
+.. figure:: img/docker-hub.png
+   :width: 600
+   :align: center
+
+You can find the ``Dockerfile`` at `<https://github.com/k2-fsa/icefall/tree/master/docker>`_.
+
+We describe the following items in this section:
+
+  - How to view available tags
+  - How to download pre-built docker images
+  - How to run the `yesno`_ recipe within a docker container on ``CPU``
+
+View available tags
+===================
+
+You can use the following command to view available tags:
+
+.. code-block:: bash
+
+   curl -s 'https://registry.hub.docker.com/v2/repositories/k2fsa/icefall/tags/'|jq '."results"[]["name"]'
+
+which will give you something like below:
+
+.. code-block:: bash
+
+  "torch2.1.0-cuda12.1"
+  "torch2.1.0-cuda11.8"
+  "torch2.0.0-cuda11.7"
+  "torch1.12.1-cuda11.3"
+  "torch1.9.0-cuda10.2"
+  "torch1.13.0-cuda11.6"
+
+.. hint::
+
+   Available tags will be updated when there are new releases of `torch`_.
+
+Please select an appropriate combination of `torch`_ and  CUDA.
+
+Download a docker image
+=======================
+
+Suppose that you select the tag ``torch1.13.0-cuda11.6``, you can use
+the following command to download it:
+
+.. code-block:: bash
+
+   sudo docker image pull k2fsa/icefall:torch1.13.0-cuda11.6
+
+Run a docker image with GPU
+===========================
+
+.. code-block:: bash
+
+  sudo docker run --gpus all --rm -it k2fsa/icefall:torch1.13.0-cuda11.6 /bin/bash
+
+Run a docker image with CPU
+===========================
+
+.. code-block:: bash
+
+  sudo docker run --rm -it k2fsa/icefall:torch1.13.0-cuda11.6 /bin/bash
+
+Run yesno within a docker container
+===================================
+
+After starting the container, the following interface is presented:
+
+.. code-block:: bash
+
+  root@60c947eac59c:/workspace/icefall#
+
+It shows the current user is ``root`` and the current working directory
+is ``/workspace/icefall``.
+
+Update the code
+---------------
+
+Please first run:
+
+.. code-block:: bash
+
+  root@60c947eac59c:/workspace/icefall# git pull
+
+so that your local copy contains the latest code.
+
+Data preparation
+----------------
+
+Now we can use
+
+.. code-block:: bash
+
+  root@60c947eac59c:/workspace/icefall# cd egs/yesno/ASR/
+
+to switch to the ``yesno`` recipe and run
+
+.. code-block:: bash
+
+  root@60c947eac59c:/workspace/icefall/egs/yesno/ASR# ./prepare.sh
+
+.. hint::
+
+   If you are running without GPU, it may report the following error:
+
+    .. code-block:: bash
+
+        File "/opt/conda/lib/python3.9/site-packages/k2/__init__.py", line 23, in <module>
+          from _k2 import DeterminizeWeightPushingType
+        ImportError: libcuda.so.1: cannot open shared object file: No such file or directory
+
+  We can use the following command to fix it:
+
+    .. code-block:: bash
+
+      root@60c947eac59c:/workspace/icefall/egs/yesno/ASR# ln -s /opt/conda/lib/stubs/libcuda.so /opt/conda/lib/stubs/libcuda.so.1
+
+The logs of running ``./prepare.sh`` are listed below:
+
+.. literalinclude:: ./log/log-preparation.txt
+
+Training
+--------
+
+After preparing the data, we can start training with the following command
+
+.. code-block:: bash
+
+  root@60c947eac59c:/workspace/icefall/egs/yesno/ASR# ./tdnn/train.py
+
+All of the training logs are given below:
+
+.. hint::
+
+   It is running on CPU and it takes only 16 seconds for this run.
+
+.. literalinclude:: ./log/log-train-2023-08-01-01-55-27
+
+
+Decoding
+--------
+
+After training, we can decode the trained model with
+
+.. code-block:: bash
+
+  root@60c947eac59c:/workspace/icefall/egs/yesno/ASR# ./tdnn/decode.py
+
+The decoding logs are given below:
+
+.. code-block:: bash
+
+  2023-08-01 02:06:22,400 INFO [decode.py:263] Decoding started
+  2023-08-01 02:06:22,400 INFO [decode.py:264] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lm_dir': PosixPath('data/lm'), 'feature_dim': 23, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'epoch': 14, 'avg': 2, 'export': False, 'feature_dir': PosixPath('data/fbank'), 'max_duration': 30.0, 'bucketing_sampler': False, 'num_buckets': 10, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': False, 'return_cuts': True, 'num_workers': 2, 'env_info': {'k2-version': '1.24.3', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '4c05309499a08454997adf500b56dcc629e35ae5', 'k2-git-date': 'Tue Jul 25 16:23:36 2023', 'lhotse-version': '1.16.0.dev+git.7640d663.clean', 'torch-version': '1.13.0', 'torch-cuda-available': False, 'torch-cuda-version': '11.6', 'python-version': '3.9', 'icefall-git-branch': 'master', 'icefall-git-sha1': '375520d-clean', 'icefall-git-date': 'Fri Jul 28 07:43:08 2023', 'icefall-path': '/workspace/icefall', 'k2-path': '/opt/conda/lib/python3.9/site-packages/k2/__init__.py', 'lhotse-path': '/opt/conda/lib/python3.9/site-packages/lhotse/__init__.py', 'hostname': '60c947eac59c', 'IP address': '172.17.0.2'}}
+  2023-08-01 02:06:22,401 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+  2023-08-01 02:06:22,403 INFO [decode.py:273] device: cpu
+  2023-08-01 02:06:22,406 INFO [decode.py:291] averaging ['tdnn/exp/epoch-13.pt', 'tdnn/exp/epoch-14.pt']
+  2023-08-01 02:06:22,424 INFO [asr_datamodule.py:218] About to get test cuts
+  2023-08-01 02:06:22,425 INFO [asr_datamodule.py:252] About to get test cuts
+  2023-08-01 02:06:22,504 INFO [decode.py:204] batch 0/?, cuts processed until now is 4
+  [W NNPACK.cpp:53] Could not initialize NNPACK! Reason: Unsupported hardware.
+  2023-08-01 02:06:22,687 INFO [decode.py:241] The transcripts are stored in tdnn/exp/recogs-test_set.txt
+  2023-08-01 02:06:22,688 INFO [utils.py:564] [test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
+  2023-08-01 02:06:22,690 INFO [decode.py:249] Wrote detailed error stats to tdnn/exp/errs-test_set.txt
+  2023-08-01 02:06:22,690 INFO [decode.py:316] Done!
+
+Congratulations! You have finished successfully running `icefall`_ within a docker container.
diff --git a/docs/source/faqs.rst b/docs/source/faqs.rst
new file mode 100644
index 000000000..72b0302d7
--- /dev/null
+++ b/docs/source/faqs.rst
@@ -0,0 +1,107 @@
+Frequently Asked Questions (FAQs)
+=================================
+
+In this section, we collect issues reported by users and post the corresponding
+solutions.
+
+
+OSError: libtorch_hip.so: cannot open shared object file: no such file or directory
+-----------------------------------------------------------------------------------
+
+One user is using the following code to install ``torch`` and ``torchaudio``:
+
+.. code-block:: bash
+
+  pip install \
+    torch==1.10.0+cu111 \
+    torchvision==0.11.0+cu111 \
+    torchaudio==0.10.0 \
+    -f https://download.pytorch.org/whl/torch_stable.html
+
+and it throws the following error when running ``tdnn/train.py``:
+
+.. code-block::
+
+  OSError: libtorch_hip.so: cannot open shared object file: no such file or directory
+
+The fix is to specify the CUDA version while installing ``torchaudio``. That
+is, change ``torchaudio==0.10.0`` to ``torchaudio==0.10.0+cu11```. Therefore,
+the correct command is:
+
+.. code-block:: bash
+
+  pip install \
+    torch==1.10.0+cu111 \
+    torchvision==0.11.0+cu111 \
+    torchaudio==0.10.0+cu111 \
+    -f https://download.pytorch.org/whl/torch_stable.html
+
+AttributeError: module 'distutils' has no attribute 'version'
+-------------------------------------------------------------
+
+The error log is:
+
+.. code-block::
+
+  Traceback (most recent call last):
+    File "./tdnn/train.py", line 14, in <module>
+      from asr_datamodule import YesNoAsrDataModule
+    File "/home/xxx/code/next-gen-kaldi/icefall/egs/yesno/ASR/tdnn/asr_datamodule.py", line 34, in <module>
+      from icefall.dataset.datamodule import DataModule
+    File "/home/xxx/code/next-gen-kaldi/icefall/icefall/__init__.py", line 3, in <module>
+      from . import (
+    File "/home/xxx/code/next-gen-kaldi/icefall/icefall/decode.py", line 23, in <module>
+      from icefall.utils import add_eos, add_sos, get_texts
+    File "/home/xxx/code/next-gen-kaldi/icefall/icefall/utils.py", line 39, in <module>
+      from torch.utils.tensorboard import SummaryWriter
+    File "/home/xxx/tool/miniconda3/envs/yyy/lib/python3.8/site-packages/torch/utils/tensorboard/__init__.py", line 4, in <module>
+      LooseVersion = distutils.version.LooseVersion
+  AttributeError: module 'distutils' has no attribute 'version'
+
+The fix is:
+
+.. code-block:: bash
+
+  pip uninstall setuptools
+
+  pip install setuptools==58.0.4
+
+ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
+--------------------------------------------------------------------------------------------
+
+If you are using ``conda`` and encounter the following issue:
+
+.. code-block::
+
+  Traceback (most recent call last):
+    File "/k2-dev/yangyifan/anaconda3/envs/icefall/lib/python3.10/site-packages/k2-1.23.3.dev20230112+cuda11.6.torch1.13.1-py3.10-linux-x86_64.egg/k2/__init__.py", line 24, in <module>
+      from _k2 import DeterminizeWeightPushingType
+  ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
+
+  During handling of the above exception, another exception occurred:
+
+  Traceback (most recent call last):
+    File "/k2-dev/yangyifan/icefall/egs/librispeech/ASR/./pruned_transducer_stateless7_ctc_bs/decode.py", line 104, in <module>
+      import k2
+    File "/k2-dev/yangyifan/anaconda3/envs/icefall/lib/python3.10/site-packages/k2-1.23.3.dev20230112+cuda11.6.torch1.13.1-py3.10-linux-x86_64.egg/k2/__init__.py", line 30, in <module>
+      raise ImportError(
+  ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
+  Note: If you're using anaconda and importing k2 on MacOS,
+        you can probably fix this by setting the environment variable:
+    export DYLD_LIBRARY_PATH=$CONDA_PREFIX/lib/python3.10/site-packages:$DYLD_LIBRARY_PATH
+
+Please first try to find where ``libpython3.10.so.1.0`` locates.
+
+For instance,
+
+.. code-block:: bash
+
+  cd $CONDA_PREFIX/lib
+  find . -name "libpython*"
+
+If you are able to find it inside ``$CODNA_PREFIX/lib``, please set the
+following environment variable:
+
+.. code-block:: bash
+
+  export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
diff --git a/docs/source/for-dummies/data-preparation.rst b/docs/source/for-dummies/data-preparation.rst
new file mode 100644
index 000000000..f03d44e79
--- /dev/null
+++ b/docs/source/for-dummies/data-preparation.rst
@@ -0,0 +1,180 @@
+.. _dummies_tutorial_data_preparation:
+
+Data Preparation
+================
+
+After :ref:`dummies_tutorial_environment_setup`, we can start preparing the
+data for training and decoding.
+
+The first step is to prepare the data for training. We have already provided
+`prepare.sh <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/prepare.sh>`_
+that would prepare everything required for training.
+
+.. code-block::
+
+   cd /tmp/icefall
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+   cd egs/yesno/ASR
+
+   ./prepare.sh
+
+Note that in each recipe from `icefall`_, there exists a file ``prepare.sh``,
+which you should run before you run anything else.
+
+That is all you need for data preparation.
+
+For the more curious
+--------------------
+
+If you are wondering how to prepare your own dataset, please refer to the following
+URLs for more details:
+
+  - `<https://github.com/lhotse-speech/lhotse/tree/master/lhotse/recipes>`_
+
+    It contains recipes for a variety of dataset. If you want to add your own
+    dataset, please read recipes in this folder first.
+
+  - `<https://github.com/lhotse-speech/lhotse/blob/master/lhotse/recipes/yesno.py>`_
+
+    The `yesno`_ recipe in `lhotse`_.
+
+If you already have a `Kaldi`_ dataset directory, which contains files like
+``wav.scp``, ``feats.scp``, then you can refer to `<https://lhotse.readthedocs.io/en/latest/kaldi.html#example>`_.
+
+A quick look to the generated files
+-----------------------------------
+
+``./prepare.sh`` puts generated files into two directories:
+
+  - ``download``
+  - ``data``
+
+download
+^^^^^^^^
+
+The ``download`` directory contains downloaded dataset files:
+
+.. code-block:: bas
+
+    tree -L 1 ./download/
+
+    ./download/
+    |-- waves_yesno
+    `-- waves_yesno.tar.gz
+
+.. hint::
+
+   Please refer to `<https://github.com/lhotse-speech/lhotse/blob/master/lhotse/recipes/yesno.py#L41>`_
+   for how the data is downloaded and extracted.
+
+data
+^^^^
+
+.. code-block:: bash
+
+    tree ./data/
+
+    ./data/
+    |-- fbank
+    |   |-- yesno_cuts_test.jsonl.gz
+    |   |-- yesno_cuts_train.jsonl.gz
+    |   |-- yesno_feats_test.lca
+    |   `-- yesno_feats_train.lca
+    |-- lang_phone
+    |   |-- HLG.pt
+    |   |-- L.pt
+    |   |-- L_disambig.pt
+    |   |-- Linv.pt
+    |   |-- lexicon.txt
+    |   |-- lexicon_disambig.txt
+    |   |-- tokens.txt
+    |   `-- words.txt
+    |-- lm
+    |   |-- G.arpa
+    |   `-- G.fst.txt
+    `-- manifests
+        |-- yesno_recordings_test.jsonl.gz
+        |-- yesno_recordings_train.jsonl.gz
+        |-- yesno_supervisions_test.jsonl.gz
+        `-- yesno_supervisions_train.jsonl.gz
+
+    4 directories, 18 files
+
+**data/manifests**:
+
+  This directory contains manifests. They are used to generate files in
+  ``data/fbank``.
+
+  To give you an idea of what it contains, we examine the first few lines of
+  the manifests related to the ``train`` dataset.
+
+  .. code-block:: bash
+
+      cd data/manifests
+      gunzip -c  yesno_recordings_train.jsonl.gz  | head -n 3
+
+  The output is given below:
+
+    .. code-block:: bash
+
+      {"id": "0_0_0_0_1_1_1_1", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_0_0_1_1_1_1.wav"}], "sampling_rate": 8000, "num_samples": 50800, "duration": 6.35, "channel_ids": [0]}
+      {"id": "0_0_0_1_0_1_1_0", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_0_1_0_1_1_0.wav"}], "sampling_rate": 8000, "num_samples": 48880, "duration": 6.11, "channel_ids": [0]}
+      {"id": "0_0_1_0_0_1_1_0", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_1_0_0_1_1_0.wav"}], "sampling_rate": 8000, "num_samples": 48160, "duration": 6.02, "channel_ids": [0]}
+
+  Please refer to `<https://github.com/lhotse-speech/lhotse/blob/master/lhotse/audio.py#L300>`_
+  for the meaning of each field per line.
+
+  .. code-block:: bash
+
+      gunzip -c  yesno_supervisions_train.jsonl.gz  | head -n 3
+
+  The output is given below:
+
+  .. code-block:: bash
+
+      {"id": "0_0_0_0_1_1_1_1", "recording_id": "0_0_0_0_1_1_1_1", "start": 0.0, "duration": 6.35, "channel": 0, "text": "NO NO NO NO YES YES YES YES", "language": "Hebrew"}
+      {"id": "0_0_0_1_0_1_1_0", "recording_id": "0_0_0_1_0_1_1_0", "start": 0.0, "duration": 6.11, "channel": 0, "text": "NO NO NO YES NO YES YES NO", "language": "Hebrew"}
+      {"id": "0_0_1_0_0_1_1_0", "recording_id": "0_0_1_0_0_1_1_0", "start": 0.0, "duration": 6.02, "channel": 0, "text": "NO NO YES NO NO YES YES NO", "language": "Hebrew"}
+
+  Please refer to `<https://github.com/lhotse-speech/lhotse/blob/master/lhotse/supervision.py#L510>`_
+  for the meaning of each field per line.
+
+**data/fbank**:
+
+  This directory contains everything from ``data/manifests``. Furthermore, it also contains features
+  for training.
+
+  ``data/fbank/yesno_feats_train.lca`` contains the features for the train dataset.
+  Features are compressed using `lilcom`_.
+
+  ``data/fbank/yesno_cuts_train.jsonl.gz`` stores the `CutSet <https://github.com/lhotse-speech/lhotse/blob/master/lhotse/cut/set.py#L72>`_,
+  which stores `RecordingSet <https://github.com/lhotse-speech/lhotse/blob/master/lhotse/audio.py#L928>`_,
+  `SupervisionSet <https://github.com/lhotse-speech/lhotse/blob/master/lhotse/supervision.py#L510>`_,
+  and `FeatureSet <https://github.com/lhotse-speech/lhotse/blob/master/lhotse/features/base.py#L593>`_.
+
+  To give you an idea about what it looks like, we can run the following command:
+
+    .. code-block:: bash
+
+        cd data/fbank
+
+        gunzip -c yesno_cuts_train.jsonl.gz | head -n 3
+
+  The output is given below:
+
+    .. code-block:: bash
+
+      {"id": "0_0_0_0_1_1_1_1-0", "start": 0, "duration": 6.35, "channel": 0, "supervisions": [{"id": "0_0_0_0_1_1_1_1", "recording_id": "0_0_0_0_1_1_1_1", "start": 0.0, "duration": 6.35, "channel": 0, "text": "NO NO NO NO YES YES YES YES", "language": "Hebrew"}], "features": {"type": "kaldi-fbank", "num_frames": 635, "num_features": 23, "frame_shift": 0.01, "sampling_rate": 8000, "start": 0, "duration": 6.35, "storage_type": "lilcom_chunky", "storage_path": "data/fbank/yesno_feats_train.lca", "storage_key": "0,13000,3570", "channels": 0}, "recording": {"id": "0_0_0_0_1_1_1_1", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_0_0_1_1_1_1.wav"}], "sampling_rate": 8000, "num_samples": 50800, "duration": 6.35, "channel_ids": [0]}, "type": "MonoCut"}
+      {"id": "0_0_0_1_0_1_1_0-1", "start": 0, "duration": 6.11, "channel": 0, "supervisions": [{"id": "0_0_0_1_0_1_1_0", "recording_id": "0_0_0_1_0_1_1_0", "start": 0.0, "duration": 6.11, "channel": 0, "text": "NO NO NO YES NO YES YES NO", "language": "Hebrew"}], "features": {"type": "kaldi-fbank", "num_frames": 611, "num_features": 23, "frame_shift": 0.01, "sampling_rate": 8000, "start": 0, "duration": 6.11, "storage_type": "lilcom_chunky", "storage_path": "data/fbank/yesno_feats_train.lca", "storage_key": "16570,12964,2929", "channels": 0}, "recording": {"id": "0_0_0_1_0_1_1_0", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_0_1_0_1_1_0.wav"}], "sampling_rate": 8000, "num_samples": 48880, "duration": 6.11, "channel_ids": [0]}, "type": "MonoCut"}
+      {"id": "0_0_1_0_0_1_1_0-2", "start": 0, "duration": 6.02, "channel": 0, "supervisions": [{"id": "0_0_1_0_0_1_1_0", "recording_id": "0_0_1_0_0_1_1_0", "start": 0.0, "duration": 6.02, "channel": 0, "text": "NO NO YES NO NO YES YES NO", "language": "Hebrew"}], "features": {"type": "kaldi-fbank", "num_frames": 602, "num_features": 23, "frame_shift": 0.01, "sampling_rate": 8000, "start": 0, "duration": 6.02, "storage_type": "lilcom_chunky", "storage_path": "data/fbank/yesno_feats_train.lca", "storage_key": "32463,12936,2696", "channels": 0}, "recording": {"id": "0_0_1_0_0_1_1_0", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_1_0_0_1_1_0.wav"}], "sampling_rate": 8000, "num_samples": 48160, "duration": 6.02, "channel_ids": [0]}, "type": "MonoCut"}
+
+  Note that ``yesno_cuts_train.jsonl.gz`` only stores the information about how to read the features.
+  The actual features are stored separately in ``data/fbank/yesno_feats_train.lca``.
+
+**data/lang**:
+
+  This directory contains the lexicon.
+
+**data/lm**:
+
+  This directory contains language models.
diff --git a/docs/source/for-dummies/decoding.rst b/docs/source/for-dummies/decoding.rst
new file mode 100644
index 000000000..3e48e8bfd
--- /dev/null
+++ b/docs/source/for-dummies/decoding.rst
@@ -0,0 +1,39 @@
+.. _dummies_tutorial_decoding:
+
+Decoding
+========
+
+After :ref:`dummies_tutorial_training`, we can start decoding.
+
+The command to start the decoding is quite simple:
+
+.. code-block:: bash
+
+   cd /tmp/icefall
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+   cd egs/yesno/ASR
+
+   # We use CPU for decoding by setting the following environment variable
+   export CUDA_VISIBLE_DEVICES=""
+
+   ./tdnn/decode.py
+
+The output logs are given below:
+
+.. literalinclude:: ./code/decoding-yesno.txt
+
+For the more curious
+--------------------
+
+.. code-block:: bash
+
+   ./tdnn/decode.py --help
+
+will print the usage information about ``./tdnn/decode.py``. For instance, you
+can specify:
+
+  - ``--epoch`` to use which checkpoint for decoding
+  - ``--avg`` to select how many checkpoints to use for model averaging
+
+You usually try different combinations of ``--epoch`` and ``--avg`` and select
+one that leads to the lowest WER (`Word Error Rate <https://en.wikipedia.org/wiki/Word_error_rate>`_).
diff --git a/docs/source/for-dummies/environment-setup.rst b/docs/source/for-dummies/environment-setup.rst
new file mode 100644
index 000000000..0cb8ecc1d
--- /dev/null
+++ b/docs/source/for-dummies/environment-setup.rst
@@ -0,0 +1,121 @@
+.. _dummies_tutorial_environment_setup:
+
+Environment setup
+=================
+
+We will create an environment for `Next-gen Kaldi`_ that runs on ``CPU``
+in this tutorial.
+
+.. note::
+
+   Since the `yesno`_ dataset used in this tutorial is very tiny, training on
+   ``CPU`` works very well for it.
+
+   If your dataset is very large, e.g., hundreds or thousands of hours of
+   training data, please follow :ref:`install icefall` to install `icefall`_
+   that works with ``GPU``.
+
+
+Create a virtual environment
+----------------------------
+
+.. code-block:: bash
+
+  virtualenv -p python3 /tmp/icefall_env
+
+The above command creates a virtual environment in the directory ``/tmp/icefall_env``.
+You can select any directory you want.
+
+The output of the above command is given below:
+
+.. code-block:: bash
+
+  Already using interpreter /usr/bin/python3
+  Using base prefix '/usr'
+  New python executable in /tmp/icefall_env/bin/python3
+  Also creating executable in /tmp/icefall_env/bin/python
+  Installing setuptools, pkg_resources, pip, wheel...done.
+
+Now we can activate the environment using:
+
+.. code-block:: bash
+
+  source /tmp/icefall_env/bin/activate
+
+Install dependencies
+--------------------
+
+.. warning::
+
+   Remeber to activate your virtual environment before you continue!
+
+After activating the virtual environment, we can use the following command
+to install dependencies of `icefall`_:
+
+.. hint::
+
+   Remeber that we will run this tutorial on ``CPU``, so we install
+   dependencies required only by running on ``CPU``.
+
+.. code-block:: bash
+
+   # Caution: Installation order matters!
+
+   # We use torch 2.0.0 and torchaduio 2.0.0 in this tutorial.
+   # Other versions should also work.
+
+   pip install torch==2.0.0+cpu torchaudio==2.0.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
+
+   # If you are using macOS or Windows, please use the following command to install torch and torchaudio
+   # pip install torch==2.0.0 torchaudio==2.0.0 -f https://download.pytorch.org/whl/torch_stable.html
+
+   # Now install k2
+   # Please refer to https://k2-fsa.github.io/k2/installation/from_wheels.html#linux-cpu-example
+
+   pip install k2==1.24.3.dev20230726+cpu.torch2.0.0 -f https://k2-fsa.github.io/k2/cpu.html
+
+   # Install the latest version of lhotse
+
+   pip install git+https://github.com/lhotse-speech/lhotse
+
+
+Install icefall
+---------------
+
+We will put the source code of `icefall`_ into the directory ``/tmp``
+You can select any directory you want.
+
+.. code-block:: bash
+
+   cd /tmp
+   git clone https://github.com/k2-fsa/icefall
+   cd icefall
+   pip install -r ./requirements.txt
+
+.. code-block:: bash
+
+   # Anytime we want to use icefall, we have to set the following
+   # environment variable
+
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+
+.. hint::
+
+   If you get the following error during this tutorial:
+
+    .. code-block:: bash
+
+      ModuleNotFoundError: No module named 'icefall'
+
+  please set the above environment variable to fix it.
+
+
+Congratulations! You have installed `icefall`_ successfully.
+
+For the more curious
+--------------------
+
+`icefall`_ contains a collection of Python scripts and you don't need to
+use ``python3 setup.py install`` or ``pip install icefall`` to install it.
+All you need to do is to download the code and set the environment variable
+``PYTHONPATH``.
diff --git a/docs/source/for-dummies/index.rst b/docs/source/for-dummies/index.rst
new file mode 100644
index 000000000..7c0a3d8ee
--- /dev/null
+++ b/docs/source/for-dummies/index.rst
@@ -0,0 +1,34 @@
+Icefall for dummies tutorial
+============================
+
+This tutorial walks you step by step about how to create a simple
+ASR (`Automatic Speech Recognition <https://en.wikipedia.org/wiki/Speech_recognition>`_)
+system with `Next-gen Kaldi`_.
+
+We use the `yesno`_ dataset for demonstration. We select it out of two reasons:
+
+  - It is quite tiny, containing only about 12 minutes of data
+  - The training can be finished within 20 seconds on ``CPU``.
+
+That also means you don't need a ``GPU`` to run this tutorial.
+
+Let's get started!
+
+Please follow items below **sequentially**.
+
+.. note::
+
+   The :ref:`dummies_tutorial_data_preparation` runs only on Linux and on macOS.
+   All other parts run on Linux, macOS, and Windows.
+
+   Help from the community is appreciated to port the :ref:`dummies_tutorial_data_preparation`
+   to Windows.
+
+.. toctree::
+   :maxdepth: 2
+
+   ./environment-setup.rst
+   ./data-preparation.rst
+   ./training.rst
+   ./decoding.rst
+   ./model-export.rst
diff --git a/docs/source/for-dummies/model-export.rst b/docs/source/for-dummies/model-export.rst
new file mode 100644
index 000000000..079ebc712
--- /dev/null
+++ b/docs/source/for-dummies/model-export.rst
@@ -0,0 +1,310 @@
+Model Export
+============
+
+There are three ways to export a pre-trained model.
+
+  - Export the model parameters via `model.state_dict() <https://pytorch.org/docs/stable/generated/torch.nn.Module.html?highlight=load_state_dict#torch.nn.Module.state_dict>`_
+  - Export via `torchscript <https://pytorch.org/docs/stable/jit.html>`_: either `torch.jit.script() <https://pytorch.org/docs/stable/generated/torch.jit.script.html#torch.jit.script>`_ or `torch.jit.trace() <https://pytorch.org/docs/stable/generated/torch.jit.trace.html>`_
+  - Export to `ONNX`_ via `torch.onnx.export() <https://pytorch.org/docs/stable/onnx.html>`_
+
+Each method is explained below in detail.
+
+Export the model parameters via model.state_dict()
+---------------------------------------------------
+
+The command for this kind of export is
+
+.. code-block:: bash
+
+   cd /tmp/icefall
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+   cd egs/yesno/ASR
+
+   # assume that "--epoch 14 --avg 2" produces the lowest WER.
+
+   ./tdnn/export.py --epoch 14 --avg 2
+
+The output logs are given below:
+
+.. code-block:: bash
+
+  2023-08-16 20:42:03,912 INFO [export.py:76] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lr': 0.01, 'feature_dim': 23, 'weight_decay': 1e-06, 'start_epoch': 0, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 10, 'reset_interval': 20, 'valid_interval': 10, 'beam_size': 10, 'reduction': 'sum', 'use_double_scores': True, 'epoch': 14, 'avg': 2, 'jit': False}
+  2023-08-16 20:42:03,913 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+  2023-08-16 20:42:03,950 INFO [export.py:93] averaging ['tdnn/exp/epoch-13.pt', 'tdnn/exp/epoch-14.pt']
+  2023-08-16 20:42:03,971 INFO [export.py:106] Not using torch.jit.script
+  2023-08-16 20:42:03,974 INFO [export.py:111] Saved to tdnn/exp/pretrained.pt
+
+We can see from the logs that the exported model is saved to the file ``tdnn/exp/pretrained.pt``.
+
+To give you an idea of what ``tdnn/exp/pretrained.pt`` contains, we can use the following command:
+
+.. code-block:: python3
+
+    >>> import torch
+    >>> m = torch.load("tdnn/exp/pretrained.pt")
+    >>> list(m.keys())
+    ['model']
+    >>> list(m["model"].keys())
+    ['tdnn.0.weight', 'tdnn.0.bias', 'tdnn.2.running_mean', 'tdnn.2.running_var', 'tdnn.2.num_batches_tracked', 'tdnn.3.weight', 'tdnn.3.bias', 'tdnn.5.running_mean', 'tdnn.5.running_var', 'tdnn.5.num_batches_tracked', 'tdnn.6.weight', 'tdnn.6.bias', 'tdnn.8.running_mean', 'tdnn.8.running_var', 'tdnn.8.num_batches_tracked', 'output_linear.weight', 'output_linear.bias']
+
+We can use ``tdnn/exp/pretrained.pt`` in the following way with ``./tdnn/decode.py``:
+
+.. code-block:: bash
+
+   cd tdnn/exp
+   ln -s pretrained.pt epoch-99.pt
+   cd ../..
+
+   ./tdnn/decode.py --epoch 99 --avg 1
+
+The output logs of the above command are given below:
+
+.. code-block:: bash
+
+    2023-08-16 20:45:48,089 INFO [decode.py:262] Decoding started
+    2023-08-16 20:45:48,090 INFO [decode.py:263] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'feature_dim': 23, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'epoch': 99, 'avg': 1, 'export': False, 'feature_dir': PosixPath('data/fbank'), 'max_duration': 30.0, 'bucketing_sampler': False, 'num_buckets': 10, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': False, 'return_cuts': True, 'num_workers': 2, 'env_info': {'k2-version': '1.24.3', 'k2-build-type': 'Release', 'k2-with-cuda': False, 'k2-git-sha1': 'ad79f1c699c684de9785ed6ca5edb805a41f78c3', 'k2-git-date': 'Wed Jul 26 09:30:42 2023', 'lhotse-version': '1.16.0.dev+git.aa073f6.clean', 'torch-version': '2.0.0', 'torch-cuda-available': False, 'torch-cuda-version': None, 'python-version': '3.1', 'icefall-git-branch': 'master', 'icefall-git-sha1': '9a47c08-clean', 'icefall-git-date': 'Mon Aug 14 22:10:50 2023', 'icefall-path': '/private/tmp/icefall', 'k2-path': '/private/tmp/icefall_env/lib/python3.11/site-packages/k2/__init__.py', 'lhotse-path': '/private/tmp/icefall_env/lib/python3.11/site-packages/lhotse/__init__.py', 'hostname': 'fangjuns-MacBook-Pro.local', 'IP address': '127.0.0.1'}}
+    2023-08-16 20:45:48,092 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+    2023-08-16 20:45:48,103 INFO [decode.py:272] device: cpu
+    2023-08-16 20:45:48,109 INFO [checkpoint.py:112] Loading checkpoint from tdnn/exp/epoch-99.pt
+    2023-08-16 20:45:48,115 INFO [asr_datamodule.py:218] About to get test cuts
+    2023-08-16 20:45:48,115 INFO [asr_datamodule.py:253] About to get test cuts
+    2023-08-16 20:45:50,386 INFO [decode.py:203] batch 0/?, cuts processed until now is 4
+    2023-08-16 20:45:50,556 INFO [decode.py:240] The transcripts are stored in tdnn/exp/recogs-test_set.txt
+    2023-08-16 20:45:50,557 INFO [utils.py:564] [test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
+    2023-08-16 20:45:50,558 INFO [decode.py:248] Wrote detailed error stats to tdnn/exp/errs-test_set.txt
+    2023-08-16 20:45:50,559 INFO [decode.py:315] Done!
+
+We can see that it produces an identical WER as before.
+
+We can also use it to decode files with the following command:
+
+.. code-block:: bash
+
+  # ./tdnn/pretrained.py requires kaldifeat
+  #
+  # Please refer to https://csukuangfj.github.io/kaldifeat/installation/from_wheels.html
+  # for how to install kaldifeat
+
+  pip install kaldifeat==1.25.0.dev20230726+cpu.torch2.0.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+
+  ./tdnn/pretrained.py \
+    --checkpoint ./tdnn/exp/pretrained.pt \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+The output is given below:
+
+.. code-block:: bash
+
+  2023-08-16 20:53:19,208 INFO [pretrained.py:136] {'feature_dim': 23, 'num_classes': 4, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './tdnn/exp/pretrained.pt', 'words_file': './data/lang_phone/words.txt', 'HLG': './data/lang_phone/HLG.pt', 'sound_files': ['download/waves_yesno/0_0_0_1_0_0_0_1.wav', 'download/waves_yesno/0_0_1_0_0_0_1_0.wav']}
+  2023-08-16 20:53:19,208 INFO [pretrained.py:142] device: cpu
+  2023-08-16 20:53:19,208 INFO [pretrained.py:144] Creating model
+  2023-08-16 20:53:19,212 INFO [pretrained.py:156] Loading HLG from ./data/lang_phone/HLG.pt
+  2023-08-16 20:53:19,213 INFO [pretrained.py:160] Constructing Fbank computer
+  2023-08-16 20:53:19,213 INFO [pretrained.py:170] Reading sound files: ['download/waves_yesno/0_0_0_1_0_0_0_1.wav', 'download/waves_yesno/0_0_1_0_0_0_1_0.wav']
+  2023-08-16 20:53:19,224 INFO [pretrained.py:176] Decoding started
+  2023-08-16 20:53:19,304 INFO [pretrained.py:212]
+  download/waves_yesno/0_0_0_1_0_0_0_1.wav:
+  NO NO NO YES NO NO NO YES
+
+  download/waves_yesno/0_0_1_0_0_0_1_0.wav:
+  NO NO YES NO NO NO YES NO
+
+
+  2023-08-16 20:53:19,304 INFO [pretrained.py:214] Decoding Done
+
+
+Export via torch.jit.script()
+-----------------------------
+
+The command for this kind of export is
+
+.. code-block:: bash
+
+   cd /tmp/icefall
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+   cd egs/yesno/ASR
+
+   # assume that "--epoch 14 --avg 2" produces the lowest WER.
+
+   ./tdnn/export.py --epoch 14 --avg 2 --jit true
+
+The output logs are given below:
+
+.. code-block:: bash
+
+  2023-08-16 20:47:44,666 INFO [export.py:76] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lr': 0.01, 'feature_dim': 23, 'weight_decay': 1e-06, 'start_epoch': 0, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 10, 'reset_interval': 20, 'valid_interval': 10, 'beam_size': 10, 'reduction': 'sum', 'use_double_scores': True, 'epoch': 14, 'avg': 2, 'jit': True}
+  2023-08-16 20:47:44,667 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+  2023-08-16 20:47:44,670 INFO [export.py:93] averaging ['tdnn/exp/epoch-13.pt', 'tdnn/exp/epoch-14.pt']
+  2023-08-16 20:47:44,677 INFO [export.py:100] Using torch.jit.script
+  2023-08-16 20:47:44,843 INFO [export.py:104] Saved to tdnn/exp/cpu_jit.pt
+
+From the output logs we can see that the generated file is saved to ``tdnn/exp/cpu_jit.pt``.
+
+Don't be confused by the name ``cpu_jit.pt``. The ``cpu`` part means the model is moved to
+CPU before exporting. That means, when you load it with:
+
+.. code-block:: bash
+
+   torch.jit.load()
+
+you don't need to specify the argument `map_location <https://pytorch.org/docs/stable/generated/torch.jit.load.html#torch.jit.load>`_
+and it resides on CPU by default.
+
+To use ``tdnn/exp/cpu_jit.pt`` with `icefall`_ to decode files, we can use:
+
+.. code-block:: bash
+
+  # ./tdnn/jit_pretrained.py requires kaldifeat
+  #
+  # Please refer to https://csukuangfj.github.io/kaldifeat/installation/from_wheels.html
+  # for how to install kaldifeat
+
+  pip install kaldifeat==1.25.0.dev20230726+cpu.torch2.0.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+
+
+  ./tdnn/jit_pretrained.py \
+    --nn-model ./tdnn/exp/cpu_jit.pt \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+The output is given below:
+
+.. code-block:: bash
+
+  2023-08-16 20:56:00,603 INFO [jit_pretrained.py:121] {'feature_dim': 23, 'num_classes': 4, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'nn_model': './tdnn/exp/cpu_jit.pt', 'words_file': './data/lang_phone/words.txt', 'HLG': './data/lang_phone/HLG.pt', 'sound_files': ['download/waves_yesno/0_0_0_1_0_0_0_1.wav', 'download/waves_yesno/0_0_1_0_0_0_1_0.wav']}
+  2023-08-16 20:56:00,603 INFO [jit_pretrained.py:127] device: cpu
+  2023-08-16 20:56:00,603 INFO [jit_pretrained.py:129] Loading torchscript model
+  2023-08-16 20:56:00,640 INFO [jit_pretrained.py:134] Loading HLG from ./data/lang_phone/HLG.pt
+  2023-08-16 20:56:00,641 INFO [jit_pretrained.py:138] Constructing Fbank computer
+  2023-08-16 20:56:00,641 INFO [jit_pretrained.py:148] Reading sound files: ['download/waves_yesno/0_0_0_1_0_0_0_1.wav', 'download/waves_yesno/0_0_1_0_0_0_1_0.wav']
+  2023-08-16 20:56:00,642 INFO [jit_pretrained.py:154] Decoding started
+  2023-08-16 20:56:00,727 INFO [jit_pretrained.py:190]
+  download/waves_yesno/0_0_0_1_0_0_0_1.wav:
+  NO NO NO YES NO NO NO YES
+
+  download/waves_yesno/0_0_1_0_0_0_1_0.wav:
+  NO NO YES NO NO NO YES NO
+
+
+  2023-08-16 20:56:00,727 INFO [jit_pretrained.py:192] Decoding Done
+
+.. hint::
+
+   We provide only code for ``torch.jit.script()``. You can try ``torch.jit.trace()``
+   if you want.
+
+Export via torch.onnx.export()
+------------------------------
+
+The command for this kind of export is
+
+.. code-block:: bash
+
+   cd /tmp/icefall
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+   cd egs/yesno/ASR
+
+   # tdnn/export_onnx.py requires onnx and onnxruntime
+   pip install onnx onnxruntime
+
+   # assume that "--epoch 14 --avg 2" produces the lowest WER.
+
+   ./tdnn/export_onnx.py \
+     --epoch 14 \
+     --avg 2
+
+The output logs are given below:
+
+.. code-block:: bash
+
+  2023-08-16 20:59:20,888 INFO [export_onnx.py:83] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lr': 0.01, 'feature_dim': 23, 'weight_decay': 1e-06, 'start_epoch': 0, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 10, 'reset_interval': 20, 'valid_interval': 10, 'beam_size': 10, 'reduction': 'sum', 'use_double_scores': True, 'epoch': 14, 'avg': 2}
+  2023-08-16 20:59:20,888 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+  2023-08-16 20:59:20,892 INFO [export_onnx.py:100] averaging ['tdnn/exp/epoch-13.pt', 'tdnn/exp/epoch-14.pt']
+  ================ Diagnostic Run torch.onnx.export version 2.0.0 ================
+  verbose: False, log level: Level.ERROR
+  ======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================
+
+  2023-08-16 20:59:21,047 INFO [export_onnx.py:127] Saved to tdnn/exp/model-epoch-14-avg-2.onnx
+  2023-08-16 20:59:21,047 INFO [export_onnx.py:136] meta_data: {'model_type': 'tdnn', 'version': '1', 'model_author': 'k2-fsa', 'comment': 'non-streaming tdnn for the yesno recipe', 'vocab_size': 4}
+  2023-08-16 20:59:21,049 INFO [export_onnx.py:140] Generate int8 quantization models
+  2023-08-16 20:59:21,075 INFO [onnx_quantizer.py:538] Quantization parameters for tensor:"/Transpose_1_output_0" not specified
+  2023-08-16 20:59:21,081 INFO [export_onnx.py:151] Saved to tdnn/exp/model-epoch-14-avg-2.int8.onnx
+
+We can see from the logs that it generates two files:
+
+  - ``tdnn/exp/model-epoch-14-avg-2.onnx`` (ONNX model with ``float32`` weights)
+  - ``tdnn/exp/model-epoch-14-avg-2.int8.onnx`` (ONNX model with ``int8`` weights)
+
+To use the generated ONNX model files for decoding with `onnxruntime`_, we can use
+
+.. code-block:: bash
+
+  # ./tdnn/onnx_pretrained.py requires kaldifeat
+  #
+  # Please refer to https://csukuangfj.github.io/kaldifeat/installation/from_wheels.html
+  # for how to install kaldifeat
+
+  pip install kaldifeat==1.25.0.dev20230726+cpu.torch2.0.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+
+  ./tdnn/onnx_pretrained.py \
+    --nn-model ./tdnn/exp/model-epoch-14-avg-2.onnx \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+The output is given below:
+
+.. code-block:: bash
+
+  2023-08-16 21:03:24,260 INFO [onnx_pretrained.py:166] {'feature_dim': 23, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'nn_model': './tdnn/exp/model-epoch-14-avg-2.onnx', 'words_file': './data/lang_phone/words.txt', 'HLG': './data/lang_phone/HLG.pt', 'sound_files': ['download/waves_yesno/0_0_0_1_0_0_0_1.wav', 'download/waves_yesno/0_0_1_0_0_0_1_0.wav']}
+  2023-08-16 21:03:24,260 INFO [onnx_pretrained.py:171] device: cpu
+  2023-08-16 21:03:24,260 INFO [onnx_pretrained.py:173] Loading onnx model ./tdnn/exp/model-epoch-14-avg-2.onnx
+  2023-08-16 21:03:24,267 INFO [onnx_pretrained.py:176] Loading HLG from ./data/lang_phone/HLG.pt
+  2023-08-16 21:03:24,270 INFO [onnx_pretrained.py:180] Constructing Fbank computer
+  2023-08-16 21:03:24,273 INFO [onnx_pretrained.py:190] Reading sound files: ['download/waves_yesno/0_0_0_1_0_0_0_1.wav', 'download/waves_yesno/0_0_1_0_0_0_1_0.wav']
+  2023-08-16 21:03:24,279 INFO [onnx_pretrained.py:196] Decoding started
+  2023-08-16 21:03:24,318 INFO [onnx_pretrained.py:232]
+  download/waves_yesno/0_0_0_1_0_0_0_1.wav:
+  NO NO NO YES NO NO NO YES
+
+  download/waves_yesno/0_0_1_0_0_0_1_0.wav:
+  NO NO YES NO NO NO YES NO
+
+
+  2023-08-16 21:03:24,318 INFO [onnx_pretrained.py:234] Decoding Done
+
+.. note::
+
+   To use the ``int8`` ONNX model for decoding, please use:
+
+   .. code-block:: bash
+
+      ./tdnn/onnx_pretrained.py \
+        --nn-model ./tdnn/exp/model-epoch-14-avg-2.onnx \
+        --HLG ./data/lang_phone/HLG.pt \
+        --words-file ./data/lang_phone/words.txt \
+        download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+        download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+For the more curious
+--------------------
+
+If you are wondering how to deploy the model without ``torch``, please
+continue reading. We will show how to use `sherpa-onnx`_ to run the
+exported ONNX models, which depends only on `onnxruntime`_ and does not
+depend on ``torch``.
+
+In this tutorial, we will only demonstrate the usage of `sherpa-onnx`_ with the
+pre-trained model of the `yesno`_ recipe. There are also other two frameworks
+available:
+
+  - `sherpa`_. It works with torchscript models.
+  - `sherpa-ncnn`_. It works with models exported using :ref:`icefall_export_to_ncnn` with `ncnn`_
+
+Please see `<https://k2-fsa.github.io/sherpa/>`_ for further details.
diff --git a/docs/source/for-dummies/training.rst b/docs/source/for-dummies/training.rst
new file mode 100644
index 000000000..816ef2d3b
--- /dev/null
+++ b/docs/source/for-dummies/training.rst
@@ -0,0 +1,39 @@
+.. _dummies_tutorial_training:
+
+Training
+========
+
+After :ref:`dummies_tutorial_data_preparation`, we can start training.
+
+The command to start the training is quite simple:
+
+.. code-block:: bash
+
+   cd /tmp/icefall
+   export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+   cd egs/yesno/ASR
+
+   # We use CPU for training by setting the following environment variable
+   export CUDA_VISIBLE_DEVICES=""
+
+   ./tdnn/train.py
+
+That's it!
+
+You can find the training logs below:
+
+.. literalinclude:: ./code/train-yesno.txt
+
+For the more curious
+--------------------
+
+.. code-block:: bash
+
+   ./tdnn/train.py --help
+
+will print the usage information about ``./tdnn/train.py``. For instance, you
+can specify the number of epochs to train and the location to save the training
+results.
+
+The training text logs are saved in ``tdnn/exp/log`` while the tensorboard
+logs are in ``tdnn/exp/tensorboard``.
diff --git a/docs/source/huggingface/index.rst b/docs/source/huggingface/index.rst
new file mode 100644
index 000000000..bd731793b
--- /dev/null
+++ b/docs/source/huggingface/index.rst
@@ -0,0 +1,13 @@
+Huggingface
+===========
+
+This section describes how to find pre-trained models.
+It also demonstrates how to try them from within your browser
+without installing anything by using
+`Huggingface spaces <https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_.
+
+.. toctree::
+   :maxdepth: 2
+
+   pretrained-models
+   spaces
diff --git a/docs/source/huggingface/pic/hugging-face-sherpa-2.png b/docs/source/huggingface/pic/hugging-face-sherpa-2.png
new file mode 100644
index 000000000..3b47bd51b
Binary files /dev/null and b/docs/source/huggingface/pic/hugging-face-sherpa-2.png differ
diff --git a/docs/source/huggingface/pic/hugging-face-sherpa-3.png b/docs/source/huggingface/pic/hugging-face-sherpa-3.png
new file mode 100644
index 000000000..1d7a2d316
Binary files /dev/null and b/docs/source/huggingface/pic/hugging-face-sherpa-3.png differ
diff --git a/docs/source/huggingface/pic/hugging-face-sherpa.png b/docs/source/huggingface/pic/hugging-face-sherpa.png
new file mode 100644
index 000000000..dea0b1d46
Binary files /dev/null and b/docs/source/huggingface/pic/hugging-face-sherpa.png differ
diff --git a/docs/source/huggingface/pretrained-models.rst b/docs/source/huggingface/pretrained-models.rst
new file mode 100644
index 000000000..8ae22f76f
--- /dev/null
+++ b/docs/source/huggingface/pretrained-models.rst
@@ -0,0 +1,17 @@
+Pre-trained models
+==================
+
+We have uploaded pre-trained models for all recipes in ``icefall``
+to `<https://huggingface.co/>`_.
+
+You can find them by visiting the following link:
+
+`<https://huggingface.co/models?search=icefall>`_.
+
+You can also find links of pre-trained models for a specific recipe
+by looking at the corresponding ``RESULTS.md``. For instance:
+
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/RESULTS.md>`_
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/gigaspeech/ASR/RESULTS.md>`_
+  - `<https://github.com/k2-fsa/icefall/blob/master/egs/wenetspeech/ASR/RESULTS.md>`_
diff --git a/docs/source/huggingface/spaces.rst b/docs/source/huggingface/spaces.rst
new file mode 100644
index 000000000..e718c3731
--- /dev/null
+++ b/docs/source/huggingface/spaces.rst
@@ -0,0 +1,65 @@
+Huggingface spaces
+==================
+
+We have integrated the server framework
+`sherpa <http://github.com/k2-fsa/sherpa>`_
+with `Huggingface spaces <https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_
+so that you can try pre-trained models from within your browser
+without the need to download or install anything.
+
+All you need is a browser, which can be run on Windows, macOS, Linux, or even on your
+iPad and your phone.
+
+Start your browser and visit the following address:
+
+`<https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_
+
+and you will see a page like the following screenshot:
+
+.. image:: ./pic/hugging-face-sherpa.png
+   :alt: screenshot of `<https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_
+   :target: https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition
+
+You can:
+
+  1. Select a language for recognition. Currently, we provide pre-trained models
+     from ``icefall`` for the following languages: ``Chinese``, ``English``, and
+     ``Chinese+English``.
+  2. After selecting the target language, you can select a pre-trained model
+     corresponding to the language.
+  3. Select the decoding method. Currently, it provides ``greedy search``
+     and ``modified_beam_search``.
+  4. If you selected ``modified_beam_search``, you can choose the number of
+     active paths during the search.
+  5. Either upload a file or record your speech for recognition.
+  6. Click the button ``Submit for recognition``.
+  7. Wait for a moment and you will get the recognition results.
+
+The following screenshot shows an example when selecting ``Chinese+English``:
+
+.. image:: ./pic/hugging-face-sherpa-3.png
+   :alt: screenshot of `<https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_
+   :target: https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition
+
+
+In the bottom part of the page, you can find a table of examples. You can click
+one of them and then click ``Submit for recognition``.
+
+.. image:: ./pic/hugging-face-sherpa-2.png
+   :alt: screenshot of `<https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_
+   :target: https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition
+
+YouTube Video
+-------------
+
+We provide the following YouTube video demonstrating how to use
+`<https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition>`_.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ElN3r9dkKE4
diff --git a/docs/source/index.rst b/docs/source/index.rst
new file mode 100644
index 000000000..fb539d3f2
--- /dev/null
+++ b/docs/source/index.rst
@@ -0,0 +1,44 @@
+.. icefall documentation master file, created by
+   sphinx-quickstart on Mon Aug 23 16:07:39 2021.
+   You can adapt this file completely to your liking, but it should at least
+   contain the root `toctree` directive.
+
+Icefall
+=======
+
+.. image:: _static/logo.png
+    :alt: icefall logo
+    :width: 168px
+    :align: center
+    :target: https://github.com/k2-fsa/icefall
+
+
+Documentation for `icefall <https://github.com/k2-fsa/icefall>`_, containing
+speech recognition recipes using `k2 <https://github.com/k2-fsa/k2>`_.
+
+.. toctree::
+   :maxdepth: 2
+   :caption: Contents:
+
+   for-dummies/index.rst
+   installation/index
+   docker/index
+   faqs
+   model-export/index
+
+
+.. toctree::
+   :maxdepth: 3
+
+   recipes/index
+
+.. toctree::
+   :maxdepth: 2
+
+   contributing/index
+   huggingface/index
+
+.. toctree::
+   :maxdepth: 2
+   
+   decoding-with-langugage-models/index
diff --git a/docs/source/installation/images/README.md b/docs/source/installation/images/README.md
new file mode 100644
index 000000000..97c1e993c
--- /dev/null
+++ b/docs/source/installation/images/README.md
@@ -0,0 +1,4 @@
+
+# Introduction
+
+<https://shields.io/> is used to generate files in this directory.
diff --git a/docs/source/installation/images/device-CPU_CUDA-orange.svg b/docs/source/installation/images/device-CPU_CUDA-orange.svg
new file mode 100644
index 000000000..a023a1283
--- /dev/null
+++ b/docs/source/installation/images/device-CPU_CUDA-orange.svg
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="122" height="20" role="img" aria-label="device: CPU | CUDA"><title>device: CPU | CUDA</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="122" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="45" height="20" fill="#555"/><rect x="45" width="77" height="20" fill="#fe7d37"/><rect width="122" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="235" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="350">device</text><text x="235" y="140" transform="scale(.1)" fill="#fff" textLength="350">device</text><text aria-hidden="true" x="825" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="670">CPU | CUDA</text><text x="825" y="140" transform="scale(.1)" fill="#fff" textLength="670">CPU | CUDA</text></g></svg>
diff --git a/docs/source/installation/images/k2-gt-v1.9-blueviolet.svg b/docs/source/installation/images/k2-gt-v1.9-blueviolet.svg
new file mode 100644
index 000000000..3019ff03d
--- /dev/null
+++ b/docs/source/installation/images/k2-gt-v1.9-blueviolet.svg
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="80" height="20" role="img" aria-label="k2: &gt;= v1.9"><title>k2: &gt;= v1.9</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="80" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="23" height="20" fill="#555"/><rect x="23" width="57" height="20" fill="blueviolet"/><rect width="80" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="125" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="130">k2</text><text x="125" y="140" transform="scale(.1)" fill="#fff" textLength="130">k2</text><text aria-hidden="true" x="505" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="470">&gt;= v1.9</text><text x="505" y="140" transform="scale(.1)" fill="#fff" textLength="470">&gt;= v1.9</text></g></svg>
diff --git a/docs/source/installation/images/os-Linux_macOS-ff69b4.svg b/docs/source/installation/images/os-Linux_macOS-ff69b4.svg
new file mode 100644
index 000000000..178813ed4
--- /dev/null
+++ b/docs/source/installation/images/os-Linux_macOS-ff69b4.svg
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="114" height="20" role="img" aria-label="os: Linux | macOS"><title>os: Linux | macOS</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="114" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="23" height="20" fill="#555"/><rect x="23" width="91" height="20" fill="#ff69b4"/><rect width="114" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="125" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="130">os</text><text x="125" y="140" transform="scale(.1)" fill="#fff" textLength="130">os</text><text aria-hidden="true" x="675" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="810">Linux | macOS</text><text x="675" y="140" transform="scale(.1)" fill="#fff" textLength="810">Linux | macOS</text></g></svg>
diff --git a/docs/source/installation/images/python-gt-v3.6-blue.svg b/docs/source/installation/images/python-gt-v3.6-blue.svg
new file mode 100644
index 000000000..df677ad09
--- /dev/null
+++ b/docs/source/installation/images/python-gt-v3.6-blue.svg
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="98" height="20" role="img" aria-label="python: &gt;= 3.6"><title>python: &gt;= 3.6</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="98" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="49" height="20" fill="#555"/><rect x="49" width="49" height="20" fill="#007ec6"/><rect width="98" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="255" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="390">python</text><text x="255" y="140" transform="scale(.1)" fill="#fff" textLength="390">python</text><text aria-hidden="true" x="725" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="390">&gt;= 3.6</text><text x="725" y="140" transform="scale(.1)" fill="#fff" textLength="390">&gt;= 3.6</text></g></svg>
diff --git a/docs/source/installation/images/torch-gt-v1.6.0-green.svg b/docs/source/installation/images/torch-gt-v1.6.0-green.svg
new file mode 100644
index 000000000..d7007d742
--- /dev/null
+++ b/docs/source/installation/images/torch-gt-v1.6.0-green.svg
@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="100" height="20" role="img" aria-label="torch: &gt;= 1.6.0"><title>torch: &gt;= 1.6.0</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="100" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="39" height="20" fill="#555"/><rect x="39" width="61" height="20" fill="#97ca00"/><rect width="100" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="205" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="290">torch</text><text x="205" y="140" transform="scale(.1)" fill="#fff" textLength="290">torch</text><text aria-hidden="true" x="685" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="510">&gt;= 1.6.0</text><text x="685" y="140" transform="scale(.1)" fill="#fff" textLength="510">&gt;= 1.6.0</text></g></svg>
diff --git a/docs/source/installation/index.rst b/docs/source/installation/index.rst
new file mode 100644
index 000000000..5a034ef5b
--- /dev/null
+++ b/docs/source/installation/index.rst
@@ -0,0 +1,550 @@
+.. _install icefall:
+
+Installation
+============
+
+.. hint::
+
+   We also provide :ref:`icefall_docker` support, which has already setup
+   the environment for you.
+
+.. hint::
+
+  We have a colab notebook guiding you step by step to setup the environment.
+
+  |yesno colab notebook|
+
+  .. |yesno colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+     :target: https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing
+
+`icefall`_ depends on `k2`_ and `lhotse`_.
+
+We recommend that you use the following steps to install the dependencies.
+
+- (0) Install CUDA toolkit and cuDNN
+- (1) Install `torch`_ and `torchaudio`_
+- (2) Install `k2`_
+- (3) Install `lhotse`_
+
+.. caution::
+
+  Installation order matters.
+
+(0) Install CUDA toolkit and cuDNN
+----------------------------------
+
+Please refer to
+`<https://k2-fsa.github.io/k2/installation/cuda-cudnn.html>`_
+to install CUDA and cuDNN.
+
+
+(1) Install torch and torchaudio
+--------------------------------
+
+Please refer `<https://pytorch.org/>`_ to install `torch`_ and `torchaudio`_.
+
+.. caution::
+
+   Please install torch and torchaudio at the same time.
+
+(2) Install k2
+--------------
+
+Please refer to `<https://k2-fsa.github.io/k2/installation/index.html>`_
+to install `k2`_.
+
+.. caution::
+
+  Please don't change your installed PyTorch after you have installed k2.
+
+.. note::
+
+   We suggest that you install k2 from pre-compiled wheels by following
+   `<https://k2-fsa.github.io/k2/installation/from_wheels.html>`_
+
+.. hint::
+
+   Please always install the latest version of `k2`_.
+
+(3) Install lhotse
+------------------
+
+Please refer to `<https://lhotse.readthedocs.io/en/latest/getting-started.html#installation>`_
+to install `lhotse`_.
+
+.. hint::
+
+    We strongly recommend you to use::
+
+      pip install git+https://github.com/lhotse-speech/lhotse
+
+    to install the latest version of `lhotse`_.
+
+(4) Download icefall
+--------------------
+
+`icefall`_ is a collection of Python scripts; what you need is to download it
+and set the environment variable ``PYTHONPATH`` to point to it.
+
+Assume you want to place `icefall`_ in the folder ``/tmp``. The
+following commands show you how to setup `icefall`_:
+
+.. code-block:: bash
+
+  cd /tmp
+  git clone https://github.com/k2-fsa/icefall
+  cd icefall
+  pip install -r requirements.txt
+  export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+
+.. HINT::
+
+  You can put several versions of `icefall`_ in the same virtual environment.
+  To switch among different versions of `icefall`_, just set ``PYTHONPATH``
+  to point to the version you want.
+
+Installation example
+--------------------
+
+The following shows an example about setting up the environment.
+
+(1) Create a virtual environment
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+   kuangfangjun:~$ virtualenv -p python3.8 test-icefall
+   created virtual environment CPython3.8.0.final.0-64 in 9422ms
+     creator CPython3Posix(dest=/star-fj/fangjun/test-icefall, clear=False, no_vcs_ignore=False, global=False)
+     seeder FromAppData(download=False, pip=bundle, setuptools=bundle, wheel=bundle, via=copy, app_data_dir=/star-fj/fangjun/.local/share/virtualenv)
+       added seed packages: pip==22.3.1, setuptools==65.6.3, wheel==0.38.4
+     activators BashActivator,CShellActivator,FishActivator,NushellActivator,PowerShellActivator,PythonActivator
+
+   kuangfangjun:~$ source test-icefall/bin/activate
+
+   (test-icefall) kuangfangjun:~$
+
+(2) Install CUDA toolkit and cuDNN
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+You need to determine the version of CUDA toolkit to install.
+
+.. code-block:: bash
+
+   (test-icefall) kuangfangjun:~$ nvidia-smi | head -n 4
+
+   Wed Jul 26 21:57:49 2023
+   +-----------------------------------------------------------------------------+
+   | NVIDIA-SMI 510.47.03    Driver Version: 510.47.03    CUDA Version: 11.6     |
+   |-------------------------------+----------------------+----------------------+
+
+You can choose any CUDA version that is ``not`` greater than the version printed by ``nvidia-smi``.
+In our case, we can choose any version ``<= 11.6``.
+
+We will use ``CUDA 11.6`` in this example. Please follow
+`<https://k2-fsa.github.io/k2/installation/cuda-cudnn.html#cuda-11-6>`_
+to install CUDA toolkit and cuDNN if you have not done that before.
+
+After installing CUDA toolkit, you can use the following command to verify it:
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ nvcc --version
+
+  nvcc: NVIDIA (R) Cuda compiler driver
+  Copyright (c) 2005-2019 NVIDIA Corporation
+  Built on Wed_Oct_23_19:24:38_PDT_2019
+  Cuda compilation tools, release 10.2, V10.2.89
+
+(3) Install torch and torchaudio
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Since we have selected CUDA toolkit ``11.6``, we have to install a version of `torch`_
+that is compiled against CUDA ``11.6``. We select ``torch 1.13.0+cu116`` in this
+example.
+
+After selecting the version of `torch`_ to install, we need to also install
+a compatible version of `torchaudio`_, which is ``0.13.0+cu116`` in our case.
+
+Please refer to `<https://pytorch.org/audio/stable/installation.html#compatibility-matrix>`_
+to select an appropriate version of `torchaudio`_ to install if you use a different
+version of `torch`_.
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ pip install torch==1.13.0+cu116 torchaudio==0.13.0+cu116 -f https://download.pytorch.org/whl/torch_stable.html
+
+  Looking in links: https://download.pytorch.org/whl/torch_stable.html
+  Collecting torch==1.13.0+cu116
+    Downloading https://download.pytorch.org/whl/cu116/torch-1.13.0%2Bcu116-cp38-cp38-linux_x86_64.whl (1983.0 MB)
+       ________________________________________ 2.0/2.0 GB 764.4 kB/s eta 0:00:00
+  Collecting torchaudio==0.13.0+cu116
+    Downloading https://download.pytorch.org/whl/cu116/torchaudio-0.13.0%2Bcu116-cp38-cp38-linux_x86_64.whl (4.2 MB)
+       ________________________________________ 4.2/4.2 MB 1.3 MB/s eta 0:00:00
+  Requirement already satisfied: typing-extensions in /star-fj/fangjun/test-icefall/lib/python3.8/site-packages (from torch==1.13.0+cu116) (4.7.1)
+  Installing collected packages: torch, torchaudio
+  Successfully installed torch-1.13.0+cu116 torchaudio-0.13.0+cu116
+
+Verify that `torch`_ and `torchaudio`_ are successfully installed:
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ python3 -c "import torch; print(torch.__version__)"
+
+  1.13.0+cu116
+
+  (test-icefall) kuangfangjun:~$ python3 -c "import torchaudio; print(torchaudio.__version__)"
+
+  0.13.0+cu116
+
+(4) Install k2
+~~~~~~~~~~~~~~
+
+We will install `k2`_ from pre-compiled wheels by following
+`<https://k2-fsa.github.io/k2/installation/from_wheels.html>`_
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ pip install k2==1.24.3.dev20230725+cuda11.6.torch1.13.0 -f https://k2-fsa.github.io/k2/cuda.html
+
+  Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
+  Looking in links: https://k2-fsa.github.io/k2/cuda.html
+  Collecting k2==1.24.3.dev20230725+cuda11.6.torch1.13.0
+    Downloading https://huggingface.co/csukuangfj/k2/resolve/main/ubuntu-cuda/k2-1.24.3.dev20230725%2Bcuda11.6.torch1.13.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (104.3 MB)
+       ________________________________________ 104.3/104.3 MB 5.1 MB/s eta 0:00:00
+  Requirement already satisfied: torch==1.13.0 in /star-fj/fangjun/test-icefall/lib/python3.8/site-packages (from k2==1.24.3.dev20230725+cuda11.6.torch1.13.0) (1.13.0+cu116)
+  Collecting graphviz
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/de/5e/fcbb22c68208d39edff467809d06c9d81d7d27426460ebc598e55130c1aa/graphviz-0.20.1-py3-none-any.whl (47 kB)
+  Requirement already satisfied: typing-extensions in /star-fj/fangjun/test-icefall/lib/python3.8/site-packages (from torch==1.13.0->k2==1.24.3.dev20230725+cuda11.6.torch1.13.0) (4.7.1)
+  Installing collected packages: graphviz, k2
+  Successfully installed graphviz-0.20.1 k2-1.24.3.dev20230725+cuda11.6.torch1.13.0
+
+.. hint::
+
+   Please refer to `<https://k2-fsa.github.io/k2/cuda.html>`_ for the available
+   pre-compiled wheels about `k2`_.
+
+Verify that `k2`_ has been installed successfully:
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ python3 -m k2.version
+
+  Collecting environment information...
+
+  k2 version: 1.24.3
+  Build type: Release
+  Git SHA1: 4c05309499a08454997adf500b56dcc629e35ae5
+  Git date: Tue Jul 25 16:23:36 2023
+  Cuda used to build k2: 11.6
+  cuDNN used to build k2: 8.3.2
+  Python version used to build k2: 3.8
+  OS used to build k2: CentOS Linux release 7.9.2009 (Core)
+  CMake version: 3.27.0
+  GCC version: 9.3.1
+  CMAKE_CUDA_FLAGS:  -Wno-deprecated-gpu-targets   -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_35,code=sm_35  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_50,code=sm_50  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_60,code=sm_60  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_61,code=sm_61  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_70,code=sm_70  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_75,code=sm_75  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_80,code=sm_80  -lineinfo --expt-extended-lambda -use_fast_math -Xptxas=-w  --expt-extended-lambda -gencode arch=compute_86,code=sm_86 -DONNX_NAMESPACE=onnx_c2 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_86,code=compute_86 -Xcudafe --diag_suppress=cc_clobber_ignored,--diag_suppress=integer_sign_change,--diag_suppress=useless_using_declaration,--diag_suppress=set_but_not_used,--diag_suppress=field_without_dll_interface,--diag_suppress=base_class_has_different_dll_interface,--diag_suppress=dll_interface_conflict_none_assumed,--diag_suppress=dll_interface_conflict_dllexport_assumed,--diag_suppress=implicit_return_from_non_void_function,--diag_suppress=unsigned_compare_with_zero,--diag_suppress=declared_but_not_referenced,--diag_suppress=bad_friend_decl --expt-relaxed-constexpr --expt-extended-lambda -D_GLIBCXX_USE_CXX11_ABI=0 --compiler-options -Wall  --compiler-options -Wno-strict-overflow  --compiler-options -Wno-unknown-pragmas
+  CMAKE_CXX_FLAGS:  -D_GLIBCXX_USE_CXX11_ABI=0 -Wno-unused-variable  -Wno-strict-overflow
+  PyTorch version used to build k2: 1.13.0+cu116
+  PyTorch is using Cuda: 11.6
+  NVTX enabled: True
+  With CUDA: True
+  Disable debug: True
+  Sync kernels : False
+  Disable checks: False
+  Max cpu memory allocate: 214748364800 bytes (or 200.0 GB)
+  k2 abort: False
+  __file__: /star-fj/fangjun/test-icefall/lib/python3.8/site-packages/k2/version/version.py
+  _k2.__file__: /star-fj/fangjun/test-icefall/lib/python3.8/site-packages/_k2.cpython-38-x86_64-linux-gnu.so
+
+(5) Install lhotse
+~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ pip install git+https://github.com/lhotse-speech/lhotse
+
+  Collecting git+https://github.com/lhotse-speech/lhotse
+    Cloning https://github.com/lhotse-speech/lhotse to /tmp/pip-req-build-vq12fd5i
+    Running command git clone --filter=blob:none --quiet https://github.com/lhotse-speech/lhotse /tmp/pip-req-build-vq12fd5i
+    Resolved https://github.com/lhotse-speech/lhotse to commit 7640d663469b22cd0b36f3246ee9b849cd25e3b7
+    Installing build dependencies ... done
+    Getting requirements to build wheel ... done
+    Preparing metadata (pyproject.toml) ... done
+  Collecting cytoolz>=0.10.1
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/1e/3b/a7828d575aa17fb7acaf1ced49a3655aa36dad7e16eb7e6a2e4df0dda76f/cytoolz-0.12.2-cp38-cp38-
+  manylinux_2_17_x86_64.manylinux2014_x86_64.whl (2.0 MB)
+       ________________________________________ 2.0/2.0 MB 33.2 MB/s eta 0:00:00
+  Collecting pyyaml>=5.3.1
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/c8/6b/6600ac24725c7388255b2f5add93f91e58a5d7efaf4af244fdbcc11a541b/PyYAML-6.0.1-cp38-cp38-ma
+  nylinux_2_17_x86_64.manylinux2014_x86_64.whl (736 kB)
+       ________________________________________ 736.6/736.6 kB 38.6 MB/s eta 0:00:00
+  Collecting dataclasses
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/26/2f/1095cdc2868052dd1e64520f7c0d5c8c550ad297e944e641dbf1ffbb9a5d/dataclasses-0.6-py3-none-
+  any.whl (14 kB)
+  Requirement already satisfied: torchaudio in ./test-icefall/lib/python3.8/site-packages (from lhotse==1.16.0.dev0+git.7640d66.clean) (0.13.0+cu116)
+  Collecting lilcom>=1.1.0
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/a8/65/df0a69c52bd085ca1ad4e5c4c1a5c680e25f9477d8e49316c4ff1e5084a4/lilcom-1.7-cp38-cp38-many
+  linux_2_17_x86_64.manylinux2014_x86_64.whl (87 kB)
+       ________________________________________ 87.1/87.1 kB 8.7 MB/s eta 0:00:00
+  Collecting tqdm
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/e6/02/a2cff6306177ae6bc73bc0665065de51dfb3b9db7373e122e2735faf0d97/tqdm-4.65.0-py3-none-any
+  .whl (77 kB)
+  Requirement already satisfied: numpy>=1.18.1 in ./test-icefall/lib/python3.8/site-packages (from lhotse==1.16.0.dev0+git.7640d66.clean) (1.24.4)
+  Collecting audioread>=2.1.9
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/5d/cb/82a002441902dccbe427406785db07af10182245ee639ea9f4d92907c923/audioread-3.0.0.tar.gz (
+  377 kB)
+    Preparing metadata (setup.py) ... done
+  Collecting tabulate>=0.8.1
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/40/44/4a5f08c96eb108af5cb50b41f76142f0afa346dfa99d5296fe7202a11854/tabulate-0.9.0-py3-none-
+  any.whl (35 kB)
+  Collecting click>=7.1.1
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/1a/70/e63223f8116931d365993d4a6b7ef653a4d920b41d03de7c59499962821f/click-8.1.6-py3-none-any.
+  whl (97 kB)
+       ________________________________________ 97.9/97.9 kB 8.4 MB/s eta 0:00:00
+  Collecting packaging
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/ab/c3/57f0601a2d4fe15de7a553c00adbc901425661bf048f2a22dfc500caf121/packaging-23.1-py3-none-
+  any.whl (48 kB)
+  Collecting intervaltree>=3.1.0
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/50/fb/396d568039d21344639db96d940d40eb62befe704ef849b27949ded5c3bb/intervaltree-3.1.0.tar.gz
+   (32 kB)
+    Preparing metadata (setup.py) ... done
+  Requirement already satisfied: torch in ./test-icefall/lib/python3.8/site-packages (from lhotse==1.16.0.dev0+git.7640d66.clean) (1.13.0+cu116)
+  Collecting SoundFile>=0.10
+    Downloading https://pypi.tuna.tsinghua.edu.cn/packages/ad/bd/0602167a213d9184fc688b1086dc6d374b7ae8c33eccf169f9b50ce6568c/soundfile-0.12.1-py2.py3-
+  none-manylinux_2_17_x86_64.whl (1.3 MB)
+       ________________________________________ 1.3/1.3 MB 46.5 MB/s eta 0:00:00
+  Collecting toolz>=0.8.0
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/7f/5c/922a3508f5bda2892be3df86c74f9cf1e01217c2b1f8a0ac4841d903e3e9/toolz-0.12.0-py3-none-any.whl (55 kB)
+  Collecting sortedcontainers<3.0,>=2.0
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/32/46/9cb0e58b2deb7f82b84065f37f3bffeb12413f947f9388e4cac22c4621ce/sortedcontainers-2.4.0-py2.py3-none-any.whl (29 kB)
+  Collecting cffi>=1.0
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/b7/8b/06f30caa03b5b3ac006de4f93478dbd0239e2a16566d81a106c322dc4f79/cffi-1.15.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (442 kB)
+  Requirement already satisfied: typing-extensions in ./test-icefall/lib/python3.8/site-packages (from torch->lhotse==1.16.0.dev0+git.7640d66.clean) (4.7.1)
+  Collecting pycparser
+    Using cached https://pypi.tuna.tsinghua.edu.cn/packages/62/d5/5f610ebe421e85889f2e55e33b7f9a6795bd982198517d912eb1c76e1a53/pycparser-2.21-py2.py3-none-any.whl (118 kB)
+  Building wheels for collected packages: lhotse, audioread, intervaltree
+    Building wheel for lhotse (pyproject.toml) ... done
+    Created wheel for lhotse: filename=lhotse-1.16.0.dev0+git.7640d66.clean-py3-none-any.whl size=687627 sha256=cbf0a4d2d0b639b33b91637a4175bc251d6a021a069644ecb1a9f2b3a83d072a
+    Stored in directory: /tmp/pip-ephem-wheel-cache-wwtk90_m/wheels/7f/7a/8e/a0bf241336e2e3cb573e1e21e5600952d49f5162454f2e612f
+    Building wheel for audioread (setup.py) ... done
+    Created wheel for audioread: filename=audioread-3.0.0-py3-none-any.whl size=23704 sha256=5e2d3537c96ce9cf0f645a654c671163707bf8cb8d9e358d0e2b0939a85ff4c2
+    Stored in directory: /star-fj/fangjun/.cache/pip/wheels/e2/c3/9c/f19ae5a03f8862d9f0776b0c0570f1fdd60a119d90954e3f39
+    Building wheel for intervaltree (setup.py) ... done
+    Created wheel for intervaltree: filename=intervaltree-3.1.0-py2.py3-none-any.whl size=26098 sha256=2604170976cfffe0d2f678cb1a6e5b525f561cd50babe53d631a186734fec9f9
+    Stored in directory: /star-fj/fangjun/.cache/pip/wheels/f3/ed/2b/c179ebfad4e15452d6baef59737f27beb9bfb442e0620f7271
+  Successfully built lhotse audioread intervaltree
+  Installing collected packages: sortedcontainers, dataclasses, tqdm, toolz, tabulate, pyyaml, pycparser, packaging, lilcom, intervaltree, click, audioread, cytoolz, cffi, SoundFile, lhotse
+  Successfully installed SoundFile-0.12.1 audioread-3.0.0 cffi-1.15.1 click-8.1.6 cytoolz-0.12.2 dataclasses-0.6 intervaltree-3.1.0 lhotse-1.16.0.dev0+git.7640d66.clean lilcom-1.7 packaging-23.1 pycparser-2.21 pyyaml-6.0.1 sortedcontainers-2.4.0 tabulate-0.9.0 toolz-0.12.0 tqdm-4.65.0
+
+
+Verify that `lhotse`_ has been installed successfully:
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ python3 -c "import lhotse; print(lhotse.__version__)"
+
+  1.16.0.dev+git.7640d66.clean
+
+(6) Download icefall
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:~$ cd /tmp/
+
+  (test-icefall) kuangfangjun:tmp$ git clone https://github.com/k2-fsa/icefall
+
+  Cloning into 'icefall'...
+  remote: Enumerating objects: 12942, done.
+  remote: Counting objects: 100% (67/67), done.
+  remote: Compressing objects: 100% (56/56), done.
+  remote: Total 12942 (delta 17), reused 35 (delta 6), pack-reused 12875
+  Receiving objects: 100% (12942/12942), 14.77 MiB | 9.29 MiB/s, done.
+  Resolving deltas: 100% (8835/8835), done.
+
+  (test-icefall) kuangfangjun:tmp$ cd icefall/
+
+  (test-icefall) kuangfangjun:icefall$ pip install -r ./requirements.txt
+
+Test Your Installation
+----------------------
+
+To test that your installation is successful, let us run
+the `yesno recipe <https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR>`_
+on ``CPU``.
+
+Data preparation
+~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  (test-icefall) kuangfangjun:icefall$ export PYTHONPATH=/tmp/icefall:$PYTHONPATH
+
+  (test-icefall) kuangfangjun:icefall$ cd /tmp/icefall
+
+  (test-icefall) kuangfangjun:icefall$ cd egs/yesno/ASR
+
+  (test-icefall) kuangfangjun:ASR$ ./prepare.sh
+
+
+The log of running ``./prepare.sh`` is:
+
+.. code-block::
+
+  2023-07-27 12:41:39 (prepare.sh:27:main) dl_dir: /tmp/icefall/egs/yesno/ASR/download
+  2023-07-27 12:41:39 (prepare.sh:30:main) Stage 0: Download data
+  /tmp/icefall/egs/yesno/ASR/download/waves_yesno.tar.gz: 100%|___________________________________________________| 4.70M/4.70M [00:00<00:00, 11.1MB/s]
+  2023-07-27 12:41:46 (prepare.sh:39:main) Stage 1: Prepare yesno manifest
+  2023-07-27 12:41:50 (prepare.sh:45:main) Stage 2: Compute fbank for yesno
+  2023-07-27 12:41:55,718 INFO [compute_fbank_yesno.py:65] Processing train
+  Extracting and storing features: 100%|_______________________________________________________________________________| 90/90 [00:01<00:00, 87.82it/s]
+  2023-07-27 12:41:56,778 INFO [compute_fbank_yesno.py:65] Processing test
+  Extracting and storing features: 100%|______________________________________________________________________________| 30/30 [00:00<00:00, 256.92it/s]
+  2023-07-27 12:41:57 (prepare.sh:51:main) Stage 3: Prepare lang
+  2023-07-27 12:42:02 (prepare.sh:66:main) Stage 4: Prepare G
+  /project/kaldilm/csrc/arpa_file_parser.cc:void kaldilm::ArpaFileParser::Read(std::istream&):79
+  [I] Reading \data\ section.
+  /project/kaldilm/csrc/arpa_file_parser.cc:void kaldilm::ArpaFileParser::Read(std::istream&):140
+  [I] Reading \1-grams: section.
+  2023-07-27 12:42:02 (prepare.sh:92:main) Stage 5: Compile HLG
+  2023-07-27 12:42:07,275 INFO [compile_hlg.py:124] Processing data/lang_phone
+  2023-07-27 12:42:07,276 INFO [lexicon.py:171] Converting L.pt to Linv.pt
+  2023-07-27 12:42:07,309 INFO [compile_hlg.py:48] Building ctc_topo. max_token_id: 3
+  2023-07-27 12:42:07,310 INFO [compile_hlg.py:52] Loading G.fst.txt
+  2023-07-27 12:42:07,314 INFO [compile_hlg.py:62] Intersecting L and G
+  2023-07-27 12:42:07,323 INFO [compile_hlg.py:64] LG shape: (4, None)
+  2023-07-27 12:42:07,323 INFO [compile_hlg.py:66] Connecting LG
+  2023-07-27 12:42:07,323 INFO [compile_hlg.py:68] LG shape after k2.connect: (4, None)
+  2023-07-27 12:42:07,323 INFO [compile_hlg.py:70] <class 'torch.Tensor'>
+  2023-07-27 12:42:07,323 INFO [compile_hlg.py:71] Determinizing LG
+  2023-07-27 12:42:07,341 INFO [compile_hlg.py:74] <class '_k2.ragged.RaggedTensor'>
+  2023-07-27 12:42:07,341 INFO [compile_hlg.py:76] Connecting LG after k2.determinize
+  2023-07-27 12:42:07,341 INFO [compile_hlg.py:79] Removing disambiguation symbols on LG
+  2023-07-27 12:42:07,354 INFO [compile_hlg.py:91] LG shape after k2.remove_epsilon: (6, None)
+  2023-07-27 12:42:07,445 INFO [compile_hlg.py:96] Arc sorting LG
+  2023-07-27 12:42:07,445 INFO [compile_hlg.py:99] Composing H and LG
+  2023-07-27 12:42:07,446 INFO [compile_hlg.py:106] Connecting LG
+  2023-07-27 12:42:07,446 INFO [compile_hlg.py:109] Arc sorting LG
+  2023-07-27 12:42:07,447 INFO [compile_hlg.py:111] HLG.shape: (8, None)
+  2023-07-27 12:42:07,447 INFO [compile_hlg.py:127] Saving HLG.pt to data/lang_phone
+
+Training
+~~~~~~~~
+
+Now let us run the training part:
+
+.. code-block::
+
+  (test-icefall) kuangfangjun:ASR$ export CUDA_VISIBLE_DEVICES=""
+
+  (test-icefall) kuangfangjun:ASR$ ./tdnn/train.py
+
+.. CAUTION::
+
+  We use ``export CUDA_VISIBLE_DEVICES=""`` so that `icefall`_ uses CPU
+  even if there are GPUs available.
+
+.. hint::
+
+   In case you get a ``Segmentation fault (core dump)`` error, please use:
+
+      .. code-block:: bash
+
+        export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+   See more at `<https://github.com/k2-fsa/icefall/issues/674>` if you are
+   interested.
+
+The training log is given below:
+
+.. code-block::
+
+    2023-07-27 12:50:51,936 INFO [train.py:481] Training started
+    2023-07-27 12:50:51,936 INFO [train.py:482] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lr': 0.01, 'feature_dim': 23, 'weight_decay': 1e-06, 'start_epoch': 0, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 10, 'reset_interval': 20, 'valid_interval': 10, 'beam_size': 10, 'reduction': 'sum', 'use_double_scores': True, 'world_size': 1, 'master_port': 12354, 'tensorboard': True, 'num_epochs': 15, 'seed': 42, 'feature_dir': PosixPath('data/fbank'), 'max_duration': 30.0, 'bucketing_sampler': False, 'num_buckets': 10, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': False, 'return_cuts': True, 'num_workers': 2, 'env_info': {'k2-version': '1.24.3', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '4c05309499a08454997adf500b56dcc629e35ae5', 'k2-git-date': 'Tue Jul 25 16:23:36 2023', 'lhotse-version': '1.16.0.dev+git.7640d66.clean', 'torch-version': '1.13.0+cu116', 'torch-cuda-available': False, 'torch-cuda-version': '11.6', 'python-version': '3.8', 'icefall-git-branch': 'master', 'icefall-git-sha1': '3fb0a43-clean', 'icefall-git-date': 'Thu Jul 27 12:36:05 2023', 'icefall-path': '/tmp/icefall', 'k2-path': '/star-fj/fangjun/test-icefall/lib/python3.8/site-packages/k2/__init__.py', 'lhotse-path': '/star-fj/fangjun/test-icefall/lib/python3.8/site-packages/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-1-1220091118-57c4d55446-sph26', 'IP address': '10.177.77.20'}}
+    2023-07-27 12:50:51,941 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+    2023-07-27 12:50:51,949 INFO [train.py:495] device: cpu
+    2023-07-27 12:50:51,965 INFO [asr_datamodule.py:146] About to get train cuts
+    2023-07-27 12:50:51,965 INFO [asr_datamodule.py:244] About to get train cuts
+    2023-07-27 12:50:51,967 INFO [asr_datamodule.py:149] About to create train dataset
+    2023-07-27 12:50:51,967 INFO [asr_datamodule.py:199] Using SingleCutSampler.
+    2023-07-27 12:50:51,967 INFO [asr_datamodule.py:205] About to create train dataloader
+    2023-07-27 12:50:51,968 INFO [asr_datamodule.py:218] About to get test cuts
+    2023-07-27 12:50:51,968 INFO [asr_datamodule.py:252] About to get test cuts
+    2023-07-27 12:50:52,565 INFO [train.py:422] Epoch 0, batch 0, loss[loss=1.065, over 2436.00 frames. ], tot_loss[loss=1.065, over 2436.00 frames. ], batch size: 4
+    2023-07-27 12:50:53,681 INFO [train.py:422] Epoch 0, batch 10, loss[loss=0.4561, over 2828.00 frames. ], tot_loss[loss=0.7076, over 22192.90 frames.], batch size: 4
+    2023-07-27 12:50:54,167 INFO [train.py:444] Epoch 0, validation loss=0.9002, over 18067.00 frames.
+    2023-07-27 12:50:55,011 INFO [train.py:422] Epoch 0, batch 20, loss[loss=0.2555, over 2695.00 frames. ], tot_loss[loss=0.484, over 34971.47 frames. ], batch size: 5
+    2023-07-27 12:50:55,331 INFO [train.py:444] Epoch 0, validation loss=0.4688, over 18067.00 frames.
+    2023-07-27 12:50:55,368 INFO [checkpoint.py:75] Saving checkpoint to tdnn/exp/epoch-0.pt
+    2023-07-27 12:50:55,633 INFO [train.py:422] Epoch 1, batch 0, loss[loss=0.2532, over 2436.00 frames. ], tot_loss[loss=0.2532, over 2436.00 frames. ],
+     batch size: 4
+    2023-07-27 12:50:56,242 INFO [train.py:422] Epoch 1, batch 10, loss[loss=0.1139, over 2828.00 frames. ], tot_loss[loss=0.1592, over 22192.90 frames.], batch size: 4
+    2023-07-27 12:50:56,522 INFO [train.py:444] Epoch 1, validation loss=0.1627, over 18067.00 frames.
+    2023-07-27 12:50:57,209 INFO [train.py:422] Epoch 1, batch 20, loss[loss=0.07055, over 2695.00 frames. ], tot_loss[loss=0.1175, over 34971.47 frames.], batch size: 5
+    2023-07-27 12:50:57,600 INFO [train.py:444] Epoch 1, validation loss=0.07091, over 18067.00 frames.
+    2023-07-27 12:50:57,640 INFO [checkpoint.py:75] Saving checkpoint to tdnn/exp/epoch-1.pt
+    2023-07-27 12:50:57,847 INFO [train.py:422] Epoch 2, batch 0, loss[loss=0.07731, over 2436.00 frames. ], tot_loss[loss=0.07731, over 2436.00 frames.], batch size: 4
+    2023-07-27 12:50:58,427 INFO [train.py:422] Epoch 2, batch 10, loss[loss=0.04391, over 2828.00 frames. ], tot_loss[loss=0.05341, over 22192.90 frames. ], batch size: 4
+    2023-07-27 12:50:58,884 INFO [train.py:444] Epoch 2, validation loss=0.04384, over 18067.00 frames.
+    2023-07-27 12:50:59,387 INFO [train.py:422] Epoch 2, batch 20, loss[loss=0.03458, over 2695.00 frames. ], tot_loss[loss=0.04616, over 34971.47 frames. ], batch size: 5
+    2023-07-27 12:50:59,707 INFO [train.py:444] Epoch 2, validation loss=0.03379, over 18067.00 frames.
+    2023-07-27 12:50:59,758 INFO [checkpoint.py:75] Saving checkpoint to tdnn/exp/epoch-2.pt
+
+      ... ...
+
+    2023-07-27 12:51:23,433 INFO [train.py:422] Epoch 13, batch 0, loss[loss=0.01054, over 2436.00 frames. ], tot_loss[loss=0.01054, over 2436.00 frames. ], batch size: 4
+    2023-07-27 12:51:23,980 INFO [train.py:422] Epoch 13, batch 10, loss[loss=0.009014, over 2828.00 frames. ], tot_loss[loss=0.009974, over 22192.90 frames. ], batch size: 4
+    2023-07-27 12:51:24,489 INFO [train.py:444] Epoch 13, validation loss=0.01085, over 18067.00 frames.
+    2023-07-27 12:51:25,258 INFO [train.py:422] Epoch 13, batch 20, loss[loss=0.01172, over 2695.00 frames. ], tot_loss[loss=0.01055, over 34971.47 frames. ], batch size: 5
+    2023-07-27 12:51:25,621 INFO [train.py:444] Epoch 13, validation loss=0.01074, over 18067.00 frames.
+    2023-07-27 12:51:25,699 INFO [checkpoint.py:75] Saving checkpoint to tdnn/exp/epoch-13.pt
+    2023-07-27 12:51:25,866 INFO [train.py:422] Epoch 14, batch 0, loss[loss=0.01044, over 2436.00 frames. ], tot_loss[loss=0.01044, over 2436.00 frames. ], batch size: 4
+    2023-07-27 12:51:26,844 INFO [train.py:422] Epoch 14, batch 10, loss[loss=0.008942, over 2828.00 frames. ], tot_loss[loss=0.01, over 22192.90 frames. ], batch size: 4
+    2023-07-27 12:51:27,221 INFO [train.py:444] Epoch 14, validation loss=0.01082, over 18067.00 frames.
+    2023-07-27 12:51:27,970 INFO [train.py:422] Epoch 14, batch 20, loss[loss=0.01169, over 2695.00 frames. ], tot_loss[loss=0.01054, over 34971.47 frames. ], batch size: 5
+    2023-07-27 12:51:28,247 INFO [train.py:444] Epoch 14, validation loss=0.01073, over 18067.00 frames.
+    2023-07-27 12:51:28,323 INFO [checkpoint.py:75] Saving checkpoint to tdnn/exp/epoch-14.pt
+    2023-07-27 12:51:28,326 INFO [train.py:555] Done!
+
+Decoding
+~~~~~~~~
+
+Let us use the trained model to decode the test set:
+
+.. code-block::
+
+  (test-icefall) kuangfangjun:ASR$ ./tdnn/decode.py
+
+  2023-07-27 12:55:12,840 INFO [decode.py:263] Decoding started
+  2023-07-27 12:55:12,840 INFO [decode.py:264] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lm_dir': PosixPath('data/lm'), 'feature_dim': 23, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'epoch': 14, 'avg': 2, 'export': False, 'feature_dir': PosixPath('data/fbank'), 'max_duration': 30.0, 'bucketing_sampler': False, 'num_buckets': 10, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': False, 'return_cuts': True, 'num_workers': 2, 'env_info': {'k2-version': '1.24.3', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '4c05309499a08454997adf500b56dcc629e35ae5', 'k2-git-date': 'Tue Jul 25 16:23:36 2023', 'lhotse-version': '1.16.0.dev+git.7640d66.clean', 'torch-version': '1.13.0+cu116', 'torch-cuda-available': False, 'torch-cuda-version': '11.6', 'python-version': '3.8', 'icefall-git-branch': 'master', 'icefall-git-sha1': '3fb0a43-clean', 'icefall-git-date': 'Thu Jul 27 12:36:05 2023', 'icefall-path': '/tmp/icefall', 'k2-path': '/star-fj/fangjun/test-icefall/lib/python3.8/site-packages/k2/__init__.py', 'lhotse-path': '/star-fj/fangjun/test-icefall/lib/python3.8/site-packages/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-1-1220091118-57c4d55446-sph26', 'IP address': '10.177.77.20'}}
+  2023-07-27 12:55:12,841 INFO [lexicon.py:168] Loading pre-compiled data/lang_phone/Linv.pt
+  2023-07-27 12:55:12,855 INFO [decode.py:273] device: cpu
+  2023-07-27 12:55:12,868 INFO [decode.py:291] averaging ['tdnn/exp/epoch-13.pt', 'tdnn/exp/epoch-14.pt']
+  2023-07-27 12:55:12,882 INFO [asr_datamodule.py:218] About to get test cuts
+  2023-07-27 12:55:12,883 INFO [asr_datamodule.py:252] About to get test cuts
+  2023-07-27 12:55:13,157 INFO [decode.py:204] batch 0/?, cuts processed until now is 4
+  2023-07-27 12:55:13,701 INFO [decode.py:241] The transcripts are stored in tdnn/exp/recogs-test_set.txt
+  2023-07-27 12:55:13,702 INFO [utils.py:564] [test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
+  2023-07-27 12:55:13,704 INFO [decode.py:249] Wrote detailed error stats to tdnn/exp/errs-test_set.txt
+  2023-07-27 12:55:13,704 INFO [decode.py:316] Done!
+
+
+**Congratulations!** You have successfully setup the environment and have run the first recipe in `icefall`_.
+
+Have fun with ``icefall``!
+
+YouTube Video
+-------------
+
+We provide the following YouTube video showing how to install `icefall`_.
+It also shows how to debug various problems that you may encounter while
+using `icefall`_.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: LVmrBD0tLfE
diff --git a/docs/source/model-export/export-model-state-dict.rst b/docs/source/model-export/export-model-state-dict.rst
new file mode 100644
index 000000000..5596bb7a6
--- /dev/null
+++ b/docs/source/model-export/export-model-state-dict.rst
@@ -0,0 +1,135 @@
+Export model.state_dict()
+=========================
+
+When to use it
+--------------
+
+During model training, we save checkpoints periodically to disk.
+
+A checkpoint contains the following information:
+
+  - ``model.state_dict()``
+  - ``optimizer.state_dict()``
+  - and some other information related to training
+
+When we need to resume the training process from some point, we need a checkpoint.
+However, if we want to publish the model for inference, then only
+``model.state_dict()`` is needed. In this case, we need to strip all other information
+except ``model.state_dict()`` to reduce the file size of the published model.
+
+How to export
+-------------
+
+Every recipe contains a file ``export.py`` that you can use to
+export ``model.state_dict()`` by taking some checkpoints as inputs.
+
+.. hint::
+
+   Each ``export.py`` contains well-documented usage information.
+
+In the following, we use
+`<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless3/export.py>`_
+as an example.
+
+.. note::
+
+   The steps for other recipes are almost the same.
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+
+  ./pruned_transducer_stateless3/export.py \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --epoch 20 \
+    --avg 10
+
+will generate a file ``pruned_transducer_stateless3/exp/pretrained.pt``, which
+is a dict containing ``{"model": model.state_dict()}`` saved by ``torch.save()``.
+
+How to use the exported model
+-----------------------------
+
+For each recipe, we provide pretrained models hosted on huggingface.
+You can find links to pretrained models in ``RESULTS.md`` of each dataset.
+
+In the following, we demonstrate how to use the pretrained model from
+`<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13>`_.
+
+.. code-block:: bash
+
+   cd egs/librispeech/ASR
+
+   git lfs install
+   git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+
+After cloning the repo with ``git lfs``, you will find several files in the folder
+``icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/exp``
+that have a prefix ``pretrained-``. Those files contain ``model.state_dict()``
+exported by the above ``export.py``.
+
+In each recipe, there is also a file ``pretrained.py``, which can use
+``pretrained-xxx.pt`` to decode waves. The following is an example:
+
+.. code-block:: bash
+
+   cd egs/librispeech/ASR
+
+   ./pruned_transducer_stateless3/pretrained.py \
+      --checkpoint ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/exp/pretrained-iter-1224000-avg-14.pt \
+      --tokens ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/data/lang_bpe_500/tokens.txt \
+      --method greedy_search \
+      ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/test_wavs/1089-134686-0001.wav \
+      ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/test_wavs/1221-135766-0001.wav \
+      ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/test_wavs/1221-135766-0002.wav
+
+The above commands show how to use the exported model with ``pretrained.py`` to
+decode multiple sound files. Its output is given as follows for reference:
+
+.. literalinclude:: ./code/export-model-state-dict-pretrained-out.txt
+
+Use the exported model to run decode.py
+---------------------------------------
+
+When we publish the model, we always note down its WERs on some test
+dataset in ``RESULTS.md``. This section describes how to use the
+pretrained model to reproduce the WER.
+
+.. code-block:: bash
+
+   cd egs/librispeech/ASR
+   git lfs install
+   git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+
+   cd icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/exp
+   ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+   cd ../..
+
+We create a symlink with name ``epoch-9999.pt`` to ``pretrained-iter-1224000-avg-14.pt``,
+so that we can pass ``--epoch 9999 --avg 1`` to ``decode.py`` in the following
+command:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless3/decode.py \
+      --epoch 9999 \
+      --avg 1 \
+      --exp-dir ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/exp \
+      --lang-dir ./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/data/lang_bpe_500 \
+      --max-duration 600 \
+      --decoding-method greedy_search
+
+You will find the decoding results in
+``./icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/exp/greedy_search``.
+
+.. caution::
+
+   For some recipes, you also need to pass ``--use-averaged-model False``
+   to ``decode.py``. The reason is that the exported pretrained model is already
+   the averaged one.
+
+.. hint::
+
+   Before running ``decode.py``, we assume that you have already run
+   ``prepare.sh`` to prepare the test dataset.
diff --git a/docs/source/model-export/export-ncnn-conv-emformer.rst b/docs/source/model-export/export-ncnn-conv-emformer.rst
new file mode 100644
index 000000000..93392aee7
--- /dev/null
+++ b/docs/source/model-export/export-ncnn-conv-emformer.rst
@@ -0,0 +1,752 @@
+.. _export_conv_emformer_transducer_models_to_ncnn:
+
+Export ConvEmformer transducer models to ncnn
+=============================================
+
+We use the pre-trained model from the following repository as an example:
+
+  - `<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>`_
+
+We will show you step by step how to export it to `ncnn`_ and run it with `sherpa-ncnn`_.
+
+.. hint::
+
+  We use ``Ubuntu 18.04``, ``torch 1.13``, and ``Python 3.8`` for testing.
+
+.. caution::
+
+  Please use a more recent version of PyTorch. For instance, ``torch 1.8``
+  may ``not`` work.
+
+1. Download the pre-trained model
+---------------------------------
+
+.. hint::
+
+  You can also refer to `<https://k2-fsa.github.io/sherpa/cpp/pretrained_models/online_transducer.html#icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>`_ to download the pre-trained model.
+
+  You have to install `git-lfs`_ before you continue.
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+
+  GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+
+  git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+  git lfs pull --include "data/lang_bpe_500/bpe.model"
+
+  cd ..
+
+.. note::
+
+  We downloaded ``exp/pretrained-xxx.pt``, not ``exp/cpu-jit_xxx.pt``.
+
+
+In the above code, we downloaded the pre-trained model into the directory
+``egs/librispeech/ASR/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05``.
+
+.. _export_for_ncnn_install_ncnn_and_pnnx:
+
+2. Install ncnn and pnnx
+------------------------
+
+.. code-block:: bash
+
+  # We put ncnn into $HOME/open-source/ncnn
+  # You can change it to anywhere you like
+
+  cd $HOME
+  mkdir -p open-source
+  cd open-source
+
+  git clone https://github.com/csukuangfj/ncnn
+  cd ncnn
+  git submodule update --recursive --init
+
+  # Note: We don't use "python setup.py install" or "pip install ." here
+
+  mkdir -p build-wheel
+  cd build-wheel
+
+  cmake \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DNCNN_PYTHON=ON \
+    -DNCNN_BUILD_BENCHMARK=OFF \
+    -DNCNN_BUILD_EXAMPLES=OFF \
+    -DNCNN_BUILD_TOOLS=ON \
+  ..
+
+  make -j4
+
+  cd ..
+
+  # Note: $PWD here is $HOME/open-source/ncnn
+
+  export PYTHONPATH=$PWD/python:$PYTHONPATH
+  export PATH=$PWD/tools/pnnx/build/src:$PATH
+  export PATH=$PWD/build-wheel/tools/quantize:$PATH
+
+  # Now build pnnx
+  cd tools/pnnx
+  mkdir build
+  cd build
+  cmake ..
+  make -j4
+
+  ./src/pnnx
+
+Congratulations! You have successfully installed the following components:
+
+  - ``pnnx``, which is an executable located in
+    ``$HOME/open-source/ncnn/tools/pnnx/build/src``. We will use
+    it to convert models exported by ``torch.jit.trace()``.
+  - ``ncnn2int8``, which is an executable located in
+    ``$HOME/open-source/ncnn/build-wheel/tools/quantize``. We will use
+    it to quantize our models to ``int8``.
+  - ``ncnn.cpython-38-x86_64-linux-gnu.so``, which is a Python module located
+    in ``$HOME/open-source/ncnn/python/ncnn``.
+
+    .. note::
+
+      I am using ``Python 3.8``, so it
+      is ``ncnn.cpython-38-x86_64-linux-gnu.so``. If you use a different
+      version, say, ``Python 3.9``, the name would be
+      ``ncnn.cpython-39-x86_64-linux-gnu.so``.
+
+      Also, if you are not using Linux, the file name would also be different.
+      But that does not matter. As long as you can compile it, it should work.
+
+We have set up ``PYTHONPATH`` so that you can use ``import ncnn`` in your
+Python code. We have also set up ``PATH`` so that you can use
+``pnnx`` and ``ncnn2int8`` later in your terminal.
+
+.. caution::
+
+  Please don't use `<https://github.com/tencent/ncnn>`_.
+  We have made some modifications to the official `ncnn`_.
+
+  We will synchronize `<https://github.com/csukuangfj/ncnn>`_ periodically
+  with the official one.
+
+3. Export the model via torch.jit.trace()
+-----------------------------------------
+
+First, let us rename our pre-trained model:
+
+.. code-block::
+
+  cd egs/librispeech/ASR
+
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp
+
+  ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-30.pt
+
+  cd ../..
+
+Next, we use the following code to export our model:
+
+.. code-block:: bash
+
+  dir=./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/
+
+  ./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+    --exp-dir $dir/exp \
+    --tokens $dir/data/lang_bpe_500/tokens.txt \
+    --epoch 30 \
+    --avg 1 \
+    --use-averaged-model 0 \
+    --num-encoder-layers 12 \
+    --chunk-length 32 \
+    --cnn-module-kernel 31 \
+    --left-context-length 32 \
+    --right-context-length 8 \
+    --memory-size 32 \
+    --encoder-dim 512
+
+.. caution::
+
+   If your model has different configuration parameters, please change them accordingly.
+
+.. hint::
+
+  We have renamed our model to ``epoch-30.pt`` so that we can use ``--epoch 30``.
+  There is only one pre-trained model, so we use ``--avg 1 --use-averaged-model 0``.
+
+  If you have trained a model by yourself and if you have all checkpoints
+  available, please first use ``decode.py`` to tune ``--epoch --avg``
+  and select the best combination with with ``--use-averaged-model 1``.
+
+.. note::
+
+  You will see the following log output:
+
+  .. literalinclude:: ./code/export-conv-emformer-transducer-for-ncnn-output.txt
+
+  The log shows the model has ``75490012`` parameters, i.e., ``~75 M``.
+
+  .. code-block::
+
+    ls -lh icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/pretrained-epoch-30-avg-10-averaged.pt
+
+    -rw-r--r-- 1 kuangfangjun root 289M Jan 11 12:05 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/pretrained-epoch-30-avg-10-averaged.pt
+
+  You can see that the file size of the pre-trained model is ``289 MB``, which
+  is roughly equal to ``75490012*4/1024/1024 = 287.97 MB``.
+
+After running ``conv_emformer_transducer_stateless2/export-for-ncnn.py``,
+we will get the following files:
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/*pnnx*
+
+  -rw-r--r-- 1 kuangfangjun root 1010K Jan 11 12:15 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  283M Jan 11 12:15 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  3.0M Jan 11 12:15 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.pt
+
+
+.. _conv-emformer-step-4-export-torchscript-model-via-pnnx:
+
+4. Export torchscript model via pnnx
+------------------------------------
+
+.. hint::
+
+  Make sure you have set up the ``PATH`` environment variable. Otherwise,
+  it will throw an error saying that ``pnnx`` could not be found.
+
+Now, it's time to export our models to `ncnn`_ via ``pnnx``.
+
+.. code-block::
+
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt
+
+It will generate the following files:
+
+.. code-block:: bash
+
+  ls -lh  icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/*ncnn*{bin,param}
+
+  -rw-r--r-- 1 kuangfangjun root 503K Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 142M Jan 11 12:36 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  79K Jan 11 12:36 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 1.5M Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param
+
+There are two types of files:
+
+- ``param``: It is a text file containing the model architectures. You can
+  use a text editor to view its content.
+- ``bin``: It is a binary file containing the model parameters.
+
+We compare the file sizes of the models below before and after converting via ``pnnx``:
+
+.. see https://tableconvert.com/restructuredtext-generator
+
++----------------------------------+------------+
+| File name                        | File size  |
++==================================+============+
+| encoder_jit_trace-pnnx.pt        | 283 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.pt        | 1010 KB    |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.pt         | 3.0 MB     |
++----------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin  | 142 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin  | 503 KB     |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin   | 1.5 MB     |
++----------------------------------+------------+
+
+You can see that the file sizes of the models after conversion are about one half
+of the models before conversion:
+
+  - encoder: 283 MB vs 142 MB
+  - decoder: 1010 KB vs 503 KB
+  - joiner: 3.0 MB vs 1.5 MB
+
+The reason is that by default ``pnnx`` converts ``float32`` parameters
+to ``float16``. A ``float32`` parameter occupies 4 bytes, while it is 2 bytes
+for ``float16``. Thus, it is ``twice smaller`` after conversion.
+
+.. hint::
+
+  If you use ``pnnx ./encoder_jit_trace-pnnx.pt fp16=0``, then ``pnnx``
+  won't convert ``float32`` to ``float16``.
+
+5. Test the exported models in icefall
+--------------------------------------
+
+.. note::
+
+  We assume you have set up the environment variable ``PYTHONPATH`` when
+  building `ncnn`_.
+
+Now we have successfully converted our pre-trained model to `ncnn`_ format.
+The generated 6 files are what we need. You can use the following code to
+test the converted models:
+
+.. code-block:: bash
+
+  ./conv_emformer_transducer_stateless2/streaming-ncnn-decode.py \
+    --tokens ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/data/lang_bpe_500/tokens.txt \
+    --encoder-param-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param \
+    --encoder-bin-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin \
+    --decoder-param-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param \
+    --decoder-bin-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin \
+    --joiner-param-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param \
+    --joiner-bin-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin \
+    ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/test_wavs/1089-134686-0001.wav
+
+.. hint::
+
+  `ncnn`_ supports only ``batch size == 1``, so ``streaming-ncnn-decode.py`` accepts
+  only 1 wave file as input.
+
+The output is given below:
+
+.. literalinclude:: ./code/test-streaming-ncnn-decode-conv-emformer-transducer-libri.txt
+
+Congratulations! You have successfully exported a model from PyTorch to `ncnn`_!
+
+
+.. _conv-emformer-modify-the-exported-encoder-for-sherpa-ncnn:
+
+6. Modify the exported encoder for sherpa-ncnn
+----------------------------------------------
+
+In order to use the exported models in `sherpa-ncnn`_, we have to modify
+``encoder_jit_trace-pnnx.ncnn.param``.
+
+Let us have a look at the first few lines of ``encoder_jit_trace-pnnx.ncnn.param``:
+
+.. code-block::
+
+  7767517
+  1060 1342
+  Input                    in0                      0 1 in0
+
+**Explanation** of the above three lines:
+
+  1. ``7767517``, it is a magic number and should not be changed.
+  2. ``1060 1342``, the first number ``1060`` specifies the number of layers
+     in this file, while ``1342`` specifies the number of intermediate outputs
+     of this file
+  3. ``Input in0 0 1 in0``, ``Input`` is the layer type of this layer; ``in0``
+     is the layer name of this layer; ``0`` means this layer has no input;
+     ``1`` means this layer has one output; ``in0`` is the output name of
+     this layer.
+
+We need to add 1 extra line and also increment the number of layers.
+The result looks like below:
+
+.. code-block:: bash
+
+  7767517
+  1061 1342
+  SherpaMetaData           sherpa_meta_data1        0 0 0=1 1=12 2=32 3=31 4=8 5=32 6=8 7=512
+  Input                    in0                      0 1 in0
+
+**Explanation**
+
+  1. ``7767517``, it is still the same
+  2. ``1061 1342``, we have added an extra layer, so we need to update ``1060`` to ``1061``.
+     We don't need to change ``1342`` since the newly added layer has no inputs or outputs.
+  3. ``SherpaMetaData  sherpa_meta_data1  0 0 0=1 1=12 2=32 3=31 4=8 5=32 6=8 7=512``
+     This line is newly added. Its explanation is given below:
+
+      - ``SherpaMetaData`` is the type of this layer. Must be ``SherpaMetaData``.
+      - ``sherpa_meta_data1`` is the name of this layer. Must be ``sherpa_meta_data1``.
+      - ``0 0`` means this layer has no inputs or output. Must be ``0 0``
+      - ``0=1``, 0 is the key and 1 is the value. MUST be ``0=1``
+      - ``1=12``, 1 is the key and 12 is the value of the
+        parameter ``--num-encoder-layers`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``2=32``, 2 is the key and 32 is the value of the
+        parameter ``--memory-size`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``3=31``, 3 is the key and 31 is the value of the
+        parameter ``--cnn-module-kernel`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``4=8``, 4 is the key and 8 is the value of the
+        parameter ``--left-context-length`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``5=32``, 5 is the key and 32 is the value of the
+        parameter ``--chunk-length`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``6=8``, 6 is the key and 8 is the value of the
+        parameter ``--right-context-length`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``7=512``, 7 is the key and 512 is the value of the
+        parameter ``--encoder-dim`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+
+      For ease of reference, we list the key-value pairs that you need to add
+      in the following table. If your model has a different setting, please
+      change the values for ``SherpaMetaData`` accordingly. Otherwise, you
+      will be ``SAD``.
+
+          +------+-----------------------------+
+          | key  | value                       |
+          +======+=============================+
+          | 0    | 1 (fixed)                   |
+          +------+-----------------------------+
+          | 1    | ``--num-encoder-layers``    |
+          +------+-----------------------------+
+          | 2    | ``--memory-size``           |
+          +------+-----------------------------+
+          | 3    | ``--cnn-module-kernel``     |
+          +------+-----------------------------+
+          | 4    | ``--left-context-length``   |
+          +------+-----------------------------+
+          | 5    | ``--chunk-length``          |
+          +------+-----------------------------+
+          | 6    | ``--right-context-length``  |
+          +------+-----------------------------+
+          | 7    | ``--encoder-dim``           |
+          +------+-----------------------------+
+
+  4. ``Input in0 0 1 in0``. No need to change it.
+
+.. caution::
+
+  When you add a new layer ``SherpaMetaData``, please remember to update the
+  number of layers. In our case, update  ``1060`` to ``1061``. Otherwise,
+  you will be SAD later.
+
+.. hint::
+
+  After adding the new layer ``SherpaMetaData``, you cannot use this model
+  with ``streaming-ncnn-decode.py`` anymore since ``SherpaMetaData`` is
+  supported only in `sherpa-ncnn`_.
+
+.. hint::
+
+  `ncnn`_ is very flexible. You can add new layers to it just by text-editing
+  the ``param`` file! You don't need to change the ``bin`` file.
+
+Now you can use this model in `sherpa-ncnn`_.
+Please refer to the following documentation:
+
+  - Linux/macOS/Windows/arm/aarch64: `<https://k2-fsa.github.io/sherpa/ncnn/install/index.html>`_
+  - ``Android``: `<https://k2-fsa.github.io/sherpa/ncnn/android/index.html>`_
+  - ``iOS``: `<https://k2-fsa.github.io/sherpa/ncnn/ios/index.html>`_
+  - Python: `<https://k2-fsa.github.io/sherpa/ncnn/python/index.html>`_
+
+We have a list of pre-trained models that have been exported for `sherpa-ncnn`_:
+
+  - `<https://k2-fsa.github.io/sherpa/ncnn/pretrained_models/index.html>`_
+
+    You can find more usages there.
+
+7. (Optional) int8 quantization with sherpa-ncnn
+------------------------------------------------
+
+This step is optional.
+
+In this step, we describe how to quantize our model with ``int8``.
+
+Change :ref:`conv-emformer-step-4-export-torchscript-model-via-pnnx` to
+disable ``fp16`` when using ``pnnx``:
+
+.. code-block::
+
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt fp16=0
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt fp16=0
+
+.. note::
+
+  We add ``fp16=0`` when exporting the encoder and joiner. `ncnn`_ does not
+  support quantizing the decoder model yet. We will update this documentation
+  once `ncnn`_ supports it. (Maybe in this year, 2023).
+
+It will generate the following files
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/*_jit_trace-pnnx.ncnn.{param,bin}
+
+  -rw-r--r-- 1 kuangfangjun root 503K Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 283M Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  79K Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 3.0M Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param
+
+Let us compare again the file sizes:
+
++----------------------------------------+------------+
+| File name                              | File size  |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.pt              | 283 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.pt              | 1010 KB    |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.pt               | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp16) | 142 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin (fp16) | 503 KB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp16) | 1.5 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp32) | 283 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp32) | 3.0 MB     |
++----------------------------------------+------------+
+
+You can see that the file sizes are doubled when we disable ``fp16``.
+
+.. note::
+
+  You can again use ``streaming-ncnn-decode.py`` to test the exported models.
+
+Next, follow :ref:`conv-emformer-modify-the-exported-encoder-for-sherpa-ncnn`
+to modify ``encoder_jit_trace-pnnx.ncnn.param``.
+
+Change
+
+.. code-block:: bash
+
+  7767517
+  1060 1342
+  Input                    in0                      0 1 in0
+
+to
+
+.. code-block:: bash
+
+  7767517
+  1061 1342
+  SherpaMetaData           sherpa_meta_data1        0 0 0=1 1=12 2=32 3=31 4=8 5=32 6=8 7=512
+  Input                    in0                      0 1 in0
+
+.. caution::
+
+  Please follow :ref:`conv-emformer-modify-the-exported-encoder-for-sherpa-ncnn`
+  to change the values for ``SherpaMetaData`` if your model uses a different setting.
+
+
+Next, let us compile `sherpa-ncnn`_ since we will quantize our models within
+`sherpa-ncnn`_.
+
+.. code-block:: bash
+
+  # We will download sherpa-ncnn to $HOME/open-source/
+  # You can change it to anywhere you like.
+  cd $HOME
+  mkdir -p open-source
+
+  cd open-source
+  git clone https://github.com/k2-fsa/sherpa-ncnn
+  cd sherpa-ncnn
+  mkdir build
+  cd build
+  cmake ..
+  make -j 4
+
+  ./bin/generate-int8-scale-table
+
+  export PATH=$HOME/open-source/sherpa-ncnn/build/bin:$PATH
+
+The output of the above commands are:
+
+.. code-block:: bash
+
+  (py38) kuangfangjun:build$ generate-int8-scale-table
+  Please provide 10 arg. Currently given: 1
+  Usage:
+  generate-int8-scale-table encoder.param encoder.bin decoder.param decoder.bin joiner.param joiner.bin encoder-scale-table.txt joiner-scale-table.txt wave_filenames.txt
+
+  Each line in wave_filenames.txt is a path to some 16k Hz mono wave file.
+
+We need to create a file ``wave_filenames.txt``, in which we need to put
+some calibration wave files. For testing purpose, we put the ``test_wavs``
+from the pre-trained model repository `<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>`_
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  cat <<EOF > wave_filenames.txt
+  ../test_wavs/1089-134686-0001.wav
+  ../test_wavs/1221-135766-0001.wav
+  ../test_wavs/1221-135766-0002.wav
+  EOF
+
+Now we can calculate the scales needed for quantization with the calibration data:
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  generate-int8-scale-table \
+    ./encoder_jit_trace-pnnx.ncnn.param \
+    ./encoder_jit_trace-pnnx.ncnn.bin \
+    ./decoder_jit_trace-pnnx.ncnn.param \
+    ./decoder_jit_trace-pnnx.ncnn.bin \
+    ./joiner_jit_trace-pnnx.ncnn.param \
+    ./joiner_jit_trace-pnnx.ncnn.bin \
+    ./encoder-scale-table.txt \
+    ./joiner-scale-table.txt \
+    ./wave_filenames.txt
+
+The output logs are in the following:
+
+.. literalinclude:: ./code/generate-int-8-scale-table-for-conv-emformer.txt
+
+It generates the following two files:
+
+.. code-block:: bash
+
+  $ ls -lh encoder-scale-table.txt joiner-scale-table.txt
+  -rw-r--r-- 1 kuangfangjun root 955K Jan 11 17:28 encoder-scale-table.txt
+  -rw-r--r-- 1 kuangfangjun root  18K Jan 11 17:28 joiner-scale-table.txt
+
+.. caution::
+
+  Definitely, you need more calibration data to compute the scale table.
+
+Finally, let us use the scale table to quantize our models into ``int8``.
+
+.. code-block:: bash
+
+  ncnn2int8
+
+  usage: ncnn2int8 [inparam] [inbin] [outparam] [outbin] [calibration table]
+
+First, we quantize the encoder model:
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  ncnn2int8 \
+    ./encoder_jit_trace-pnnx.ncnn.param \
+    ./encoder_jit_trace-pnnx.ncnn.bin \
+    ./encoder_jit_trace-pnnx.ncnn.int8.param \
+    ./encoder_jit_trace-pnnx.ncnn.int8.bin \
+    ./encoder-scale-table.txt
+
+Next, we quantize the joiner model:
+
+.. code-block:: bash
+
+  ncnn2int8 \
+    ./joiner_jit_trace-pnnx.ncnn.param \
+    ./joiner_jit_trace-pnnx.ncnn.bin \
+    ./joiner_jit_trace-pnnx.ncnn.int8.param \
+    ./joiner_jit_trace-pnnx.ncnn.int8.bin \
+    ./joiner-scale-table.txt
+
+The above two commands generate the following 4 files:
+
+.. code-block:: bash
+
+  -rw-r--r-- 1 kuangfangjun root  99M Jan 11 17:34 encoder_jit_trace-pnnx.ncnn.int8.bin
+  -rw-r--r-- 1 kuangfangjun root  78K Jan 11 17:34 encoder_jit_trace-pnnx.ncnn.int8.param
+  -rw-r--r-- 1 kuangfangjun root 774K Jan 11 17:35 joiner_jit_trace-pnnx.ncnn.int8.bin
+  -rw-r--r-- 1 kuangfangjun root  496 Jan 11 17:35 joiner_jit_trace-pnnx.ncnn.int8.param
+
+Congratulations! You have successfully quantized your model from ``float32`` to ``int8``.
+
+.. caution::
+
+  ``ncnn.int8.param`` and ``ncnn.int8.bin`` must be used in pairs.
+
+  You can replace ``ncnn.param`` and ``ncnn.bin`` with ``ncnn.int8.param``
+  and ``ncnn.int8.bin`` in `sherpa-ncnn`_ if you like.
+
+  For instance, to use only the ``int8`` encoder in ``sherpa-ncnn``, you can
+  replace the following invocation:
+
+    .. code-block:: bash
+
+      cd egs/librispeech/ASR
+      cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+      sherpa-ncnn \
+        ../data/lang_bpe_500/tokens.txt \
+        ./encoder_jit_trace-pnnx.ncnn.param \
+        ./encoder_jit_trace-pnnx.ncnn.bin \
+        ./decoder_jit_trace-pnnx.ncnn.param \
+        ./decoder_jit_trace-pnnx.ncnn.bin \
+        ./joiner_jit_trace-pnnx.ncnn.param \
+        ./joiner_jit_trace-pnnx.ncnn.bin \
+        ../test_wavs/1089-134686-0001.wav
+
+  with
+
+    .. code-block::
+
+      cd egs/librispeech/ASR
+      cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+      sherpa-ncnn \
+        ../data/lang_bpe_500/tokens.txt \
+        ./encoder_jit_trace-pnnx.ncnn.int8.param \
+        ./encoder_jit_trace-pnnx.ncnn.int8.bin \
+        ./decoder_jit_trace-pnnx.ncnn.param \
+        ./decoder_jit_trace-pnnx.ncnn.bin \
+        ./joiner_jit_trace-pnnx.ncnn.param \
+        ./joiner_jit_trace-pnnx.ncnn.bin \
+        ../test_wavs/1089-134686-0001.wav
+
+
+The following table compares again the file sizes:
+
+
++----------------------------------------+------------+
+| File name                              | File size  |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.pt              | 283 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.pt              | 1010 KB    |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.pt               | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp16) | 142 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin (fp16) | 503 KB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp16) | 1.5 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp32) | 283 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp32) | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.int8.bin   | 99 MB      |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.int8.bin    | 774 KB     |
++----------------------------------------+------------+
+
+You can see that the file sizes of the model after ``int8`` quantization
+are much smaller.
+
+.. hint::
+
+    Currently, only linear layers and convolutional layers are quantized
+    with ``int8``, so you don't see an exact ``4x`` reduction in file sizes.
+
+.. note::
+
+  You need to test the recognition accuracy after ``int8`` quantization.
+
+You can find the speed comparison at `<https://github.com/k2-fsa/sherpa-ncnn/issues/44>`_.
+
+
+That's it! Have fun with `sherpa-ncnn`_!
diff --git a/docs/source/model-export/export-ncnn-lstm.rst b/docs/source/model-export/export-ncnn-lstm.rst
new file mode 100644
index 000000000..310c3d8e4
--- /dev/null
+++ b/docs/source/model-export/export-ncnn-lstm.rst
@@ -0,0 +1,644 @@
+.. _export_lstm_transducer_models_to_ncnn:
+
+Export LSTM transducer models to ncnn
+-------------------------------------
+
+We use the pre-trained model from the following repository as an example:
+
+`<https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03>`_
+
+We will show you step by step how to export it to `ncnn`_ and run it with `sherpa-ncnn`_.
+
+.. hint::
+
+  We use ``Ubuntu 18.04``, ``torch 1.13``, and ``Python 3.8`` for testing.
+
+.. caution::
+
+  Please use a more recent version of PyTorch. For instance, ``torch 1.8``
+  may ``not`` work.
+
+1. Download the pre-trained model
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. hint::
+
+  You have to install `git-lfs`_ before you continue.
+
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+
+  git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+  git lfs pull --include "data/lang_bpe_500/bpe.model"
+
+  cd ..
+
+.. note::
+
+  We downloaded ``exp/pretrained-xxx.pt``, not ``exp/cpu-jit_xxx.pt``.
+
+In the above code, we downloaded the pre-trained model into the directory
+``egs/librispeech/ASR/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03``.
+
+2. Install ncnn and pnnx
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+Please refer to :ref:`export_for_ncnn_install_ncnn_and_pnnx` .
+
+
+3. Export the model via torch.jit.trace()
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+First, let us rename our pre-trained model:
+
+.. code-block::
+
+  cd egs/librispeech/ASR
+
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp
+
+  ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+
+  cd ../..
+
+Next, we use the following code to export our model:
+
+.. code-block:: bash
+
+  dir=./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+
+  ./lstm_transducer_stateless2/export-for-ncnn.py \
+    --exp-dir $dir/exp \
+    --tokens $dir/data/lang_bpe_500/tokens.txt \
+    --epoch 99 \
+    --avg 1 \
+    --use-averaged-model 0 \
+    --num-encoder-layers 12 \
+    --encoder-dim 512 \
+    --rnn-hidden-size 1024
+
+.. hint::
+
+  We have renamed our model to ``epoch-99.pt`` so that we can use ``--epoch 99``.
+  There is only one pre-trained model, so we use ``--avg 1 --use-averaged-model 0``.
+
+  If you have trained a model by yourself and if you have all checkpoints
+  available, please first use ``decode.py`` to tune ``--epoch --avg``
+  and select the best combination with with ``--use-averaged-model 1``.
+
+.. note::
+
+  You will see the following log output:
+
+  .. literalinclude:: ./code/export-lstm-transducer-for-ncnn-output.txt
+
+  The log shows the model has ``84176356`` parameters, i.e., ``~84 M``.
+
+  .. code-block::
+
+    ls -lh icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/pretrained-iter-468000-avg-16.pt
+
+    -rw-r--r-- 1 kuangfangjun root 324M Feb 17 10:34 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/pretrained-iter-468000-avg-16.pt
+
+  You can see that the file size of the pre-trained model is ``324 MB``, which
+  is roughly equal to ``84176356*4/1024/1024 = 321.107 MB``.
+
+After running ``lstm_transducer_stateless2/export-for-ncnn.py``,
+we will get the following files:
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/*pnnx.pt
+
+  -rw-r--r-- 1 kuangfangjun root 1010K Feb 17 11:22 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  318M Feb 17 11:22 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  3.0M Feb 17 11:22 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.pt
+
+
+.. _lstm-transducer-step-4-export-torchscript-model-via-pnnx:
+
+4. Export torchscript model via pnnx
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. hint::
+
+  Make sure you have set up the ``PATH`` environment variable
+  in :ref:`export_for_ncnn_install_ncnn_and_pnnx`. Otherwise,
+  it will throw an error saying that ``pnnx`` could not be found.
+
+Now, it's time to export our models to `ncnn`_ via ``pnnx``.
+
+.. code-block::
+
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt
+
+It will generate the following files:
+
+.. code-block:: bash
+
+  ls -lh  icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/*ncnn*{bin,param}
+
+  -rw-r--r-- 1 kuangfangjun root 503K Feb 17 11:32 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Feb 17 11:32 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 159M Feb 17 11:32 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  21K Feb 17 11:32 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 1.5M Feb 17 11:33 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Feb 17 11:33 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.ncnn.param
+
+
+There are two types of files:
+
+- ``param``: It is a text file containing the model architectures. You can
+  use a text editor to view its content.
+- ``bin``: It is a binary file containing the model parameters.
+
+We compare the file sizes of the models below before and after converting via ``pnnx``:
+
+.. see https://tableconvert.com/restructuredtext-generator
+
++----------------------------------+------------+
+| File name                        | File size  |
++==================================+============+
+| encoder_jit_trace-pnnx.pt        | 318 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.pt        | 1010 KB    |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.pt         | 3.0 MB     |
++----------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin  | 159 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin  | 503 KB     |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin   | 1.5 MB     |
++----------------------------------+------------+
+
+You can see that the file sizes of the models after conversion are about one half
+of the models before conversion:
+
+  - encoder: 318 MB vs 159 MB
+  - decoder: 1010 KB vs 503 KB
+  - joiner: 3.0 MB vs 1.5 MB
+
+The reason is that by default ``pnnx`` converts ``float32`` parameters
+to ``float16``. A ``float32`` parameter occupies 4 bytes, while it is 2 bytes
+for ``float16``. Thus, it is ``twice smaller`` after conversion.
+
+.. hint::
+
+  If you use ``pnnx ./encoder_jit_trace-pnnx.pt fp16=0``, then ``pnnx``
+  won't convert ``float32`` to ``float16``.
+
+5. Test the exported models in icefall
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. note::
+
+  We assume you have set up the environment variable ``PYTHONPATH`` when
+  building `ncnn`_.
+
+Now we have successfully converted our pre-trained model to `ncnn`_ format.
+The generated 6 files are what we need. You can use the following code to
+test the converted models:
+
+.. code-block:: bash
+
+  python3 ./lstm_transducer_stateless2/streaming-ncnn-decode.py \
+    --tokens ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/data/lang_bpe_500/tokens.txt \
+    --encoder-param-filename ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.ncnn.param \
+    --encoder-bin-filename ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.ncnn.bin \
+    --decoder-param-filename ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.ncnn.param \
+    --decoder-bin-filename ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.ncnn.bin \
+    --joiner-param-filename ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.ncnn.param \
+    --joiner-bin-filename ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.ncnn.bin \
+    ./icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/test_wavs/1089-134686-0001.wav
+
+.. hint::
+
+  `ncnn`_ supports only ``batch size == 1``, so ``streaming-ncnn-decode.py`` accepts
+  only 1 wave file as input.
+
+The output is given below:
+
+.. literalinclude:: ./code/test-streaming-ncnn-decode-lstm-transducer-libri.txt
+
+Congratulations! You have successfully exported a model from PyTorch to `ncnn`_!
+
+.. _lstm-modify-the-exported-encoder-for-sherpa-ncnn:
+
+6. Modify the exported encoder for sherpa-ncnn
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+In order to use the exported models in `sherpa-ncnn`_, we have to modify
+``encoder_jit_trace-pnnx.ncnn.param``.
+
+Let us have a look at the first few lines of ``encoder_jit_trace-pnnx.ncnn.param``:
+
+.. code-block::
+
+  7767517
+  267 379
+  Input                    in0                      0 1 in0
+
+**Explanation** of the above three lines:
+
+  1. ``7767517``, it is a magic number and should not be changed.
+  2. ``267 379``, the first number ``267`` specifies the number of layers
+     in this file, while ``379`` specifies the number of intermediate outputs
+     of this file
+  3. ``Input in0 0 1 in0``, ``Input`` is the layer type of this layer; ``in0``
+     is the layer name of this layer; ``0`` means this layer has no input;
+     ``1`` means this layer has one output; ``in0`` is the output name of
+     this layer.
+
+We need to add 1 extra line and also increment the number of layers.
+The result looks like below:
+
+.. code-block:: bash
+
+  7767517
+  268 379
+  SherpaMetaData           sherpa_meta_data1        0 0 0=3 1=12 2=512 3=1024
+  Input                    in0                      0 1 in0
+
+**Explanation**
+
+  1. ``7767517``, it is still the same
+  2. ``268 379``, we have added an extra layer, so we need to update ``267`` to ``268``.
+     We don't need to change ``379`` since the newly added layer has no inputs or outputs.
+  3. ``SherpaMetaData  sherpa_meta_data1  0 0 0=3 1=12 2=512 3=1024``
+     This line is newly added. Its explanation is given below:
+
+      - ``SherpaMetaData`` is the type of this layer. Must be ``SherpaMetaData``.
+      - ``sherpa_meta_data1`` is the name of this layer. Must be ``sherpa_meta_data1``.
+      - ``0 0`` means this layer has no inputs or output. Must be ``0 0``
+      - ``0=3``, 0 is the key and 3 is the value. MUST be ``0=3``
+      - ``1=12``, 1 is the key and 12 is the value of the
+        parameter ``--num-encoder-layers`` that you provided when running
+        ``./lstm_transducer_stateless2/export-for-ncnn.py``.
+      - ``2=512``, 2 is the key and 512 is the value of the
+        parameter ``--encoder-dim`` that you provided when running
+        ``./lstm_transducer_stateless2/export-for-ncnn.py``.
+      - ``3=1024``, 3 is the key and 1024 is the value of the
+        parameter ``--rnn-hidden-size`` that you provided when running
+        ``./lstm_transducer_stateless2/export-for-ncnn.py``.
+
+      For ease of reference, we list the key-value pairs that you need to add
+      in the following table. If your model has a different setting, please
+      change the values for ``SherpaMetaData`` accordingly. Otherwise, you
+      will be ``SAD``.
+
+          +------+-----------------------------+
+          | key  | value                       |
+          +======+=============================+
+          | 0    | 3 (fixed)                   |
+          +------+-----------------------------+
+          | 1    | ``--num-encoder-layers``    |
+          +------+-----------------------------+
+          | 2    | ``--encoder-dim``           |
+          +------+-----------------------------+
+          | 3    | ``--rnn-hidden-size``       |
+          +------+-----------------------------+
+
+  4. ``Input in0 0 1 in0``. No need to change it.
+
+.. caution::
+
+  When you add a new layer ``SherpaMetaData``, please remember to update the
+  number of layers. In our case, update  ``267`` to ``268``. Otherwise,
+  you will be SAD later.
+
+.. hint::
+
+  After adding the new layer ``SherpaMetaData``, you cannot use this model
+  with ``streaming-ncnn-decode.py`` anymore since ``SherpaMetaData`` is
+  supported only in `sherpa-ncnn`_.
+
+.. hint::
+
+  `ncnn`_ is very flexible. You can add new layers to it just by text-editing
+  the ``param`` file! You don't need to change the ``bin`` file.
+
+Now you can use this model in `sherpa-ncnn`_.
+Please refer to the following documentation:
+
+  - Linux/macOS/Windows/arm/aarch64: `<https://k2-fsa.github.io/sherpa/ncnn/install/index.html>`_
+  - ``Android``: `<https://k2-fsa.github.io/sherpa/ncnn/android/index.html>`_
+  - ``iOS``: `<https://k2-fsa.github.io/sherpa/ncnn/ios/index.html>`_
+  - Python: `<https://k2-fsa.github.io/sherpa/ncnn/python/index.html>`_
+
+We have a list of pre-trained models that have been exported for `sherpa-ncnn`_:
+
+  - `<https://k2-fsa.github.io/sherpa/ncnn/pretrained_models/index.html>`_
+
+    You can find more usages there.
+
+7. (Optional) int8 quantization with sherpa-ncnn
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This step is optional.
+
+In this step, we describe how to quantize our model with ``int8``.
+
+Change :ref:`lstm-transducer-step-4-export-torchscript-model-via-pnnx` to
+disable ``fp16`` when using ``pnnx``:
+
+.. code-block::
+
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt fp16=0
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt fp16=0
+
+.. note::
+
+  We add ``fp16=0`` when exporting the encoder and joiner. `ncnn`_ does not
+  support quantizing the decoder model yet. We will update this documentation
+  once `ncnn`_ supports it. (Maybe in this year, 2023).
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/*_jit_trace-pnnx.ncnn.{param,bin}
+
+  -rw-r--r-- 1 kuangfangjun root 503K Feb 17 11:32 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Feb 17 11:32 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 317M Feb 17 11:54 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  21K Feb 17 11:54 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 3.0M Feb 17 11:54 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Feb 17 11:54 icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/joiner_jit_trace-pnnx.ncnn.param
+
+
+Let us compare again the file sizes:
+
++----------------------------------------+------------+
+| File name                              | File size  |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.pt              | 318 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.pt              | 1010 KB    |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.pt               | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp16) | 159 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin (fp16) | 503 KB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp16) | 1.5 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp32) | 317 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp32) | 3.0 MB     |
++----------------------------------------+------------+
+
+You can see that the file sizes are doubled when we disable ``fp16``.
+
+.. note::
+
+  You can again use ``streaming-ncnn-decode.py`` to test the exported models.
+
+Next, follow :ref:`lstm-modify-the-exported-encoder-for-sherpa-ncnn`
+to modify ``encoder_jit_trace-pnnx.ncnn.param``.
+
+Change
+
+.. code-block:: bash
+
+  7767517
+  267 379
+  Input                    in0                      0 1 in0
+
+to
+
+.. code-block:: bash
+
+  7767517
+  268 379
+  SherpaMetaData           sherpa_meta_data1        0 0 0=3 1=12 2=512 3=1024
+  Input                    in0                      0 1 in0
+
+.. caution::
+
+  Please follow :ref:`lstm-modify-the-exported-encoder-for-sherpa-ncnn`
+  to change the values for ``SherpaMetaData`` if your model uses a different setting.
+
+Next, let us compile `sherpa-ncnn`_ since we will quantize our models within
+`sherpa-ncnn`_.
+
+.. code-block:: bash
+
+  # We will download sherpa-ncnn to $HOME/open-source/
+  # You can change it to anywhere you like.
+  cd $HOME
+  mkdir -p open-source
+
+  cd open-source
+  git clone https://github.com/k2-fsa/sherpa-ncnn
+  cd sherpa-ncnn
+  mkdir build
+  cd build
+  cmake ..
+  make -j 4
+
+  ./bin/generate-int8-scale-table
+
+  export PATH=$HOME/open-source/sherpa-ncnn/build/bin:$PATH
+
+The output of the above commands are:
+
+.. code-block:: bash
+
+  (py38) kuangfangjun:build$ generate-int8-scale-table
+  Please provide 10 arg. Currently given: 1
+  Usage:
+  generate-int8-scale-table encoder.param encoder.bin decoder.param decoder.bin joiner.param joiner.bin encoder-scale-table.txt joiner-scale-table.txt wave_filenames.txt
+
+  Each line in wave_filenames.txt is a path to some 16k Hz mono wave file.
+
+We need to create a file ``wave_filenames.txt``, in which we need to put
+some calibration wave files. For testing purpose, we put the ``test_wavs``
+from the pre-trained model repository
+`<https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03>`_
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/
+
+  cat <<EOF > wave_filenames.txt
+  ../test_wavs/1089-134686-0001.wav
+  ../test_wavs/1221-135766-0001.wav
+  ../test_wavs/1221-135766-0002.wav
+  EOF
+
+Now we can calculate the scales needed for quantization with the calibration data:
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/
+
+  generate-int8-scale-table \
+    ./encoder_jit_trace-pnnx.ncnn.param \
+    ./encoder_jit_trace-pnnx.ncnn.bin \
+    ./decoder_jit_trace-pnnx.ncnn.param \
+    ./decoder_jit_trace-pnnx.ncnn.bin \
+    ./joiner_jit_trace-pnnx.ncnn.param \
+    ./joiner_jit_trace-pnnx.ncnn.bin \
+    ./encoder-scale-table.txt \
+    ./joiner-scale-table.txt \
+    ./wave_filenames.txt
+
+The output logs are in the following:
+
+.. literalinclude:: ./code/generate-int-8-scale-table-for-lstm.txt
+
+It generates the following two files:
+
+.. code-block:: bash
+
+  ls -lh encoder-scale-table.txt joiner-scale-table.txt
+
+  -rw-r--r-- 1 kuangfangjun root 345K Feb 17 12:13 encoder-scale-table.txt
+  -rw-r--r-- 1 kuangfangjun root  17K Feb 17 12:13 joiner-scale-table.txt
+
+.. caution::
+
+  Definitely, you need more calibration data to compute the scale table.
+
+Finally, let us use the scale table to quantize our models into ``int8``.
+
+.. code-block:: bash
+
+  ncnn2int8
+
+  usage: ncnn2int8 [inparam] [inbin] [outparam] [outbin] [calibration table]
+
+First, we quantize the encoder model:
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/
+
+  ncnn2int8 \
+    ./encoder_jit_trace-pnnx.ncnn.param \
+    ./encoder_jit_trace-pnnx.ncnn.bin \
+    ./encoder_jit_trace-pnnx.ncnn.int8.param \
+    ./encoder_jit_trace-pnnx.ncnn.int8.bin \
+    ./encoder-scale-table.txt
+
+Next, we quantize the joiner model:
+
+.. code-block:: bash
+
+  ncnn2int8 \
+    ./joiner_jit_trace-pnnx.ncnn.param \
+    ./joiner_jit_trace-pnnx.ncnn.bin \
+    ./joiner_jit_trace-pnnx.ncnn.int8.param \
+    ./joiner_jit_trace-pnnx.ncnn.int8.bin \
+    ./joiner-scale-table.txt
+
+The above two commands generate the following 4 files:
+
+.. code-block::
+
+  -rw-r--r-- 1 kuangfangjun root 218M Feb 17 12:19 encoder_jit_trace-pnnx.ncnn.int8.bin
+  -rw-r--r-- 1 kuangfangjun root  21K Feb 17 12:19 encoder_jit_trace-pnnx.ncnn.int8.param
+  -rw-r--r-- 1 kuangfangjun root 774K Feb 17 12:19 joiner_jit_trace-pnnx.ncnn.int8.bin
+  -rw-r--r-- 1 kuangfangjun root  496 Feb 17 12:19 joiner_jit_trace-pnnx.ncnn.int8.param
+
+Congratulations! You have successfully quantized your model from ``float32`` to ``int8``.
+
+.. caution::
+
+  ``ncnn.int8.param`` and ``ncnn.int8.bin`` must be used in pairs.
+
+  You can replace ``ncnn.param`` and ``ncnn.bin`` with ``ncnn.int8.param``
+  and ``ncnn.int8.bin`` in `sherpa-ncnn`_ if you like.
+
+  For instance, to use only the ``int8`` encoder in ``sherpa-ncnn``, you can
+  replace the following invocation:
+
+    .. code-block::
+
+      cd egs/librispeech/ASR
+      cd icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/
+
+      sherpa-ncnn \
+        ../data/lang_bpe_500/tokens.txt \
+        ./encoder_jit_trace-pnnx.ncnn.param \
+        ./encoder_jit_trace-pnnx.ncnn.bin \
+        ./decoder_jit_trace-pnnx.ncnn.param \
+        ./decoder_jit_trace-pnnx.ncnn.bin \
+        ./joiner_jit_trace-pnnx.ncnn.param \
+        ./joiner_jit_trace-pnnx.ncnn.bin \
+        ../test_wavs/1089-134686-0001.wav
+
+  with
+
+    .. code-block:: bash
+
+      cd egs/librispeech/ASR
+      cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+      sherpa-ncnn \
+        ../data/lang_bpe_500/tokens.txt \
+        ./encoder_jit_trace-pnnx.ncnn.int8.param \
+        ./encoder_jit_trace-pnnx.ncnn.int8.bin \
+        ./decoder_jit_trace-pnnx.ncnn.param \
+        ./decoder_jit_trace-pnnx.ncnn.bin \
+        ./joiner_jit_trace-pnnx.ncnn.param \
+        ./joiner_jit_trace-pnnx.ncnn.bin \
+        ../test_wavs/1089-134686-0001.wav
+
+The following table compares again the file sizes:
+
++----------------------------------------+------------+
+| File name                              | File size  |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.pt              | 318 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.pt              | 1010 KB    |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.pt               | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp16) | 159 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin (fp16) | 503 KB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp16) | 1.5 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp32) | 317 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp32) | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.int8.bin   | 218 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.int8.bin    | 774 KB     |
++----------------------------------------+------------+
+
+You can see that the file size of the joiner model after ``int8`` quantization
+is much smaller. However, the size of the encoder model is even larger than
+the ``fp16`` counterpart. The reason is that `ncnn`_ currently does not support
+quantizing ``LSTM`` layers into ``8-bit``. Please see
+`<https://github.com/Tencent/ncnn/issues/4532>`_
+
+.. hint::
+
+    Currently, only linear layers and convolutional layers are quantized
+    with ``int8``, so you don't see an exact ``4x`` reduction in file sizes.
+
+.. note::
+
+  You need to test the recognition accuracy after ``int8`` quantization.
+
+
+That's it! Have fun with `sherpa-ncnn`_!
diff --git a/docs/source/model-export/export-ncnn-zipformer.rst b/docs/source/model-export/export-ncnn-zipformer.rst
new file mode 100644
index 000000000..a5845b0e4
--- /dev/null
+++ b/docs/source/model-export/export-ncnn-zipformer.rst
@@ -0,0 +1,387 @@
+.. _export_streaming_zipformer_transducer_models_to_ncnn:
+
+Export streaming Zipformer transducer models to ncnn
+----------------------------------------------------
+
+We use the pre-trained model from the following repository as an example:
+
+`<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>`_
+
+We will show you step by step how to export it to `ncnn`_ and run it with `sherpa-ncnn`_.
+
+.. hint::
+
+  We use ``Ubuntu 18.04``, ``torch 1.13``, and ``Python 3.8`` for testing.
+
+.. caution::
+
+  Please use a more recent version of PyTorch. For instance, ``torch 1.8``
+  may ``not`` work.
+
+1. Download the pre-trained model
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. hint::
+
+  You have to install `git-lfs`_ before you continue.
+
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+  cd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+
+  git lfs pull --include "exp/pretrained.pt"
+  git lfs pull --include "data/lang_bpe_500/bpe.model"
+
+  cd ..
+
+.. note::
+
+  We downloaded ``exp/pretrained-xxx.pt``, not ``exp/cpu-jit_xxx.pt``.
+
+In the above code, we downloaded the pre-trained model into the directory
+``egs/librispeech/ASR/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29``.
+
+2. Install ncnn and pnnx
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+Please refer to :ref:`export_for_ncnn_install_ncnn_and_pnnx` .
+
+
+3. Export the model via torch.jit.trace()
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+First, let us rename our pre-trained model:
+
+.. code-block::
+
+  cd egs/librispeech/ASR
+
+  cd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+
+  ln -s pretrained.pt epoch-99.pt
+
+  cd ../..
+
+Next, we use the following code to export our model:
+
+.. code-block:: bash
+
+  dir=./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+
+  ./pruned_transducer_stateless7_streaming/export-for-ncnn.py \
+    --tokens $dir/data/lang_bpe_500/tokens.txt \
+    --exp-dir $dir/exp \
+    --use-averaged-model 0 \
+    --epoch 99 \
+    --avg 1 \
+    --decode-chunk-len 32 \
+    --num-left-chunks 4 \
+    --num-encoder-layers "2,4,3,2,4" \
+    --feedforward-dims "1024,1024,2048,2048,1024" \
+    --nhead "8,8,8,8,8" \
+    --encoder-dims "384,384,384,384,384" \
+    --attention-dims "192,192,192,192,192" \
+    --encoder-unmasked-dims "256,256,256,256,256" \
+    --zipformer-downsampling-factors "1,2,4,8,2" \
+    --cnn-module-kernels "31,31,31,31,31" \
+    --decoder-dim 512 \
+    --joiner-dim 512
+
+.. caution::
+
+   If your model has different configuration parameters, please change them accordingly.
+
+.. hint::
+
+  We have renamed our model to ``epoch-99.pt`` so that we can use ``--epoch 99``.
+  There is only one pre-trained model, so we use ``--avg 1 --use-averaged-model 0``.
+
+  If you have trained a model by yourself and if you have all checkpoints
+  available, please first use ``decode.py`` to tune ``--epoch --avg``
+  and select the best combination with with ``--use-averaged-model 1``.
+
+.. note::
+
+  You will see the following log output:
+
+  .. literalinclude:: ./code/export-zipformer-transducer-for-ncnn-output.txt
+
+  The log shows the model has ``69920376`` parameters, i.e., ``~69.9 M``.
+
+  .. code-block:: bash
+
+   ls -lh icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/pretrained.pt
+   -rw-r--r-- 1 kuangfangjun root 269M Jan 12 12:53 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/pretrained.pt
+
+  You can see that the file size of the pre-trained model is ``269 MB``, which
+  is roughly equal to ``69920376*4/1024/1024 = 266.725 MB``.
+
+After running ``pruned_transducer_stateless7_streaming/export-for-ncnn.py``,
+we will get the following files:
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/*pnnx.pt
+
+  -rw-r--r-- 1 kuangfangjun root 1022K Feb 27 20:23 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/decoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  266M Feb 27 20:23 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/encoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  2.8M Feb 27 20:23 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/joiner_jit_trace-pnnx.pt
+
+.. _zipformer-transducer-step-4-export-torchscript-model-via-pnnx:
+
+4. Export torchscript model via pnnx
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. hint::
+
+  Make sure you have set up the ``PATH`` environment variable
+  in :ref:`export_for_ncnn_install_ncnn_and_pnnx`. Otherwise,
+  it will throw an error saying that ``pnnx`` could not be found.
+
+Now, it's time to export our models to `ncnn`_ via ``pnnx``.
+
+.. code-block::
+
+  cd icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt
+
+It will generate the following files:
+
+.. code-block:: bash
+
+  ls -lh  icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/*ncnn*{bin,param}
+
+  -rw-r--r-- 1 kuangfangjun root 509K Feb 27 20:31 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Feb 27 20:31 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 133M Feb 27 20:30 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root 152K Feb 27 20:30 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 1.4M Feb 27 20:31 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Feb 27 20:31 icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/joiner_jit_trace-pnnx.ncnn.param
+
+There are two types of files:
+
+- ``param``: It is a text file containing the model architectures. You can
+  use a text editor to view its content.
+- ``bin``: It is a binary file containing the model parameters.
+
+We compare the file sizes of the models below before and after converting via ``pnnx``:
+
+.. see https://tableconvert.com/restructuredtext-generator
+
++----------------------------------+------------+
+| File name                        | File size  |
++==================================+============+
+| encoder_jit_trace-pnnx.pt        | 266 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.pt        | 1022 KB    |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.pt         | 2.8 MB     |
++----------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin  | 133 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin  | 509 KB     |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin   | 1.4 MB     |
++----------------------------------+------------+
+
+You can see that the file sizes of the models after conversion are about one half
+of the models before conversion:
+
+  - encoder: 266 MB vs 133 MB
+  - decoder: 1022 KB vs 509 KB
+  - joiner: 2.8 MB vs 1.4 MB
+
+The reason is that by default ``pnnx`` converts ``float32`` parameters
+to ``float16``. A ``float32`` parameter occupies 4 bytes, while it is 2 bytes
+for ``float16``. Thus, it is ``twice smaller`` after conversion.
+
+.. hint::
+
+  If you use ``pnnx ./encoder_jit_trace-pnnx.pt fp16=0``, then ``pnnx``
+  won't convert ``float32`` to ``float16``.
+
+5. Test the exported models in icefall
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. note::
+
+  We assume you have set up the environment variable ``PYTHONPATH`` when
+  building `ncnn`_.
+
+Now we have successfully converted our pre-trained model to `ncnn`_ format.
+The generated 6 files are what we need. You can use the following code to
+test the converted models:
+
+.. code-block:: bash
+
+  python3 ./pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py \
+    --tokens ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/data/lang_bpe_500/tokens.txt \
+    --encoder-param-filename ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/encoder_jit_trace-pnnx.ncnn.param \
+    --encoder-bin-filename ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/encoder_jit_trace-pnnx.ncnn.bin \
+    --decoder-param-filename ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/decoder_jit_trace-pnnx.ncnn.param \
+    --decoder-bin-filename ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/decoder_jit_trace-pnnx.ncnn.bin \
+    --joiner-param-filename ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/joiner_jit_trace-pnnx.ncnn.param \
+    --joiner-bin-filename ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp/joiner_jit_trace-pnnx.ncnn.bin \
+    ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/test_wavs/1089-134686-0001.wav
+
+.. hint::
+
+  `ncnn`_ supports only ``batch size == 1``, so ``streaming-ncnn-decode.py`` accepts
+  only 1 wave file as input.
+
+The output is given below:
+
+.. literalinclude:: ./code/test-streaming-ncnn-decode-zipformer-transducer-libri.txt
+
+Congratulations! You have successfully exported a model from PyTorch to `ncnn`_!
+
+.. _zipformer-modify-the-exported-encoder-for-sherpa-ncnn:
+
+6. Modify the exported encoder for sherpa-ncnn
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+In order to use the exported models in `sherpa-ncnn`_, we have to modify
+``encoder_jit_trace-pnnx.ncnn.param``.
+
+Let us have a look at the first few lines of ``encoder_jit_trace-pnnx.ncnn.param``:
+
+.. code-block::
+
+  7767517
+  2028 2547
+  Input                    in0                      0 1 in0
+
+**Explanation** of the above three lines:
+
+  1. ``7767517``, it is a magic number and should not be changed.
+  2. ``2028 2547``, the first number ``2028`` specifies the number of layers
+     in this file, while ``2547`` specifies the number of intermediate outputs
+     of this file
+  3. ``Input in0 0 1 in0``, ``Input`` is the layer type of this layer; ``in0``
+     is the layer name of this layer; ``0`` means this layer has no input;
+     ``1`` means this layer has one output; ``in0`` is the output name of
+     this layer.
+
+We need to add 1 extra line and also increment the number of layers.
+The result looks like below:
+
+.. code-block:: bash
+
+  7767517
+  2029 2547
+  SherpaMetaData           sherpa_meta_data1        0 0 0=2 1=32 2=4 3=7 15=1 -23316=5,2,4,3,2,4 -23317=5,384,384,384,384,384 -23318=5,192,192,192,192,192 -23319=5,1,2,4,8,2 -23320=5,31,31,31,31,31
+  Input                    in0                      0 1 in0
+
+**Explanation**
+
+  1. ``7767517``, it is still the same
+  2. ``2029 2547``, we have added an extra layer, so we need to update ``2028`` to ``2029``.
+     We don't need to change ``2547`` since the newly added layer has no inputs or outputs.
+  3. ``SherpaMetaData  sherpa_meta_data1  0 0 0=2 1=32 2=4 3=7 -23316=5,2,4,3,2,4 -23317=5,384,384,384,384,384 -23318=5,192,192,192,192,192 -23319=5,1,2,4,8,2 -23320=5,31,31,31,31,31``
+     This line is newly added. Its explanation is given below:
+
+      - ``SherpaMetaData`` is the type of this layer. Must be ``SherpaMetaData``.
+      - ``sherpa_meta_data1`` is the name of this layer. Must be ``sherpa_meta_data1``.
+      - ``0 0`` means this layer has no inputs or output. Must be ``0 0``
+      - ``0=2``, 0 is the key and 2 is the value. MUST be ``0=2``
+      - ``1=32``, 1 is the key and 32 is the value of the
+        parameter ``--decode-chunk-len`` that you provided when running
+        ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+      - ``2=4``, 2 is the key and 4 is the value of the
+        parameter ``--num-left-chunks`` that you provided when running
+        ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+      - ``3=7``, 3 is the key and 7 is the value of for the amount of padding
+        used in the Conv2DSubsampling layer. It should be 7 for zipformer
+        if you don't change zipformer.py.
+      - ``15=1``, attribute 15, this is the model version. Starting from
+        `sherpa-ncnn`_ v2.0, we require that the model version has to
+        be >= 1.
+      - ``-23316=5,2,4,3,2,4``, attribute 16, this is an array attribute.
+        It is attribute 16 since -23300 - (-23316) = 16.
+        The first element of the array is the length of the array, which is 5 in our case.
+        ``2,4,3,2,4`` is the value of ``--num-encoder-layers``that you provided
+        when running ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+      - ``-23317=5,384,384,384,384,384``, attribute 17.
+        The first element of the array is the length of the array, which is 5 in our case.
+        ``384,384,384,384,384`` is the value of ``--encoder-dims``that you provided
+        when running ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+      - ``-23318=5,192,192,192,192,192``, attribute 18.
+        The first element of the array is the length of the array, which is 5 in our case.
+        ``192,192,192,192,192`` is the value of ``--attention-dims`` that you provided
+        when running ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+      - ``-23319=5,1,2,4,8,2``, attribute 19.
+        The first element of the array is the length of the array, which is 5 in our case.
+        ``1,2,4,8,2`` is the value of ``--zipformer-downsampling-factors`` that you provided
+        when running ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+      - ``-23320=5,31,31,31,31,31``, attribute 20.
+        The first element of the array is the length of the array, which is 5 in our case.
+        ``31,31,31,31,31`` is the value of ``--cnn-module-kernels`` that you provided
+        when running ``./pruned_transducer_stateless7_streaming/export-for-ncnn.py``.
+
+      For ease of reference, we list the key-value pairs that you need to add
+      in the following table. If your model has a different setting, please
+      change the values for ``SherpaMetaData`` accordingly. Otherwise, you
+      will be ``SAD``.
+
+          +----------+--------------------------------------------+
+          | key      | value                                      |
+          +==========+============================================+
+          | 0        | 2 (fixed)                                  |
+          +----------+--------------------------------------------+
+          | 1        | ``-decode-chunk-len``                      |
+          +----------+--------------------------------------------+
+          | 2        | ``--num-left-chunks``                      |
+          +----------+--------------------------------------------+
+          | 3        | 7 (if you don't change code)               |
+          +----------+--------------------------------------------+
+          | 15       | 1 (The model version)                      |
+          +----------+--------------------------------------------+
+          |-23316    | ``--num-encoder-layer``                    |
+          +----------+--------------------------------------------+
+          |-23317    | ``--encoder-dims``                         |
+          +----------+--------------------------------------------+
+          |-23318    | ``--attention-dims``                       |
+          +----------+--------------------------------------------+
+          |-23319    | ``--zipformer-downsampling-factors``       |
+          +----------+--------------------------------------------+
+          |-23320    | ``--cnn-module-kernels``                   |
+          +----------+--------------------------------------------+
+
+  4. ``Input in0 0 1 in0``. No need to change it.
+
+.. caution::
+
+  When you add a new layer ``SherpaMetaData``, please remember to update the
+  number of layers. In our case, update  ``2028`` to ``2029``. Otherwise,
+  you will be SAD later.
+
+.. hint::
+
+  After adding the new layer ``SherpaMetaData``, you cannot use this model
+  with ``streaming-ncnn-decode.py`` anymore since ``SherpaMetaData`` is
+  supported only in `sherpa-ncnn`_.
+
+.. hint::
+
+  `ncnn`_ is very flexible. You can add new layers to it just by text-editing
+  the ``param`` file! You don't need to change the ``bin`` file.
+
+Now you can use this model in `sherpa-ncnn`_.
+Please refer to the following documentation:
+
+  - Linux/macOS/Windows/arm/aarch64: `<https://k2-fsa.github.io/sherpa/ncnn/install/index.html>`_
+  - ``Android``: `<https://k2-fsa.github.io/sherpa/ncnn/android/index.html>`_
+  - ``iOS``: `<https://k2-fsa.github.io/sherpa/ncnn/ios/index.html>`_
+  - Python: `<https://k2-fsa.github.io/sherpa/ncnn/python/index.html>`_
+
+We have a list of pre-trained models that have been exported for `sherpa-ncnn`_:
+
+  - `<https://k2-fsa.github.io/sherpa/ncnn/pretrained_models/index.html>`_
+
+    You can find more usages there.
diff --git a/docs/source/model-export/export-ncnn.rst b/docs/source/model-export/export-ncnn.rst
new file mode 100644
index 000000000..634fb1e59
--- /dev/null
+++ b/docs/source/model-export/export-ncnn.rst
@@ -0,0 +1,39 @@
+.. _icefall_export_to_ncnn:
+
+Export to ncnn
+==============
+
+We support exporting the following models
+to `ncnn <https://github.com/tencent/ncnn>`_:
+
+  - `Zipformer transducer models <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming>`_
+
+  - `LSTM transducer models <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/lstm_transducer_stateless2>`_
+
+  - `ConvEmformer transducer models <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conv_emformer_transducer_stateless2>`_
+
+We also provide `sherpa-ncnn`_
+for performing speech recognition using `ncnn`_ with exported models.
+It has been tested on the following platforms:
+
+  - Linux
+  - macOS
+  - Windows
+  - ``Android``
+  - ``iOS``
+  - ``Raspberry Pi``
+  - `爱芯派 <https://wiki.sipeed.com/hardware/zh/>`_ (`MAIX-III AXera-Pi <https://wiki.sipeed.com/hardware/en/maixIII/ax-pi/axpi.html>`_).
+  - `RV1126 <https://www.rock-chips.com/a/en/products/RV11_Series/2020/0427/1076.html>`_
+
+`sherpa-ncnn`_ is self-contained and can be statically linked to produce
+a binary containing everything needed. Please refer
+to its documentation for details:
+
+ - `<https://k2-fsa.github.io/sherpa/ncnn/index.html>`_
+
+
+.. toctree::
+
+   export-ncnn-zipformer
+   export-ncnn-conv-emformer
+   export-ncnn-lstm
diff --git a/docs/source/model-export/export-onnx.rst b/docs/source/model-export/export-onnx.rst
new file mode 100644
index 000000000..d95f2acfe
--- /dev/null
+++ b/docs/source/model-export/export-onnx.rst
@@ -0,0 +1,104 @@
+Export to ONNX
+==============
+
+In this section, we describe how to export models to `ONNX`_.
+
+.. hint::
+
+   Before you continue, please run:
+
+    .. code-block:: bash
+
+        pip install onnx
+
+
+In each recipe, there is a file called ``export-onnx.py``, which is used
+to export trained models to `ONNX`_.
+
+There is also a file named ``onnx_pretrained.py``, which you can use
+the exported `ONNX`_ model in Python with `onnxruntime`_ to decode sound files.
+
+sherpa-onnx
+-----------
+
+We have a separate repository `sherpa-onnx`_ for deploying your exported models
+on various platforms such as:
+
+  - iOS
+  - Android
+  - Raspberry Pi
+  - Linux/macOS/Windows
+
+
+Please see the documentation of `sherpa-onnx`_ for details:
+
+  `<https://k2-fsa.github.io/sherpa/onnx/index.html>`_
+
+Example
+-------
+
+In the following, we demonstrate how to export a streaming Zipformer pre-trained
+model from
+`<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11>`_
+to `ONNX`_.
+
+Download the pre-trained model
+------------------------------
+
+.. hint::
+
+   We assume you have installed `git-lfs`_.
+
+.. code-block:: bash
+
+
+  cd egs/librispeech/ASR
+
+  repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+  GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+  repo=$(basename $repo_url)
+
+  pushd $repo
+  git lfs pull --include "data/lang_bpe_500/bpe.model"
+  git lfs pull --include "exp/pretrained.pt"
+  cd exp
+  ln -s pretrained.pt epoch-99.pt
+  popd
+
+Export the model to ONNX
+------------------------
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/export-onnx.py \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --use-averaged-model 0 \
+    --epoch 99 \
+    --avg 1 \
+    --decode-chunk-len 32 \
+    --exp-dir $repo/exp/
+
+.. warning::
+
+   ``export-onnx.py`` from different recipes has different options.
+
+   In the above example, ``--decode-chunk-len`` is specific for the
+   streaming Zipformer. Other models won't have such an option.
+
+It will generate the following 3 files in ``$repo/exp``
+
+  - ``encoder-epoch-99-avg-1.onnx``
+  - ``decoder-epoch-99-avg-1.onnx``
+  - ``joiner-epoch-99-avg-1.onnx``
+
+Decode sound files with exported ONNX models
+--------------------------------------------
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/onnx_pretrained.py \
+    --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+    --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+    --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav
diff --git a/docs/source/model-export/export-with-torch-jit-script.rst b/docs/source/model-export/export-with-torch-jit-script.rst
new file mode 100644
index 000000000..31c8f0bf5
--- /dev/null
+++ b/docs/source/model-export/export-with-torch-jit-script.rst
@@ -0,0 +1,58 @@
+.. _export-model-with-torch-jit-script:
+
+Export model with torch.jit.script()
+====================================
+
+In this section, we describe how to export a model via
+``torch.jit.script()``.
+
+When to use it
+--------------
+
+If we want to use our trained model with torchscript,
+we can use ``torch.jit.script()``.
+
+.. hint::
+
+  See :ref:`export-model-with-torch-jit-trace`
+  if you want to use ``torch.jit.trace()``.
+
+How to export
+-------------
+
+We use
+`<https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless3>`_
+as an example in the following.
+
+.. code-block:: bash
+
+    cd egs/librispeech/ASR
+    epoch=14
+    avg=1
+
+    ./pruned_transducer_stateless3/export.py \
+      --exp-dir ./pruned_transducer_stateless3/exp \
+      --tokens data/lang_bpe_500/tokens.txt \
+      --epoch $epoch \
+      --avg $avg \
+      --jit 1
+
+It will generate a file ``cpu_jit.pt`` in ``pruned_transducer_stateless3/exp``.
+
+.. caution::
+
+   Don't be confused by ``cpu`` in ``cpu_jit.pt``. We move all parameters
+   to CPU before saving it into a ``pt`` file; that's why we use ``cpu``
+   in the filename.
+
+How to use the exported model
+-----------------------------
+
+Please refer to the following pages for usage:
+
+- `<https://k2-fsa.github.io/sherpa/python/streaming_asr/emformer/index.html>`_
+- `<https://k2-fsa.github.io/sherpa/python/streaming_asr/conv_emformer/index.html>`_
+- `<https://k2-fsa.github.io/sherpa/python/streaming_asr/conformer/index.html>`_
+- `<https://k2-fsa.github.io/sherpa/python/offline_asr/conformer/index.html>`_
+- `<https://k2-fsa.github.io/sherpa/cpp/offline_asr/gigaspeech.html>`_
+- `<https://k2-fsa.github.io/sherpa/cpp/offline_asr/wenetspeech.html>`_
diff --git a/docs/source/model-export/export-with-torch-jit-trace.rst b/docs/source/model-export/export-with-torch-jit-trace.rst
new file mode 100644
index 000000000..be7876ab5
--- /dev/null
+++ b/docs/source/model-export/export-with-torch-jit-trace.rst
@@ -0,0 +1,69 @@
+.. _export-model-with-torch-jit-trace:
+
+Export model with torch.jit.trace()
+===================================
+
+In this section, we describe how to export a model via
+``torch.jit.trace()``.
+
+When to use it
+--------------
+
+If we want to use our trained model with torchscript,
+we can use ``torch.jit.trace()``.
+
+.. hint::
+
+  See :ref:`export-model-with-torch-jit-script`
+  if you want to use ``torch.jit.script()``.
+
+How to export
+-------------
+
+We use
+`<https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/lstm_transducer_stateless2>`_
+as an example in the following.
+
+.. code-block:: bash
+
+    iter=468000
+    avg=16
+
+    cd egs/librispeech/ASR
+
+    ./lstm_transducer_stateless2/export.py \
+      --exp-dir ./lstm_transducer_stateless2/exp \
+      --tokens data/lang_bpe_500/tokens.txt \
+      --iter $iter \
+      --avg  $avg \
+      --jit-trace 1
+
+It will generate three files inside ``lstm_transducer_stateless2/exp``:
+
+  - ``encoder_jit_trace.pt``
+  - ``decoder_jit_trace.pt``
+  - ``joiner_jit_trace.pt``
+
+You can use
+`<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/lstm_transducer_stateless2/jit_pretrained.py>`_
+to decode sound files with the following commands:
+
+.. code-block:: bash
+
+    cd egs/librispeech/ASR
+    ./lstm_transducer_stateless2/jit_pretrained.py \
+      --bpe-model ./data/lang_bpe_500/bpe.model \
+      --encoder-model-filename ./lstm_transducer_stateless2/exp/encoder_jit_trace.pt \
+      --decoder-model-filename ./lstm_transducer_stateless2/exp/decoder_jit_trace.pt \
+      --joiner-model-filename ./lstm_transducer_stateless2/exp/joiner_jit_trace.pt \
+      /path/to/foo.wav \
+      /path/to/bar.wav \
+      /path/to/baz.wav
+
+How to use the exported models
+------------------------------
+
+Please refer to
+`<https://k2-fsa.github.io/sherpa/python/streaming_asr/lstm/index.html>`_
+for its usage in `sherpa <https://k2-fsa.github.io/sherpa/python/streaming_asr/lstm/index.html>`_.
+You can also find pretrained models there.
diff --git a/docs/source/model-export/index.rst b/docs/source/model-export/index.rst
new file mode 100644
index 000000000..9b7a2ee2d
--- /dev/null
+++ b/docs/source/model-export/index.rst
@@ -0,0 +1,14 @@
+Model export
+============
+
+In this section, we describe various ways to export models.
+
+
+
+.. toctree::
+
+   export-model-state-dict
+   export-with-torch-jit-trace
+   export-with-torch-jit-script
+   export-onnx
+   export-ncnn
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/conformer_ctc.rst b/docs/source/recipes/Non-streaming-ASR/aishell/conformer_ctc.rst
new file mode 100644
index 000000000..aad90f9d0
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/aishell/conformer_ctc.rst
@@ -0,0 +1,747 @@
+Conformer CTC
+=============
+
+This tutorial shows you how to run a conformer ctc model
+with the `Aishell <https://www.openslr.org/33>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+In this tutorial, you will learn:
+
+  - (1) How to prepare data for training and decoding
+  - (2) How to start the training, either with a single GPU or multiple GPUs
+  - (3) How to do decoding after training, with ctc-decoding, 1best and attention decoder rescoring
+  - (4) How to use a pre-trained model, provided by us
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `Aishell <https://www.openslr.org/33>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/aishell`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. HINT::
+
+  A 3-gram language model will be downloaded from huggingface, we assume you have
+  installed and initialized ``git-lfs``. If not, you could install ``git-lfs`` by
+
+  .. code-block:: bash
+
+    $ sudo apt-get install git-lfs
+    $ git-lfs install
+
+  If you don't have the ``sudo`` permission, you could download the
+  `git-lfs binary <https://github.com/git-lfs/git-lfs/releases>`_ here, then add it to you ``PATH``.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The experiment folder to save logs and model checkpoints,
+    default ``./conformer_ctc/exp``.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./conformer_ctc/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder set by ``--exp-dir``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./conformer_ctc/train.py --start-epoch 10`` loads the
+    checkpoint ``./conformer_ctc/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./conformer_ctc/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ ./conformer_ctc/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./conformer_ctc/train.py --world-size 1
+
+    .. CAUTION::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it. For instance, if
+    your are using V100 NVIDIA GPU, we recommend you to set it to ``200``.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`conformer_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/conformer_ctc/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./conformer_ctc/train.py`` directly.
+
+
+.. CAUTION::
+
+  The training set is perturbed by speed with two factors: 0.9 and 1.1.
+  Each epoch actually processes ``3x150 == 450`` hours of data.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in the folder set by ``--exp-dir``
+(default ``conformer_ctc/exp``). You will find the following files in that directory:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./conformer_ctc/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd conformer_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --name "Aishell conformer ctc training with icefall" --description "Training with new LabelSmoothing loss, see https://github.com/k2-fsa/icefall/pull/109"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/engw8KSkTZqS24zBV5dgCg/
+
+        [2021-11-22T11:09:27] Started scanning logdir.
+        [2021-11-22T11:10:14] Total uploaded: 116068 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output, click it and you will see
+    the following screenshot:
+
+      .. figure:: images/aishell-conformer-ctc-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/WE1DocDqRRCOSAgmGyClhg/
+
+         TensorBoard screenshot.
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage examples
+~~~~~~~~~~~~~~
+
+The following shows typical use cases:
+
+**Case 1**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/train.py --max-duration 200
+
+It uses ``--max-duration`` of 200 to avoid OOM.
+
+
+**Case 2**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,3"
+  $ ./conformer_ctc/train.py --world-size 2
+
+It uses GPU 0 and GPU 3 for DDP training.
+
+**Case 3**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/train.py --num-epochs 10 --start-epoch 3
+
+It loads checkpoint ``./conformer_ctc/exp/epoch-2.pt`` and starts
+training from epoch 3. Also, it trains for 10 epochs.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/decode.py --help
+
+shows the options for decoding.
+
+The commonly used options are:
+
+  - ``--method``
+
+    This specifies the decoding method.
+
+    The following command uses attention decoder for rescoring:
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./conformer_ctc/decode.py --method attention-decoder --max-duration 30 --nbest-scale 0.5
+
+  - ``--nbest-scale``
+
+    It is used to scale down lattice scores so that there are more unique
+    paths for rescoring.
+
+  - ``--max-duration``
+
+    It has the same meaning as the one during training. A larger
+    value may cause OOM.
+
+Pre-trained Model
+-----------------
+
+We have uploaded a pre-trained model to
+`<https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc>`_.
+
+We describe how to use the pre-trained model to transcribe a sound file or
+multiple sound files in the following.
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following commands describe how to download the pre-trained model:
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ tree tmp
+
+.. code-block:: bash
+
+  tmp/
+  `-- icefall_asr_aishell_conformer_ctc
+      |-- README.md
+      |-- data
+      |   `-- lang_char
+      |       |-- HLG.pt
+      |       |-- tokens.txt
+      |       `-- words.txt
+      |-- exp
+      |   `-- pretrained.pt
+      `-- test_waves
+          |-- BAC009S0764W0121.wav
+          |-- BAC009S0764W0122.wav
+          |-- BAC009S0764W0123.wav
+          `-- trans.txt
+
+  5 directories, 9 files
+
+**File descriptions**:
+
+  - ``data/lang_char/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_char/tokens.txt``
+
+      It contains tokens and their IDs.
+      Provided only for convenience so that you can look up the SOS/EOS ID easily.
+
+  - ``data/lang_char/words.txt``
+
+      It contains words and their IDs.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-25.pt`` to ``epoch-84.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.wav``
+
+      It contains some test sound files from Aishell ``test`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in `test_waves/`.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall_asr_aishell_conformer_ctc/test_waves/*.wav
+
+  Input File     : 'tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.20 = 67263 samples ~ 315.295 CDDA sectors
+  File Size      : 135k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.12 = 65840 samples ~ 308.625 CDDA sectors
+  File Size      : 132k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.00 = 64000 samples ~ 300 CDDA sectors
+  File Size      : 128k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+  Total Duration of 3 files: 00:00:12.32
+
+Usage
+~~~~~
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py --help
+
+displays the help information.
+
+It supports three decoding methods:
+
+  - CTC decoding
+  - HLG decoding
+  - HLG + attention decoder rescoring
+
+CTC decoding
+^^^^^^^^^^^^
+
+CTC decoding only uses the ctc topology for decoding without a lexicon and language model
+
+The command to run CTC decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt \
+    --tokens-file ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/tokens.txt \
+    --method ctc-decoding \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav
+
+The output is given below:
+
+.. code-block::
+
+  2021-11-18 07:53:41,707 INFO [pretrained.py:229] {'sample_rate': 16000, 'subsampling_factor': 4, 'feature_dim': 80, 'nhead': 4, 'attention_dim': 512, 'num_decoder_layers': 6, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'f2fd997f752ed11bbef4c306652c433e83f9cf12', 'k2-git-date': 'Sun Sep 19 09:41:46 2021', 'lhotse-version': '0.11.0.dev+git.33cfe45.clean', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'aishell', 'icefall-git-sha1': 'd57a873-dirty', 'icefall-git-date': 'Wed Nov 17 19:53:25 2021', 'icefall-path': '/ceph-hw/kangwei/code/icefall_aishell3', 'k2-path': '/ceph-hw/kangwei/code/k2_release/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-hw/kangwei/code/lhotse/lhotse/__init__.py'}, 'checkpoint': './tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt', 'tokens_file': './tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/tokens.txt', 'words_file': None, 'HLG': None, 'method': 'ctc-decoding', 'num_paths': 100, 'ngram_lm_scale': 0.3, 'attention_decoder_scale': 0.9, 'nbest_scale': 0.5, 'sos_id': 1, 'eos_id': 1, 'num_classes': 4336, 'sound_files': ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']}
+  2021-11-18 07:53:41,708 INFO [pretrained.py:240] device: cuda:0
+  2021-11-18 07:53:41,708 INFO [pretrained.py:242] Creating model
+  2021-11-18 07:53:51,131 INFO [pretrained.py:259] Constructing Fbank computer
+  2021-11-18 07:53:51,134 INFO [pretrained.py:269] Reading sound files: ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-11-18 07:53:51,138 INFO [pretrained.py:275] Decoding started
+  2021-11-18 07:53:51,241 INFO [pretrained.py:293] Use CTC decoding
+  2021-11-18 07:53:51,704 INFO [pretrained.py:369]
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav:
+  甚 至 出 现 交 易 几 乎 停 止 的 情 况
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav:
+  一 二 线 城 市 虽 然 也 处 于 调 整 中
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因 为 聚 集 了 过 多 公 共 资 源
+
+
+  2021-11-18 07:53:51,704 INFO [pretrained.py:371] Decoding Done
+
+
+HLG decoding
+^^^^^^^^^^^^
+
+HLG decoding uses the best path of the decoding lattice as the decoding result.
+
+The command to run HLG decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt \
+    --words-file ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt \
+    --HLG ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt \
+    --method 1best \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav
+
+The output is given below:
+
+.. code-block::
+
+  2021-11-18 07:37:38,683 INFO [pretrained.py:229] {'sample_rate': 16000, 'subsampling_factor': 4, 'feature_dim': 80, 'nhead': 4, 'attention_dim': 512, 'num_decoder_layers': 6, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'f2fd997f752ed11bbef4c306652c433e83f9cf12', 'k2-git-date': 'Sun Sep 19 09:41:46 2021', 'lhotse-version': '0.11.0.dev+git.33cfe45.clean', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'aishell', 'icefall-git-sha1': 'd57a873-clean', 'icefall-git-date': 'Wed Nov 17 19:53:25 2021', 'icefall-path': '/ceph-hw/kangwei/code/icefall_aishell3', 'k2-path': '/ceph-hw/kangwei/code/k2_release/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-hw/kangwei/code/lhotse/lhotse/__init__.py'}, 'checkpoint': './tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt', 'tokens_file': None, 'words_file': './tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt', 'HLG': './tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt', 'method': '1best', 'num_paths': 100, 'ngram_lm_scale': 0.3, 'attention_decoder_scale': 0.9, 'nbest_scale': 0.5, 'sos_id': 1, 'eos_id': 1, 'num_classes': 4336, 'sound_files': ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']}
+  2021-11-18 07:37:38,684 INFO [pretrained.py:240] device: cuda:0
+  2021-11-18 07:37:38,684 INFO [pretrained.py:242] Creating model
+  2021-11-18 07:37:47,651 INFO [pretrained.py:259] Constructing Fbank computer
+  2021-11-18 07:37:47,654 INFO [pretrained.py:269] Reading sound files: ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-11-18 07:37:47,659 INFO [pretrained.py:275] Decoding started
+  2021-11-18 07:37:47,752 INFO [pretrained.py:321] Loading HLG from ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt
+  2021-11-18 07:37:51,887 INFO [pretrained.py:340] Use HLG decoding
+  2021-11-18 07:37:52,102 INFO [pretrained.py:370]
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav:
+  甚至 出现 交易 几乎 停止 的 情况
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav:
+  一二 线 城市 虽然 也 处于 调整 中
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因为 聚集 了 过多 公共 资源
+
+
+  2021-11-18 07:37:52,102 INFO [pretrained.py:372] Decoding Done
+
+
+HLG decoding + attention decoder rescoring
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+It extracts n paths from the lattice, recores the extracted paths with
+an attention decoder. The path with the highest score is the decoding result.
+
+The command to run HLG decoding + attention decoder rescoring is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt \
+    --words-file ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt \
+    --HLG ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt \
+    --method attention-decoder \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav
+
+The output is below:
+
+.. code-block::
+
+  2021-11-18 07:42:05,965 INFO [pretrained.py:229] {'sample_rate': 16000, 'subsampling_factor': 4, 'feature_dim': 80, 'nhead': 4, 'attention_dim': 512, 'num_decoder_layers': 6, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'f2fd997f752ed11bbef4c306652c433e83f9cf12', 'k2-git-date': 'Sun Sep 19 09:41:46 2021', 'lhotse-version': '0.11.0.dev+git.33cfe45.clean', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'aishell', 'icefall-git-sha1': 'd57a873-dirty', 'icefall-git-date': 'Wed Nov 17 19:53:25 2021', 'icefall-path': '/ceph-hw/kangwei/code/icefall_aishell3', 'k2-path': '/ceph-hw/kangwei/code/k2_release/k2/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-hw/kangwei/code/lhotse/lhotse/__init__.py'}, 'checkpoint': './tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt', 'tokens_file': None, 'words_file': './tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt', 'HLG': './tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt', 'method': 'attention-decoder', 'num_paths': 100, 'ngram_lm_scale': 0.3, 'attention_decoder_scale': 0.9, 'nbest_scale': 0.5, 'sos_id': 1, 'eos_id': 1, 'num_classes': 4336, 'sound_files': ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']}
+  2021-11-18 07:42:05,966 INFO [pretrained.py:240] device: cuda:0
+  2021-11-18 07:42:05,966 INFO [pretrained.py:242] Creating model
+  2021-11-18 07:42:16,821 INFO [pretrained.py:259] Constructing Fbank computer
+  2021-11-18 07:42:16,822 INFO [pretrained.py:269] Reading sound files: ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-11-18 07:42:16,826 INFO [pretrained.py:275] Decoding started
+  2021-11-18 07:42:16,916 INFO [pretrained.py:321] Loading HLG from ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt
+  2021-11-18 07:42:21,115 INFO [pretrained.py:345] Use HLG + attention decoder rescoring
+  2021-11-18 07:42:21,888 INFO [pretrained.py:370]
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav:
+  甚至 出现 交易 几乎 停止 的 情况
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav:
+  一二 线 城市 虽然 也 处于 调整 中
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因为 聚集 了 过多 公共 资源
+
+
+  2021-11-18 07:42:21,889 INFO [pretrained.py:372] Decoding Done
+
+
+Colab notebook
+--------------
+
+We do provide a colab notebook for this recipe showing how to use a pre-trained model.
+
+|aishell asr conformer ctc colab notebook|
+
+.. |aishell asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1WnG17io5HEZ0Gn_cnh_VzK5QYOoiiklC
+
+.. HINT::
+
+  Due to limited memory provided by Colab, you have to upgrade to Colab Pro to
+  run ``HLG decoding + attention decoder rescoring``.
+  Otherwise, you can only run ``HLG decoding`` with Colab.
+
+**Congratulations!** You have finished the aishell ASR recipe with
+conformer CTC models in ``icefall``.
+
+
+If you want to deploy your trained model in C++, please read the following section.
+
+Deployment with C++
+-------------------
+
+This section describes how to deploy the pre-trained model in C++, without
+Python dependencies.
+
+.. HINT::
+
+  At present, it does NOT support streaming decoding.
+
+First, let us compile k2 from source:
+
+.. code-block:: bash
+
+  $ cd $HOME
+  $ git clone https://github.com/k2-fsa/k2
+  $ cd k2
+  $ git checkout v2.0-pre
+
+.. CAUTION::
+
+  You have to switch to the branch ``v2.0-pre``!
+
+.. code-block:: bash
+
+  $ mkdir build-release
+  $ cd build-release
+  $ cmake -DCMAKE_BUILD_TYPE=Release ..
+  $ make -j hlg_decode
+
+  # You will find four binaries in `./bin`, i.e. ./bin/hlg_decode,
+
+Now you are ready to go!
+
+Assume you have run:
+
+  .. code-block:: bash
+
+    $ cd k2/build-release
+    $ ln -s /path/to/icefall-asr-aishell-conformer-ctc ./
+
+To view the usage of ``./bin/hlg_decode``, run:
+
+.. code-block::
+
+  $ ./bin/hlg_decode
+
+It will show you the following message:
+
+.. code-block:: bash
+
+  Please provide --nn_model
+
+  This file implements decoding with an HLG decoding graph.
+
+  Usage:
+    ./bin/hlg_decode \
+      --use_gpu true \
+      --nn_model <path to torch scripted pt file> \
+      --hlg <path to HLG.pt> \
+      --word_table <path to words.txt> \
+      <path to foo.wav> \
+      <path to bar.wav> \
+      <more waves if any>
+
+  To see all possible options, use
+    ./bin/hlg_decode --help
+
+  Caution:
+   - Only sound files (*.wav) with single channel are supported.
+   - It assumes the model is conformer_ctc/transformer.py from icefall.
+     If you use a different model, you have to change the code
+     related to `model.forward` in this file.
+
+
+HLG decoding
+~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./bin/hlg_decode \
+    --use_gpu true \
+    --nn_model icefall_asr_aishell_conformer_ctc/exp/cpu_jit.pt \
+    --hlg icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt \
+    --word_table icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt \
+    icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav \
+    icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav \
+    icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav
+
+The output is:
+
+.. code-block::
+
+  2021-11-18 14:48:20.89 [I] k2/torch/bin/hlg_decode.cu:115:int main(int, char**) Device: cpu
+  2021-11-18 14:48:20.89 [I] k2/torch/bin/hlg_decode.cu:124:int main(int, char**) Load wave files
+  2021-11-18 14:48:20.97 [I] k2/torch/bin/hlg_decode.cu:131:int main(int, char**) Build Fbank computer
+  2021-11-18 14:48:20.98 [I] k2/torch/bin/hlg_decode.cu:142:int main(int, char**) Compute features
+  2021-11-18 14:48:20.115 [I] k2/torch/bin/hlg_decode.cu:150:int main(int, char**) Load neural network model
+  2021-11-18 14:48:20.693 [I] k2/torch/bin/hlg_decode.cu:165:int main(int, char**) Compute nnet_output
+  2021-11-18 14:48:23.182 [I] k2/torch/bin/hlg_decode.cu:180:int main(int, char**) Load icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt
+  2021-11-18 14:48:33.489 [I] k2/torch/bin/hlg_decode.cu:185:int main(int, char**) Decoding
+  2021-11-18 14:48:45.217 [I] k2/torch/bin/hlg_decode.cu:216:int main(int, char**)
+  Decoding result:
+
+  icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav
+  甚至 出现 交易 几乎 停止 的 情况
+
+  icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav
+  一二 线 城市 虽然 也 处于 调整 中
+
+  icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav
+  但 因为 聚集 了 过多 公共 资源
+
+There is a Colab notebook showing you how to run a torch scripted model in C++.
+Please see |aishell asr conformer ctc torch script colab notebook|
+
+.. |aishell asr conformer ctc torch script colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1Vh7RER7saTW01DtNbvr7CY7ovNZgmfWz?usp=sharing
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-conformer-ctc-tensorboard-log.jpg b/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-conformer-ctc-tensorboard-log.jpg
new file mode 100644
index 000000000..c8b26f741
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-conformer-ctc-tensorboard-log.jpg differ
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-tdnn-lstm-ctc-tensorboard-log.jpg b/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-tdnn-lstm-ctc-tensorboard-log.jpg
new file mode 100644
index 000000000..b31db3ab5
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-tdnn-lstm-ctc-tensorboard-log.jpg differ
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-transducer_stateless_modified-tensorboard-log.png b/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-transducer_stateless_modified-tensorboard-log.png
new file mode 100644
index 000000000..6c84b28f2
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/aishell/images/aishell-transducer_stateless_modified-tensorboard-log.png differ
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/index.rst b/docs/source/recipes/Non-streaming-ASR/aishell/index.rst
new file mode 100644
index 000000000..b77d59bca
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/aishell/index.rst
@@ -0,0 +1,21 @@
+aishell
+=======
+
+Aishell is an open-source Chinese Mandarin speech corpus published by Beijing
+Shell Shell Technology Co.,Ltd.
+
+400 people from different accent areas in China are invited to participate in
+the recording, which is conducted in a quiet indoor environment using high
+fidelity microphone and downsampled to 16kHz. The manual transcription accuracy
+is above 95%, through professional speech annotation and strict quality
+inspection. The data is free for academic use. We hope to provide moderate
+amount of data for new researchers in the field of speech recognition.
+
+It can be downloaded from `<https://www.openslr.org/33/>`_
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_lstm_ctc
+   conformer_ctc
+   stateless_transducer
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/stateless_transducer.rst b/docs/source/recipes/Non-streaming-ASR/aishell/stateless_transducer.rst
new file mode 100644
index 000000000..e8137b8c1
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/aishell/stateless_transducer.rst
@@ -0,0 +1,714 @@
+Stateless Transducer
+====================
+
+This tutorial shows you how to do transducer training in ``icefall``.
+
+.. HINT::
+
+  Instead of using RNN-T or RNN transducer, we only use transducer
+  here. As you will see, there are no RNNs in the model.
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+In this tutorial, you will learn:
+
+  - (1) What does the transducer model look like
+  - (2) How to prepare data for training and decoding
+  - (3) How to start the training, either with a single GPU or with multiple GPUs
+  - (4) How to do decoding after training, with greedy search, beam search and, **modified beam search**
+  - (5) How to use a pre-trained model provided by us to transcribe sound files
+
+
+The Model
+---------
+
+The transducer model consists of 3 parts:
+
+- **Encoder**: It is a conformer encoder with the following parameters
+
+    - Number of heads: 8
+    - Attention dim: 512
+    - Number of layers: 12
+    - Feedforward dim: 2048
+
+- **Decoder**: We use a stateless model consisting of:
+
+    - An embedding layer with embedding dim 512
+    - A Conv1d layer with a default kernel size 2 (i.e. it sees 2
+      symbols of left-context by default)
+
+- **Joiner**: It consists of a ``nn.tanh()`` and a ``nn.Linear()``.
+
+.. Caution::
+
+  The decoder is stateless and very simple. It is borrowed from
+  `<https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419>`_
+  (Rnn-Transducer with Stateless Prediction Network)
+
+  We make one modification to it: Place a Conv1d layer right after
+  the embedding layer.
+
+When using Chinese characters as modelling unit, whose vocabulary size
+is 4336 in this specific dataset,
+the number of parameters of the model is ``87939824``, i.e., about ``88 M``.
+
+The Loss
+--------
+
+We are using `<https://github.com/csukuangfj/optimized_transducer>`_
+to compute the transducer loss, which removes extra paddings
+in loss computation to save memory.
+
+.. Hint::
+
+  ``optimized_transducer`` implements the technqiues proposed
+  in `Improving RNN Transducer Modeling for End-to-End Speech Recognition <https://arxiv.org/abs/1909.12415>`_ to save memory.
+
+  Furthermore, it supports ``modified transducer``, limiting the maximum
+  number of symbols that can be emitted per frame to 1, which simplifies
+  the decoding process significantly. Also, the experiment results
+  show that it does not degrade the performance.
+
+  See `<https://github.com/csukuangfj/optimized_transducer#modified-transducer>`_
+  for what exactly modified transducer is.
+
+  `<https://github.com/csukuangfj/transducer-loss-benchmarking>`_ shows that
+  in the unpruned case ``optimized_transducer`` has the advantage about minimizing
+  memory usage.
+
+.. todo::
+
+  Add tutorial about ``pruned_transducer_stateless`` that uses k2
+  pruned transducer loss.
+
+.. hint::
+
+  You can use::
+
+    pip install optimized_transducer
+
+  to install ``optimized_transducer``. Refer to
+  `<https://github.com/csukuangfj/optimized_transducer>`_ for other
+  alternatives.
+
+Data Preparation
+----------------
+
+To prepare the data for training, please use the following commands:
+
+.. code-block:: bash
+
+  cd egs/aishell/ASR
+  ./prepare.sh --stop-stage 4
+  ./prepare.sh --stage 6 --stop-stage 6
+
+.. note::
+
+  You can use ``./prepare.sh``, though it will generate FSTs that
+  are not used in transducer training.
+
+When you finish running the script, you will get the following two folders:
+
+  - ``data/fbank``: It saves the pre-computed features
+  - ``data/lang_char``: It contains tokens that will be used in the training
+
+Training
+--------
+
+.. code-block:: bash
+
+  cd egs/aishell/ASR
+  ./transducer_stateless_modified/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The experiment folder to save logs and model checkpoints,
+    defaults to ``./transducer_stateless_modified/exp``.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./transducer_stateless_modified/train.py --num-epochs 30`` trains for 30
+    epochs and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder set by ``--exp-dir``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./transducer_stateless_modified/train.py --start-epoch 10`` loads the
+    checkpoint from ``exp_dir/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for single-machine multi-GPU DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./transducer_stateless_modified/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ ./transducer_stateless_modified/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./transducer_stateless_modified/train.py --world-size 1
+
+    .. CAUTION::
+
+      Only single-machine multi-GPU DDP training is implemented at present.
+      There is an on-going PR `<https://github.com/k2-fsa/icefall/pull/63>`_
+      that adds support for multi-machine multi-GPU DDP training.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch **before padding**.
+    If you encounter CUDA OOM, please reduce it. For instance, if
+    your are using V100 NVIDIA GPU with 32 GB RAM, we recommend you
+    to set it to ``300`` when the vocabulary size is 500.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--lr-factor``
+
+    It controls the learning rate. If you use a single GPU for training, you
+    may want to use a small value for it. If you use multiple GPUs for training,
+    you may increase it.
+
+  - ``--context-size``
+
+    It specifies the kernel size in the decoder. The default value 2 means it
+    functions as a tri-gram LM.
+
+  - ``--modified-transducer-prob``
+
+    It specifies the probability to use modified transducer loss.
+    If it is 0, then no modified transducer is used; if it is 1,
+    then it uses modified transducer loss for all batches. If it is
+    ``p``, it applies modified transducer with probability ``p``.
+
+There are some training options, e.g.,
+number of warmup steps,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`transducer_stateless_modified/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/transducer_stateless_modified/train.py#L162>`_
+
+If you need to change them, please modify ``./transducer_stateless_modified/train.py`` directly.
+
+.. CAUTION::
+
+  The training set is perturbed by speed with two factors: 0.9 and 1.1.
+  Each epoch actually processes ``3x150 == 450`` hours of data.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in the folder set by ``--exp-dir``
+(defaults to ``transducer_stateless_modified/exp``). You will find the following files in that directory:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./transducer_stateless_modified/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd transducer_stateless_modified/exp/tensorboard
+        $ tensorboard dev upload --logdir . --name "Aishell transducer training with icefall" --description "Training modified transducer, see https://github.com/k2-fsa/icefall/pull/219"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/laGZ6HrcQxOigbFD5E0Y3Q/
+
+        [2022-03-03T14:29:45] Started scanning logdir.
+        [2022-03-03T14:29:48] Total uploaded: 8477 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a `URL <https://tensorboard.dev/experiment/laGZ6HrcQxOigbFD5E0Y3Q/>`_ in the
+    above output, click it and you will see the following screenshot:
+
+      .. figure:: images/aishell-transducer_stateless_modified-tensorboard-log.png
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/laGZ6HrcQxOigbFD5E0Y3Q
+
+         TensorBoard screenshot.
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage examples
+~~~~~~~~~~~~~~
+
+The following shows typical use cases:
+
+**Case 1**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./transducer_stateless_modified/train.py --max-duration 250
+
+It uses ``--max-duration`` of 250 to avoid OOM.
+
+
+**Case 2**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,3"
+  $ ./transducer_stateless_modified/train.py --world-size 2
+
+It uses GPU 0 and GPU 3 for DDP training.
+
+**Case 3**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./transducer_stateless_modified/train.py --num-epochs 10 --start-epoch 3
+
+It loads checkpoint ``./transducer_stateless_modified/exp/epoch-2.pt`` and starts
+training from epoch 3. Also, it trains for 10 epochs.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./transducer_stateless_modified/decode.py --help
+
+shows the options for decoding.
+
+The commonly used options are:
+
+  - ``--method``
+
+    This specifies the decoding method. Currently, it supports:
+
+      - **greedy_search**. You can provide the commandline option ``--max-sym-per-frame``
+        to limit the maximum number of symbols that can be emitted per frame.
+
+      - **beam_search**. You can provide the commandline option ``--beam-size``.
+
+      - **modified_beam_search**. You can also provide the commandline option ``--beam-size``.
+        To use this method, we assume that you have trained your model with modified transducer,
+        i.e., used the option ``--modified-transducer-prob`` in the training.
+
+    The following command uses greedy search for decoding
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./transducer_stateless_modified/decode.py \
+              --epoch 64 \
+              --avg 33 \
+              --exp-dir ./transducer_stateless_modified/exp \
+              --max-duration 100 \
+              --decoding-method greedy_search \
+              --max-sym-per-frame 1
+
+    The following command uses beam search for decoding
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./transducer_stateless_modified/decode.py \
+              --epoch 64 \
+              --avg 33 \
+              --exp-dir ./transducer_stateless_modified/exp \
+              --max-duration 100 \
+              --decoding-method beam_search \
+              --beam-size 4
+
+    The following command uses ``modified`` beam search for decoding
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./transducer_stateless_modified/decode.py \
+              --epoch 64 \
+              --avg 33 \
+              --exp-dir ./transducer_stateless_modified/exp \
+              --max-duration 100 \
+              --decoding-method modified_beam_search \
+              --beam-size 4
+
+  - ``--max-duration``
+
+    It has the same meaning as the one used in training. A larger
+    value may cause OOM.
+
+  - ``--epoch``
+
+    It specifies the checkpoint from which epoch that should be used for decoding.
+
+  - ``--avg``
+
+    It specifies the number of models to average. For instance, if it is 3 and if
+    ``--epoch=10``, then it averages the checkpoints ``epoch-8.pt``, ``epoch-9.pt``,
+    and ``epoch-10.pt`` and the averaged checkpoint is used for decoding.
+
+After decoding, you can find the decoding logs and results in `exp_dir/log/<decoding_method>`, e.g.,
+``exp_dir/log/greedy_search``.
+
+Pre-trained Model
+-----------------
+
+We have uploaded a pre-trained model to
+`<https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2022-03-01>`_
+
+We describe how to use the pre-trained model to transcribe a sound file or
+multiple sound files in the following.
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following commands describe how to download the pre-trained model:
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2022-03-01
+
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ tree tmp/icefall-aishell-transducer-stateless-modified-2022-03-01
+
+
+.. code-block:: bash
+
+  tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/
+  |-- README.md
+  |-- data
+  |   `-- lang_char
+  |       |-- L.pt
+  |       |-- lexicon.txt
+  |       |-- tokens.txt
+  |       `-- words.txt
+  |-- exp
+  |   `-- pretrained.pt
+  |-- log
+  |   |-- errs-test-beam_4-epoch-64-avg-33-beam-4.txt
+  |   |-- errs-test-greedy_search-epoch-64-avg-33-context-2-max-sym-per-frame-1.txt
+  |   |-- log-decode-epoch-64-avg-33-beam-4-2022-03-02-12-05-03
+  |   |-- log-decode-epoch-64-avg-33-context-2-max-sym-per-frame-1-2022-02-28-18-13-07
+  |   |-- recogs-test-beam_4-epoch-64-avg-33-beam-4.txt
+  |   `-- recogs-test-greedy_search-epoch-64-avg-33-context-2-max-sym-per-frame-1.txt
+  `-- test_wavs
+      |-- BAC009S0764W0121.wav
+      |-- BAC009S0764W0122.wav
+      |-- BAC009S0764W0123.wav
+      `-- transcript.txt
+
+  5 directories, 16 files
+
+
+**File descriptions**:
+
+  - ``data/lang_char``
+
+    It contains language related files. You can find the vocabulary size in ``tokens.txt``.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-32.pt`` to ``epoch-64.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``log``
+
+      It contains decoding logs and decoded results.
+
+  - ``test_wavs``
+
+      It contains some test sound files from Aishell ``test`` dataset.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/*.wav
+
+  Input File     : 'tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.20 = 67263 samples ~ 315.295 CDDA sectors
+  File Size      : 135k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.12 = 65840 samples ~ 308.625 CDDA sectors
+  File Size      : 132k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.00 = 64000 samples ~ 300 CDDA sectors
+  File Size      : 128k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+  Total Duration of 3 files: 00:00:12.32
+
+Usage
+~~~~~
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ ./transducer_stateless_modified/pretrained.py --help
+
+displays the help information.
+
+It supports three decoding methods:
+
+  - greedy search
+  - beam search
+  - modified beam search
+
+.. note::
+
+  In modified beam search, it limits the maximum number of symbols that can be
+  emitted per frame to 1. To use this method, you have to ensure that your model
+  has been trained with the option ``--modified-transducer-prob``. Otherwise,
+  it may give you poor results.
+
+Greedy search
+^^^^^^^^^^^^^
+
+The command to run greedy search is given below:
+
+.. code-block:: bash
+
+
+  $ cd egs/aishell/ASR
+  $ ./transducer_stateless_modified/pretrained.py \
+      --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \
+      --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \
+      --method greedy_search \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav
+
+The output is as follows:
+
+.. code-block::
+
+  2022-03-03 15:35:26,531 INFO [pretrained.py:239] device: cuda:0
+  2022-03-03 15:35:26,994 INFO [lexicon.py:176] Loading pre-compiled tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char/Linv.pt
+  2022-03-03 15:35:27,027 INFO [pretrained.py:246] {'feature_dim': 80, 'encoder_out_dim': 512, 'subsampling_factor': 4, 'attention_dim': 512, 'nhead': 8, 'dim_feedforward': 2048, 'num_encoder_layers': 12, 'vgg_frontend': False, 'env_info': {'k2-version': '1.13', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'f4fefe4882bc0ae59af951da3f47335d5495ef71', 'k2-git-date': 'Thu Feb 10 15:16:02 2022', 'lhotse-version': '1.0.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'master', 'icefall-git-sha1': '50d2281-clean', 'icefall-git-date': 'Wed Mar 2 16:02:38 2022', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-aishell', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-multi-datasets/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-aishell/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-2-0815224919-75d558775b-mmnv8', 'IP address': '10.177.72.138'}, 'sample_rate': 16000, 'checkpoint': './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt', 'lang_dir': PosixPath('tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char'), 'method': 'greedy_search', 'sound_files': ['./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav'], 'beam_size': 4, 'context_size': 2, 'max_sym_per_frame': 3, 'blank_id': 0, 'vocab_size': 4336}
+  2022-03-03 15:35:27,027 INFO [pretrained.py:248] About to create model
+  2022-03-03 15:35:36,878 INFO [pretrained.py:257] Constructing Fbank computer
+  2022-03-03 15:35:36,880 INFO [pretrained.py:267] Reading sound files: ['./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav']
+  2022-03-03 15:35:36,891 INFO [pretrained.py:273] Decoding started
+  /ceph-fj/fangjun/open-source-2/icefall-aishell/egs/aishell/ASR/transducer_stateless_modified/conformer.py:113: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').
+    lengths = ((x_lens - 1) // 2 - 1) // 2
+  2022-03-03 15:35:37,163 INFO [pretrained.py:320]
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav:
+  甚 至 出 现 交 易 几 乎 停 滞 的 情 况
+
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav:
+  一 二 线 城 市 虽 然 也 处 于 调 整 中
+
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav:
+  但 因 为 聚 集 了 过 多 公 共 资 源
+
+  2022-03-03 15:35:37,163 INFO [pretrained.py:322] Decoding Done
+
+Beam search
+^^^^^^^^^^^
+
+The command to run beam search is given below:
+
+.. code-block:: bash
+
+
+  $ cd egs/aishell/ASR
+
+  $ ./transducer_stateless_modified/pretrained.py \
+      --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \
+      --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \
+      --method beam_search \
+      --beam-size 4 \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav
+
+The output is as follows:
+
+.. code-block::
+
+  2022-03-03 15:39:09,285 INFO [pretrained.py:239] device: cuda:0
+  2022-03-03 15:39:09,708 INFO [lexicon.py:176] Loading pre-compiled tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char/Linv.pt
+  2022-03-03 15:39:09,759 INFO [pretrained.py:246] {'feature_dim': 80, 'encoder_out_dim': 512, 'subsampling_factor': 4, 'attention_dim': 512, 'nhead': 8, 'dim_feedforward': 2048, 'num_encoder_layers': 12, 'vgg_frontend': False, 'env_info': {'k2-version': '1.13', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'f4fefe4882bc0ae59af951da3f47335d5495ef71', 'k2-git-date': 'Thu Feb 10 15:16:02 2022', 'lhotse-version': '1.0.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'master', 'icefall-git-sha1': '50d2281-clean', 'icefall-git-date': 'Wed Mar 2 16:02:38 2022', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-aishell', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-multi-datasets/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-aishell/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-2-0815224919-75d558775b-mmnv8', 'IP address': '10.177.72.138'}, 'sample_rate': 16000, 'checkpoint': './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt', 'lang_dir': PosixPath('tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char'), 'method': 'beam_search', 'sound_files': ['./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav'], 'beam_size': 4, 'context_size': 2, 'max_sym_per_frame': 3, 'blank_id': 0, 'vocab_size': 4336}
+  2022-03-03 15:39:09,760 INFO [pretrained.py:248] About to create model
+  2022-03-03 15:39:18,919 INFO [pretrained.py:257] Constructing Fbank computer
+  2022-03-03 15:39:18,922 INFO [pretrained.py:267] Reading sound files: ['./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav']
+  2022-03-03 15:39:18,929 INFO [pretrained.py:273] Decoding started
+  /ceph-fj/fangjun/open-source-2/icefall-aishell/egs/aishell/ASR/transducer_stateless_modified/conformer.py:113: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').
+    lengths = ((x_lens - 1) // 2 - 1) // 2
+  2022-03-03 15:39:21,046 INFO [pretrained.py:320]
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav:
+  甚 至 出 现 交 易 几 乎 停 滞 的 情 况
+
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav:
+  一 二 线 城 市 虽 然 也 处 于 调 整 中
+
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav:
+  但 因 为 聚 集 了 过 多 公 共 资 源
+
+  2022-03-03 15:39:21,047 INFO [pretrained.py:322] Decoding Done
+
+Modified Beam search
+^^^^^^^^^^^^^^^^^^^^
+
+The command to run modified beam search is given below:
+
+.. code-block:: bash
+
+
+  $ cd egs/aishell/ASR
+
+  $ ./transducer_stateless_modified/pretrained.py \
+      --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \
+      --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \
+      --method modified_beam_search \
+      --beam-size 4 \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \
+      ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav
+
+The output is as follows:
+
+.. code-block::
+
+  2022-03-03 15:41:23,319 INFO [pretrained.py:239] device: cuda:0
+  2022-03-03 15:41:23,798 INFO [lexicon.py:176] Loading pre-compiled tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char/Linv.pt
+  2022-03-03 15:41:23,831 INFO [pretrained.py:246] {'feature_dim': 80, 'encoder_out_dim': 512, 'subsampling_factor': 4, 'attention_dim': 512, 'nhead': 8, 'dim_feedforward': 2048, 'num_encoder_layers': 12, 'vgg_frontend': False, 'env_info': {'k2-version': '1.13', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'f4fefe4882bc0ae59af951da3f47335d5495ef71', 'k2-git-date': 'Thu Feb 10 15:16:02 2022', 'lhotse-version': '1.0.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'master', 'icefall-git-sha1': '50d2281-clean', 'icefall-git-date': 'Wed Mar 2 16:02:38 2022', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-aishell', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-multi-datasets/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-aishell/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-2-0815224919-75d558775b-mmnv8', 'IP address': '10.177.72.138'}, 'sample_rate': 16000, 'checkpoint': './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt', 'lang_dir': PosixPath('tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char'), 'method': 'modified_beam_search', 'sound_files': ['./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav'], 'beam_size': 4, 'context_size': 2, 'max_sym_per_frame': 3, 'blank_id': 0, 'vocab_size': 4336}
+  2022-03-03 15:41:23,831 INFO [pretrained.py:248] About to create model
+  2022-03-03 15:41:32,214 INFO [pretrained.py:257] Constructing Fbank computer
+  2022-03-03 15:41:32,215 INFO [pretrained.py:267] Reading sound files: ['./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav', './tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav']
+  2022-03-03 15:41:32,220 INFO [pretrained.py:273] Decoding started
+  /ceph-fj/fangjun/open-source-2/icefall-aishell/egs/aishell/ASR/transducer_stateless_modified/conformer.py:113: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').
+    lengths = ((x_lens - 1) // 2 - 1) // 2
+  /ceph-fj/fangjun/open-source-2/icefall-aishell/egs/aishell/ASR/transducer_stateless_modified/beam_search.py:402: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').
+    topk_hyp_indexes = topk_indexes // logits.size(-1)
+  2022-03-03 15:41:32,583 INFO [pretrained.py:320]
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav:
+  甚 至 出 现 交 易 几 乎 停 滞 的 情 况
+
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav:
+  一 二 线 城 市 虽 然 也 处 于 调 整 中
+
+  ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav:
+  但 因 为 聚 集 了 过 多 公 共 资 源
+
+  2022-03-03 15:41:32,583 INFO [pretrained.py:322] Decoding Done
+
+Colab notebook
+--------------
+
+We provide a colab notebook for this recipe showing how to use a pre-trained model to
+transcribe sound files.
+
+|aishell asr stateless modified transducer colab notebook|
+
+.. |aishell asr stateless modified transducer colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/12jpTxJB44vzwtcmJl2DTdznW0OawPb9H?usp=sharing
diff --git a/docs/source/recipes/Non-streaming-ASR/aishell/tdnn_lstm_ctc.rst b/docs/source/recipes/Non-streaming-ASR/aishell/tdnn_lstm_ctc.rst
new file mode 100644
index 000000000..8e56deb6a
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/aishell/tdnn_lstm_ctc.rst
@@ -0,0 +1,504 @@
+TDNN-LSTM CTC
+=============
+
+This tutorial shows you how to run a tdnn-lstm ctc model
+with the `Aishell <https://www.openslr.org/33>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+In this tutorial, you will learn:
+
+  - (1) How to prepare data for training and decoding
+  - (2) How to start the training, either with a single GPU or multiple GPUs
+  - (3) How to do decoding after training.
+  - (4) How to use a pre-trained model, provided by us
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `Aishell <https://www.openslr.org/33>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/aishell`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. HINT::
+
+  A 3-gram language model will be downloaded from huggingface, we assume you have
+  installed and initialized ``git-lfs``. If not, you could install ``git-lfs`` by
+
+  .. code-block:: bash
+
+    $ sudo apt-get install git-lfs
+    $ git-lfs install
+
+  If you don't have the ``sudo`` permission, you could download the
+  `git-lfs binary <https://github.com/git-lfs/git-lfs/releases>`_ here, then add it to you ``PATH``.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./tdnn_lstm_ctc/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder ``./tdnn_lstm_ctc/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./tdnn_lstm_ctc/train.py --start-epoch 10`` loads the
+    checkpoint ``./tdnn_lstm_ctc/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./tdnn_lstm_ctc/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ ./tdnn_lstm_ctc/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./tdnn_lstm_ctc/train.py --world-size 1
+
+    .. CAUTION::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it. For instance, if
+    your are using V100 NVIDIA GPU, we recommend you to set it to ``2000``.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`tdnn_lstm_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/tdnn_lstm_ctc/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./tdnn_lstm_ctc/train.py`` directly.
+
+
+.. CAUTION::
+
+  The training set is perturbed by speed with two factors: 0.9 and 1.1.
+  Each epoch actually processes ``3x150 == 450`` hours of data.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``tdnn_lstm_ctc/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./tdnn_lstm_ctc/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd tdnn_lstm_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "TDNN-LSTM CTC training for Aishell with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/LJI9MWUORLOw3jkdhxwk8A/
+
+        [2021-09-13T11:59:23] Started scanning logdir.
+        [2021-09-13T11:59:24] Total uploaded: 4454 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output, click it and you will see
+    the following screenshot:
+
+      .. figure:: images/aishell-tdnn-lstm-ctc-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/LJI9MWUORLOw3jkdhxwk8A/
+
+         TensorBoard screenshot.
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage examples
+~~~~~~~~~~~~~~
+
+The following shows typical use cases:
+
+**Case 1**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,3"
+  $ ./tdnn_lstm_ctc/train.py --world-size 2
+
+It uses GPU 0 and GPU 3 for DDP training.
+
+**Case 2**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/train.py --num-epochs 10 --start-epoch 3
+
+It loads checkpoint ``./tdnn_lstm_ctc/exp/epoch-2.pt`` and starts
+training from epoch 3. Also, it trains for 10 epochs.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/decode.py --help
+
+shows the options for decoding.
+
+The commonly used options are:
+
+  - ``--method``
+
+    This specifies the decoding method.
+
+    The following command uses attention decoder for rescoring:
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./tdnn_lstm_ctc/decode.py --method 1best --max-duration 100
+
+  - ``--max-duration``
+
+    It has the same meaning as the one during training. A larger
+    value may cause OOM.
+
+Pre-trained Model
+-----------------
+
+We have uploaded a pre-trained model to
+`<https://huggingface.co/pkufool/icefall_asr_aishell_tdnn_lstm_ctc>`_.
+
+We describe how to use the pre-trained model to transcribe a sound file or
+multiple sound files in the following.
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following commands describe how to download the pre-trained model:
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/pkufool/icefall_asr_aishell_tdnn_lstm_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ tree tmp
+
+.. code-block:: bash
+
+  tmp/
+  `-- icefall_asr_aishell_tdnn_lstm_ctc
+      |-- README.md
+      |-- data
+      |   `-- lang_phone
+      |       |-- HLG.pt
+      |       |-- tokens.txt
+      |       `-- words.txt
+      |-- exp
+      |   `-- pretrained.pt
+      `-- test_waves
+          |-- BAC009S0764W0121.wav
+          |-- BAC009S0764W0122.wav
+          |-- BAC009S0764W0123.wav
+          `-- trans.txt
+
+  5 directories, 9 files
+
+**File descriptions**:
+
+  - ``data/lang_phone/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_phone/tokens.txt``
+
+      It contains tokens and their IDs.
+      Provided only for convenience so that you can look up the SOS/EOS ID easily.
+
+  - ``data/lang_phone/words.txt``
+
+      It contains words and their IDs.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-18.pt`` to ``epoch-40.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.wav``
+
+      It contains some test sound files from Aishell ``test`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in `test_waves/`.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/*.wav
+
+  Input File     : 'tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.20 = 67263 samples ~ 315.295 CDDA sectors
+  File Size      : 135k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.12 = 65840 samples ~ 308.625 CDDA sectors
+  File Size      : 132k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.00 = 64000 samples ~ 300 CDDA sectors
+  File Size      : 128k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+  Total Duration of 3 files: 00:00:12.32
+
+Usage
+~~~~~
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/pretrained.py --help
+
+displays the help information.
+
+
+HLG decoding
+^^^^^^^^^^^^
+
+HLG decoding uses the best path of the decoding lattice as the decoding result.
+
+The command to run HLG decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/exp/pretrained.pt \
+    --words-file ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/data/lang_phone/words.txt \
+    --HLG ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/data/lang_phone/HLG.pt \
+    --method 1best \
+    ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav
+
+The output is given below:
+
+.. code-block::
+
+  2021-09-13 15:00:55,858 INFO [pretrained.py:140] device: cuda:0
+  2021-09-13 15:00:55,858 INFO [pretrained.py:142] Creating model
+  2021-09-13 15:01:05,389 INFO [pretrained.py:154] Loading HLG from ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/data/lang_phone/HLG.pt
+  2021-09-13 15:01:06,531 INFO [pretrained.py:161] Constructing Fbank computer
+  2021-09-13 15:01:06,536 INFO [pretrained.py:171] Reading sound files: ['./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-09-13 15:01:06,539 INFO [pretrained.py:177] Decoding started
+  2021-09-13 15:01:06,917 INFO [pretrained.py:207] Use HLG decoding
+  2021-09-13 15:01:07,129 INFO [pretrained.py:220]
+  ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav:
+  甚至 出现 交易 几乎 停滞 的 情况
+
+  ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav:
+  一二 线 城市 虽然 也 处于 调整 中
+
+  ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因为 聚集 了 过多 公共 资源
+
+
+  2021-09-13 15:01:07,129 INFO [pretrained.py:222] Decoding Done
+
+
+Colab notebook
+--------------
+
+We do provide a colab notebook for this recipe showing how to use a pre-trained model.
+
+|aishell asr conformer ctc colab notebook|
+
+.. |aishell asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1jbyzYq3ytm6j2nlEt-diQm-6QVWyDDEa?usp=sharing
+
+**Congratulations!** You have finished the aishell ASR recipe with
+TDNN-LSTM CTC models in ``icefall``.
diff --git a/docs/source/recipes/Non-streaming-ASR/index.rst b/docs/source/recipes/Non-streaming-ASR/index.rst
new file mode 100644
index 000000000..67123a648
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/index.rst
@@ -0,0 +1,10 @@
+Non Streaming ASR
+=================
+
+.. toctree::
+   :maxdepth: 2
+
+   aishell/index
+   librispeech/index
+   timit/index
+   yesno/index
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/conformer_ctc.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/conformer_ctc.rst
new file mode 100644
index 000000000..b7f89c89f
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/conformer_ctc.rst
@@ -0,0 +1,1070 @@
+Conformer CTC
+=============
+
+This tutorial shows you how to run a conformer ctc model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+In this tutorial, you will learn:
+
+  - (1) How to prepare data for training and decoding
+  - (2) How to start the training, either with a single GPU or multiple GPUs
+  - (3) How to do decoding after training, with n-gram LM rescoring and attention decoder rescoring
+  - (4) How to use a pre-trained model, provided by us
+  - (5) How to deploy your trained model in C++, without Python dependencies
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./conformer_ctc/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder ``./conformer_ctc/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./conformer_ctc/train.py --start-epoch 10`` loads the
+    checkpoint ``./conformer_ctc/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./conformer_ctc/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./conformer_ctc/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./conformer_ctc/train.py --world-size 1
+
+    .. CAUTION::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it. For instance, if
+    your are using V100 NVIDIA GPU, we recommend you to set it to ``200``.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`conformer_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/conformer_ctc/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./conformer_ctc/train.py`` directly.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``conformer_ctc/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./conformer_ctc/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd conformer_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Conformer CTC training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/lzGnETjwRxC3yghNMd4kPw/
+
+        [2021-08-24T16:42:43] Started scanning logdir.
+        Uploading 4540 scalars...
+
+    Note there is a URL in the above output, click it and you will see
+    the following screenshot:
+
+      .. figure:: images/librispeech-conformer-ctc-tensorboard-log.png
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/lzGnETjwRxC3yghNMd4kPw/
+
+         TensorBoard screenshot.
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage examples
+~~~~~~~~~~~~~~
+
+The following shows typical use cases:
+
+**Case 1**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/train.py --max-duration 200 --full-libri 0
+
+It uses ``--max-duration`` of 200 to avoid OOM.  Also, it uses only
+a subset of the LibriSpeech data for training.
+
+
+**Case 2**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,3"
+  $ ./conformer_ctc/train.py --world-size 2
+
+It uses GPU 0 and GPU 3 for DDP training.
+
+**Case 3**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/train.py --num-epochs 10 --start-epoch 3
+
+It loads checkpoint ``./conformer_ctc/exp/epoch-2.pt`` and starts
+training from epoch 3. Also, it trains for 10 epochs.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/decode.py --help
+
+shows the options for decoding.
+
+The commonly used options are:
+
+  - ``--method``
+
+    This specifies the decoding method. This script supports 7 decoding methods.
+    As for ctc decoding, it uses a sentence piece model to convert word pieces to words.
+    And it needs neither a lexicon nor an n-gram LM.
+
+    For example, the following command uses CTC topology for decoding:
+
+    .. code-block::
+
+      $ cd egs/librispeech/ASR
+      $ ./conformer_ctc/decode.py --method ctc-decoding --max-duration 300
+      # Caution: The above command is tested with a model with vocab size 500.
+
+    And the following command uses attention decoder for rescoring:
+
+    .. code-block::
+
+      $ cd egs/librispeech/ASR
+      $ ./conformer_ctc/decode.py --method attention-decoder --max-duration 30 --nbest-scale 0.5
+
+  - ``--nbest-scale``
+
+    It is used to scale down lattice scores so that there are more unique
+    paths for rescoring.
+
+  - ``--max-duration``
+
+    It has the same meaning as the one during training. A larger
+    value may cause OOM.
+
+Here are some results for CTC decoding with a vocab size of 500:
+
+Usage:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  # NOTE: Tested with a model with vocab size 500.
+  # It won't work for a model with vocab size 5000.
+  $ ./conformer_ctc/decode.py \
+      --epoch 25 \
+      --avg 1 \
+      --max-duration 300 \
+      --exp-dir conformer_ctc/exp \
+      --lang-dir data/lang_bpe_500 \
+      --method ctc-decoding
+
+The output is given below:
+
+.. code-block:: bash
+
+  2021-09-26 12:44:31,033 INFO [decode.py:537] Decoding started
+  2021-09-26 12:44:31,033 INFO [decode.py:538]
+  {'lm_dir': PosixPath('data/lm'), 'subsampling_factor': 4, 'vgg_frontend': False, 'use_feat_batchnorm': True,
+  'feature_dim': 80, 'nhead': 8, 'attention_dim': 512, 'num_decoder_layers': 6, 'search_beam': 20, 'output_beam': 8,
+  'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True,
+  'epoch': 25, 'avg': 1, 'method': 'ctc-decoding', 'num_paths': 100, 'nbest_scale': 0.5,
+  'export': False, 'exp_dir': PosixPath('conformer_ctc/exp'), 'lang_dir': PosixPath('data/lang_bpe_500'), 'full_libri': False,
+  'feature_dir': PosixPath('data/fbank'), 'max_duration': 100, 'bucketing_sampler': False, 'num_buckets': 30,
+  'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False,
+  'shuffle': True, 'return_cuts': True, 'num_workers': 2}
+  2021-09-26 12:44:31,406 INFO [lexicon.py:113] Loading pre-compiled data/lang_bpe_500/Linv.pt
+  2021-09-26 12:44:31,464 INFO [decode.py:548] device: cuda:0
+  2021-09-26 12:44:36,171 INFO [checkpoint.py:92] Loading checkpoint from conformer_ctc/exp/epoch-25.pt
+  2021-09-26 12:44:36,776 INFO [decode.py:652] Number of model parameters: 109226120
+  2021-09-26 12:44:37,714 INFO [decode.py:473] batch 0/206, cuts processed until now is 12
+  2021-09-26 12:45:15,944 INFO [decode.py:473] batch 100/206, cuts processed until now is 1328
+  2021-09-26 12:45:54,443 INFO [decode.py:473] batch 200/206, cuts processed until now is 2563
+  2021-09-26 12:45:56,411 INFO [decode.py:494] The transcripts are stored in conformer_ctc/exp/recogs-test-clean-ctc-decoding.txt
+  2021-09-26 12:45:56,592 INFO [utils.py:331] [test-clean-ctc-decoding] %WER 3.26% [1715 / 52576, 163 ins, 128 del, 1424 sub ]
+  2021-09-26 12:45:56,807 INFO [decode.py:506] Wrote detailed error stats to conformer_ctc/exp/errs-test-clean-ctc-decoding.txt
+  2021-09-26 12:45:56,808 INFO [decode.py:522]
+  For test-clean, WER of different settings are:
+  ctc-decoding    3.26    best for test-clean
+
+  2021-09-26 12:45:57,362 INFO [decode.py:473] batch 0/203, cuts processed until now is 15
+  2021-09-26 12:46:35,565 INFO [decode.py:473] batch 100/203, cuts processed until now is 1477
+  2021-09-26 12:47:15,106 INFO [decode.py:473] batch 200/203, cuts processed until now is 2922
+  2021-09-26 12:47:16,131 INFO [decode.py:494] The transcripts are stored in conformer_ctc/exp/recogs-test-other-ctc-decoding.txt
+  2021-09-26 12:47:16,208 INFO [utils.py:331] [test-other-ctc-decoding] %WER 8.21% [4295 / 52343, 396 ins, 315 del, 3584 sub ]
+  2021-09-26 12:47:16,432 INFO [decode.py:506] Wrote detailed error stats to conformer_ctc/exp/errs-test-other-ctc-decoding.txt
+  2021-09-26 12:47:16,432 INFO [decode.py:522]
+  For test-other, WER of different settings are:
+  ctc-decoding    8.21    best for test-other
+
+  2021-09-26 12:47:16,433 INFO [decode.py:680] Done!
+
+Pre-trained Model
+-----------------
+
+We have uploaded a pre-trained model to
+`<https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09>`_
+
+We describe how to use the pre-trained model to transcribe a sound file or
+multiple sound files in the following.
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following commands describe how to download the pre-trained model:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+  $ cd icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+  $ git lfs pull
+
+.. CAUTION::
+
+  You have to use ``git lfs pull`` to download the pre-trained model.
+  Otherwise, you will have the following issue when running ``decode.py``:
+
+    .. code-block::
+
+       _pickle.UnpicklingError: invalid load key, 'v'
+
+  To fix that issue, please use:
+
+     .. code-block:: bash
+
+        cd icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+        git lfs pull
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.9 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ tree icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+
+.. code-block:: bash
+
+  icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+  |-- README.md
+  |-- data
+  |   |-- lang_bpe_500
+  |   |   |-- HLG.pt
+  |   |   |-- HLG_modified.pt
+  |   |   |-- bpe.model
+  |   |   |-- tokens.txt
+  |   |   `-- words.txt
+  |   `-- lm
+  |       `-- G_4_gram.pt
+  |-- exp
+  |   |-- cpu_jit.pt
+  |   `-- pretrained.pt
+  |-- log
+  |   `-- log-decode-2021-11-09-17-38-28
+  `-- test_wavs
+      |-- 1089-134686-0001.wav
+      |-- 1221-135766-0001.wav
+      |-- 1221-135766-0002.wav
+      `-- trans.txt
+
+
+**File descriptions**:
+  - ``data/lang_bpe_500/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_bpe_500/HLG_modified.pt``
+
+      It uses a modified CTC topology while building HLG.
+
+  - ``data/lang_bpe_500/bpe.model``
+
+      It is a sentencepiece model. You can use it to reproduce our results.
+
+  - ``data/lang_bpe_500/tokens.txt``
+
+      It contains tokens and their IDs, generated from ``bpe.model``.
+      Provided only for convenience so that you can look up the SOS/EOS ID easily.
+
+  - ``data/lang_bpe_500/words.txt``
+
+      It contains words and their IDs.
+
+  - ``data/lm/G_4_gram.pt``
+
+      It is a 4-gram LM, used for n-gram LM rescoring.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-23.pt`` to ``epoch-77.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``exp/cpu_jit.pt``
+
+      It contains torch scripted model that can be deployed in C++.
+
+  - ``test_wavs/*.wav``
+
+      It contains some test sound files from LibriSpeech ``test-clean`` dataset.
+
+  - ``test_wavs/trans.txt``
+
+      It contains the reference transcripts for the sound files in ``test_wavs/``.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/*.wav
+
+  Input File     : 'icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:06.62 = 106000 samples ~ 496.875 CDDA sectors
+  File Size      : 212k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:16.71 = 267440 samples ~ 1253.62 CDDA sectors
+  File Size      : 535k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.83 = 77200 samples ~ 361.875 CDDA sectors
+  File Size      : 154k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+  Total Duration of 3 files: 00:00:28.16
+
+Usage
+~~~~~
+
+.. code-block::
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/pretrained.py --help
+
+displays the help information.
+
+It supports 4 decoding methods:
+
+  - CTC decoding
+  - HLG decoding
+  - HLG + n-gram LM rescoring
+  - HLG + n-gram LM rescoring + attention decoder rescoring
+
+CTC decoding
+^^^^^^^^^^^^
+
+CTC decoding uses the best path of the decoding lattice as the decoding result
+without any LM or lexicon.
+
+The command to run CTC decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/pretrained.py \
+     --checkpoint ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt \
+     --bpe-model ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/bpe.model \
+     --method ctc-decoding \
+     --num-classes 500 \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+The output is given below:
+
+.. code-block::
+
+  2021-11-10 12:12:29,554 INFO [pretrained.py:260] {'sample_rate': 16000, 'subsampling_factor': 4, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'feature_dim': 80, 'nhead': 8, 'attention_dim': 512, 'num_decoder_layers': 0, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt', 'words_file': None, 'HLG': None, 'bpe_model': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/bpe.model', 'method': 'ctc-decoding', 'G': None, 'num_paths': 100, 'ngram_lm_scale': 1.3, 'attention_decoder_scale': 1.2, 'nbest_scale': 0.5, 'sos_id': 1, 'num_classes': 500, 'eos_id': 1, 'sound_files': ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav'], 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '7178d67e594bc7fa89c2b331ad7bd1c62a6a9eb4', 'k2-git-date': 'Tue Oct 26 22:12:54 2021', 'lhotse-version': '0.11.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'bpe-500', 'icefall-git-sha1': '8d93169-dirty', 'icefall-git-date': 'Wed Nov 10 11:52:44 2021', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-fix', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-bpe-500/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-bpe-500/lhotse/__init__.py'}}
+  2021-11-10 12:12:29,554 INFO [pretrained.py:266] device: cuda:0
+  2021-11-10 12:12:29,554 INFO [pretrained.py:268] Creating model
+  2021-11-10 12:12:35,600 INFO [pretrained.py:285] Constructing Fbank computer
+  2021-11-10 12:12:35,601 INFO [pretrained.py:295] Reading sound files: ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav']
+  2021-11-10 12:12:35,758 INFO [pretrained.py:301] Decoding started
+  2021-11-10 12:12:36,025 INFO [pretrained.py:319] Use CTC decoding
+  2021-11-10 12:12:36,204 INFO [pretrained.py:425]
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav:
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROFFELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav:
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED B
+  OSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav:
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+  2021-11-10 12:12:36,204 INFO [pretrained.py:427] Decoding Done
+
+HLG decoding
+^^^^^^^^^^^^
+
+HLG decoding uses the best path of the decoding lattice as the decoding result.
+
+The command to run HLG decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/pretrained.py \
+     --checkpoint ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt \
+     --words-file ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt \
+     --method 1best \
+     --num-classes 500 \
+     --HLG ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+The output is given below:
+
+.. code-block::
+
+  2021-11-10 13:33:03,723 INFO [pretrained.py:260] {'sample_rate': 16000, 'subsampling_factor': 4, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'feature_dim': 80, 'nhead': 8, 'attention_dim': 512, 'num_decoder_layers': 0, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt', 'words_file': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt', 'HLG': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt', 'bpe_model': None, 'method': '1best', 'G': None, 'num_paths': 100, 'ngram_lm_scale': 1.3, 'attention_decoder_scale': 1.2, 'nbest_scale': 0.5, 'sos_id': 1, 'num_classes': 500, 'eos_id': 1, 'sound_files': ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav'], 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '7178d67e594bc7fa89c2b331ad7bd1c62a6a9eb4', 'k2-git-date': 'Tue Oct 26 22:12:54 2021', 'lhotse-version': '0.11.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'bpe-500', 'icefall-git-sha1': '8d93169-dirty', 'icefall-git-date': 'Wed Nov 10 11:52:44 2021', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-fix', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-bpe-500/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-bpe-500/lhotse/__init__.py'}}
+  2021-11-10 13:33:03,723 INFO [pretrained.py:266] device: cuda:0
+  2021-11-10 13:33:03,723 INFO [pretrained.py:268] Creating model
+  2021-11-10 13:33:09,775 INFO [pretrained.py:285] Constructing Fbank computer
+  2021-11-10 13:33:09,776 INFO [pretrained.py:295] Reading sound files: ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav']
+  2021-11-10 13:33:09,881 INFO [pretrained.py:301] Decoding started
+  2021-11-10 13:33:09,951 INFO [pretrained.py:352] Loading HLG from ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt
+  2021-11-10 13:33:13,234 INFO [pretrained.py:384] Use HLG decoding
+  2021-11-10 13:33:13,571 INFO [pretrained.py:425]
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav:
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav:
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav:
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+  2021-11-10 13:33:13,571 INFO [pretrained.py:427] Decoding Done
+
+
+HLG decoding + LM rescoring
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+It uses an n-gram LM to rescore the decoding lattice and the best
+path of the rescored lattice is the decoding result.
+
+The command to run HLG decoding + LM rescoring is:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  ./conformer_ctc/pretrained.py \
+     --checkpoint ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt \
+     --words-file ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt \
+     --method whole-lattice-rescoring \
+     --num-classes 500 \
+     --HLG ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt \
+     --G ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt \
+     --ngram-lm-scale 1.0 \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+Its output is:
+
+.. code-block::
+
+  2021-11-10 13:39:55,857 INFO [pretrained.py:260] {'sample_rate': 16000, 'subsampling_factor': 4, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'feature_dim': 80, 'nhead': 8, 'attention_dim': 512, 'num_decoder_layers': 0, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt', 'words_file': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt', 'HLG': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt', 'bpe_model': None, 'method': 'whole-lattice-rescoring', 'G': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt', 'num_paths': 100, 'ngram_lm_scale': 1.0, 'attention_decoder_scale': 1.2, 'nbest_scale': 0.5, 'sos_id': 1, 'num_classes': 500, 'eos_id': 1, 'sound_files': ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav'], 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-$it-sha1': '7178d67e594bc7fa89c2b331ad7bd1c62a6a9eb4', 'k2-git-date': 'Tue Oct 26 22:12:54 2021', 'lhotse-version': '0.11.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'bpe-500', 'icefall-git-sha1': '8d93169-dirty', 'icefall-git-date': 'Wed Nov 10 11:52:44 2021', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-fix', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-bpe-500/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-bpe-500/lhotse/__init__.py'}}
+  2021-11-10 13:39:55,858 INFO [pretrained.py:266] device: cuda:0
+  2021-11-10 13:39:55,858 INFO [pretrained.py:268] Creating model
+  2021-11-10 13:40:01,979 INFO [pretrained.py:285] Constructing Fbank computer
+  2021-11-10 13:40:01,980 INFO [pretrained.py:295] Reading sound files: ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav']
+  2021-11-10 13:40:02,055 INFO [pretrained.py:301] Decoding started
+  2021-11-10 13:40:02,117 INFO [pretrained.py:352] Loading HLG from ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt
+  2021-11-10 13:40:05,051 INFO [pretrained.py:363] Loading G from ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt
+  2021-11-10 13:40:18,959 INFO [pretrained.py:389] Use HLG decoding + LM rescoring
+  2021-11-10 13:40:19,546 INFO [pretrained.py:425]
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav:
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav:
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav:
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+  2021-11-10 13:40:19,546 INFO [pretrained.py:427] Decoding Done
+
+
+HLG decoding + LM rescoring + attention decoder rescoring
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+It uses an n-gram LM to rescore the decoding lattice, extracts
+n paths from the rescored lattice, recores the extracted paths with
+an attention decoder. The path with the highest score is the decoding result.
+
+The command to run HLG decoding + LM rescoring + attention decoder rescoring is:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./conformer_ctc/pretrained.py \
+     --checkpoint ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt \
+     --words-file ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt \
+     --method attention-decoder \
+     --num-classes 500 \
+     --HLG ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt \
+     --G ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt \
+     --ngram-lm-scale 2.0 \
+     --attention-decoder-scale 2.0 \
+     --nbest-scale 0.5 \
+     --num-paths 100 \
+     --sos-id 1 \
+     --eos-id 1 \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+     ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+The output is below:
+
+.. code-block::
+
+  2021-11-10 13:43:45,598 INFO [pretrained.py:260] {'sample_rate': 16000, 'subsampling_factor': 4, 'vgg_frontend': False, 'use_feat_batchnorm': True, 'feature_dim': 80, 'nhead': 8, 'attention_dim': 512, 'num_decoder_layers': 6, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/pretrained.pt', 'words_file': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt', 'HLG': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt', 'bpe_model': None, 'method': 'attention-decoder', 'G': './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt', 'num_paths': 100, 'ngram_lm_scale': 2.0, 'attention_decoder_scale': 2.0, 'nbest_scale': 0.5, 'sos_id': 1, 'num_classes': 500, 'eos_id': 1, 'sound_files': ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav'], 'env_info': {'k2-version': '1.9', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '7178d67e594bc7fa89c2b331ad7bd1c62a6a9eb4', 'k2-git-date': 'Tue Oct 26 22:12:54 2021', 'lhotse-version': '0.11.0.dev+missing.version.file', 'torch-cuda-available': True, 'torch-cuda-version': '10.1', 'python-version': '3.8', 'icefall-git-branch': 'bpe-500', 'icefall-git-sha1': '8d93169-dirty', 'icefall-git-date': 'Wed Nov 10 11:52:44 2021', 'icefall-path': '/ceph-fj/fangjun/open-source-2/icefall-fix', 'k2-path': '/ceph-fj/fangjun/open-source-2/k2-bpe-500/k2/python/k2/__init__.py', 'lhotse-path': '/ceph-fj/fangjun/open-source-2/lhotse-bpe-500/lhotse/__init__.py'}}
+  2021-11-10 13:43:45,599 INFO [pretrained.py:266] device: cuda:0
+  2021-11-10 13:43:45,599 INFO [pretrained.py:268] Creating model
+  2021-11-10 13:43:51,833 INFO [pretrained.py:285] Constructing Fbank computer
+  2021-11-10 13:43:51,834 INFO [pretrained.py:295] Reading sound files: ['./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav', './icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav']
+  2021-11-10 13:43:51,915 INFO [pretrained.py:301] Decoding started
+  2021-11-10 13:43:52,076 INFO [pretrained.py:352] Loading HLG from ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt
+  2021-11-10 13:43:55,110 INFO [pretrained.py:363] Loading G from ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt
+  2021-11-10 13:44:09,329 INFO [pretrained.py:397] Use HLG + LM rescoring + attention decoder rescoring
+  2021-11-10 13:44:10,192 INFO [pretrained.py:425]
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav:
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav:
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav:
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+  2021-11-10 13:44:10,192 INFO [pretrained.py:427] Decoding Done
+
+
+Compute WER with the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+To check the WER of the pre-trained model on the test datasets, run:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ cd icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/
+  $ ln -s pretrained.pt epoch-999.pt
+  $ cd ../..
+  $ ./conformer_ctc/decode.py \
+      --exp-dir ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp \
+      --lang-dir ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500 \
+      --lm-dir ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm \
+      --epoch 999 \
+      --avg 1 \
+      --concatenate-cuts 0 \
+      --bucketing-sampler 1 \
+      --max-duration 30 \
+      --num-paths 1000 \
+      --method attention-decoder \
+      --nbest-scale 0.5
+
+
+Colab notebook
+--------------
+
+We do provide a colab notebook for this recipe showing how to use a pre-trained model.
+
+|librispeech asr conformer ctc colab notebook|
+
+.. |librispeech asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1huyupXAcHsUrKaWfI83iMEJ6J0Nh0213?usp=sharing
+
+.. HINT::
+
+  Due to limited memory provided by Colab, you have to upgrade to Colab Pro to
+  run ``HLG decoding + LM rescoring`` and
+  ``HLG decoding + LM rescoring + attention decoder rescoring``.
+  Otherwise, you can only run ``HLG decoding`` with Colab.
+
+**Congratulations!** You have finished the LibriSpeech ASR recipe with
+conformer CTC models in ``icefall``.
+
+If you want to deploy your trained model in C++, please read the following section.
+
+Deployment with C++
+-------------------
+
+This section describes how to deploy the pre-trained model in C++, without
+Python dependencies.
+
+.. HINT::
+
+  At present, it does NOT support streaming decoding.
+
+First, let us compile k2 from source:
+
+.. code-block:: bash
+
+  $ cd $HOME
+  $ git clone https://github.com/k2-fsa/k2
+  $ cd k2
+  $ git checkout v2.0-pre
+
+.. CAUTION::
+
+  You have to switch to the branch ``v2.0-pre``!
+
+.. code-block:: bash
+
+  $ mkdir build-release
+  $ cd build-release
+  $ cmake -DCMAKE_BUILD_TYPE=Release ..
+  $ make -j ctc_decode hlg_decode ngram_lm_rescore attention_rescore
+
+  # You will find four binaries in `./bin`, i.e.,
+  # ./bin/ctc_decode, ./bin/hlg_decode,
+  # ./bin/ngram_lm_rescore, and ./bin/attention_rescore
+
+Now you are ready to go!
+
+Assume you have run:
+
+  .. code-block:: bash
+
+    $ cd k2/build-release
+    $ ln -s /path/to/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09 ./
+
+To view the usage of ``./bin/ctc_decode``, run:
+
+.. code-block::
+
+  $ ./bin/ctc_decode
+
+It will show you the following message:
+
+.. code-block:: bash
+
+  Please provide --nn_model
+
+  This file implements decoding with a CTC topology, without any
+  kinds of LM or lexicons.
+
+  Usage:
+    ./bin/ctc_decode \
+      --use_gpu true \
+      --nn_model <path to torch scripted pt file> \
+      --bpe_model <path to pre-trained BPE model> \
+      <path to foo.wav> \
+      <path to bar.wav> \
+      <more waves if any>
+
+  To see all possible options, use
+    ./bin/ctc_decode --help
+
+  Caution:
+   - Only sound files (*.wav) with single channel are supported.
+   - It assumes the model is conformer_ctc/transformer.py from icefall.
+     If you use a different model, you have to change the code
+     related to `model.forward` in this file.
+
+
+CTC decoding
+~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./bin/ctc_decode \
+    --use_gpu true \
+    --nn_model ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/cpu_jit.pt \
+    --bpe_model ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/bpe.model \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+Its output is:
+
+.. code-block::
+
+  2021-11-10 13:57:55.316 [I] k2/torch/bin/ctc_decode.cu:105:int main(int, char**) Use GPU
+  2021-11-10 13:57:55.316 [I] k2/torch/bin/ctc_decode.cu:109:int main(int, char**) Device: cuda:0
+  2021-11-10 13:57:55.316 [I] k2/torch/bin/ctc_decode.cu:118:int main(int, char**) Load wave files
+  2021-11-10 13:58:01.221 [I] k2/torch/bin/ctc_decode.cu:125:int main(int, char**) Build Fbank computer
+  2021-11-10 13:58:01.222 [I] k2/torch/bin/ctc_decode.cu:136:int main(int, char**) Compute features
+  2021-11-10 13:58:01.228 [I] k2/torch/bin/ctc_decode.cu:144:int main(int, char**) Load neural network model
+  2021-11-10 13:58:02.19 [I] k2/torch/bin/ctc_decode.cu:159:int main(int, char**) Compute nnet_output
+  2021-11-10 13:58:02.543 [I] k2/torch/bin/ctc_decode.cu:174:int main(int, char**) Build CTC topo
+  2021-11-10 13:58:02.547 [I] k2/torch/bin/ctc_decode.cu:177:int main(int, char**) Decoding
+  2021-11-10 13:58:02.708 [I] k2/torch/bin/ctc_decode.cu:207:int main(int, char**)
+  Decoding result:
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROFFELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+HLG decoding
+~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./bin/hlg_decode \
+    --use_gpu true \
+    --nn_model ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/cpu_jit.pt \
+    --hlg ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt \
+    --word_table ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+The output is:
+
+.. code-block::
+
+  2021-11-10 13:59:04.729 [I] k2/torch/bin/hlg_decode.cu:111:int main(int, char**) Use GPU
+  2021-11-10 13:59:04.729 [I] k2/torch/bin/hlg_decode.cu:115:int main(int, char**) Device: cuda:0
+  2021-11-10 13:59:04.729 [I] k2/torch/bin/hlg_decode.cu:124:int main(int, char**) Load wave files
+  2021-11-10 13:59:10.702 [I] k2/torch/bin/hlg_decode.cu:131:int main(int, char**) Build Fbank computer
+  2021-11-10 13:59:10.703 [I] k2/torch/bin/hlg_decode.cu:142:int main(int, char**) Compute features
+  2021-11-10 13:59:10.707 [I] k2/torch/bin/hlg_decode.cu:150:int main(int, char**) Load neural network model
+  2021-11-10 13:59:11.545 [I] k2/torch/bin/hlg_decode.cu:165:int main(int, char**) Compute nnet_output
+  2021-11-10 13:59:12.72 [I] k2/torch/bin/hlg_decode.cu:180:int main(int, char**) Load ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt
+  2021-11-10 13:59:12.994 [I] k2/torch/bin/hlg_decode.cu:185:int main(int, char**) Decoding
+  2021-11-10 13:59:13.268 [I] k2/torch/bin/hlg_decode.cu:216:int main(int, char**)
+  Decoding result:
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+
+HLG decoding + n-gram LM rescoring
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./bin/ngram_lm_rescore \
+    --use_gpu true \
+    --nn_model ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/cpu_jit.pt \
+    --hlg ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt \
+    --g ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt \
+    --ngram_lm_scale 1.0 \
+    --word_table ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+The output is:
+
+.. code-block::
+
+  2021-11-10 14:00:55.279 [I] k2/torch/bin/ngram_lm_rescore.cu:122:int main(int, char**) Use GPU
+  2021-11-10 14:00:55.280 [I] k2/torch/bin/ngram_lm_rescore.cu:126:int main(int, char**) Device: cuda:0
+  2021-11-10 14:00:55.280 [I] k2/torch/bin/ngram_lm_rescore.cu:135:int main(int, char**) Load wave files
+  2021-11-10 14:01:01.214 [I] k2/torch/bin/ngram_lm_rescore.cu:142:int main(int, char**) Build Fbank computer
+  2021-11-10 14:01:01.215 [I] k2/torch/bin/ngram_lm_rescore.cu:153:int main(int, char**) Compute features
+  2021-11-10 14:01:01.219 [I] k2/torch/bin/ngram_lm_rescore.cu:161:int main(int, char**) Load neural network model
+  2021-11-10 14:01:01.945 [I] k2/torch/bin/ngram_lm_rescore.cu:176:int main(int, char**) Compute nnet_output
+  2021-11-10 14:01:02.475 [I] k2/torch/bin/ngram_lm_rescore.cu:191:int main(int, char**) Load ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt
+  2021-11-10 14:01:03.398 [I] k2/torch/bin/ngram_lm_rescore.cu:199:int main(int, char**) Decoding
+  2021-11-10 14:01:03.515 [I] k2/torch/bin/ngram_lm_rescore.cu:205:int main(int, char**) Load n-gram LM: ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt
+  2021-11-10 14:01:07.432 [W] k2/torch/csrc/deserialization.cu:441:k2::FsaClass k2::LoadFsa(const string&, c10::optional<c10::Device>)
+  Ignore non tensor attribute: 'dummy' of type: Int
+  2021-11-10 14:01:07.589 [I] k2/torch/bin/ngram_lm_rescore.cu:214:int main(int, char**) Rescore with an n-gram LM
+  2021-11-10 14:01:08.68 [I] k2/torch/bin/ngram_lm_rescore.cu:242:int main(int, char**)
+  Decoding result:
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+
+HLG decoding + n-gram LM rescoring + attention decoder rescoring
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./bin/attention_rescore \
+    --use_gpu true \
+    --nn_model ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/exp/cpu_jit.pt \
+    --hlg ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt \
+    --g ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt \
+    --ngram_lm_scale 2.0 \
+    --attention_scale 2.0 \
+    --num_paths 100 \
+    --nbest_scale 0.5 \
+    --word_table ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/words.txt \
+    --sos_id 1 \
+    --eos_id 1 \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav \
+    ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+
+The output is:
+
+.. code-block::
+
+  2021-11-10 14:02:43.656 [I] k2/torch/bin/attention_rescore.cu:149:int main(int, char**) Use GPU
+  2021-11-10 14:02:43.656 [I] k2/torch/bin/attention_rescore.cu:153:int main(int, char**) Device: cuda:0
+  2021-11-10 14:02:43.656 [I] k2/torch/bin/attention_rescore.cu:162:int main(int, char**) Load wave files
+  2021-11-10 14:02:49.216 [I] k2/torch/bin/attention_rescore.cu:169:int main(int, char**) Build Fbank computer
+  2021-11-10 14:02:49.217 [I] k2/torch/bin/attention_rescore.cu:180:int main(int, char**) Compute features
+  2021-11-10 14:02:49.222 [I] k2/torch/bin/attention_rescore.cu:188:int main(int, char**) Load neural network model
+  2021-11-10 14:02:49.984 [I] k2/torch/bin/attention_rescore.cu:203:int main(int, char**) Compute nnet_output
+  2021-11-10 14:02:50.624 [I] k2/torch/bin/attention_rescore.cu:220:int main(int, char**) Load ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lang_bpe_500/HLG.pt
+  2021-11-10 14:02:51.519 [I] k2/torch/bin/attention_rescore.cu:228:int main(int, char**) Decoding
+  2021-11-10 14:02:51.632 [I] k2/torch/bin/attention_rescore.cu:234:int main(int, char**) Load n-gram LM: ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/data/lm/G_4_gram.pt
+  2021-11-10 14:02:55.537 [W] k2/torch/csrc/deserialization.cu:441:k2::FsaClass k2::LoadFsa(const string&, c10::optional<c10::Device>) Ignore non tensor attribute: 'dummy' of type: Int
+  2021-11-10 14:02:55.645 [I] k2/torch/bin/attention_rescore.cu:243:int main(int, char**) Rescore with an n-gram LM
+  2021-11-10 14:02:55.970 [I] k2/torch/bin/attention_rescore.cu:246:int main(int, char**) Sample 100 paths
+  2021-11-10 14:02:56.215 [I] k2/torch/bin/attention_rescore.cu:293:int main(int, char**) Run attention decoder
+  2021-11-10 14:02:57.35 [I] k2/torch/bin/attention_rescore.cu:303:int main(int, char**) Rescoring
+  2021-11-10 14:02:57.179 [I] k2/torch/bin/attention_rescore.cu:369:int main(int, char**)
+  Decoding result:
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1089-134686-0001.wav
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0001.wav
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/test_wavs/1221-135766-0002.wav
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+There is a Colab notebook showing you how to run a torch scripted model in C++.
+Please see |librispeech asr conformer ctc torch script colab notebook|
+
+.. |librispeech asr conformer ctc torch script colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1BIGLWzS36isskMXHKcqC9ysN6pspYXs_?usp=sharing
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst
new file mode 100644
index 000000000..37edf7de9
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst
@@ -0,0 +1,223 @@
+Distillation with HuBERT
+========================
+
+This tutorial shows you how to perform knowledge distillation in `icefall <https://github.com/k2-fsa/icefall>`_
+with the `LibriSpeech`_ dataset. The distillation method
+used here is called "Multi Vector Quantization Knowledge Distillation" (MVQ-KD).
+Please have a look at our paper `Predicting Multi-Codebook Vector Quantization Indexes for Knowledge Distillation <https://arxiv.org/abs/2211.00508>`_
+for more details about MVQ-KD.
+
+.. note::
+
+    This tutorial is based on recipe
+    `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`_.
+    Currently, we only implement MVQ-KD in this recipe. However, MVQ-KD is theoretically applicable to all recipes
+    with only minor changes needed. Feel free to try out MVQ-KD in different recipes. If you
+    encounter any problems, please open an issue here `icefall <https://github.com/k2-fsa/icefall/issues>`__.
+
+.. note::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for `icefall`_.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+Data preparation
+----------------
+
+We first prepare necessary training data for `LibriSpeech`_.
+This is the same as in :ref:`non_streaming_librispeech_pruned_transducer_stateless`.
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to :ref:`codebook_index_preparation` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop_stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop_stage 0 # run only stage 0
+  $ ./prepare.sh --stage 2 --stop_stage 5 # run from stage 2 to stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech`_
+  dataset and the `musan`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+.. _codebook_index_preparation:
+
+Codebook index preparation
+--------------------------
+
+Here, we prepare necessary data for MVQ-KD. This requires the generation
+of codebook indexes (please read our `paper <https://arxiv.org/abs/2211.00508>`_.
+if you are interested in details). In this tutorial, we use the pre-computed
+codebook indexes for convenience. The only thing you need to do is to
+run `./distillation_with_hubert.sh <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/distillation_with_hubert.sh>`_.
+
+.. note::
+
+  There are 5 stages in total, the first and second stage will be automatically skipped
+  when choosing to downloaded codebook indexes prepared by `icefall`_.
+  Of course, you can extract and compute the codebook indexes by yourself. This
+  will require you downloading a HuBERT-XL model and it can take a while for
+  the extraction of codebook indexes.
+
+
+As usual, you can control the stages you want to run by specifying the following
+two options:
+
+  - ``--stage``
+  - ``--stop_stage``
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./distillation_with_hubert.sh --stage 0 --stop_stage 0 # run only stage 0
+  $ ./distillation_with_hubert.sh --stage 2 --stop_stage 4 # run from stage 2 to stage 5
+
+Here are a few options in `./distillation_with_hubert.sh <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/distillation_with_hubert.sh>`_
+you need to know before you proceed.
+
+- ``--full_libri`` If True, use full 960h data. Otherwise only ``train-clean-100`` will be used
+- ``--use_extracted_codebook`` If True, the first two stages will be skipped and the codebook
+  indexes uploaded by us will be downloaded.
+
+Since we are using the pre-computed codebook indexes, we set
+``use_extracted_codebook=True``. If you want to do full `LibriSpeech`_
+experiments, please set ``full_libri=True``.
+
+The following command downloads the pre-computed codebook indexes
+and prepares MVQ-augmented training manifests.
+
+.. code-block:: bash
+
+  $ ./distillation_with_hubert.sh --stage 2 --stop_stage 2 # run only stage 2
+
+Please see the
+following screenshot for the output of an example execution.
+
+.. figure:: ./images/distillation_codebook.png
+  :width: 800
+  :alt: Downloading codebook indexes and preparing training manifest.
+  :align: center
+
+  Downloading codebook indexes and preparing training manifest.
+
+.. hint::
+
+  The codebook indexes we prepared for you in this tutorial
+  are extracted from the 36-th layer of a fine-tuned HuBERT-XL model
+  with 8 codebooks. If you want to try other configurations, please
+  set ``use_extracted_codebook=False`` and set ``embedding_layer`` and
+  ``num_codebooks`` by yourself.
+
+Now, you should see the following files under the directory ``./data/vq_fbank_layer36_cb8``.
+
+.. figure:: ./images/distillation_directory.png
+  :width: 800
+  :alt: MVQ-augmented training manifests
+  :align: center
+
+  MVQ-augmented training manifests.
+
+Whola! You are ready to perform knowledge distillation training now!
+
+Training
+--------
+
+To perform training, please run stage 3 by executing the following command.
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 3 --stop_stage 3 # run MVQ training
+
+Here is the code snippet for training:
+
+.. code-block:: bash
+
+  WORLD_SIZE=$(echo ${CUDA_VISIBLE_DEVICES} | awk '{n=split($1, _, ","); print n}')
+
+  ./pruned_transducer_stateless6/train.py \
+    --manifest-dir ./data/vq_fbank_layer36_cb8 \
+    --master-port 12359 \
+    --full-libri $full_libri \
+    --spec-aug-time-warp-factor -1 \
+    --max-duration 300 \
+    --world-size ${WORLD_SIZE} \
+    --num-epochs 30 \
+    --exp-dir $exp_dir \
+    --enable-distillation True \
+    --codebook-loss-scale 0.01
+
+There are a few training arguments in the following
+training commands that should be paid attention to.
+
+  - ``--enable-distillation`` If True, knowledge distillation training is enabled.
+  - ``--codebook-loss-scale`` The scale of the knowledge distillation loss.
+  - ``--manifest-dir`` The path to the MVQ-augmented manifest.
+
+
+Decoding
+--------
+
+After training finished, you can test the performance on using
+the following command.
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES=0
+  ./pruned_transducer_stateless6/train.py \
+    --decoding-method "modified_beam_search" \
+    --epoch 30 \
+    --avg 10 \
+    --max-duration 200 \
+    --exp-dir $exp_dir \
+    --enable-distillation True
+
+You should get similar results as `here <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS-100hours.md#distillation-with-hubert>`__.
+
+That's all! Feel free to experiment with your own setups and report your results.
+If you encounter any problems during training, please open up an issue `here <https://github.com/k2-fsa/icefall/issues>`__.
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/images/distillation_codebook.png b/docs/source/recipes/Non-streaming-ASR/librispeech/images/distillation_codebook.png
new file mode 100644
index 000000000..1a40d6c6e
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/librispeech/images/distillation_codebook.png differ
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/images/distillation_directory.png b/docs/source/recipes/Non-streaming-ASR/librispeech/images/distillation_directory.png
new file mode 100644
index 000000000..30763046f
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/librispeech/images/distillation_directory.png differ
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/images/librispeech-conformer-ctc-tensorboard-log.png b/docs/source/recipes/Non-streaming-ASR/librispeech/images/librispeech-conformer-ctc-tensorboard-log.png
new file mode 100644
index 000000000..4e8c2ea7c
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/librispeech/images/librispeech-conformer-ctc-tensorboard-log.png differ
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/images/librispeech-pruned-transducer-tensorboard-log.jpg b/docs/source/recipes/Non-streaming-ASR/librispeech/images/librispeech-pruned-transducer-tensorboard-log.jpg
new file mode 100644
index 000000000..800835749
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/librispeech/images/librispeech-pruned-transducer-tensorboard-log.jpg differ
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/index.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/index.rst
new file mode 100644
index 000000000..bf439861a
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/index.rst
@@ -0,0 +1,12 @@
+LibriSpeech
+===========
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_lstm_ctc
+   conformer_ctc
+   pruned_transducer_stateless
+   zipformer_mmi
+   zipformer_ctc_blankskip
+   distillation
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst
new file mode 100644
index 000000000..f356e97e7
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst
@@ -0,0 +1,548 @@
+.. _non_streaming_librispeech_pruned_transducer_stateless:
+
+Pruned transducer statelessX
+============================
+
+This tutorial shows you how to run a conformer transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. Note::
+
+   The tutorial is suitable for `pruned_transducer_stateless <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless>`__,
+   `pruned_transducer_stateless2 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless2>`__,
+   `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`__,
+   `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless5>`__,
+   We will take pruned_transducer_stateless4 as an example in this tutorial.
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use a Conformer model (the reworked version by Daniel Povey)
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+
+Data preparation
+----------------
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to ``Training`` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/train.py --help
+
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The directory to save checkpoints, training logs and tensorboard.
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless4/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless4/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless4/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless4/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless4/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless4/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless4/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--use-fp16``
+
+    If it is True, the model will train with half precision, from our experiment
+    results, by using half precision you can train with two times larger ``--max-duration``
+    so as to get almost 2X speed up.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., number of encoder layers,
+encoder dimension, decoder dimension, number of warmup steps etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless4/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless4/train.py`` directly.
+
+
+.. NOTE::
+
+  The options for `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless5/train.py>`__ are a little different from
+  other recipes. It allows you to configure ``--num-encoder-layers``, ``--dim-feedforward``, ``--nhead``, ``--encoder-dim``, ``--decoder-dim``, ``--joiner-dim`` from commandline, so that you can train models with different size with pruned_transducer_stateless5.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``--exp-dir`` (e.g. ``pruned_transducer_stateless4/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless4/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "pruned transducer training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/QOGSPBgsR8KzcRMmie9JGw/
+
+        [2022-11-20T15:50:50] Started scanning logdir.
+        Uploading 4468 scalars...
+        [2022-11-20T15:53:02] Total uploaded: 210171 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output. Click it and you will see
+    the following screenshot:
+
+      .. figure:: images/librispeech-pruned-transducer-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/QOGSPBgsR8KzcRMmie9JGw/
+
+         TensorBoard screenshot.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless4/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 6 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5"
+  ./pruned_transducer_stateless4/train.py \
+     --world-size 6 \
+     --num-epochs 30 \
+     --start-epoch 1 \
+     --exp-dir pruned_transducer_stateless4/exp \
+     --full-libri 1 \
+     --max-duration 300
+
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/decode.py --help
+
+shows the options for decoding.
+
+The following shows two examples (for two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 25 20; do
+      for avg in 7 5 3 1; do
+        ./pruned_transducer_stateless4/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless4/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. Note::
+
+  Supporting decoding methods are as follows:
+
+    - ``greedy_search`` : It takes the symbol with largest posterior probability
+      of each frame as the decoding result.
+
+    - ``beam_search`` :  It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf and
+      `espnet/nets/beam_search_transducer.py <https://github.com/espnet/espnet/blob/master/espnet/nets/beam_search_transducer.py#L247>`_
+      is used as a reference. Basically, it keeps topk states for each frame, and expands the kept states with their own contexts to
+      next frame.
+
+    - ``modified_beam_search`` : It implements the same algorithm as ``beam_search`` above, but it
+      runs in batch mode with ``--max-sym-per-frame=1`` being hardcoded.
+
+    - ``fast_beam_search`` : It implements graph composition between the output ``log_probs`` and
+      given ``FSAs``. It is hard to describe the details in several lines of texts, you can read
+      our paper in https://arxiv.org/pdf/2211.00484.pdf or our `rnnt decode code in k2 <https://github.com/k2-fsa/k2/blob/master/k2/csrc/rnnt_decode.h>`_. ``fast_beam_search`` can decode with ``FSAs`` on GPU efficiently.
+
+    - ``fast_beam_search_LG`` : The same as ``fast_beam_search`` above, ``fast_beam_search`` uses
+      an trivial graph that has only one state, while ``fast_beam_search_LG`` uses an LG graph
+      (with N-gram LM).
+
+    - ``fast_beam_search_nbest`` : It produces the decoding results as follows:
+
+      - (1) Use ``fast_beam_search`` to get a lattice
+      - (2) Select ``num_paths`` paths from the lattice using ``k2.random_paths()``
+      - (3) Unique the selected paths
+      - (4) Intersect the selected paths with the lattice and compute the
+            shortest path from the intersection result
+      - (5) The path with the largest score is used as the decoding output.
+
+    - ``fast_beam_search_nbest_LG`` : It implements same logic as ``fast_beam_search_nbest``, the
+      only difference is that it uses ``fast_beam_search_LG`` to generate the lattice.
+
+
+Export Model
+------------
+
+`pruned_transducer_stateless4/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/export.py>`_ supports exporting checkpoints from ``pruned_transducer_stateless4/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless4/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --epoch 25 --avg 3 produces the smallest WER
+  # (You can get such information after running ./pruned_transducer_stateless4/decode.py)
+
+  epoch=25
+  avg=3
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch $epoch \
+    --avg  $avg
+
+It will generate a file ``./pruned_transducer_stateless4/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless4/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless4/exp
+      ln -s pretrained.pt epoch-999.pt
+
+   And then pass ``--epoch 999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless4/decode.py``.
+
+To use the exported model with ``./pruned_transducer_stateless4/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless4/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 25 \
+    --avg 3 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+.. NOTE::
+
+   You will need this ``cpu_jit.pt`` when deploying with Sherpa framework.
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `pruned_transducer_stateless <https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12>`__
+
+  - `pruned_transducer_stateless2 <https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless2-2022-04-29>`__
+
+  - `pruned_transducer_stateless4 <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless4-2022-06-03>`__
+
+  - `pruned_transducer_stateless5 <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-2022-07-07>`__
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+
+Deploy with Sherpa
+------------------
+
+Please see `<https://k2-fsa.github.io/sherpa/python/offline_asr/conformer/librispeech.html#>`_
+for how to deploy the models in ``sherpa``.
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/tdnn_lstm_ctc.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/tdnn_lstm_ctc.rst
new file mode 100644
index 000000000..aa380396a
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/tdnn_lstm_ctc.rst
@@ -0,0 +1,404 @@
+TDNN-LSTM-CTC
+=============
+
+This tutorial shows you how to run a TDNN-LSTM-CTC model with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+Now describing the training of TDNN-LSTM-CTC model, contained in
+the `tdnn_lstm_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/tdnn_lstm_ctc>`_
+folder.
+
+The command to run the training part is:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  $ ./tdnn_lstm_ctc/train.py --world-size 4
+
+By default, it will run ``20`` epochs. Training logs and checkpoints are saved
+in ``tdnn_lstm_ctc/exp``.
+
+In ``tdnn_lstm_ctc/exp``, you will find the following files:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-19.pt``
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./tdnn_lstm_ctc/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd tdnn_lstm_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "TDNN LSTM training for librispeech with icefall"
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+
+To see available training options, you can use:
+
+.. code-block:: bash
+
+  $ ./tdnn_lstm_ctc/train.py --help
+
+Other training options, e.g., learning rate, results dir, etc., are
+pre-configured in the function ``get_params()``
+in `tdnn_lstm_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/tdnn_lstm_ctc/train.py>`_.
+Normally, you don't need to change them. You can change them by modifying the code, if
+you want.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+The command for decoding is:
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./tdnn_lstm_ctc/decode.py
+
+You will see the WER in the output log.
+
+Decoded results are saved in ``tdnn_lstm_ctc/exp``.
+
+.. code-block:: bash
+
+  $ ./tdnn_lstm_ctc/decode.py --help
+
+shows you the available decoding options.
+
+Some commonly used options are:
+
+  - ``--epoch``
+
+    You can select which checkpoint to be used for decoding.
+    For instance, ``./tdnn_lstm_ctc/decode.py --epoch 10`` means to use
+    ``./tdnn_lstm_ctc/exp/epoch-10.pt`` for decoding.
+
+  - ``--avg``
+
+    It's related to model averaging. It specifies number of checkpoints
+    to be averaged. The averaged model is used for decoding.
+    For example, the following command:
+
+      .. code-block:: bash
+
+        $ ./tdnn_lstm_ctc/decode.py --epoch 10 --avg 3
+
+    uses the average of ``epoch-8.pt``, ``epoch-9.pt`` and ``epoch-10.pt``
+    for decoding.
+
+  - ``--export``
+
+    If it is ``True``, i.e., ``./tdnn_lstm_ctc/decode.py --export 1``, the code
+    will save the averaged model to ``tdnn_lstm_ctc/exp/pretrained.pt``.
+    See :ref:`tdnn_lstm_ctc use a pre-trained model` for how to use it.
+
+
+.. _tdnn_lstm_ctc use a pre-trained model:
+
+Pre-trained Model
+-----------------
+
+We have uploaded the pre-trained model to
+`<https://huggingface.co/pkufool/icefall_asr_librispeech_tdnn-lstm_ctc>`_.
+
+The following shows you how to use the pre-trained model.
+
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/pkufool/icefall_asr_librispeech_tdnn-lstm_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ tree tmp
+
+.. code-block:: bash
+
+  tmp/
+  `-- icefall_asr_librispeech_tdnn-lstm_ctc
+      |-- README.md
+      |-- data
+      |   |-- lang_phone
+      |   |   |-- HLG.pt
+      |   |   |-- tokens.txt
+      |   |   `-- words.txt
+      |   `-- lm
+      |       `-- G_4_gram.pt
+      |-- exp
+      |   `-- pretrained.pt
+      `-- test_wavs
+          |-- 1089-134686-0001.flac
+          |-- 1221-135766-0001.flac
+          |-- 1221-135766-0002.flac
+          `-- trans.txt
+
+  6 directories, 10 files
+
+**File descriptions**:
+
+  - ``data/lang_phone/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_phone/tokens.txt``
+
+      It contains tokens and their IDs.
+
+  - ``data/lang_phone/words.txt``
+
+      It contains words and their IDs.
+
+  - ``data/lm/G_4_gram.pt``
+
+      It is a 4-gram LM, useful for LM rescoring.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-14.pt`` to ``epoch-19.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.flac``
+
+      It contains some test sound files from LibriSpeech ``test-clean`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in ``test_waves/``.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/*.flac
+
+  Input File     : 'tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:06.62 = 106000 samples ~ 496.875 CDDA sectors
+  File Size      : 116k
+  Bit Rate       : 140k
+  Sample Encoding: 16-bit FLAC
+
+
+  Input File     : 'tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:16.71 = 267440 samples ~ 1253.62 CDDA sectors
+  File Size      : 343k
+  Bit Rate       : 164k
+  Sample Encoding: 16-bit FLAC
+
+
+  Input File     : 'tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.83 = 77200 samples ~ 361.875 CDDA sectors
+  File Size      : 105k
+  Bit Rate       : 174k
+  Sample Encoding: 16-bit FLAC
+
+  Total Duration of 3 files: 00:00:28.16
+
+
+Inference with a pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./tdnn_lstm_ctc/pretrained.py --help
+
+shows the usage information of ``./tdnn_lstm_ctc/pretrained.py``.
+
+To decode with ``1best`` method, we can use:
+
+.. code-block:: bash
+
+  ./tdnn_lstm_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/exp/pretraind.pt \
+    --words-file ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/words.txt \
+    --HLG ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt \
+    ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac \
+    ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac \
+    ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac
+
+The output is:
+
+.. code-block::
+
+  2021-08-24 16:57:13,315 INFO [pretrained.py:168] device: cuda:0
+  2021-08-24 16:57:13,315 INFO [pretrained.py:170] Creating model
+  2021-08-24 16:57:18,331 INFO [pretrained.py:182] Loading HLG from ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt
+  2021-08-24 16:57:27,581 INFO [pretrained.py:199] Constructing Fbank computer
+  2021-08-24 16:57:27,584 INFO [pretrained.py:209] Reading sound files: ['./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac']
+  2021-08-24 16:57:27,599 INFO [pretrained.py:215] Decoding started
+  2021-08-24 16:57:27,791 INFO [pretrained.py:245] Use HLG decoding
+  2021-08-24 16:57:28,098 INFO [pretrained.py:266]
+  ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac:
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac:
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac:
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+
+  2021-08-24 16:57:28,099 INFO [pretrained.py:268] Decoding Done
+
+
+To decode with ``whole-lattice-rescoring`` methond, you can use
+
+.. code-block:: bash
+
+  ./tdnn_lstm_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/exp/pretraind.pt \
+    --words-file ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/words.txt \
+    --HLG ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt \
+    --method whole-lattice-rescoring \
+    --G ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lm/G_4_gram.pt \
+    --ngram-lm-scale 0.8 \
+    ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac \
+    ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac \
+    ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac
+
+The decoding output is:
+
+.. code-block::
+
+  2021-08-24 16:39:24,725 INFO [pretrained.py:168] device: cuda:0
+  2021-08-24 16:39:24,725 INFO [pretrained.py:170] Creating model
+  2021-08-24 16:39:29,403 INFO [pretrained.py:182] Loading HLG from ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt
+  2021-08-24 16:39:40,631 INFO [pretrained.py:190] Loading G from ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lm/G_4_gram.pt
+  2021-08-24 16:39:53,098 INFO [pretrained.py:199] Constructing Fbank computer
+  2021-08-24 16:39:53,107 INFO [pretrained.py:209] Reading sound files: ['./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac']
+  2021-08-24 16:39:53,121 INFO [pretrained.py:215] Decoding started
+  2021-08-24 16:39:53,443 INFO [pretrained.py:250] Use HLG decoding + LM rescoring
+  2021-08-24 16:39:54,010 INFO [pretrained.py:266]
+  ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac:
+  AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
+
+  ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac:
+  GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
+
+  ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac:
+  YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
+
+
+  2021-08-24 16:39:54,010 INFO [pretrained.py:268] Decoding Done
+
+
+Colab notebook
+--------------
+
+We provide a colab notebook for decoding with pre-trained model.
+
+|librispeech tdnn_lstm_ctc colab notebook|
+
+.. |librispeech tdnn_lstm_ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1-iSfQMp2So-We_Uu49N4AAcMInB72u9z?usp=sharing
+
+
+**Congratulations!** You have finished the TDNN-LSTM-CTC recipe on librispeech in ``icefall``.
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.rst
new file mode 100644
index 000000000..aa73bfe33
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.rst
@@ -0,0 +1,454 @@
+Zipformer CTC Blank Skip
+========================
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+This tutorial shows you how to train a Zipformer model based on the guidance from 
+a co-trained CTC model using `blank skip method <https://arxiv.org/pdf/2210.16481.pdf>`_
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. note::
+
+    We use both CTC and RNN-T loss to train. During the forward pass, the encoder output
+    is first used to calculate the CTC posterior probability; then for each output frame,
+    if its blank posterior is bigger than some threshold, it will be simply discarded
+    from the encoder output. To prevent information loss, we also put a convolution module
+    similar to the one used in conformer (referred to as “LConv”) before the frame reduction.
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+.. note::
+
+   We encourage you to read ``./prepare.sh``.
+
+The data preparation contains several stages. You can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. hint::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. note::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+For stability, it doesn`t use blank skip method until model warm-up.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_ctc_bs/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless7_ctc_bs/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless7_ctc_bs/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless7_ctc_bs/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless7_ctc_bs/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless7_ctc_bs/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless7_ctc_bs/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless7_ctc_bs/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless7_ctc_bs/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless7_ctc_bs/train.py`` directly.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``pruned_transducer_stateless7_ctc_bs/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_ctc_bs/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_ctc_bs/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless7_ctc_bs/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Zipformer-CTC co-training using blank skip for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/
+
+    Note there is a URL in the above output. Click it and you will see
+    tensorboard.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless7_ctc_bs/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./pruned_transducer_stateless7_ctc_bs/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --full-libri 1 \
+    --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --use-fp16 1
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py --help
+
+shows the options for decoding.
+
+The following shows the example using ``epoch-*.pt``:
+
+.. code-block:: bash
+
+    for m in greedy_search fast_beam_search modified_beam_search; do
+        ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+            --epoch 30 \
+            --avg 13 \
+            --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+            --max-duration 600 \
+            --decoding-method $m
+    done
+
+To test CTC branch, you can use the following command:
+
+.. code-block:: bash
+
+    for m in ctc-decoding 1best; do
+        ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+            --epoch 30 \
+            --avg 13 \
+            --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+            --max-duration 600 \
+            --decoding-method $m
+    done
+
+Export models
+-------------
+
+`pruned_transducer_stateless7_ctc_bs/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export.py>`_ supports exporting checkpoints from ``pruned_transducer_stateless7_ctc_bs/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless7_ctc_bs/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/export.py \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 13 \
+    --jit 0
+
+It will generate a file ``./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless7_ctc_bs/exp
+      ln -s pretrained epoch-9999.pt
+
+   And then pass ``--epoch 9999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py``.
+
+To use the exported model with ``./pruned_transducer_stateless7_ctc_bs/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+To test CTC branch using the exported model with ``./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --method ctc-decoding \
+    --sample-rate 16000 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/export.py \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 13 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+To use the generated files with ``./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py \
+    --nn-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+To test CTC branch using the generated files with ``./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+    --model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --method ctc-decoding \
+    --sample-rate 16000 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - trained on LibriSpeech 100h: `<https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2022-12-14>`_
+  - trained on LibriSpeech 960h: `<https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_mmi.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_mmi.rst
new file mode 100644
index 000000000..a7b59a992
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_mmi.rst
@@ -0,0 +1,422 @@
+Zipformer MMI
+===============
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+This tutorial shows you how to train an Zipformer MMI model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+We use LF-MMI to compute the loss.
+
+.. note::
+
+   You can find the document about LF-MMI training at the following address:
+
+   `<https://github.com/k2-fsa/next-gen-kaldi-wechat/blob/master/pdf/LF-MMI-training-and-decoding-in-k2-Part-I.pdf>`_
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+.. note::
+
+   We encourage you to read ``./prepare.sh``.
+
+The data preparation contains several stages. You can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. hint::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. note::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+For stability, it uses CTC loss for model warm-up and then switches to MMI loss.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./zipformer_mmi/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./zipformer_mmi/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./zipformer_mmi/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./zipformer_mmi/train.py --start-epoch 10`` loads the
+    checkpoint ``./zipformer_mmi/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./zipformer_mmi/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./zipformer_mmi/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./zipformer_mmi/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`zipformer_mmi/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer_mmi/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./zipformer_mmi/train.py`` directly.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``zipformer_mmi/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./zipformer_mmi/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./zipformer_mmi/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd zipformer_mmi/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Zipformer MMI training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/
+
+    Note there is a URL in the above output. Click it and you will see
+    tensorboard.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd zipformer_mmi/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./zipformer_mmi/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --full-libri 1 \
+    --exp-dir zipformer_mmi/exp \
+    --max-duration 500 \
+    --use-fp16 1 \
+    --num-workers 2
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``zipformer_mmi/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``zipformer_mmi/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./zipformer_mmi/decode.py --help
+
+shows the options for decoding.
+
+The following shows the example using ``epoch-*.pt``:
+
+.. code-block:: bash
+
+  for m in nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+    ./zipformer_mmi/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir ./zipformer_mmi/exp/ \
+      --max-duration 100 \
+      --lang-dir data/lang_bpe_500 \
+      --nbest-scale 1.2 \
+      --hp-scale 1.0 \
+      --decoding-method $m
+  done
+
+
+Export models
+-------------
+
+`zipformer_mmi/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer_mmi/export.py>`_ supports exporting checkpoints from ``zipformer_mmi/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``zipformer_mmi/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  ./zipformer_mmi/export.py \
+    --exp-dir ./zipformer_mmi/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --jit 0
+
+It will generate a file ``./zipformer_mmi/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``zipformer_mmi/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd zipformer_mmi/exp
+      ln -s pretrained epoch-9999.pt
+
+   And then pass ``--epoch 9999 --avg 1 --use-averaged-model 0`` to
+   ``./zipformer_mmi/decode.py``.
+
+To use the exported model with ``./zipformer_mmi/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./zipformer_mmi/export.py \
+    --exp-dir ./zipformer_mmi/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+To use the generated files with ``./zipformer_mmi/jit_pretrained.py``:
+
+.. code-block:: bash
+
+  ./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
diff --git a/docs/source/recipes/Non-streaming-ASR/timit/index.rst b/docs/source/recipes/Non-streaming-ASR/timit/index.rst
new file mode 100644
index 000000000..5ee147be7
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/timit/index.rst
@@ -0,0 +1,8 @@
+TIMIT
+=====
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_ligru_ctc
+   tdnn_lstm_ctc
diff --git a/docs/source/recipes/Non-streaming-ASR/timit/tdnn_ligru_ctc.rst b/docs/source/recipes/Non-streaming-ASR/timit/tdnn_ligru_ctc.rst
new file mode 100644
index 000000000..3d7aefe02
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/timit/tdnn_ligru_ctc.rst
@@ -0,0 +1,406 @@
+TDNN-LiGRU-CTC
+==============
+
+This tutorial shows you how to run a TDNN-LiGRU-CTC model with the `TIMIT <https://data.deepai.org/timit.zip>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+
+Training
+--------
+
+Now describing the training of TDNN-LiGRU-CTC model, contained in
+the `tdnn_ligru_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/timit/ASR/tdnn_ligru_ctc>`_
+folder.
+
+.. HINT::
+
+  TIMIT is a very small dataset. So one GPU is enough.
+
+The command to run the training part is:
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./tdnn_ligru_ctc/train.py
+
+By default, it will run ``25`` epochs. Training logs and checkpoints are saved
+in ``tdnn_ligru_ctc/exp``.
+
+In ``tdnn_ligru_ctc/exp``, you will find the following files:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./tdnn_ligru_ctc/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd tdnn_ligru_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "TDNN ligru training for timit with icefall"
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+
+To see available training options, you can use:
+
+.. code-block:: bash
+
+  $ ./tdnn_ligru_ctc/train.py --help
+
+Other training options, e.g., learning rate, results dir, etc., are
+pre-configured in the function ``get_params()``
+in `tdnn_ligru_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/timit/ASR/tdnn_ligru_ctc/train.py>`_.
+Normally, you don't need to change them. You can change them by modifying the code, if
+you want.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+The command for decoding is:
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./tdnn_ligru_ctc/decode.py
+
+You will see the WER in the output log.
+
+Decoded results are saved in ``tdnn_ligru_ctc/exp``.
+
+.. code-block:: bash
+
+  $ ./tdnn_ligru_ctc/decode.py --help
+
+shows you the available decoding options.
+
+Some commonly used options are:
+
+  - ``--epoch``
+
+    You can select which checkpoint to be used for decoding.
+    For instance, ``./tdnn_ligru_ctc/decode.py --epoch 10`` means to use
+    ``./tdnn_ligru_ctc/exp/epoch-10.pt`` for decoding.
+
+  - ``--avg``
+
+    It's related to model averaging. It specifies number of checkpoints
+    to be averaged. The averaged model is used for decoding.
+    For example, the following command:
+
+      .. code-block:: bash
+
+        $ ./tdnn_ligru_ctc/decode.py --epoch 25 --avg 17
+
+    uses the average of ``epoch-9.pt``, ``epoch-10.pt``, ``epoch-11.pt``,
+    ``epoch-12.pt``, ``epoch-13.pt``, ``epoch-14.pt``, ``epoch-15.pt``,
+    ``epoch-16.pt``, ``epoch-17.pt``, ``epoch-18.pt``, ``epoch-19.pt``,
+    ``epoch-20.pt``, ``epoch-21.pt``, ``epoch-22.pt``, ``epoch-23.pt``,
+    ``epoch-24.pt`` and ``epoch-25.pt``
+    for decoding.
+
+  - ``--export``
+
+    If it is ``True``, i.e., ``./tdnn_ligru_ctc/decode.py --export 1``, the code
+    will save the averaged model to ``tdnn_ligru_ctc/exp/pretrained.pt``.
+    See :ref:`tdnn_ligru_ctc use a pre-trained model` for how to use it.
+
+
+.. _tdnn_ligru_ctc use a pre-trained model:
+
+Pre-trained Model
+-----------------
+
+We have uploaded the pre-trained model to
+`<https://huggingface.co/luomingshuang/icefall_asr_timit_tdnn_ligru_ctc>`_.
+
+The following shows you how to use the pre-trained model.
+
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ mkdir tmp-ligru
+  $ cd tmp-ligru
+  $ git lfs install
+  $ git clone https://huggingface.co/luomingshuang/icefall_asr_timit_tdnn_ligru_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ tree tmp-ligru
+
+.. code-block:: bash
+
+  tmp-ligru/
+  `-- icefall_asr_timit_tdnn_ligru_ctc
+      |-- README.md
+      |-- data
+      |   |-- lang_phone
+      |   |   |-- HLG.pt
+      |   |   |-- tokens.txt
+      |   |   `-- words.txt
+      |   `-- lm
+      |       `-- G_4_gram.pt
+      |-- exp
+      |   `-- pretrained_average_9_25.pt
+      `-- test_wavs
+          |-- FDHC0_SI1559.WAV
+          |-- FELC0_SI756.WAV
+          |-- FMGD0_SI1564.WAV
+          `-- trans.txt
+
+  6 directories, 10 files
+
+**File descriptions**:
+
+  - ``data/lang_phone/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_phone/tokens.txt``
+
+      It contains tokens and their IDs.
+
+  - ``data/lang_phone/words.txt``
+
+      It contains words and their IDs.
+
+  - ``data/lm/G_4_gram.pt``
+
+      It is a 4-gram LM, useful for LM rescoring.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-9.pt`` to ``epoch-25.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.WAV``
+
+      It contains some test sound files from timit ``TEST`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in ``test_waves/``.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ ffprobe -show_format tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV
+
+  Input #0, nistsphere, from 'tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV':
+  Metadata:
+    database_id     : TIMIT
+    database_version: 1.0
+    utterance_id    : dhc0_si1559
+    sample_min      : -4176
+    sample_max      : 5984
+  Duration: 00:00:03.40, bitrate: 258 kb/s
+    Stream #0:0: Audio: pcm_s16le, 16000 Hz, 1 channels, s16, 256 kb/s
+
+  $ ffprobe -show_format tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV
+
+  Input #0, nistsphere, from 'tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV':
+  Metadata:
+    database_id     : TIMIT
+    database_version: 1.0
+    utterance_id    : elc0_si756
+    sample_min      : -1546
+    sample_max      : 1989
+  Duration: 00:00:04.19, bitrate: 257 kb/s
+    Stream #0:0: Audio: pcm_s16le, 16000 Hz, 1 channels, s16, 256 kb/s
+
+  $ ffprobe -show_format tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV
+
+  Input #0, nistsphere, from 'tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV':
+  Metadata:
+    database_id     : TIMIT
+    database_version: 1.0
+    utterance_id    : mgd0_si1564
+    sample_min      : -7626
+    sample_max      : 10573
+  Duration: 00:00:04.44, bitrate: 257 kb/s
+    Stream #0:0: Audio: pcm_s16le, 16000 Hz, 1 channels, s16, 256 kb/s
+
+
+Inference with a pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ ./tdnn_ligru_ctc/pretrained.py --help
+
+shows the usage information of ``./tdnn_ligru_ctc/pretrained.py``.
+
+To decode with ``1best`` method, we can use:
+
+.. code-block:: bash
+
+  ./tdnn_ligru_ctc/pretrained.py
+    --method 1best
+    --checkpoint ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/exp/pretrained_average_9_25.pt
+    --words-file ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lang_phone/words.txt
+    --HLG ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lang_phone/HLG.pt
+    ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV
+    ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV
+    ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV
+
+The output is:
+
+.. code-block::
+
+  2021-11-08 20:41:33,660 INFO [pretrained.py:169] device: cuda:0
+  2021-11-08 20:41:33,660 INFO [pretrained.py:171] Creating model
+  2021-11-08 20:41:38,680 INFO [pretrained.py:183] Loading HLG from ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lang_phone/HLG.pt
+  2021-11-08 20:41:38,695 INFO [pretrained.py:200] Constructing Fbank computer
+  2021-11-08 20:41:38,697 INFO [pretrained.py:210] Reading sound files: ['./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV', './tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV', './tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV']
+  2021-11-08 20:41:38,704 INFO [pretrained.py:216] Decoding started
+  2021-11-08 20:41:39,819 INFO [pretrained.py:246] Use HLG decoding
+  2021-11-08 20:41:39,829 INFO [pretrained.py:267]
+  ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV:
+  sil dh ih sh uw ah l iy v iy z ih sil p r aa sil k s ih m ey dx ih sil d w uh dx ih w ih s f iy l ih ng w ih th ih n ih m s eh l f sil jh
+
+  ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV:
+  sil m ih sil t ih r iy s sil s er r ih m ih sil m aa l ih sil k l ey sil r eh sil d w ay sil d aa r sil b ah f sil jh
+
+  ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV:
+  sil hh ah z sil b ih sil g r iy w ah z sil d aw n ih sil b ay s sil n ey sil w eh l f eh n s ih z eh n dh eh r w er sil g r ey z ih ng sil k ae dx l sil
+
+
+  2021-11-08 20:41:39,829 INFO [pretrained.py:269] Decoding Done
+
+
+To decode with ``whole-lattice-rescoring`` methond, you can use
+
+.. code-block:: bash
+
+  ./tdnn_ligru_ctc/pretrained.py \
+    --method whole-lattice-rescoring \
+    --checkpoint ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/exp/pretrained_average_9_25.pt \
+    --words-file ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lang_phone/words.txt \
+    --HLG ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lang_phone/HLG.pt \
+    --G ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lm/G_4_gram.pt \
+    --ngram-lm-scale 0.1 \
+    ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV
+    ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV
+    ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV
+
+The decoding output is:
+
+.. code-block::
+
+  2021-11-08 20:37:50,693 INFO [pretrained.py:169] device: cuda:0
+  2021-11-08 20:37:50,693 INFO [pretrained.py:171] Creating model
+  2021-11-08 20:37:54,693 INFO [pretrained.py:183] Loading HLG from ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lang_phone/HLG.pt
+  2021-11-08 20:37:54,705 INFO [pretrained.py:191] Loading G from ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/data/lm/G_4_gram.pt
+  2021-11-08 20:37:54,714 INFO [pretrained.py:200] Constructing Fbank computer
+  2021-11-08 20:37:54,715 INFO [pretrained.py:210] Reading sound files: ['./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV', './tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV', './tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV']
+  2021-11-08 20:37:54,720 INFO [pretrained.py:216] Decoding started
+  2021-11-08 20:37:55,808 INFO [pretrained.py:251] Use HLG decoding + LM rescoring
+  2021-11-08 20:37:56,348 INFO [pretrained.py:267]
+  ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FDHC0_SI1559.WAV:
+  sil dh ih sh uw ah l iy v iy z ah sil p r aa sil k s ih m ey dx ih sil d w uh dx iy w ih s f iy l iy ng w ih th ih n ih m s eh l f sil jh
+
+  ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FELC0_SI756.WAV:
+  sil m ih sil t ih r iy l s sil s er r eh m ih sil m aa l ih ng sil k l ey sil r eh sil d w ay sil d aa r sil b ah f sil jh ch
+
+  ./tmp-ligru/icefall_asr_timit_tdnn_ligru_ctc/test_waves/FMGD0_SI1564.WAV:
+  sil hh ah z sil b ih n sil g r iy w ah z sil b aw n ih sil b ay s sil n ey sil w er l f eh n s ih z eh n dh eh r w er sil g r ey z ih ng sil k ae dx l sil
+
+
+  2021-11-08 20:37:56,348 INFO [pretrained.py:269] Decoding Done
+
+
+Colab notebook
+--------------
+
+We provide a colab notebook for decoding with pre-trained model.
+
+|timit tdnn_ligru_ctc colab notebook|
+
+.. |timit tdnn_ligru_ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/11IT-k4HQIgQngXz1uvWsEYktjqQt7Tmb
+
+
+**Congratulations!** You have finished the TDNN-LiGRU-CTC recipe on timit in ``icefall``.
diff --git a/docs/source/recipes/Non-streaming-ASR/timit/tdnn_lstm_ctc.rst b/docs/source/recipes/Non-streaming-ASR/timit/tdnn_lstm_ctc.rst
new file mode 100644
index 000000000..ee67a6edc
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/timit/tdnn_lstm_ctc.rst
@@ -0,0 +1,404 @@
+TDNN-LSTM-CTC
+=============
+
+This tutorial shows you how to run a TDNN-LSTM-CTC model with the `TIMIT <https://data.deepai.org/timit.zip>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+
+Training
+--------
+
+Now describing the training of TDNN-LSTM-CTC model, contained in
+the `tdnn_lstm_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/timit/ASR/tdnn_lstm_ctc>`_
+folder.
+
+.. HINT::
+
+  TIMIT is a very small dataset. So one GPU for training is enough.
+
+The command to run the training part is:
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./tdnn_lstm_ctc/train.py
+
+By default, it will run ``25`` epochs. Training logs and checkpoints are saved
+in ``tdnn_lstm_ctc/exp``.
+
+In ``tdnn_lstm_ctc/exp``, you will find the following files:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./tdnn_lstm_ctc/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd tdnn_lstm_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "TDNN LSTM training for timit with icefall"
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+
+To see available training options, you can use:
+
+.. code-block:: bash
+
+  $ ./tdnn_lstm_ctc/train.py --help
+
+Other training options, e.g., learning rate, results dir, etc., are
+pre-configured in the function ``get_params()``
+in `tdnn_lstm_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/timit/ASR/tdnn_lstm_ctc/train.py>`_.
+Normally, you don't need to change them. You can change them by modifying the code, if
+you want.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+The command for decoding is:
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./tdnn_lstm_ctc/decode.py
+
+You will see the WER in the output log.
+
+Decoded results are saved in ``tdnn_lstm_ctc/exp``.
+
+.. code-block:: bash
+
+  $ ./tdnn_lstm_ctc/decode.py --help
+
+shows you the available decoding options.
+
+Some commonly used options are:
+
+  - ``--epoch``
+
+    You can select which checkpoint to be used for decoding.
+    For instance, ``./tdnn_lstm_ctc/decode.py --epoch 10`` means to use
+    ``./tdnn_lstm_ctc/exp/epoch-10.pt`` for decoding.
+
+  - ``--avg``
+
+    It's related to model averaging. It specifies number of checkpoints
+    to be averaged. The averaged model is used for decoding.
+    For example, the following command:
+
+      .. code-block:: bash
+
+        $ ./tdnn_lstm_ctc/decode.py --epoch 25 --avg 10
+
+    uses the average of ``epoch-16.pt``, ``epoch-17.pt``, ``epoch-18.pt``,
+    ``epoch-19.pt``, ``epoch-20.pt``, ``epoch-21.pt``, ``epoch-22.pt``,
+    ``epoch-23.pt``, ``epoch-24.pt`` and ``epoch-25.pt``
+    for decoding.
+
+  - ``--export``
+
+    If it is ``True``, i.e., ``./tdnn_lstm_ctc/decode.py --export 1``, the code
+    will save the averaged model to ``tdnn_lstm_ctc/exp/pretrained.pt``.
+    See :ref:`tdnn_lstm_ctc use a pre-trained model` for how to use it.
+
+
+.. _tdnn_lstm_ctc use a pre-trained model:
+
+Pre-trained Model
+-----------------
+
+We have uploaded the pre-trained model to
+`<https://huggingface.co/luomingshuang/icefall_asr_timit_tdnn_lstm_ctc>`_.
+
+The following shows you how to use the pre-trained model.
+
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ mkdir tmp-lstm
+  $ cd tmp-lstm
+  $ git lfs install
+  $ git clone https://huggingface.co/luomingshuang/icefall_asr_timit_tdnn_lstm_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ tree tmp-lstm
+
+.. code-block:: bash
+
+  tmp-lstm/
+  `-- icefall_asr_timit_tdnn_lstm_ctc
+      |-- README.md
+      |-- data
+      |   |-- lang_phone
+      |   |   |-- HLG.pt
+      |   |   |-- tokens.txt
+      |   |   `-- words.txt
+      |   `-- lm
+      |       `-- G_4_gram.pt
+      |-- exp
+      |   `-- pretrained_average_16_25.pt
+      `-- test_wavs
+          |-- FDHC0_SI1559.WAV
+          |-- FELC0_SI756.WAV
+          |-- FMGD0_SI1564.WAV
+          `-- trans.txt
+
+  6 directories, 10 files
+
+**File descriptions**:
+
+  - ``data/lang_phone/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_phone/tokens.txt``
+
+      It contains tokens and their IDs.
+
+  - ``data/lang_phone/words.txt``
+
+      It contains words and their IDs.
+
+  - ``data/lm/G_4_gram.pt``
+
+      It is a 4-gram LM, useful for LM rescoring.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-16.pt`` to ``epoch-25.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.WAV``
+
+      It contains some test sound files from timit ``TEST`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in ``test_waves/``.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ ffprobe -show_format tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV
+
+  Input #0, nistsphere, from 'tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV':
+  Metadata:
+    database_id     : TIMIT
+    database_version: 1.0
+    utterance_id    : dhc0_si1559
+    sample_min      : -4176
+    sample_max      : 5984
+  Duration: 00:00:03.40, bitrate: 258 kb/s
+    Stream #0:0: Audio: pcm_s16le, 16000 Hz, 1 channels, s16, 256 kb/s
+
+  $ ffprobe -show_format tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV
+
+  Input #0, nistsphere, from 'tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV':
+  Metadata:
+    database_id     : TIMIT
+    database_version: 1.0
+    utterance_id    : elc0_si756
+    sample_min      : -1546
+    sample_max      : 1989
+  Duration: 00:00:04.19, bitrate: 257 kb/s
+    Stream #0:0: Audio: pcm_s16le, 16000 Hz, 1 channels, s16, 256 kb/s
+
+  $ ffprobe -show_format tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV
+
+  Input #0, nistsphere, from 'tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV':
+  Metadata:
+    database_id     : TIMIT
+    database_version: 1.0
+    utterance_id    : mgd0_si1564
+    sample_min      : -7626
+    sample_max      : 10573
+  Duration: 00:00:04.44, bitrate: 257 kb/s
+    Stream #0:0: Audio: pcm_s16le, 16000 Hz, 1 channels, s16, 256 kb/s
+
+
+Inference with a pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/timit/ASR
+  $ ./tdnn_lstm_ctc/pretrained.py --help
+
+shows the usage information of ``./tdnn_lstm_ctc/pretrained.py``.
+
+To decode with ``1best`` method, we can use:
+
+.. code-block:: bash
+
+  ./tdnn_lstm_ctc/pretrained.py
+    --method 1best
+    --checkpoint ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/exp/pretrained_average_16_25.pt
+    --words-file ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lang_phone/words.txt
+    --HLG ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lang_phone/HLG.pt
+    ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV
+    ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV
+    ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV
+
+The output is:
+
+.. code-block::
+
+  2021-11-08 21:02:49,583 INFO [pretrained.py:169] device: cuda:0
+  2021-11-08 21:02:49,584 INFO [pretrained.py:171] Creating model
+  2021-11-08 21:02:53,816 INFO [pretrained.py:183] Loading HLG from ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lang_phone/HLG.pt
+  2021-11-08 21:02:53,827 INFO [pretrained.py:200] Constructing Fbank computer
+  2021-11-08 21:02:53,827 INFO [pretrained.py:210] Reading sound files: ['./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV', './tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV', './tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV']
+  2021-11-08 21:02:53,831 INFO [pretrained.py:216] Decoding started
+  2021-11-08 21:02:54,380 INFO [pretrained.py:246] Use HLG decoding
+  2021-11-08 21:02:54,387 INFO [pretrained.py:267]
+  ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV:
+  sil dh ih sh uw ah l iy v iy z ih sil p r aa sil k s ih m ey dx ih sil d w uh dx iy w ih s f iy l iy w ih th ih n ih m s eh l f sil jh
+
+  ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV:
+  sil dh ih sil t ih r ih s sil s er r ih m ih sil m aa l ih ng sil k l ey sil r eh sil d w ay sil d aa r sil b ah f sil <UNK> jh
+
+  ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV:
+  sil hh ae z sil b ih n iy w ah z sil b ae n ih sil b ay s sil n ey sil k eh l f eh n s ih z eh n dh eh r w er sil g r ey z ih ng sil k ae dx l sil
+
+
+  2021-11-08 21:02:54,387 INFO [pretrained.py:269] Decoding Done
+
+
+To decode with ``whole-lattice-rescoring`` methond, you can use
+
+.. code-block:: bash
+
+  ./tdnn_lstm_ctc/pretrained.py \
+    --method whole-lattice-rescoring \
+    --checkpoint ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/exp/pretrained_average_16_25.pt \
+    --words-file ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lang_phone/words.txt \
+    --HLG ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lang_phone/HLG.pt \
+    --G ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lm/G_4_gram.pt \
+    --ngram-lm-scale 0.08 \
+    ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV
+    ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV
+    ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV
+
+The decoding output is:
+
+.. code-block::
+
+  2021-11-08 20:05:22,739 INFO [pretrained.py:169] device: cuda:0
+  2021-11-08 20:05:22,739 INFO [pretrained.py:171] Creating model
+  2021-11-08 20:05:26,959 INFO [pretrained.py:183] Loading HLG from ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lang_phone/HLG.pt
+  2021-11-08 20:05:26,971 INFO [pretrained.py:191] Loading G from ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/data/lm/G_4_gram.pt
+  2021-11-08 20:05:26,977 INFO [pretrained.py:200] Constructing Fbank computer
+  2021-11-08 20:05:26,978 INFO [pretrained.py:210] Reading sound files: ['./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV', './tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV', './tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV']
+  2021-11-08 20:05:26,981 INFO [pretrained.py:216] Decoding started
+  2021-11-08 20:05:27,519 INFO [pretrained.py:251] Use HLG decoding + LM rescoring
+  2021-11-08 20:05:27,878 INFO [pretrained.py:267]
+  ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FDHC0_SI1559.WAV:
+  sil dh ih sh uw l iy v iy z ih sil p r aa sil k s ah m ey dx ih sil w uh dx iy w ih s f iy l ih ng w ih th ih n ih m s eh l f sil jh
+
+  ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FELC0_SI756.WAV:
+  sil dh ih sil t ih r iy ih s sil s er r eh m ih sil n ah l ih ng sil k l ey sil r eh sil d w ay sil d aa r sil b ow f sil jh
+
+  ./tmp-lstm/icefall_asr_timit_tdnn_lstm_ctc/test_waves/FMGD0_SI1564.WAV:
+  sil hh ah z sil b ih n iy w ah z sil b ae n ih sil b ay s sil n ey sil k ih l f eh n s ih z eh n dh eh r w er sil g r ey z ih n sil k ae dx l sil
+
+
+2021-11-08 20:05:27,878 INFO [pretrained.py:269] Decoding Done
+
+
+Colab notebook
+--------------
+
+We provide a colab notebook for decoding with pre-trained model.
+
+|timit tdnn_lstm_ctc colab notebook|
+
+.. |timit tdnn_lstm_ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1Hs9DA4V96uapw_30uNp32OMJgkuR5VVd
+
+
+**Congratulations!** You have finished the TDNN-LSTM-CTC recipe on timit in ``icefall``.
diff --git a/docs/source/recipes/Non-streaming-ASR/yesno/images/tdnn-tensorboard-log.png b/docs/source/recipes/Non-streaming-ASR/yesno/images/tdnn-tensorboard-log.png
new file mode 100644
index 000000000..3d2612c9c
Binary files /dev/null and b/docs/source/recipes/Non-streaming-ASR/yesno/images/tdnn-tensorboard-log.png differ
diff --git a/docs/source/recipes/Non-streaming-ASR/yesno/index.rst b/docs/source/recipes/Non-streaming-ASR/yesno/index.rst
new file mode 100644
index 000000000..d68523a97
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/yesno/index.rst
@@ -0,0 +1,7 @@
+YesNo
+=====
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn
diff --git a/docs/source/recipes/Non-streaming-ASR/yesno/tdnn.rst b/docs/source/recipes/Non-streaming-ASR/yesno/tdnn.rst
new file mode 100644
index 000000000..e8b748e6b
--- /dev/null
+++ b/docs/source/recipes/Non-streaming-ASR/yesno/tdnn.rst
@@ -0,0 +1,445 @@
+TDNN-CTC
+========
+
+This page shows you how to run the `yesno <https://www.openslr.org/1>`_ recipe. It contains:
+
+  - (1) Prepare data for training
+  - (2) Train a TDNN model
+
+    - (a) View text format logs and visualize TensorBoard logs
+    - (b) Select device type, i.e., CPU and GPU, for training
+    - (c) Change training options
+    - (d) Resume training from a checkpoint
+
+  - (3) Decode with a trained model
+
+    - (a) Select a checkpoint for decoding
+    - (b) Model averaging
+
+  - (4) Colab notebook
+
+    - (a) It shows you step by step how to setup the environment, how to do training,
+          and how to do decoding
+    - (b) How to use a pre-trained model
+
+  - (5) Inference with a pre-trained model
+
+    - (a) Download a pre-trained model, provided by us
+    - (b) Decode a single sound file with a pre-trained model
+    - (c) Decode multiple sound files at the same time
+
+It does **NOT** show you:
+
+  - (1) How to train with multiple GPUs
+
+    The ``yesno`` dataset is so small that CPU is more than enough
+    for training as well as for decoding.
+
+  - (2) How to use LM rescoring for decoding
+
+    The dataset does not have an LM for rescoring.
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  You **don't** need a **GPU** to run this recipe. It can be run on a **CPU**.
+  The training part takes less than 30 **seconds** on a CPU and you will get
+  the following WER at the end::
+
+    [test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/yesno/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/yesno/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+
+Training
+--------
+
+We provide only a TDNN model, contained in
+the `tdnn <https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR/tdnn>`_
+folder, for ``yesno``.
+
+The command to run the training part is:
+
+.. code-block:: bash
+
+  $ cd egs/yesno/ASR
+  $ export CUDA_VISIBLE_DEVICES=""
+  $ ./tdnn/train.py
+
+By default, it will run ``15`` epochs. Training logs and checkpoints are saved
+in ``tdnn/exp``.
+
+In ``tdnn/exp``, you will find the following files:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./tdnn/train.py --start-epoch 11
+
+  - ``tensorboard/``
+
+    This folder contains TensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd tdnn/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "TDNN training for yesno with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/yKUbhb5wRmOSXYkId1z9eg/
+
+        [2021-08-23T23:49:41] Started scanning logdir.
+        [2021-08-23T23:49:42] Total uploaded: 135 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output, click it and you will see
+    the following screenshot:
+
+      .. figure:: images/tdnn-tensorboard-log.png
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/yKUbhb5wRmOSXYkId1z9eg/
+
+         TensorBoard screenshot.
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+
+
+.. NOTE::
+
+  By default, ``./tdnn/train.py`` uses GPU 0 for training if GPUs are available.
+  If you have two GPUs, say, GPU 0 and GPU 1, and you want to use GPU 1 for
+  training, you can run:
+
+    .. code-block:: bash
+
+      $ export CUDA_VISIBLE_DEVICES="1"
+      $ ./tdnn/train.py
+
+  Since the ``yesno`` dataset is very small, containing only 30 sound files
+  for training, and the model in use is also very small, we use:
+
+    .. code-block:: bash
+
+      $ export CUDA_VISIBLE_DEVICES=""
+
+  so that ``./tdnn/train.py`` uses CPU during training.
+
+  If you don't have GPUs, then you don't need to
+  run ``export CUDA_VISIBLE_DEVICES=""``.
+
+To see available training options, you can use:
+
+.. code-block:: bash
+
+  $ ./tdnn/train.py --help
+
+Other training options, e.g., learning rate, results dir, etc., are
+pre-configured in the function ``get_params()``
+in `tdnn/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/tdnn/train.py>`_.
+Normally, you don't need to change them. You can change them by modifying the code, if
+you want.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+The command for decoding is:
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES=""
+  $ ./tdnn/decode.py
+
+You will see the WER in the output log.
+
+Decoded results are saved in ``tdnn/exp``.
+
+.. code-block:: bash
+
+  $ ./tdnn/decode.py --help
+
+shows you the available decoding options.
+
+Some commonly used options are:
+
+  - ``--epoch``
+
+    You can select which checkpoint to be used for decoding.
+    For instance, ``./tdnn/decode.py --epoch 10`` means to use
+    ``./tdnn/exp/epoch-10.pt`` for decoding.
+
+  - ``--avg``
+
+    It's related to model averaging. It specifies number of checkpoints
+    to be averaged. The averaged model is used for decoding.
+    For example, the following command:
+
+      .. code-block:: bash
+
+        $ ./tdnn/decode.py --epoch 10 --avg 3
+
+    uses the average of ``epoch-8.pt``, ``epoch-9.pt`` and ``epoch-10.pt``
+    for decoding.
+
+  - ``--export``
+
+    If it is ``True``, i.e., ``./tdnn/decode.py --export 1``, the code
+    will save the averaged model to ``tdnn/exp/pretrained.pt``.
+    See :ref:`yesno use a pre-trained model` for how to use it.
+
+
+.. _yesno use a pre-trained model:
+
+Pre-trained Model
+-----------------
+
+We have uploaded the pre-trained model to
+`<https://huggingface.co/csukuangfj/icefall_asr_yesno_tdnn>`_.
+
+The following shows you how to use the pre-trained model.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/yesno/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/csukuangfj/icefall_asr_yesno_tdnn
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/yesno/ASR
+  $ tree tmp
+
+.. code-block:: bash
+
+  tmp/
+  `-- icefall_asr_yesno_tdnn
+      |-- README.md
+      |-- lang_phone
+      |   |-- HLG.pt
+      |   |-- L.pt
+      |   |-- L_disambig.pt
+      |   |-- Linv.pt
+      |   |-- lexicon.txt
+      |   |-- lexicon_disambig.txt
+      |   |-- tokens.txt
+      |   `-- words.txt
+      |-- lm
+      |   |-- G.arpa
+      |   `-- G.fst.txt
+      |-- pretrained.pt
+      `-- test_waves
+          |-- 0_0_0_1_0_0_0_1.wav
+          |-- 0_0_1_0_0_0_1_0.wav
+          |-- 0_0_1_0_0_1_1_1.wav
+          |-- 0_0_1_0_1_0_0_1.wav
+          |-- 0_0_1_1_0_0_0_1.wav
+          |-- 0_0_1_1_0_1_1_0.wav
+          |-- 0_0_1_1_1_0_0_0.wav
+          |-- 0_0_1_1_1_1_0_0.wav
+          |-- 0_1_0_0_0_1_0_0.wav
+          |-- 0_1_0_0_1_0_1_0.wav
+          |-- 0_1_0_1_0_0_0_0.wav
+          |-- 0_1_0_1_1_1_0_0.wav
+          |-- 0_1_1_0_0_1_1_1.wav
+          |-- 0_1_1_1_0_0_1_0.wav
+          |-- 0_1_1_1_1_0_1_0.wav
+          |-- 1_0_0_0_0_0_0_0.wav
+          |-- 1_0_0_0_0_0_1_1.wav
+          |-- 1_0_0_1_0_1_1_1.wav
+          |-- 1_0_1_1_0_1_1_1.wav
+          |-- 1_0_1_1_1_1_0_1.wav
+          |-- 1_1_0_0_0_1_1_1.wav
+          |-- 1_1_0_0_1_0_1_1.wav
+          |-- 1_1_0_1_0_1_0_0.wav
+          |-- 1_1_0_1_1_0_0_1.wav
+          |-- 1_1_0_1_1_1_1_0.wav
+          |-- 1_1_1_0_0_1_0_1.wav
+          |-- 1_1_1_0_1_0_1_0.wav
+          |-- 1_1_1_1_0_0_1_0.wav
+          |-- 1_1_1_1_1_0_0_0.wav
+          `-- 1_1_1_1_1_1_1_1.wav
+
+  4 directories, 42 files
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav
+
+  Input File     : 'tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav'
+  Channels       : 1
+  Sample Rate    : 8000
+  Precision      : 16-bit
+  Duration       : 00:00:06.76 = 54080 samples ~ 507 CDDA sectors
+  File Size      : 108k
+  Bit Rate       : 128k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+- ``0_0_1_0_1_0_0_1.wav``
+
+    0 means No; 1 means Yes. No and Yes are not in English,
+    but in `Hebrew <https://en.wikipedia.org/wiki/Hebrew_language>`_.
+    So this file contains ``NO NO YES NO YES NO NO YES``.
+
+Download kaldifeat
+~~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used for extracting
+features from a single or multiple sound files. Please refer to
+`<https://github.com/csukuangfj/kaldifeat>`_ to install ``kaldifeat`` first.
+
+Inference with a pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/yesno/ASR
+  $ ./tdnn/pretrained.py --help
+
+shows the usage information of ``./tdnn/pretrained.py``.
+
+To decode a single file, we can use:
+
+.. code-block:: bash
+
+  ./tdnn/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_yesno_tdnn/pretrained.pt \
+    --words-file ./tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt \
+    --HLG ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt \
+    ./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav
+
+The output is:
+
+.. code-block::
+
+  2021-08-24 12:22:51,621 INFO [pretrained.py:119] {'feature_dim': 23, 'num_classes': 4, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './tmp/icefall_asr_yesno_tdnn/pretrained.pt', 'words_file': './tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt', 'HLG': './tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt', 'sound_files': ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav']}
+  2021-08-24 12:22:51,645 INFO [pretrained.py:125] device: cpu
+  2021-08-24 12:22:51,645 INFO [pretrained.py:127] Creating model
+  2021-08-24 12:22:51,650 INFO [pretrained.py:139] Loading HLG from ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt
+  2021-08-24 12:22:51,651 INFO [pretrained.py:143] Constructing Fbank computer
+  2021-08-24 12:22:51,652 INFO [pretrained.py:153] Reading sound files: ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav']
+  2021-08-24 12:22:51,684 INFO [pretrained.py:159] Decoding started
+  2021-08-24 12:22:51,708 INFO [pretrained.py:198]
+  ./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav:
+  NO NO YES NO YES NO NO YES
+
+
+  2021-08-24 12:22:51,708 INFO [pretrained.py:200] Decoding Done
+
+You can see that for the sound file ``0_0_1_0_1_0_0_1.wav``, the decoding result is
+``NO NO YES NO YES NO NO YES``.
+
+To decode **multiple** files at the same time, you can use
+
+.. code-block:: bash
+
+  ./tdnn/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_yesno_tdnn/pretrained.pt \
+    --words-file ./tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt \
+    --HLG ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt \
+    ./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav \
+    ./tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav
+
+The decoding output is:
+
+.. code-block::
+
+  2021-08-24 12:25:20,159 INFO [pretrained.py:119] {'feature_dim': 23, 'num_classes': 4, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './tmp/icefall_asr_yesno_tdnn/pretrained.pt', 'words_file': './tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt', 'HLG': './tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt', 'sound_files': ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav', './tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav']}
+  2021-08-24 12:25:20,181 INFO [pretrained.py:125] device: cpu
+  2021-08-24 12:25:20,181 INFO [pretrained.py:127] Creating model
+  2021-08-24 12:25:20,185 INFO [pretrained.py:139] Loading HLG from ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt
+  2021-08-24 12:25:20,186 INFO [pretrained.py:143] Constructing Fbank computer
+  2021-08-24 12:25:20,187 INFO [pretrained.py:153] Reading sound files: ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav',
+  './tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav']
+  2021-08-24 12:25:20,213 INFO [pretrained.py:159] Decoding started
+  2021-08-24 12:25:20,287 INFO [pretrained.py:198]
+  ./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav:
+  NO NO YES NO YES NO NO YES
+
+  ./tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav:
+  YES NO YES YES NO YES YES YES
+
+  2021-08-24 12:25:20,287 INFO [pretrained.py:200] Decoding Done
+
+You can see again that it decodes correctly.
+
+Colab notebook
+--------------
+
+We do provide a colab notebook for this recipe.
+
+|yesno colab notebook|
+
+.. |yesno colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing
+
+
+**Congratulations!** You have finished the simplest speech recognition recipe in ``icefall``.
diff --git a/docs/source/recipes/RNN-LM/index.rst b/docs/source/recipes/RNN-LM/index.rst
new file mode 100644
index 000000000..4b74e64c7
--- /dev/null
+++ b/docs/source/recipes/RNN-LM/index.rst
@@ -0,0 +1,7 @@
+RNN-LM
+======
+
+.. toctree::
+   :maxdepth: 2
+
+   librispeech/lm-training
\ No newline at end of file
diff --git a/docs/source/recipes/RNN-LM/librispeech/lm-training.rst b/docs/source/recipes/RNN-LM/librispeech/lm-training.rst
new file mode 100644
index 000000000..46499a374
--- /dev/null
+++ b/docs/source/recipes/RNN-LM/librispeech/lm-training.rst
@@ -0,0 +1,104 @@
+.. _train_nnlm:
+
+Train an RNN language model
+======================================
+
+If you have enough text data, you can train a neural network language model (NNLM) to improve
+the WER of your E2E ASR system. This tutorial shows you how to train an RNNLM from 
+scratch.
+
+.. HINT::
+
+    For how to use an NNLM during decoding, please refer to the following tutorials:
+    :ref:`shallow_fusion`, :ref:`LODR`, :ref:`rescoring`
+
+.. note::
+
+    This tutorial is based on the LibriSpeech recipe. Please check it out for the necessary
+    python scripts for this tutorial. We use the LibriSpeech LM-corpus as the LM training set 
+    for illustration purpose. You can also collect your own data. The data format is quite simple:
+    each line should contain a complete sentence, and words should be separated by space.
+
+First, let's download the training data for the RNNLM. This can be done via the 
+following command:
+
+.. code-block:: bash
+
+    $ wget https://www.openslr.org/resources/11/librispeech-lm-norm.txt.gz 
+    $ gzip -d librispeech-lm-norm.txt.gz
+
+As we are training a BPE-level RNNLM, we need to tokenize the training text, which requires a
+BPE tokenizer. This can be achieved by executing the following command:
+
+.. code-block:: bash
+    
+    $ # if you don't have the BPE
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+    $ cd icefall-asr-librispeech-zipformer-2023-05-15/data/lang_bpe_500
+    $ git lfs pull --include bpe.model
+    $ cd ../../..
+
+    $ ./local/prepare_lm_training_data.py \
+        --bpe-model icefall-asr-librispeech-zipformer-2023-05-15/data/lang_bpe_500/bpe.model \
+        --lm-data librispeech-lm-norm.txt \
+        --lm-archive data/lang_bpe_500/lm_data.pt
+
+Now, you should have a file name ``lm_data.pt`` file store under the directory ``data/lang_bpe_500``.
+This is the packed training data for the RNNLM. We then sort the training data according to its
+sentence length.
+
+.. code-block:: bash
+
+    $ # This could take a while (~ 20 minutes), feel free to grab a cup of coffee :)
+    $ ./local/sort_lm_training_data.py \
+        --in-lm-data data/lang_bpe_500/lm_data.pt \
+        --out-lm-data data/lang_bpe_500/sorted_lm_data.pt \
+        --out-statistics data/lang_bpe_500/lm_data_stats.txt
+
+
+The aforementioned steps can be repeated to create a a validation set for you RNNLM. Let's say 
+you have a validation set in ``valid.txt``, you can just set ``--lm-data valid.txt`` 
+and ``--lm-archive data/lang_bpe_500/lm-data-valid.pt`` when calling ``./local/prepare_lm_training_data.py``.
+
+After completing the previous steps, the training and testing sets for training RNNLM are ready. 
+The next step is to train the RNNLM model. The training command is as follows:
+
+.. code-block:: bash
+
+    $ # assume you are in the icefall root directory
+    $ cd rnn_lm
+    $ ln -s ../../egs/librispeech/ASR/data .
+    $ cd ..
+    $ ./rnn_lm/train.py \
+        --world-size 4 \
+        --exp-dir ./rnn_lm/exp \
+        --start-epoch 0 \
+        --num-epochs 10 \
+        --use-fp16 0 \
+        --tie-weights 1 \
+        --embedding-dim 2048 \
+        --hidden_dim 2048 \
+        --num-layers 3 \
+        --batch-size 300 \
+        --lm-data rnn_lm/data/lang_bpe_500/sorted_lm_data.pt \
+        --lm-data-valid rnn_lm/data/lang_bpe_500/sorted_lm_data.pt
+
+
+.. note::
+
+    You can adjust the RNNLM hyper parameters to control the size of the RNNLM,
+    such as embedding dimension and hidden state dimension. For more details, please
+    run ``./rnn_lm/train.py --help``.
+
+.. note::
+
+    The training of RNNLM can take a long time (usually a couple of days).
+
+
+
+
+
+
+
+
+
diff --git a/docs/source/recipes/Streaming-ASR/index.rst b/docs/source/recipes/Streaming-ASR/index.rst
new file mode 100644
index 000000000..8c0ffe447
--- /dev/null
+++ b/docs/source/recipes/Streaming-ASR/index.rst
@@ -0,0 +1,12 @@
+Streaming ASR
+=============
+
+.. toctree::
+   :maxdepth: 1
+
+   introduction
+
+.. toctree::
+   :maxdepth: 2
+
+   librispeech/index
diff --git a/docs/source/recipes/Streaming-ASR/introduction.rst b/docs/source/recipes/Streaming-ASR/introduction.rst
new file mode 100644
index 000000000..28f5b8fbf
--- /dev/null
+++ b/docs/source/recipes/Streaming-ASR/introduction.rst
@@ -0,0 +1,53 @@
+Introduction
+============
+
+This page shows you how we implement streaming **X-former transducer** models for ASR.
+
+.. HINT::
+   X-former transducer here means the encoder of the transducer model uses Multi-Head Attention,
+   like `Conformer <https://arxiv.org/pdf/2005.08100.pdf>`_, `EmFormer <https://arxiv.org/pdf/2010.10759.pdf>`_ etc.
+
+Currently we have implemented two types of streaming models, one uses Conformer as encoder, the other uses Emformer as encoder.
+
+Streaming Conformer
+-------------------
+
+The main idea of training a streaming model is to make the model see limited contexts
+in training time, we can achieve this by applying a mask to the output of self-attention.
+In icefall, we implement the streaming conformer the way just like what `WeNet <https://arxiv.org/pdf/2012.05481.pdf>`_ did.
+
+.. NOTE::
+   The conformer-transducer recipes in LibriSpeech datasets, like, `pruned_transducer_stateless <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless>`_,
+   `pruned_transducer_stateless2 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless2>`_,
+   `pruned_transducer_stateless3 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless3>`_,
+   `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`_,
+   `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless5>`_
+   all support streaming.
+
+.. NOTE::
+   Training a streaming conformer model in ``icefall`` is almost the same as training a
+   non-streaming model, all you need to do is passing several extra arguments.
+   See :doc:`Pruned transducer statelessX <librispeech/pruned_transducer_stateless>` for more details.
+
+.. HINT::
+   If you want to modify a non-streaming conformer recipe to support both streaming and non-streaming, please refer
+   to `this pull request <https://github.com/k2-fsa/icefall/pull/454>`_.  After adding the code needed by streaming training,
+   you have to re-train it with the extra arguments mentioned in the docs above to get a streaming model.
+
+
+Streaming Emformer
+------------------
+
+The Emformer model proposed `here <https://arxiv.org/pdf/2010.10759.pdf>`_ uses more
+complicated techniques. It has a memory bank component to memorize history information,
+what' more, it also introduces right context in training time by hard-copying part of
+the input features.
+
+We have three variants of Emformer models in ``icefall``.
+
+ - ``pruned_stateless_emformer_rnnt2`` using Emformer from torchaudio, see `LibriSpeech recipe <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2>`__.
+ - ``conv_emformer_transducer_stateless`` using ConvEmformer implemented by ourself. Different from the Emformer in torchaudio,
+   ConvEmformer has a convolution in each layer and uses the mechanisms in our reworked conformer model.
+   See `LibriSpeech recipe <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conv_emformer_transducer_stateless>`__.
+ - ``conv_emformer_transducer_stateless2`` using ConvEmformer implemented by ourself. The only difference from the above one is that
+   it uses a simplified memory bank. See `LibriSpeech recipe <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conv_emformer_transducer_stateless2>`_.
diff --git a/docs/source/recipes/Streaming-ASR/librispeech/images/librispeech-lstm-transducer-tensorboard-log.png b/docs/source/recipes/Streaming-ASR/librispeech/images/librispeech-lstm-transducer-tensorboard-log.png
new file mode 100644
index 000000000..cc475a45f
Binary files /dev/null and b/docs/source/recipes/Streaming-ASR/librispeech/images/librispeech-lstm-transducer-tensorboard-log.png differ
diff --git a/docs/source/recipes/Streaming-ASR/librispeech/images/streaming-librispeech-pruned-transducer-tensorboard-log.jpg b/docs/source/recipes/Streaming-ASR/librispeech/images/streaming-librispeech-pruned-transducer-tensorboard-log.jpg
new file mode 100644
index 000000000..9c77b8bae
Binary files /dev/null and b/docs/source/recipes/Streaming-ASR/librispeech/images/streaming-librispeech-pruned-transducer-tensorboard-log.jpg differ
diff --git a/docs/source/recipes/Streaming-ASR/librispeech/index.rst b/docs/source/recipes/Streaming-ASR/librispeech/index.rst
new file mode 100644
index 000000000..d52e08058
--- /dev/null
+++ b/docs/source/recipes/Streaming-ASR/librispeech/index.rst
@@ -0,0 +1,11 @@
+LibriSpeech
+===========
+
+.. toctree::
+   :maxdepth: 1
+
+   pruned_transducer_stateless
+
+   lstm_pruned_stateless_transducer
+
+   zipformer_transducer
diff --git a/docs/source/recipes/Streaming-ASR/librispeech/lstm_pruned_stateless_transducer.rst b/docs/source/recipes/Streaming-ASR/librispeech/lstm_pruned_stateless_transducer.rst
new file mode 100644
index 000000000..911e84656
--- /dev/null
+++ b/docs/source/recipes/Streaming-ASR/librispeech/lstm_pruned_stateless_transducer.rst
@@ -0,0 +1,532 @@
+LSTM Transducer
+===============
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+This tutorial shows you how to train an LSTM transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use an LSTM model
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+.. hint::
+
+   Since the encoder model is an LSTM, not Transformer/Conformer, the
+   resulting model is suitable for streaming/online ASR.
+
+
+Which model to use
+------------------
+
+Currently, there are two folders about LSTM stateless transducer training:
+
+  - ``(1)`` `<https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/lstm_transducer_stateless>`_
+
+    This recipe uses only LibriSpeech during training.
+
+  - ``(2)`` `<https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/lstm_transducer_stateless2>`_
+
+    This recipe uses GigaSpeech + LibriSpeech during training.
+
+``(1)`` and ``(2)`` use the same model architecture. The only difference is that ``(2)`` supports
+multi-dataset. Since ``(2)`` uses more data, it has a lower WER than ``(1)`` but it needs
+more training time.
+
+We use ``lstm_transducer_stateless2`` as an example below.
+
+.. note::
+
+   You need to download the `GigaSpeech <https://github.com/SpeechColab/GigaSpeech>`_ dataset
+   to run ``(2)``. If you have only ``LibriSpeech`` dataset available, feel free to use ``(1)``.
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+  # If you use (1), you can **skip** the following command
+  $ ./prepare_giga_speech.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+.. note::
+
+   We encourage you to read ``./prepare.sh``.
+
+The data preparation contains several stages. You can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. hint::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. note::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./lstm_transducer_stateless2/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./lstm_transducer_stateless2/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./lstm_transducer_stateless2/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./lstm_transducer_stateless2/train.py --start-epoch 10`` loads the
+    checkpoint ``./lstm_transducer_stateless2/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./lstm_transducer_stateless2/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./lstm_transducer_stateless2/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./lstm_transducer_stateless2/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--giga-prob``
+
+    The probability to select a batch from the ``GigaSpeech`` dataset.
+    Note: It is available only for ``(2)``.
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`lstm_transducer_stateless2/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/lstm_transducer_stateless2/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./lstm_transducer_stateless2/train.py`` directly.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``lstm_transducer_stateless2/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./lstm_transducer_stateless2/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./lstm_transducer_stateless2/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd lstm_transducer_stateless2/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "LSTM transducer training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/cj2vtPiwQHKN9Q1tx6PTpg/
+
+        [2022-09-20T15:50:50] Started scanning logdir.
+        Uploading 4468 scalars...
+        [2022-09-20T15:53:02] Total uploaded: 210171 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output. Click it and you will see
+    the following screenshot:
+
+      .. figure:: images/librispeech-lstm-transducer-tensorboard-log.png
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/lzGnETjwRxC3yghNMd4kPw/
+
+         TensorBoard screenshot.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd lstm_transducer_stateless2/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 8 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+  ./lstm_transducer_stateless2/train.py \
+    --world-size 8 \
+    --num-epochs 35 \
+    --start-epoch 1 \
+    --full-libri 1 \
+    --exp-dir lstm_transducer_stateless2/exp \
+    --max-duration 500 \
+    --use-fp16 0 \
+    --lr-epochs 10 \
+    --num-workers 2 \
+    --giga-prob 0.9
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``lstm_transducer_stateless2/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``lstm_transducer_stateless2/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./lstm_transducer_stateless2/decode.py --help
+
+shows the options for decoding.
+
+The following shows two examples:
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 17; do
+      for avg in 1 2; do
+        ./lstm_transducer_stateless2/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --exp-dir lstm_transducer_stateless2/exp \
+          --max-duration 600 \
+          --num-encoder-layers 12 \
+          --rnn-hidden-size 1024 \
+          --decoding-method $m \
+          --use-averaged-model True \
+          --beam 4 \
+          --max-contexts 4 \
+          --max-states 8 \
+          --beam-size 4
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./lstm_transducer_stateless2/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --exp-dir lstm_transducer_stateless2/exp \
+          --max-duration 600 \
+          --num-encoder-layers 12 \
+          --rnn-hidden-size 1024 \
+          --decoding-method $m \
+          --use-averaged-model True \
+          --beam 4 \
+          --max-contexts 4 \
+          --max-states 8 \
+          --beam-size 4
+      done
+    done
+  done
+
+Export models
+-------------
+
+`lstm_transducer_stateless2/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/lstm_transducer_stateless2/export.py>`_ supports exporting checkpoints from ``lstm_transducer_stateless2/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``lstm_transducer_stateless2/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --iter 468000 --avg 16 produces the smallest WER
+  # (You can get such information after running ./lstm_transducer_stateless2/decode.py)
+
+  iter=468000
+  avg=16
+
+  ./lstm_transducer_stateless2/export.py \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --iter $iter \
+    --avg  $avg
+
+It will generate a file ``./lstm_transducer_stateless2/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``lstm_transducer_stateless2/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd lstm_transducer_stateless2/exp
+      ln -s pretrained epoch-9999.pt
+
+   And then pass ``--epoch 9999 --avg 1 --use-averaged-model 0`` to
+   ``./lstm_transducer_stateless2/decode.py``.
+
+To use the exported model with ``./lstm_transducer_stateless2/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./lstm_transducer_stateless2/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless2/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Export model using ``torch.jit.trace()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  iter=468000
+  avg=16
+
+  ./lstm_transducer_stateless2/export.py \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --iter $iter \
+    --avg  $avg \
+    --jit-trace 1
+
+It will generate 3 files:
+
+  - ``./lstm_transducer_stateless2/exp/encoder_jit_trace.pt``
+  - ``./lstm_transducer_stateless2/exp/decoder_jit_trace.pt``
+  - ``./lstm_transducer_stateless2/exp/joiner_jit_trace.pt``
+
+To use the generated files with ``./lstm_transducer_stateless2/jit_pretrained``:
+
+.. code-block:: bash
+
+  ./lstm_transducer_stateless2/jit_pretrained.py \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --encoder-model-filename ./lstm_transducer_stateless2/exp/encoder_jit_trace.pt \
+    --decoder-model-filename ./lstm_transducer_stateless2/exp/decoder_jit_trace.pt \
+    --joiner-model-filename ./lstm_transducer_stateless2/exp/joiner_jit_trace.pt \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+.. hint::
+
+   Please see `<https://k2-fsa.github.io/sherpa/python/streaming_asr/lstm/english/server.html>`_
+   for how to use the exported models in ``sherpa``.
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `<https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03>`_
+
+  - `<https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+You can find more usages of the pretrained models in
+`<https://k2-fsa.github.io/sherpa/python/streaming_asr/lstm/index.html>`_
diff --git a/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst b/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst
new file mode 100644
index 000000000..d6e424e2f
--- /dev/null
+++ b/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst
@@ -0,0 +1,735 @@
+Pruned transducer statelessX
+============================
+
+This tutorial shows you how to run a **streaming** conformer transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. Note::
+
+   The tutorial is suitable for `pruned_transducer_stateless <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless>`__,
+   `pruned_transducer_stateless2 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless2>`__,
+   `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`__,
+   `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless5>`__,
+   We will take pruned_transducer_stateless4 as an example in this tutorial.
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use a Conformer model (the reworked version by Daniel Povey)
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+
+Data preparation
+----------------
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to ``Training`` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+.. NOTE::
+
+   We put the streaming and non-streaming model in one recipe, to train a streaming model you only
+   need to add **4** extra options comparing with training a non-streaming model. These options are
+   ``--dynamic-chunk-training``, ``--num-left-chunks``, ``--causal-convolution``, ``--short-chunk-size``.
+   You can see the configurable options below for their meanings or read https://arxiv.org/pdf/2012.05481.pdf for more details.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/train.py --help
+
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The directory to save checkpoints, training logs and tensorboard.
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless4/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless4/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless4/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless4/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless4/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless4/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless4/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--use-fp16``
+
+    If it is True, the model will train with half precision, from our experiment
+    results, by using half precision you can train with two times larger ``--max-duration``
+    so as to get almost 2X speed up.
+
+  - ``--dynamic-chunk-training``
+
+    The flag that indicates whether to train a streaming model or not, it
+    **MUST** be True if you want to train a streaming model.
+
+  - ``--short-chunk-size``
+
+    When training a streaming attention model with chunk masking, the chunk size
+    would be either max sequence length of current batch or uniformly sampled from
+    (1, short_chunk_size). The default value is 25, you don't have to change it most of the time.
+
+  - ``--num-left-chunks``
+
+    It indicates how many left context (in chunks) that can be seen when calculating attention.
+    The default value is 4, you don't have to change it most of the time.
+
+
+  - ``--causal-convolution``
+
+    Whether to use causal convolution in conformer encoder layer, this requires
+    to be True when training a streaming model.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., number of encoder layers,
+encoder dimension, decoder dimension, number of warmup steps etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless4/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless4/train.py`` directly.
+
+
+.. NOTE::
+
+  The options for `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless5/train.py>`__ are a little different from
+  other recipes. It allows you to configure ``--num-encoder-layers``, ``--dim-feedforward``, ``--nhead``, ``--encoder-dim``, ``--decoder-dim``, ``--joiner-dim`` from commandline, so that you can train models with different size with pruned_transducer_stateless5.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``--exp-dir`` (e.g. ``pruned_transducer_stateless4/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless4/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "pruned transducer training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/97VKXf80Ru61CnP2ALWZZg/
+
+        [2022-11-20T15:50:50] Started scanning logdir.
+        Uploading 4468 scalars...
+        [2022-11-20T15:53:02] Total uploaded: 210171 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output. Click it and you will see
+    the following screenshot:
+
+      .. figure:: images/streaming-librispeech-pruned-transducer-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/97VKXf80Ru61CnP2ALWZZg/
+
+         TensorBoard screenshot.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless4/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./pruned_transducer_stateless4/train.py \
+     --world-size 4 \
+     --dynamic-chunk-training 1 \
+     --causal-convolution 1 \
+     --num-epochs 30 \
+     --start-epoch 1 \
+     --exp-dir pruned_transducer_stateless4/exp \
+     --full-libri 1 \
+     --max-duration 300
+
+.. NOTE::
+
+   Comparing with training a non-streaming model, you only need to add two extra options,
+   ``--dynamic-chunk-training 1``  and ``--causal-convolution 1`` .
+
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. tip::
+
+    To decode a streaming model, you can use either ``simulate streaming decoding`` in ``decode.py`` or
+    ``real streaming decoding`` in ``streaming_decode.py``, the difference between ``decode.py`` and
+    ``streaming_decode.py`` is that, ``decode.py`` processes the whole acoustic frames at one time with masking (i.e. same as training),
+    but ``streaming_decode.py`` processes the acoustic frames chunk by chunk (so it can only see limited context).
+
+.. NOTE::
+
+   ``simulate streaming decoding`` in ``decode.py`` and ``real streaming decoding`` in ``streaming_decode.py`` should
+   produce almost the same results given the same ``--decode-chunk-size`` and ``--left-context``.
+
+
+Simulate streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--simulate-streaming``
+
+    If you want to decode a streaming model with ``decode.py``, you **MUST** set
+    ``--simulate-streaming`` to ``True``. ``simulate`` here means the acoustic frames
+    are not processed frame by frame (or chunk by chunk), instead, the whole sequence
+    is processed at one time with masking (the same as training).
+
+  ``--causal-convolution``
+
+    If True, the convolution module in encoder layers will be causal convolution.
+    This is **MUST** be True when decoding with a streaming model.
+
+  ``--decode-chunk-size``
+
+    For streaming models, we will calculate the chunk-wise attention, ``--decode-chunk-size``
+    indicates the chunk length (in frames after subsampling) for chunk-wise attention.
+    For ``simulate streaming decoding`` the ``decode-chunk-size`` is used to generate
+    the attention mask.
+
+  ``--left-context``
+
+    ``--left-context`` indicates how many left context frames (after subsampling) can be seen
+    for current chunk when calculating chunk-wise attention. Normally, ``left-context`` should equal
+    to ``decode-chunk-size * num-left-chunks``, where ``num-left-chunks`` is the option used
+    to train this model. For ``simulate streaming decoding`` the ``left-context`` is used to generate
+    the attention mask.
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 25 20; do
+      for avg in 7 5 3 1; do
+        ./pruned_transducer_stateless4/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --simulate-streaming 1 \
+          --causal-convolution 1 \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless4/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --simulate-streaming 1 \
+          --causal-convolution 1 \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+Real streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/streaming_decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--decode-chunk-size``
+
+    For streaming models, we will calculate the chunk-wise attention, ``--decode-chunk-size``
+    indicates the chunk length (in frames after subsampling) for chunk-wise attention.
+    For ``real streaming decoding``, we will process ``decode-chunk-size`` acoustic frames at each time.
+
+  ``--left-context``
+
+    ``--left-context`` indicates how many left context frames (after subsampling) can be seen
+    for current chunk when calculating chunk-wise attention. Normally, ``left-context`` should equal
+    to ``decode-chunk-size * num-left-chunks``, where ``num-left-chunks`` is the option used
+    to train this model.
+
+  ``--num-decode-streams``
+
+    The number of decoding streams that can be run in parallel (very similar to the ``bath size``).
+    For ``real streaming decoding``, the batches will be packed dynamically, for example, if the
+    ``num-decode-streams`` equals to 10, then, sequence 1 to 10 will be decoded at first, after a while,
+    suppose sequence 1 and 2 are done, so, sequence 3 to 12 will be processed parallelly in a batch.
+
+
+.. NOTE::
+
+   We also try adding ``--right-context`` in the real streaming decoding, but it seems not to benefit
+   the performance for all the models, the reasons might be the training and decoding mismatch. You
+   can try decoding with ``--right-context`` to see if it helps. The default value is 0.
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 25 20; do
+      for avg in 7 5 3 1; do
+        ./pruned_transducer_stateless4/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless4/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. tip::
+
+  Supporting decoding methods are as follows:
+
+    - ``greedy_search`` : It takes the symbol with largest posterior probability
+      of each frame as the decoding result.
+
+    - ``beam_search`` :  It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf and
+      `espnet/nets/beam_search_transducer.py <https://github.com/espnet/espnet/blob/master/espnet/nets/beam_search_transducer.py#L247>`_
+      is used as a reference. Basically, it keeps topk states for each frame, and expands the kept states with their own contexts to
+      next frame.
+
+    - ``modified_beam_search`` : It implements the same algorithm as ``beam_search`` above, but it
+      runs in batch mode with ``--max-sym-per-frame=1`` being hardcoded.
+
+    - ``fast_beam_search`` : It implements graph composition between the output ``log_probs`` and
+      given ``FSAs``. It is hard to describe the details in several lines of texts, you can read
+      our paper in https://arxiv.org/pdf/2211.00484.pdf or our `rnnt decode code in k2 <https://github.com/k2-fsa/k2/blob/master/k2/csrc/rnnt_decode.h>`_. ``fast_beam_search`` can decode with ``FSAs`` on GPU efficiently.
+
+    - ``fast_beam_search_LG`` : The same as ``fast_beam_search`` above, ``fast_beam_search`` uses
+      an trivial graph that has only one state, while ``fast_beam_search_LG`` uses an LG graph
+      (with N-gram LM).
+
+    - ``fast_beam_search_nbest`` : It produces the decoding results as follows:
+
+      - (1) Use ``fast_beam_search`` to get a lattice
+      - (2) Select ``num_paths`` paths from the lattice using ``k2.random_paths()``
+      - (3) Unique the selected paths
+      - (4) Intersect the selected paths with the lattice and compute the
+            shortest path from the intersection result
+      - (5) The path with the largest score is used as the decoding output.
+
+    - ``fast_beam_search_nbest_LG`` : It implements same logic as ``fast_beam_search_nbest``, the
+      only difference is that it uses ``fast_beam_search_LG`` to generate the lattice.
+
+.. NOTE::
+
+  The supporting decoding methods in ``streaming_decode.py`` might be less than that in ``decode.py``, if needed,
+  you can implement them by yourself or file a issue in `icefall <https://github.com/k2-fsa/icefall/issues>`_ .
+
+
+Export Model
+------------
+
+`pruned_transducer_stateless4/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/export.py>`_ supports exporting checkpoints from ``pruned_transducer_stateless4/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless4/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --epoch 25 --avg 3 produces the smallest WER
+  # (You can get such information after running ./pruned_transducer_stateless4/decode.py)
+
+  epoch=25
+  avg=3
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --streaming-model 1 \
+    --causal-convolution 1 \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch $epoch \
+    --avg  $avg
+
+.. caution::
+
+   ``--streaming-model`` and ``--causal-convolution`` require to be True to export
+   a streaming model.
+
+It will generate a file ``./pruned_transducer_stateless4/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless4/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless4/exp
+      ln -s pretrained.pt epoch-999.pt
+
+   And then pass ``--epoch 999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless4/decode.py``.
+
+To use the exported model with ``./pruned_transducer_stateless4/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless4/exp/pretrained.pt \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --streaming-model 1 \
+    --causal-convolution 1 \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 25 \
+    --avg 3 \
+    --jit 1
+
+.. caution::
+
+   ``--streaming-model`` and ``--causal-convolution`` require to be True to export
+   a streaming model.
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+.. NOTE::
+
+   You will need this ``cpu_jit.pt`` when deploying with Sherpa framework.
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `pruned_transducer_stateless <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless_20220625>`_
+
+  - `pruned_transducer_stateless2 <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless2_20220625>`_
+
+  - `pruned_transducer_stateless4 <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless4_20220625>`_
+
+  - `pruned_transducer_stateless5 <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+
+Deploy with Sherpa
+------------------
+
+Please see `<https://k2-fsa.github.io/sherpa/python/streaming_asr/conformer/index.html#>`_
+for how to deploy the models in ``sherpa``.
diff --git a/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst b/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst
new file mode 100644
index 000000000..8b75473c6
--- /dev/null
+++ b/docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst
@@ -0,0 +1,654 @@
+Zipformer Transducer
+====================
+
+This tutorial shows you how to run a **streaming** zipformer transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. Note::
+
+   The tutorial is suitable for `pruned_transducer_stateless7_streaming <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming>`__,
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use a Zipformer model (proposed by Daniel Povey)
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+
+Data preparation
+----------------
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to ``Training`` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_streaming/train.py --help
+
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The directory to save checkpoints, training logs and tensorboard.
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless7_streaming/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless7_streaming/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless7_streaming/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless7_streaming/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless7_streaming/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless7_streaming/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless7_streaming/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--use-fp16``
+
+    If it is True, the model will train with half precision, from our experiment
+    results, by using half precision you can train with two times larger ``--max-duration``
+    so as to get almost 2X speed up.
+
+    We recommend using ``--use-fp16 True``.
+
+  - ``--short-chunk-size``
+
+    When training a streaming attention model with chunk masking, the chunk size
+    would be either max sequence length of current batch or uniformly sampled from
+    (1, short_chunk_size). The default value is 50, you don't have to change it most of the time.
+
+  - ``--num-left-chunks``
+
+    It indicates how many left context (in chunks) that can be seen when calculating attention.
+    The default value is 4, you don't have to change it most of the time.
+
+
+  - ``--decode-chunk-len``
+
+    The chunk size for decoding (in frames before subsampling). It is used for validation.
+    The default value is 32 (i.e., 320ms).
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., number of encoder layers,
+encoder dimension, decoder dimension, number of warmup steps etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless7_streaming/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless7_streaming/train.py`` directly.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``--exp-dir`` (e.g. ``pruned_transducer_stateless7_streaming/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_streaming/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_streaming/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless7_streaming/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "pruned transducer training for LibriSpeech with icefall"
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless7_streaming/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./pruned_transducer_stateless7_streaming/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --use-fp16 1 \
+    --exp-dir pruned_transducer_stateless7_streaming/exp \
+    --full-libri 1 \
+    --max-duration 550
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless7_streaming/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless7_streaming/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. tip::
+
+    To decode a streaming model, you can use either ``simulate streaming decoding`` in ``decode.py`` or
+    ``real chunk-wise streaming decoding`` in ``streaming_decode.py``. The difference between ``decode.py`` and
+    ``streaming_decode.py`` is that, ``decode.py`` processes the whole acoustic frames at one time with masking (i.e. same as training),
+    but ``streaming_decode.py`` processes the acoustic frames chunk by chunk.
+
+.. NOTE::
+
+   ``simulate streaming decoding`` in ``decode.py`` and ``real chunk-size streaming decoding`` in ``streaming_decode.py`` should
+   produce almost the same results given the same ``--decode-chunk-len``.
+
+
+Simulate streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_streaming/decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--decode-chunk-len``
+
+    It is same as in ``train.py``, which specifies the chunk size for decoding (in frames before subsampling).
+    The default value is 32 (i.e., 320ms).
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 30; do
+      for avg in 12 11 10 9 8; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --decode-chunk-len 32 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --decode-chunk-len 32 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+Real streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_streaming/streaming_decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--decode-chunk-len``
+
+    It is same as in ``train.py``, which specifies the chunk size for decoding (in frames before subsampling).
+    The default value is 32 (i.e., 320ms).
+    For ``real streaming decoding``, we will process ``decode-chunk-len`` acoustic frames at each time.
+
+  ``--num-decode-streams``
+
+    The number of decoding streams that can be run in parallel (very similar to the ``bath size``).
+    For ``real streaming decoding``, the batches will be packed dynamically, for example, if the
+    ``num-decode-streams`` equals to 10, then, sequence 1 to 10 will be decoded at first, after a while,
+    suppose sequence 1 and 2 are done, so, sequence 3 to 12 will be processed parallelly in a batch.
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 30; do
+      for avg in 12 11 10 9 8; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --decode-chunk-len 32 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --decode-chunk-len 16 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. tip::
+
+  Supporting decoding methods are as follows:
+
+    - ``greedy_search`` : It takes the symbol with largest posterior probability
+      of each frame as the decoding result.
+
+    - ``beam_search`` :  It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf and
+      `espnet/nets/beam_search_transducer.py <https://github.com/espnet/espnet/blob/master/espnet/nets/beam_search_transducer.py#L247>`_
+      is used as a reference. Basicly, it keeps topk states for each frame, and expands the kept states with their own contexts to
+      next frame.
+
+    - ``modified_beam_search`` : It implements the same algorithm as ``beam_search`` above, but it
+      runs in batch mode with ``--max-sym-per-frame=1`` being hardcoded.
+
+    - ``fast_beam_search`` : It implements graph composition between the output ``log_probs`` and
+      given ``FSAs``. It is hard to describe the details in several lines of texts, you can read
+      our paper in https://arxiv.org/pdf/2211.00484.pdf or our `rnnt decode code in k2 <https://github.com/k2-fsa/k2/blob/master/k2/csrc/rnnt_decode.h>`_. ``fast_beam_search`` can decode with ``FSAs`` on GPU efficiently.
+
+    - ``fast_beam_search_LG`` : The same as ``fast_beam_search`` above, ``fast_beam_search`` uses
+      an trivial graph that has only one state, while ``fast_beam_search_LG`` uses an LG graph
+      (with N-gram LM).
+
+    - ``fast_beam_search_nbest`` : It produces the decoding results as follows:
+
+      - (1) Use ``fast_beam_search`` to get a lattice
+      - (2) Select ``num_paths`` paths from the lattice using ``k2.random_paths()``
+      - (3) Unique the selected paths
+      - (4) Intersect the selected paths with the lattice and compute the
+            shortest path from the intersection result
+      - (5) The path with the largest score is used as the decoding output.
+
+    - ``fast_beam_search_nbest_LG`` : It implements same logic as ``fast_beam_search_nbest``, the
+      only difference is that it uses ``fast_beam_search_LG`` to generate the lattice.
+
+.. NOTE::
+
+  The supporting decoding methods in ``streaming_decode.py`` might be less than that in ``decode.py``, if needed,
+  you can implement them by yourself or file a issue in `icefall <https://github.com/k2-fsa/icefall/issues>`_ .
+
+
+Export Model
+------------
+
+Currently it supports exporting checkpoints from ``pruned_transducer_stateless7_streaming/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless7_streaming/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --epoch 30 --avg 9 produces the smallest WER
+  # (You can get such information after running ./pruned_transducer_stateless7_streaming/decode.py)
+
+  epoch=30
+  avg=9
+
+  ./pruned_transducer_stateless7_streaming/export.py \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch $epoch \
+    --avg  $avg \
+    --use-averaged-model=True \
+    --decode-chunk-len 32
+
+It will generate a file ``./pruned_transducer_stateless7_streaming/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless7_streaming/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless7_streaming/exp
+      ln -s pretrained.pt epoch-999.pt
+
+   And then pass ``--epoch 999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless7_streaming/decode.py``.
+
+To use the exported model with ``./pruned_transducer_stateless7_streaming/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    --decode-chunk-len 32 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/export.py \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --decode-chunk-len 32 \
+    --jit 1
+
+.. caution::
+
+   ``--decode-chunk-len`` is required to export a ScriptModule.
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+Export model using ``torch.jit.trace()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  epoch=30
+  avg=9
+
+  ./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --use-averaged-model=True \
+    --decode-chunk-len 32 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --epoch $epoch \
+    --avg $avg
+
+.. caution::
+
+   ``--decode-chunk-len`` is required to export a ScriptModule.
+
+It will generate 3 files:
+
+  - ``./pruned_transducer_stateless7_streaming/exp/encoder_jit_trace.pt``
+  - ``./pruned_transducer_stateless7_streaming/exp/decoder_jit_trace.pt``
+  - ``./pruned_transducer_stateless7_streaming/exp/joiner_jit_trace.pt``
+
+To use the generated files with ``./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py \
+    --encoder-model-filename ./pruned_transducer_stateless7_streaming/exp/encoder_jit_trace.pt \
+    --decoder-model-filename ./pruned_transducer_stateless7_streaming/exp/decoder_jit_trace.pt \
+    --joiner-model-filename ./pruned_transducer_stateless7_streaming/exp/joiner_jit_trace.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --decode-chunk-len 32 \
+    /path/to/foo.wav
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `pruned_transducer_stateless7_streaming <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>`__
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+Deploy with Sherpa
+------------------
+
+Please see `<https://k2-fsa.github.io/sherpa/python/streaming_asr/conformer/index.html#>`_
+for how to deploy the models in ``sherpa``.
diff --git a/docs/source/recipes/TTS/index.rst b/docs/source/recipes/TTS/index.rst
new file mode 100644
index 000000000..80d67a2f3
--- /dev/null
+++ b/docs/source/recipes/TTS/index.rst
@@ -0,0 +1,8 @@
+TTS
+======
+
+.. toctree::
+   :maxdepth: 2
+
+   ljspeech/vits
+   vctk/vits
\ No newline at end of file
diff --git a/docs/source/recipes/TTS/ljspeech/vits.rst b/docs/source/recipes/TTS/ljspeech/vits.rst
new file mode 100644
index 000000000..d08aa0f47
--- /dev/null
+++ b/docs/source/recipes/TTS/ljspeech/vits.rst
@@ -0,0 +1,123 @@
+VITS
+===============
+
+This tutorial shows you how to train an VITS model
+with the `LJSpeech <https://keithito.com/LJ-Speech-Dataset/>`_ dataset.
+
+.. note::
+  
+   TTS related recipes require packages in ``requirements-tts.txt``.
+
+.. note::
+
+   The VITS paper: `Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech <https://arxiv.org/pdf/2106.06103.pdf>`_
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/ljspeech/TTS
+  $ ./prepare.sh
+
+To run stage 1 to stage 5, use
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 1 --stop_stage 5
+
+
+Build Monotonic Alignment Search
+--------------------------------
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage -1 --stop_stage -1
+
+or
+
+.. code-block:: bash
+
+  $ cd vits/monotonic_align
+  $ python setup.py build_ext --inplace
+  $ cd ../../
+
+
+Training
+--------
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  $ ./vits/train.py \
+      --world-size 4 \
+      --num-epochs 1000 \
+      --start-epoch 1 \
+      --use-fp16 1 \
+      --exp-dir vits/exp \
+      --tokens data/tokens.txt
+      --max-duration 500
+
+.. note::
+
+    You can adjust the hyper-parameters to control the size of the VITS model and
+    the training configurations. For more details, please run ``./vits/train.py --help``.
+
+.. note::
+
+    The training can take a long time (usually a couple of days).
+
+Training logs, checkpoints and tensorboard logs are saved in ``vits/exp``.
+
+
+Inference
+---------
+
+The inference part uses checkpoints saved by the training part, so you have to run the
+training part first. It will save the ground-truth and generated wavs to the directory
+``vits/exp/infer/epoch-*/wav``, e.g., ``vits/exp/infer/epoch-1000/wav``.
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./vits/infer.py \
+      --epoch 1000 \
+      --exp-dir vits/exp \
+      --tokens data/tokens.txt \
+      --max-duration 500
+
+.. note::
+
+    For more details, please run ``./vits/infer.py --help``.
+
+
+Export models
+-------------
+
+Currently we only support ONNX model exporting. It will generate two files in the given ``exp-dir``:
+``vits-epoch-*.onnx`` and ``vits-epoch-*.int8.onnx``.
+
+.. code-block:: bash
+
+  $ ./vits/export-onnx.py \
+      --epoch 1000 \
+      --exp-dir vits/exp \
+      --tokens data/tokens.txt
+
+You can test the exported ONNX model with:
+
+.. code-block:: bash
+
+  $ ./vits/test_onnx.py \
+      --model-filename vits/exp/vits-epoch-1000.onnx \
+      --tokens data/tokens.txt
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following link:
+
+  - `<https://huggingface.co/Zengwei/icefall-tts-ljspeech-vits-2023-11-29>`_
diff --git a/docs/source/recipes/TTS/vctk/vits.rst b/docs/source/recipes/TTS/vctk/vits.rst
new file mode 100644
index 000000000..34024a5ea
--- /dev/null
+++ b/docs/source/recipes/TTS/vctk/vits.rst
@@ -0,0 +1,125 @@
+VITS
+===============
+
+This tutorial shows you how to train an VITS model
+with the `VCTK <https://datashare.ed.ac.uk/handle/10283/3443>`_ dataset.
+
+.. note::
+  
+   TTS related recipes require packages in ``requirements-tts.txt``.
+
+.. note::
+
+   The VITS paper: `Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech <https://arxiv.org/pdf/2106.06103.pdf>`_
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/vctk/TTS
+  $ ./prepare.sh
+
+To run stage 1 to stage 6, use
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 1 --stop_stage 6
+
+
+Build Monotonic Alignment Search
+--------------------------------
+
+To build the monotonic alignment search, use the following commands:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage -1 --stop_stage -1
+
+or
+
+.. code-block:: bash
+
+  $ cd vits/monotonic_align
+  $ python setup.py build_ext --inplace
+  $ cd ../../
+
+
+Training
+--------
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  $ ./vits/train.py \
+      --world-size 4 \
+      --num-epochs 1000 \
+      --start-epoch 1 \
+      --use-fp16 1 \
+      --exp-dir vits/exp \
+      --tokens data/tokens.txt
+      --max-duration 350
+
+.. note::
+
+    You can adjust the hyper-parameters to control the size of the VITS model and
+    the training configurations. For more details, please run ``./vits/train.py --help``.
+
+.. note::
+
+    The training can take a long time (usually a couple of days).
+
+Training logs, checkpoints and tensorboard logs are saved in ``vits/exp``.
+
+
+Inference
+---------
+
+The inference part uses checkpoints saved by the training part, so you have to run the
+training part first. It will save the ground-truth and generated wavs to the directory
+``vits/exp/infer/epoch-*/wav``, e.g., ``vits/exp/infer/epoch-1000/wav``.
+
+.. code-block:: bash
+
+  $ export CUDA_VISIBLE_DEVICES="0"
+  $ ./vits/infer.py \
+      --epoch 1000 \
+      --exp-dir vits/exp \
+      --tokens data/tokens.txt \
+      --max-duration 500
+
+.. note::
+
+    For more details, please run ``./vits/infer.py --help``.
+
+
+Export models
+-------------
+
+Currently we only support ONNX model exporting. It will generate two files in the given ``exp-dir``:
+``vits-epoch-*.onnx`` and ``vits-epoch-*.int8.onnx``.
+
+.. code-block:: bash
+
+  $ ./vits/export-onnx.py \
+      --epoch 1000 \
+      --exp-dir vits/exp \
+      --tokens data/tokens.txt
+
+You can test the exported ONNX model with:
+
+.. code-block:: bash
+
+  $ ./vits/test_onnx.py \
+      --model-filename vits/exp/vits-epoch-1000.onnx \
+      --tokens data/tokens.txt
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following link:
+
+  - `<https://huggingface.co/zrjin/icefall-tts-vctk-vits-2023-12-05>`_
diff --git a/docs/source/recipes/index.rst b/docs/source/recipes/index.rst
new file mode 100644
index 000000000..8df61f0d0
--- /dev/null
+++ b/docs/source/recipes/index.rst
@@ -0,0 +1,19 @@
+Recipes
+=======
+
+This page contains various recipes in ``icefall``.
+Currently, we provide recipes for speech recognition, language model, and speech synthesis.
+
+We may add recipes for other tasks as well in the future.
+
+.. we put the yesno recipe as the first recipe since it is the simplest one.
+.. Other recipes are listed in a alphabetical order.
+
+.. toctree::
+   :maxdepth: 2
+   :caption: Table of Contents
+
+   Non-streaming-ASR/index
+   Streaming-ASR/index
+   RNN-LM/index
+   TTS/index
diff --git a/egs/aidatatang_200zh/ASR/README.md b/egs/aidatatang_200zh/ASR/README.md
new file mode 100644
index 000000000..b85895a09
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/README.md
@@ -0,0 +1,38 @@
+Note: This recipe is trained with the codes from this PR https://github.com/k2-fsa/icefall/pull/375
+# Pre-trained Transducer-Stateless2 models for the Aidatatang_200zh dataset with icefall.
+The model was trained on full [Aidatatang_200zh](https://www.openslr.org/62) with the scripts in [icefall](https://github.com/k2-fsa/icefall) based on the latest version k2.
+## Training procedure
+The main repositories are list below, we will update the training and decoding scripts with the update of version.
+k2: https://github.com/k2-fsa/k2
+icefall: https://github.com/k2-fsa/icefall
+lhotse: https://github.com/lhotse-speech/lhotse
+* Install k2 and lhotse, k2 installation guide refers to https://k2.readthedocs.io/en/latest/installation/index.html, lhotse refers to https://lhotse.readthedocs.io/en/latest/getting-started.html#installation. I think the latest version would be ok. And please also install the requirements listed in icefall.
+* Clone icefall(https://github.com/k2-fsa/icefall) and check to the commit showed above.
+```
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+```
+* Preparing data.
+```
+cd egs/aidatatang_200zh/ASR
+bash ./prepare.sh
+```
+* Training
+```
+export CUDA_VISIBLE_DEVICES="0,1"
+./pruned_transducer_stateless2/train.py \
+                  --world-size 2 \
+                  --num-epochs 30 \
+                  --start-epoch 0 \
+                  --exp-dir pruned_transducer_stateless2/exp \
+                  --lang-dir data/lang_char \
+                  --max-duration 250
+```
+## Evaluation results
+The decoding results (WER%) on Aidatatang_200zh(dev and test) are listed below, we got this result by averaging models from epoch 11 to 29.
+The WERs are
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 5.53       | 6.59       | --epoch 29, --avg 19, --max-duration 100 |
+| modified beam search (beam size 4) | 5.27       | 6.33       | --epoch 29, --avg 19, --max-duration 100 |
+| fast beam search (set as default)  | 5.30       | 6.34       | --epoch 29, --avg 19, --max-duration 1500|
diff --git a/egs/aidatatang_200zh/ASR/RESULTS.md b/egs/aidatatang_200zh/ASR/RESULTS.md
new file mode 100644
index 000000000..5b82fb61f
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/RESULTS.md
@@ -0,0 +1,72 @@
+## Results
+
+### Aidatatang_200zh Char training results (Pruned Transducer Stateless2)
+
+#### 2022-05-16
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/375.
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 5.53       | 6.59       | --epoch 29, --avg 19, --max-duration 100 |
+| modified beam search (beam size 4) | 5.27       | 6.33       | --epoch 29, --avg 19, --max-duration 100 |
+| fast beam search (set as default)  | 5.30       | 6.34       | --epoch 29, --avg 19, --max-duration 1500|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 2 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 250 \
+  --save-every-n 1000
+
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/xS7kgYf2RwyDpQAOdS8rAA/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=19
+
+## greedy search
+./pruned_transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 100
+
+## modified beam search
+./pruned_transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir ./data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_aidatatang-200zh_pruned_transducer_stateless2>
diff --git a/egs/aidatatang_200zh/ASR/local/__init__.py b/egs/aidatatang_200zh/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aidatatang_200zh/ASR/local/compute_fbank_aidatatang_200zh.py b/egs/aidatatang_200zh/ASR/local/compute_fbank_aidatatang_200zh.py
new file mode 100755
index 000000000..9caacb78b
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/compute_fbank_aidatatang_200zh.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the aidatatang_200zh dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_aidatatang_200zh(num_mel_bins: int = 80, perturb_speed: bool = False):
+    src_dir = Path("data/manifests/aidatatang_200zh")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+    prefix = "aidatatang"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+
+            for sup in m["supervisions"]:
+                sup.custom = {"origin": "aidatatang_200zh"}
+
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and perturb_speed:
+                logging.info(f"Doing speed perturb")
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_aidatatang_200zh(
+        num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed
+    )
diff --git a/egs/aidatatang_200zh/ASR/local/compute_fbank_musan.py b/egs/aidatatang_200zh/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/local/display_manifest_statistics.py b/egs/aidatatang_200zh/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..d66e5cfca
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,96 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    paths = [
+        "./data/fbank/aidatatang_cuts_train.jsonl.gz",
+        "./data/fbank/aidatatang_cuts_dev.jsonl.gz",
+        "./data/fbank/aidatatang_cuts_test.jsonl.gz",
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest_lazy(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/aidatatang_cuts_train.jsonl.gz
+Cuts count: 494715
+Total duration (hours): 422.6
+Speech duration (hours): 422.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.1
+std     1.2
+min     1.0
+25%     2.3
+50%     2.7
+75%     3.5
+99%     7.2
+99.5%   8.0
+99.9%   9.5
+max     18.1
+Starting display the statistics for ./data/fbank/aidatatang_cuts_dev.jsonl.gz
+Cuts count: 24216
+Total duration (hours): 20.2
+Speech duration (hours): 20.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.0
+std     1.0
+min     1.2
+25%     2.3
+50%     2.7
+75%     3.4
+99%     6.7
+99.5%   7.3
+99.9%   8.8
+max     11.3
+Starting display the statistics for ./data/fbank/aidatatang_cuts_test.jsonl.gz
+Cuts count: 48144
+Total duration (hours): 40.2
+Speech duration (hours): 40.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.0
+std     1.1
+min     0.9
+25%     2.3
+50%     2.6
+75%     3.4
+99%     6.9
+99.5%   7.5
+99.9%   9.0
+max     21.8
+"""
diff --git a/egs/aidatatang_200zh/ASR/local/prepare_char.py b/egs/aidatatang_200zh/ASR/local/prepare_char.py
new file mode 100755
index 000000000..6b440dfb3
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/prepare_char.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = list(line)
+            for char in chars:
+                if char not in tokens:
+                    tokens[char] = len(tokens)
+    return tokens
+
+
+def main():
+    lang_dir = Path("data/lang_char")
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/local/prepare_lang.py b/egs/aidatatang_200zh/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..c8cf9b881
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/prepare_lang.py
@@ -0,0 +1,388 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone")
+    return parser.parse_args()
+
+
+def main():
+    out_dir = Path(get_args().lang_dir)
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/local/prepare_words.py b/egs/aidatatang_200zh/ASR/local/prepare_words.py
new file mode 100755
index 000000000..65aca2983
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/prepare_words.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input words.txt without ids:
+    - words_no_ids.txt
+and generates the new words.txt with related ids.
+    - words.txt
+"""
+
+
+import argparse
+import logging
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare words.txt",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/words_no_ids.txt",
+        type=str,
+        help="the words file without ids for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/words.txt",
+        type=str,
+        help="the words file with ids for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    add_words = ["<eps> 0", "!SIL 1", "<SPOKEN_NOISE> 2", "<UNK> 3"]
+    new_lines.extend(add_words)
+
+    logging.info("Starting reading the input file")
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        idx = 4 + i
+        new_line = str(x.strip("\n")) + " " + str(idx)
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the words.txt")
+    f_out = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/local/test_prepare_lang.py b/egs/aidatatang_200zh/ASR/local/test_prepare_lang.py
new file mode 100755
index 000000000..74e025ad7
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/test_prepare_lang.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/local/text2token.py b/egs/aidatatang_200zh/ASR/local/text2token.py
new file mode 100755
index 000000000..85047c367
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/local/text2token.py
@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument("--space", default="<space>", type=str, help="space symbol")
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument("text", type=str, default=False, nargs="?", help="input text")
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer)
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = "".join(a_flat)
+        print(a_chars)
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/prepare.sh b/egs/aidatatang_200zh/ASR/prepare.sh
new file mode 100755
index 000000000..40ee2eb97
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/prepare.sh
@@ -0,0 +1,120 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+perturb_speed=true
+
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/aidatatang_200zh
+#      You can find "corpus" and "transcript" inside it.
+#      You can download it at
+#       https://openslr.org/62/
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  if [ ! -f $dl_dir/aidatatang_200zh/transcript/aidatatang_200_zh_transcript.txt ]; then
+    lhotse download aidatatang-200zh $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare aidatatang_200zh manifest"
+  # We assume that you have downloaded the aidatatang_200zh corpus
+  # to $dl_dir/aidatatang_200zh
+  if [ ! -f data/manifests/aidatatang_200zh/.manifests.done ]; then
+    mkdir -p data/manifests/aidatatang_200zh
+    lhotse prepare aidatatang-200zh $dl_dir data/manifests/aidatatang_200zh
+    touch data/manifests/aidatatang_200zh/.manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests/
+    lhotse prepare musan $dl_dir/musan data/manifests/
+    touch data/manifests/.manifests.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for aidatatang_200zh"
+  if [ ! -f data/fbank/.aidatatang_200zh.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_aidatatang_200zh.py --perturb-speed ${perturb_speed}
+    touch data/fbank/.aidatatang_200zh.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+  # Prepare text.
+  # Note: in Linux, you can install jq with the following command:
+  # 1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+  # 2. chmod +x ./jq
+  # 3. cp jq /usr/bin
+  if [ ! -f $lang_char_dir/text ]; then
+    gunzip -c data/manifests/aidatatang_200zh/aidatatang_supervisions_train.jsonl.gz \
+      |jq '.text' |sed -e 's/["text:\t ]*//g' | sed 's/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text
+  fi
+  # Prepare words.txt
+  if [ ! -f $lang_char_dir/text_words ]; then
+    gunzip -c data/manifests/aidatatang_200zh/aidatatang_supervisions_train.jsonl.gz \
+      | jq '.text' | sed -e 's/["text:\t]*//g' | sed 's/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text_words
+  fi
+
+  cat $lang_char_dir/text_words | sed 's/ /\n/g' | sort -u | sed '/^$/d' \
+    | uniq > $lang_char_dir/words_no_ids.txt
+
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    ./local/prepare_char.py
+  fi
+fi
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/__init__.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 100644
index 000000000..d491996b2
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1,411 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    load_manifest,
+    load_manifest_lazy,
+    set_caching_enabled,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+set_caching_enabled(False)
+torch.set_num_threads(1)
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class Aidatatang_200zhAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/dev/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=False,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=300,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+                buffer_size=50000,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        dev_iter_dataset = IterableDatasetWrapper(
+            dataset=validate,
+            sampler=valid_sampler,
+        )
+        valid_dl = DataLoader(
+            dev_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        test_iter_dataset = IterableDatasetWrapper(
+            dataset=test,
+            sampler=sampler,
+        )
+        test_dl = DataLoader(
+            test_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aidatatang_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aidatatang_cuts_dev.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aidatatang_cuts_test.jsonl.gz"
+        )
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/beam_search.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/conformer.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/conformer.py
new file mode 120000
index 000000000..a65957180
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/decode.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..2512f233f
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,536 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When training with the L subset, usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import Aidatatang_200zhAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--batch",
+        type=int,
+        default=None,
+        help="It specifies the batch checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 50
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text).replace(" ", "")) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Aidatatang_200zhAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    elif params.batch is not None:
+        filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints([filenames], device=device))
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aidatatang_200zh = Aidatatang_200zhAsrDataModule(args)
+
+    dev_cuts = aidatatang_200zh.valid_cuts()
+    test_cuts = aidatatang_200zh.test_cuts()
+    dev_dl = aidatatang_200zh.valid_dataloaders(dev_cuts)
+    test_dl = aidatatang_200zh.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/decoder.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py
new file mode 100644
index 000000000..e348f7b2b
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,178 @@
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 29 \
+  --avg 19
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aidatatang_200zh/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/joiner.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/model.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/optim.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/pretrained.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/pretrained.py
new file mode 100644
index 000000000..75c316eaf
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/pretrained.py
@@ -0,0 +1,338 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		         2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(3) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
+./pruned_transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and modified_beam_search ",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/scaling.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py
new file mode 100644
index 000000000..fa809b768
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,958 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 2 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 250 \
+  --save-every-n 1000
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 2 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 250 \
+  --save-every-n 1000
+  --use-fp16 True
+
+"""
+
+import argparse
+import logging
+import os
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import Aidatatang_200zhAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12359,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+    Explanation of options saved in `params`:
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+        - best_train_epoch: It is the epoch that has the best training loss.
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+        - subsampling_factor:  The subsampling factor for the model.
+        - encoder_dim: Hidden dim for multi-head attention model.
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 10,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 400,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 200,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aidatatang_200zh = Aidatatang_200zhAsrDataModule(args)
+
+    train_cuts = aidatatang_200zh.train_cuts()
+    valid_cuts = aidatatang_200zh.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 10.0 seconds
+        #
+        # Caution: There is a reason to select 10.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 10.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    valid_dl = aidatatang_200zh.valid_dataloaders(valid_cuts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aidatatang_200zh.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if not params.print_diagnostics and params.start_batch == 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    Aidatatang_200zhAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aidatatang_200zh/ASR/shared b/egs/aidatatang_200zh/ASR/shared
new file mode 120000
index 000000000..3a3b28f96
--- /dev/null
+++ b/egs/aidatatang_200zh/ASR/shared
@@ -0,0 +1 @@
+../../../egs/aishell/ASR/shared
\ No newline at end of file
diff --git a/egs/aishell/ASR/README.md b/egs/aishell/ASR/README.md
new file mode 100644
index 000000000..176f065e5
--- /dev/null
+++ b/egs/aishell/ASR/README.md
@@ -0,0 +1,26 @@
+
+# Introduction
+
+Please refer to <https://k2-fsa.github.io/icefall/recipes/Non-streaming-ASR/aishell/index.html> for how to run models in this recipe.
+
+Aishell is an open-source Chinese Mandarin speech corpus published by Beijing Shell Shell Technology Co., Ltd.
+400 people from different accent areas in China are invited to participate in the recording, which is conducted in a quiet indoor environment using high fidelity microphone and downsampled to 16kHz. The manual transcription accuracy is above 95%, through professional speech annotation and strict quality inspection. The data is free for academic use. We hope to provide moderate amount of data for new researchers in the field of speech recognition.
+
+(From [Open Speech and Language Resources](https://www.openslr.org/33/))
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                    | Encoder   | Decoder            | Comment                                                                           |
+|------------------------------------|-----------|--------------------|-----------------------------------------------------------------------------------|
+| `transducer_stateless`             | Conformer | Embedding + Conv1d | with `k2.rnnt_loss`                                                               |
+| `transducer_stateless_modified`    | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer`                     |
+| `transducer_stateless_modified-2`  | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` + extra data      |
+| `pruned_transducer_stateless3`     | Conformer (reworked) | Embedding + Conv1d | pruned RNN-T + reworked model with random combiner + using aidatatang_20zh as extra data|
+| `pruned_transducer_stateless7`     | Zipformer | Embedding | pruned RNN-T + zipformer encoder + stateless decoder with context-size 1 |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/aishell/ASR/RESULTS.md b/egs/aishell/ASR/RESULTS.md
new file mode 100644
index 000000000..0b22f41a1
--- /dev/null
+++ b/egs/aishell/ASR/RESULTS.md
@@ -0,0 +1,834 @@
+## Results
+
+### Aishell training result (Stateless Transducer)
+
+#### Zipformer (Non-streaming)
+
+[./zipformer](./zipformer)
+
+It's reworked Zipformer with Pruned RNNT loss.
+**Caution**: It uses `--context-size=1`.
+
+##### normal-scaled model, number of model parameters: 73412551, i.e., 73.41 M
+
+|                        | test | dev  | comment                                 |
+|------------------------|------|------|-----------------------------------------|
+| greedy search          | 4.67 | 4.37 | --epoch 55 --avg 17                     |
+| modified beam search   | 4.40 | 4.13 | --epoch 55 --avg 17                     |
+| fast beam search       | 4.60 | 4.31 | --epoch 55 --avg 17                     |
+
+Command for training is:
+```bash
+./prepare.sh 
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./zipformer/train.py \
+  --world-size 2 \
+  --num-epochs 60 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --context-size 1 \
+  --enable-musan 0 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000 \
+  --enable-musan 0 \
+  --base-lr 0.045 \
+  --lr-batches 7500 \
+  --lr-epochs 18 \
+  --spec-aug-time-warp-factor 20
+```
+
+Command for decoding is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./zipformer/decode.py \
+    --epoch 55 \
+    --avg 17 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --context-size 1 \
+    --decoding-method $m
+done
+```
+Pretrained models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-2023-10-24>
+
+
+##### small-scaled model, number of model parameters: 30167139, i.e., 30.17 M
+
+|                        | test | dev  | comment                                 |
+|------------------------|------|------|-----------------------------------------|
+| greedy search          | 4.97 | 4.67 | --epoch 55 --avg 21                     |
+| modified beam search   | 4.67 | 4.40 | --epoch 55 --avg 21                     |
+| fast beam search       | 4.85 | 4.61 | --epoch 55 --avg 21                     |
+
+Command for training is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./zipformer/train.py \
+  --world-size 2 \
+  --num-epochs 60 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --context-size 1 \
+  --exp-dir zipformer/exp-small \
+  --enable-musan 0 \
+  --base-lr 0.045 \
+  --lr-batches 7500 \
+  --lr-epochs 18 \
+  --spec-aug-time-warp-factor 20 \
+  --num-encoder-layers 2,2,2,2,2,2 \
+  --feedforward-dim 512,768,768,768,768,768 \
+  --encoder-dim 192,256,256,256,256,256 \
+  --encoder-unmasked-dim 192,192,192,192,192,192 \
+  --max-duration 1200 
+```
+
+Command for decoding is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./zipformer/decode.py \
+    --epoch 55 \
+    --avg 21 \
+    --exp-dir ./zipformer/exp-small \
+    --lang-dir data/lang_char \
+    --context-size 1 \
+    --decoding-method $m \
+    --num-encoder-layers 2,2,2,2,2,2 \
+    --feedforward-dim 512,768,768,768,768,768 \
+    --encoder-dim 192,256,256,256,256,256 \
+    --encoder-unmasked-dim 192,192,192,192,192,192
+done
+```
+
+Pretrained models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-small-2023-10-24/>
+
+##### large-scaled model, number of model parameters: 157285130, i.e., 157.29 M
+
+|                        | test | dev  | comment                                 |
+|------------------------|------|------|-----------------------------------------|
+| greedy search          | 4.49 | 4.22 | --epoch 56 --avg 23                     |
+| modified beam search   | 4.28 | 4.03 | --epoch 56 --avg 23                     |
+| fast beam search       | 4.44 | 4.18 | --epoch 56 --avg 23                     |
+
+Command for training is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./zipformer/train.py \
+  --world-size 2 \
+  --num-epochs 60 \
+  --use-fp16 1 \
+  --context-size 1 \
+  --exp-dir ./zipformer/exp-large \
+  --enable-musan 0 \
+  --lr-batches 7500 \
+  --lr-epochs 18 \
+  --spec-aug-time-warp-factor 20 \
+  --num-encoder-layers 2,2,4,5,4,2 \
+  --feedforward-dim 512,768,1536,2048,1536,768 \
+  --encoder-dim 192,256,512,768,512,256 \
+  --encoder-unmasked-dim 192,192,256,320,256,192 \
+  --max-duration 800 
+```
+
+Command for decoding is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./zipformer/decode.py \
+    --epoch 56 \
+    --avg 23 \
+    --exp-dir ./zipformer/exp-large \
+    --lang-dir data/lang_char \
+    --context-size 1 \
+    --decoding-method $m \
+    --num-encoder-layers 2,2,4,5,4,2 \
+    --feedforward-dim 512,768,1536,2048,1536,768 \
+    --encoder-dim 192,256,512,768,512,256 \
+    --encoder-unmasked-dim 192,192,256,320,256,192 
+done
+```
+
+Pretrained models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-large-2023-10-24/>
+
+#### Pruned transducer stateless 7 streaming
+[./pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
+
+It's Streaming version of Zipformer1 with Pruned RNNT loss.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 6.95 | 6.29 | --epoch 44 --avg 15 --max-duration 600 |
+| modified beam search   | 6.51 | 5.90 | --epoch 44 --avg 15 --max-duration 600 |
+| fast beam search       | 6.73 | 6.09 | --epoch 44 --avg 15 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless7_streaming/train.py \
+    --world-size 2 \
+    --num-epochs 50 \
+    --use-fp16 1 \
+    --context-size 1 \
+    --max-duration 800 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --enable-musan 0 \
+    --spec-aug-time-warp-factor 20
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The decoding command is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 44 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --lang-dir data/lang_char \
+    --context-size 1 \
+    --decoding-method $m
+done
+```
+
+Pretrained models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-pruned-transducer-stateless7-streaming-2023-10-16/>
+
+
+
+#### Pruned transducer stateless 7
+
+[./pruned_transducer_stateless7](./pruned_transducer_stateless7)
+
+It's Zipformer with Pruned RNNT loss.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 5.02 | 4.61 | --epoch 42 --avg 6 --max-duration 600 |
+| modified beam search   | 4.81 | 4.4 | --epoch 42 --avg 6 --max-duration 600 |
+| fast beam search       | 4.91 | 4.52 | --epoch 42 --avg 6 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 2 \
+  --num-epochs 50 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --context-size 1 \
+  --max-duration 300
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/MHYo3ApfQxaCdYLr38cQOQ>
+
+The decoding command is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./pruned_transducer_stateless7/decode.py \
+    --epoch 42 \
+    --avg 6 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --lang-dir data/lang_char \
+    --max-duration 300 \
+    --context-size 1 \
+    --decoding-method $m
+
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/marcoyang/icefall-asr-aishell-zipformer-pruned-transducer-stateless7-2023-03-21>
+#### Pruned transducer stateless 7 (zipformer)
+
+See <https://github.com/k2-fsa/icefall/pull/986>
+
+[./pruned_transducer_stateless7_bbpe](./pruned_transducer_stateless7_bbpe)
+
+**Note**: The modeling units are byte level BPEs
+
+The best results I have gotten are:
+
+Vocab size | Greedy search(dev & test) | Modified beam search(dev & test) | Fast beam search (dev & test)  | Fast beam search LG (dev & test) | comments
+-- | -- | -- | -- | -- | --
+500 | 4.31 & 4.59 | 4.25 & 4.54 | 4.27 & 4.55 |  4.07 & 4.38 | --epoch 48 --avg 29
+
+The training command:
+
+```
+export CUDA_VISIBLE_DEVICES="4,5,6,7"
+
+./pruned_transducer_stateless7_bbpe/train.py \
+  --world-size 4 \
+  --num-epochs 50 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --max-duration 800 \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --exp-dir pruned_transducer_stateless7_bbpe/exp \
+  --lr-epochs 6 \
+  --master-port 12535
+```
+
+The decoding command:
+
+```
+for m in greedy_search modified_beam_search fast_beam_search fast_beam_search_LG; do
+    ./pruned_transducer_stateless7_bbpe/decode.py \
+      --epoch 48 \
+      --avg 29 \
+      --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+      --max-sym-per-frame 1 \
+      --ngram-lm-scale 0.25 \
+      --ilme-scale 0.2 \
+      --bpe-model data/lang_bbpe_500/bbpe.model \
+      --max-duration 2000 \
+      --decoding-method $m
+done
+```
+
+The pretrained model is available at: https://huggingface.co/pkufool/icefall_asr_aishell_pruned_transducer_stateless7_bbpe
+
+
+#### Pruned transducer stateless 3
+
+See <https://github.com/k2-fsa/icefall/pull/436>
+
+
+[./pruned_transducer_stateless3](./pruned_transducer_stateless3)
+
+It uses pruned RNN-T.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 5.39 | 5.09 | --epoch 29 --avg 5 --max-duration 600 |
+| modified beam search   | 5.05 | 4.79 | --epoch 29 --avg 5 --max-duration 600 |
+| modified beam search + RNNLM shallow fusion   | 4.73 | 4.53 | --epoch 29 --avg 5 --max-duration 600 |
+| modified beam search + LODR   | 4.57 | 4.37 | --epoch 29 --avg 5 --max-duration 600 |
+| fast beam search       | 5.13 | 4.91 | --epoch 29 --avg 5 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="4,5,6,7"
+
+./pruned_transducer_stateless3/train.py \
+  --exp-dir ./pruned_transducer_stateless3/exp-context-size-1 \
+  --world-size 4 \
+  --max-duration 200 \
+  --datatang-prob 0.5 \
+  --start-epoch 1 \
+  --num-epochs 30 \
+  --use-fp16 1 \
+  --num-encoder-layers 12 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --context-size 1 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --master-port 12356
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/OKKacljwR6ik7rbDr5gMqQ>
+
+The decoding command is:
+
+```bash
+for epoch in 29; do
+  for avg in 5; do
+    for m in greedy_search modified_beam_search fast_beam_search; do
+      ./pruned_transducer_stateless3/decode.py \
+        --exp-dir ./pruned_transducer_stateless3/exp-context-size-1 \
+        --epoch $epoch \
+        --avg $avg \
+        --use-averaged-model 1 \
+        --max-duration 600 \
+        --decoding-method $m \
+        --num-encoder-layers 12 \
+        --dim-feedforward 2048 \
+        --nhead 8 \
+        --context-size 1 \
+        --encoder-dim 512 \
+        --decoder-dim 512 \
+        --joiner-dim 512
+    done
+  done
+done
+```
+
+We provide the option of shallow fusion with a RNN language model. The pre-trained language model is
+available at <https://huggingface.co/marcoyang/icefall-aishell-rnn-lm>. To decode with the language model,
+please use the following command:
+
+```bash
+# download pre-trained model
+git lfs install
+git clone https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+
+aishell_exp=icefall-aishell-pruned-transducer-stateless3-2022-06-20/
+
+pushd ${aishell_exp}/exp
+ln -s pretrained-epoch-29-avg-5-torch-1.10.0.pt epoch-99.pt
+popd
+
+# download RNN LM
+git lfs install
+git clone https://huggingface.co/marcoyang/icefall-aishell-rnn-lm
+rnnlm_dir=icefall-aishell-rnn-lm
+
+# RNNLM shallow fusion
+for lm_scale in $(seq 0.26 0.02 0.34); do
+  python ./pruned_transducer_stateless3/decode.py \
+      --epoch 99 \
+      --avg 1 \
+      --lang-dir ${aishell_exp}/data/lang_char \
+      --exp-dir ${aishell_exp}/exp \
+      --use-averaged-model False \
+      --decoding-method modified_beam_search_lm_shallow_fusion \
+      --use-shallow-fusion 1 \
+      --lm-type rnn \
+      --lm-exp-dir ${rnnlm_dir}/exp \
+      --lm-epoch 99 \
+      --lm-scale $lm_scale \
+      --lm-avg 1 \
+      --rnn-lm-embedding-dim 2048 \
+      --rnn-lm-hidden-dim 2048 \
+      --rnn-lm-num-layers 2 \
+      --lm-vocab-size 4336
+done
+
+# RNNLM Low-order density ratio (LODR) with a 2-gram
+
+cp ${rnnlm_dir}/2gram.fst.txt ${aishell_exp}/data/lang_char/2gram.fst.txt
+
+for lm_scale in 0.48; do
+  for LODR_scale in -0.28; do
+    python ./pruned_transducer_stateless3/decode.py \
+        --epoch 99 \
+        --avg 1 \
+        --lang-dir ${aishell_exp}/data/lang_char \
+        --exp-dir ${aishell_exp}/exp \
+        --use-averaged-model False \
+        --decoding-method modified_beam_search_LODR \
+        --use-shallow-fusion 1 \
+        --lm-type rnn \
+        --lm-exp-dir ${rnnlm_dir}/exp \
+        --lm-epoch 99 \
+        --lm-scale $lm_scale \
+        --lm-avg 1 \
+        --rnn-lm-embedding-dim 2048 \
+        --rnn-lm-hidden-dim 2048 \
+        --rnn-lm-num-layers 2 \
+        --lm-vocab-size 4336 \
+        --tokens-ngram 2 \
+        --backoff-id 4336 \
+        --ngram-lm-scale $LODR_scale
+  done
+done
+
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20>
+
+We have a tutorial in [sherpa](https://github.com/k2-fsa/sherpa) about how
+to use the pre-trained model for non-streaming ASR. See
+<https://k2-fsa.github.io/sherpa/offline_asr/conformer/aishell.html>
+
+
+#### Pruned transducer stateless 2
+
+See https://github.com/k2-fsa/icefall/pull/536
+
+[./pruned_transducer_stateless2](./pruned_transducer_stateless2)
+
+It uses pruned RNN-T.
+
+|                      | test | dev  | comment                                |
+| -------------------- | ---- | ---- | -------------------------------------- |
+| greedy search        | 5.20 | 4.78 | --epoch 72 --avg 14 --max-duration 200 |
+| modified beam search | 5.07 | 4.63 | --epoch 72 --avg 14 --max-duration 200 |
+| fast beam search     | 5.13 | 4.70 | --epoch 72 --avg 14 --max-duration 200 |
+
+Training command is:
+
+```bash
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless2/train.py \
+        --world-size 2 \
+        --num-epochs 90 \
+        --start-epoch 0 \
+        --exp-dir pruned_transducer_stateless2/exp \
+        --max-duration 200 \
+```
+
+The tensorboard log is available at
+https://tensorboard.dev/experiment/QI3PVzrGRrebxpbWUPwmkA/
+
+The decoding command is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./pruned_transducer_stateless2/decode.py \
+    --epoch 72 \
+    --avg 14 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --lang-dir data/lang_char \
+    --max-duration 200 \
+    --decoding-method $m
+
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/teapoly/icefall-aishell-pruned-transducer-stateless2-2022-08-18>
+
+
+#### 2022-03-01
+
+[./transducer_stateless_modified-2](./transducer_stateless_modified-2)
+
+It uses [optimized_transducer](https://github.com/csukuangfj/optimized_transducer)
+for computing RNN-T loss.
+
+Stateless transducer + modified transducer + using [aidatatang_200zh](http://www.openslr.org/62/) as extra training data.
+
+
+|                        | test |comment                                                         |
+|------------------------|------|----------------------------------------------------------------|
+| greedy search          | 4.94 |--epoch 89, --avg 38, --max-duration 100, --max-sym-per-frame 1 |
+| modified beam search   | 4.68 |--epoch 89, --avg 38, --max-duration 100  --beam-size 4         |
+
+The training commands are:
+
+```bash
+cd egs/aishell/ASR
+./prepare.sh --stop-stage 6
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_stateless_modified-2/train.py \
+  --world-size 3 \
+  --num-epochs 90 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_modified-2/exp-2 \
+  --max-duration 250 \
+  --lr-factor 2.0 \
+  --context-size 2 \
+  --modified-transducer-prob 0.25 \
+  --datatang-prob 0.2
+```
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/oG72ZlWaSGua6fXkcGRRjA/>
+
+The commands for decoding are
+
+```bash
+# greedy search
+for epoch in 89; do
+  for avg in 38; do
+  ./transducer_stateless_modified-2/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless_modified-2/exp-2 \
+    --max-duration 100 \
+    --context-size 2 \
+    --decoding-method greedy_search \
+    --max-sym-per-frame 1
+  done
+done
+
+# modified beam search
+for epoch in 89; do
+  for avg in 38; do
+    ./transducer_stateless_modified-2/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless_modified-2/exp-2 \
+    --max-duration 100 \
+    --context-size 2 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+  done
+done
+```
+
+You can find a pre-trained model, decoding logs, and decoding results at
+<https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2-2022-03-01>
+
+#### 2022-03-01
+
+[./transducer_stateless_modified](./transducer_stateless_modified)
+
+Stateless transducer + modified transducer.
+
+|                        | test |comment                                                         |
+|------------------------|------|----------------------------------------------------------------|
+| greedy search          | 5.22 |--epoch 64, --avg 33, --max-duration 100, --max-sym-per-frame 1 |
+| modified beam search   | 5.02 |--epoch 64, --avg 33, --max-duration 100  --beam-size 4         |
+
+The training commands are:
+
+```bash
+cd egs/aishell/ASR
+./prepare.sh --stop-stage 6
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_stateless_modified/train.py \
+  --world-size 3 \
+  --num-epochs 90 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_modified/exp-4 \
+  --max-duration 250 \
+  --lr-factor 2.0 \
+  --context-size 2 \
+  --modified-transducer-prob 0.25
+```
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/C27M8YxRQCa1t2XglTqlWg/>
+
+The commands for decoding are
+
+```bash
+# greedy search
+for epoch in 64; do
+  for avg in 33; do
+  ./transducer_stateless_modified/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless_modified/exp-4 \
+    --max-duration 100 \
+    --context-size 2 \
+    --decoding-method greedy_search \
+    --max-sym-per-frame 1
+  done
+done
+
+# modified beam search
+for epoch in 64; do
+  for avg in 33; do
+    ./transducer_stateless_modified/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless_modified/exp-4 \
+    --max-duration 100 \
+    --context-size 2 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+  done
+done
+```
+
+You can find a pre-trained model, decoding logs, and decoding results at
+<https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2022-03-01>
+
+
+#### 2022-2-19
+(Duo Ma): The tensorboard log for training is available at https://tensorboard.dev/experiment/25PmX3MxSVGTdvIdhOwllw/#scalars
+You can find a pretrained model by visiting https://huggingface.co/shuanguanma/icefall_aishell_transducer_stateless_context_size2_epoch60_2022_2_19
+|                           | test |comment                                  |
+|---------------------------|------|-----------------------------------------|
+| greedy search             | 5.4 |--epoch 59, --avg 10, --max-duration 100  |
+| beam search               | 5.05|--epoch 59, --avg 10, --max-duration 100  |
+
+You can use the following commands to reproduce our results:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+python3 ./transducer_stateless/train.py \
+      --world-size 4 \
+      --num-epochs 60 \
+      --start-epoch 0 \
+      --exp-dir exp/transducer_stateless_context_size2 \
+      --max-duration 100 \
+      --lr-factor 2.5 \
+      --context-size 2
+
+lang_dir=data/lang_char
+dir=exp/transducer_stateless_context_size2
+python3 ./transducer_stateless/decode.py \
+       --epoch 59 \
+       --avg 10 \
+       --exp-dir $dir \
+       --lang-dir $lang_dir \
+       --decoding-method greedy_search \
+       --context-size 2 \
+       --max-sym-per-frame 3
+
+lang_dir=data/lang_char
+dir=exp/transducer_stateless_context_size2
+python3 ./transducer_stateless/decode.py \
+       --epoch 59 \
+       --avg 10 \
+       --exp-dir $dir \
+       --lang-dir $lang_dir \
+       --decoding-method beam_search \
+       --context-size 2 \
+       --max-sym-per-frame 3
+```
+
+### Aishell training results (Transducer-stateless)
+#### 2022-02-18
+(Pingfeng Luo) : The tensorboard log for training is available at <https://tensorboard.dev/experiment/k3QL6QMhRbCwCKYKM9po9w/>
+And pretrained model is available at <https://huggingface.co/pfluo/icefall-aishell-transducer-stateless-char-2021-12-29>
+
+||test|
+|--|--|
+|CER| 5.05% |
+
+You can use the following commands to reproduce our results:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
+./transducer_stateless/train.py \
+  --bucketing-sampler True \
+  --world-size 8 \
+  --lang-dir data/lang_char \
+  --num-epochs 60 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp_rnnt_k2 \
+  --max-duration 80 \
+  --lr-factor 3
+
+./transducer_stateless/decode.py \
+  --epoch 59 \
+  --avg 10 \
+  --lang-dir data/lang_char \
+  --exp-dir transducer_stateless/exp_rnnt_k2 \
+  --max-duration 100 \
+  --decoding-method beam_search \
+  --beam-size 4
+```
+
+### Aishell training results (Conformer-MMI)
+#### 2021-12-04
+(Pingfeng Luo): Result of <https://github.com/k2-fsa/icefall/pull/140>
+
+The tensorboard log for training is available at <https://tensorboard.dev/experiment/PSRYVbptRGynqpPRSykp1g>
+
+And pretrained model is available at <https://huggingface.co/pfluo/icefall_aishell_mmi_model>
+
+The best decoding results (CER) are listed below, we got this results by averaging models from epoch 61 to 85, and using `attention-decoder` decoder with num_paths equals to 100.
+
+||test|
+|--|--|
+|CER| 4.94% |
+
+||lm_scale|attention_scale|
+|--|--|--|
+|test|1.1|0.3|
+
+You can use the following commands to reproduce our results:
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+
+cd egs/aishell/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
+python conformer_mmi/train.py --bucketing-sampler True \
+                              --max-duration 200 \
+                              --start-epoch 0 \
+                              --num-epochs 90 \
+                              --world-size 8
+
+python conformer_mmi/decode.py --nbest-scale 0.5 \
+                               --epoch 85 \
+                               --avg 25 \
+                               --method attention-decoder \
+                               --max-duration 20 \
+                               --num-paths 100
+```
+
+### Aishell training results (Conformer-CTC)
+#### 2021-11-16
+(Wei Kang): Result of https://github.com/k2-fsa/icefall/pull/30
+
+Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc
+
+The best decoding results (CER) are listed below, we got this results by averaging models from epoch 60 to 84, and using `attention-decoder` decoder with num_paths equals to 100.
+
+||test|
+|--|--|
+|CER| 4.26% |
+
+To get more unique paths, we scaled the lattice.scores with 0.5 (see https://github.com/k2-fsa/icefall/pull/10#discussion_r690951662 for more details), we searched the lm_score_scale and attention_score_scale for best results, the scales that produced the CER above are also listed below.
+
+||lm_scale|attention_scale|
+|--|--|--|
+|test|0.3|0.9|
+
+You can use the following commands to reproduce our results:
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+
+cd egs/aishell/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+python conformer_ctc/train.py --bucketing-sampler True \
+                              --max-duration 200 \
+                              --start-epoch 0 \
+                              --num-epochs 90 \
+                              --world-size 4
+
+python conformer_ctc/decode.py --nbest-scale 0.5 \
+                               --epoch 84 \
+                               --avg 25 \
+                               --method attention-decoder \
+                               --max-duration 20 \
+                               --num-paths 100
+```
+
+### Aishell training results (Tdnn-Lstm)
+#### 2021-09-13
+
+(Wei Kang): Result of phone based Tdnn-Lstm model, https://github.com/k2-fsa/icefall/pull/30
+
+Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc_lstm_ctc
+
+The best decoding results (CER) are listed below, we got this results by averaging models from epoch 19 to 8, and using `1best` decoding method.
+
+||test|
+|--|--|
+|CER| 10.16% |
diff --git a/egs/aishell/ASR/conformer_ctc/README.md b/egs/aishell/ASR/conformer_ctc/README.md
new file mode 100644
index 000000000..50596ee92
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/README.md
@@ -0,0 +1,4 @@
+
+Please visit
+<https://icefall.readthedocs.io/en/latest/recipes/aishell/conformer_ctc.html>
+for how to run this recipe.
diff --git a/egs/aishell/ASR/conformer_ctc/__init__.py b/egs/aishell/ASR/conformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/conformer_ctc/asr_datamodule.py b/egs/aishell/ASR/conformer_ctc/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/conformer.py b/egs/aishell/ASR/conformer_ctc/conformer.py
new file mode 100644
index 000000000..ab1cbbae4
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/conformer.py
@@ -0,0 +1,895 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+        use_feat_batchnorm(bool): whether to use batch-normalize the input.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self, x: Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is [N, T, C].
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        x = self.activation(self.norm(x))
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/aishell/ASR/conformer_ctc/decode.py b/egs/aishell/ASR/conformer_ctc/decode.py
new file mode 100755
index 000000000..74a7b5933
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/decode.py
@@ -0,0 +1,568 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=49,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It maps the tokens ids to
+              tokens using token symbol tabel directly.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) attention-decoder. Extract n paths from the lattice,
+              the path with the highest score is the decoding result.
+            - (4) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The LM dir.
+        It should contain either G_3_gram.pt or G_3_gram.fst.txt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "nhead": 4,
+            "attention_dim": 512,
+            "num_encoder_layers": 12,
+            "num_decoder_layers": 6,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            # parameters for decoder
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    batch: dict,
+    lexicon: Lexicon,
+    sos_id: int,
+    eos_id: int,
+) -> Dict[str, List[List[int]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if decoding method is 1best, the key is the string `no_rescore`.
+               If attention rescoring is used, the key is the string
+               `ngram_lm_scale_xxx_attention_scale_xxx`, where `xxx` is the
+               value of `lm_scale` and `attention_scale`. An example key is
+               `ngram_lm_scale_0.7_attention_scale_0.5`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "attention-decoder", it uses attention rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains the token symbol table and the word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        key = "ctc-decoding"
+        hyps = [[lexicon.token_table[i] for i in ids] for ids in token_ids]
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=lexicon.word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method == "attention-decoder"
+
+    best_path_dict = rescore_with_attention_decoder(
+        lattice=lattice,
+        num_paths=params.num_paths,
+        model=model,
+        memory=memory,
+        memory_key_padding_mask=memory_key_padding_mask,
+        sos_id=sos_id,
+        eos_id=eos_id,
+        nbest_scale=params.nbest_scale,
+    )
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    lexicon: Lexicon,
+    sos_id: int,
+    eos_id: int,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      lexicon:
+        It contains the token symbol table and the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if the decoding method is
+      1best or it may be "ngram_lm_scale_0.7_attention_scale_0.5" if attention
+      rescoring is used. Its value is a list of tuples. Each tuple contains two
+      elements: The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            batch=batch,
+            lexicon=lexicon,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"cer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+    else:
+        H = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dls = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            lexicon=lexicon,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/conformer_ctc/export.py b/egs/aishell/ASR/conformer_ctc/export.py
new file mode 100755
index 000000000..49871d437
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/export.py
@@ -0,0 +1,162 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from conformer import Conformer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import AttributeDict, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=84,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=25,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 4,
+            "num_decoder_layers": 6,
+        }
+    )
+    return params
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    num_classes = num_tokens(token_table) + 1  # +1 for the blank
+
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
new file mode 120000
index 000000000..896b78aef
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
new file mode 120000
index 000000000..aa1b6073d
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
new file mode 120000
index 000000000..0cf42ce30
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/label_smoothing.py b/egs/aishell/ASR/conformer_ctc/label_smoothing.py
new file mode 120000
index 000000000..e9d239fff
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/label_smoothing.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/pretrained.py b/egs/aishell/ASR/conformer_ctc/pretrained.py
new file mode 100755
index 000000000..66d583396
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/pretrained.py
@@ -0,0 +1,374 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from conformer import Conformer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+)
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens-file",
+        type=str,
+        help="Path to tokens.txt" "Used only when method is ctc-decoding",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="Path to words.txt" "Used when method is NOT ctc-decoding",
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="Path to HLG.pt." "Used when method is NOT ctc-decoding",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use ctc decoding. It maps the tokens ids to tokens
+            using the token symbol table directly.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) attention-decoder - Extract n paths from the rescored
+            lattice and use the transformer attention decoder for
+            rescoring.
+            We call it HLG decoding + n-gram LM rescoring + attention
+            decoder rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.3,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--attention-decoder-scale",
+        type=float,
+        default=0.9,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for attention decoder scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sos-id",
+        type=int,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the SOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "--eos-id",
+        type=int,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the EOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "--num_classes",
+        type=int,
+        default=4336,
+        help="The Vocab size.",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "nhead": 4,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            # parameters for deocding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    if args.method != "attention-decoder":
+        # to save memory as the attention decoder
+        # will not be used
+        params.num_decoder_layers = 0
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    # Note: We don't use key padding mask for attention during decoding
+    with torch.no_grad():
+        nnet_output, memory, memory_key_padding_mask = model(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        token_sym_table = k2.SymbolTable.from_file(params.tokens_file)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=True,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = [[token_sym_table[i] for i in ids] for ids in token_ids]
+    elif params.method in ["1best", "attention-decoder"]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        elif params.method == "attention-decoder":
+            logging.info("Use HLG + attention decoder rescoring")
+            best_path_dict = rescore_with_attention_decoder(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                model=model,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                sos_id=params.sos_id,
+                eos_id=params.eos_id,
+                nbest_scale=params.nbest_scale,
+                ngram_lm_scale=params.ngram_lm_scale,
+                attention_scale=params.attention_decoder_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/conformer_ctc/subsampling.py b/egs/aishell/ASR/conformer_ctc/subsampling.py
new file mode 100644
index 000000000..8e0f73d05
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/subsampling.py
@@ -0,0 +1,153 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x
diff --git a/egs/aishell/ASR/conformer_ctc/test_subsampling.py b/egs/aishell/ASR/conformer_ctc/test_subsampling.py
new file mode 100755
index 000000000..81fa234dd
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/test_subsampling.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+
+def test_conv2d_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = Conv2dSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
+
+
+def test_vgg_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = VggSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
diff --git a/egs/aishell/ASR/conformer_ctc/test_transformer.py b/egs/aishell/ASR/conformer_ctc/test_transformer.py
new file mode 100755
index 000000000..7c0695683
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/test_transformer.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from torch.nn.utils.rnn import pad_sequence
+from transformer import (
+    Transformer,
+    add_eos,
+    add_sos,
+    decoder_padding_mask,
+    encoder_padding_mask,
+    generate_square_subsequent_mask,
+)
+
+
+def test_encoder_padding_mask():
+    supervisions = {
+        "sequence_idx": torch.tensor([0, 1, 2]),
+        "start_frame": torch.tensor([0, 0, 0]),
+        "num_frames": torch.tensor([18, 7, 13]),
+    }
+
+    max_len = ((18 - 1) // 2 - 1) // 2
+    mask = encoder_padding_mask(max_len, supervisions)
+    expected_mask = torch.tensor(
+        [
+            [False, False, False],  # ((18 - 1)//2 - 1)//2 = 3,
+            [False, True, True],  # ((7 - 1)//2 - 1)//2 = 1,
+            [False, False, True],  # ((13 - 1)//2 - 1)//2 = 2,
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_transformer():
+    num_features = 40
+    num_classes = 87
+    model = Transformer(num_features=num_features, num_classes=num_classes)
+
+    N = 31
+
+    for T in range(7, 30):
+        x = torch.rand(N, T, num_features)
+        y, _, _ = model(x)
+        assert y.shape == (N, (((T - 1) // 2) - 1) // 2, num_classes)
+
+
+def test_generate_square_subsequent_mask():
+    s = 5
+    mask = generate_square_subsequent_mask(s)
+    inf = float("inf")
+    expected_mask = torch.tensor(
+        [
+            [0.0, -inf, -inf, -inf, -inf],
+            [0.0, 0.0, -inf, -inf, -inf],
+            [0.0, 0.0, 0.0, -inf, -inf],
+            [0.0, 0.0, 0.0, 0.0, -inf],
+            [0.0, 0.0, 0.0, 0.0, 0.0],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_decoder_padding_mask():
+    x = [torch.tensor([1, 2]), torch.tensor([3]), torch.tensor([2, 5, 8])]
+    y = pad_sequence(x, batch_first=True, padding_value=-1)
+    mask = decoder_padding_mask(y, ignore_id=-1)
+    expected_mask = torch.tensor(
+        [[False, False, True], [False, True, True], [False, False, False]]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_add_sos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_sos(x, sos_id=0)
+    expected_y = [[0, 1, 2], [0, 3], [0, 2, 5, 8]]
+    assert y == expected_y
+
+
+def test_add_eos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_eos(x, eos_id=0)
+    expected_y = [[1, 2, 0], [3, 0], [2, 5, 8, 0]]
+    assert y == expected_y
diff --git a/egs/aishell/ASR/conformer_ctc/train.py b/egs/aishell/ASR/conformer_ctc/train.py
new file mode 100755
index 000000000..c2cbe6e3b
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/train.py
@@ -0,0 +1,681 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+from lhotse.utils import fix_random_seed
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=90,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.7,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - attention_dim: Attention dimension.
+
+        - nhead: Number of heads in multi-head attention.
+                 Must satisfy attention_dim // nhead == 0.
+
+        - num_encoder_layers: Number of attention encoder layers.
+
+        - num_decoder_layers: Number of attention decoder layers.
+
+        - use_feat_batchnorm: Whether to do normalization in the input layer.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - lr_factor: The lr_factor for the optimizer.
+
+        - warm_step: The warm_step for the optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 2000,
+            # parameters for k2.ctc_loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "attention_dim": 512,
+            "nhead": 4,
+            "num_encoder_layers": 12,
+            "num_decoder_layers": 6,
+            "use_feat_batchnorm": True,
+            # parameters for Noam
+            "weight_decay": 1e-5,
+            "lr_factor": 5.0,
+            "warm_step": 36000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[torch.Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    token_ids = graph_compiler.texts_to_ids(texts)
+
+    decoding_graph = graph_compiler.compile(token_ids)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AishellAsrDataModule(args)
+    train_dl = aishell.train_dataloaders(aishell.train_cuts())
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/conformer_ctc/transformer.py b/egs/aishell/ASR/conformer_ctc/transformer.py
new file mode 100644
index 000000000..a3e50e385
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/transformer.py
@@ -0,0 +1,924 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from label_smoothing import LabelSmoothingLoss
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        if use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision)
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/aishell/ASR/conformer_mmi/__init__.py b/egs/aishell/ASR/conformer_mmi/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/conformer_mmi/asr_datamodule.py b/egs/aishell/ASR/conformer_mmi/asr_datamodule.py
new file mode 120000
index 000000000..a73848de9
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/asr_datamodule.py
@@ -0,0 +1 @@
+../conformer_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_mmi/conformer.py b/egs/aishell/ASR/conformer_mmi/conformer.py
new file mode 100644
index 000000000..ab1cbbae4
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/conformer.py
@@ -0,0 +1,895 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+        use_feat_batchnorm(bool): whether to use batch-normalize the input.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self, x: Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is [N, T, C].
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        x = self.activation(self.norm(x))
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/aishell/ASR/conformer_mmi/decode.py b/egs/aishell/ASR/conformer_mmi/decode.py
new file mode 100755
index 000000000..20a855e7f
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/decode.py
@@ -0,0 +1,585 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+# Copyright 2021                    Pingfeng Luo
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=49,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It maps the tokens ids to
+              tokens using token symbol tabel directly.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) attention-decoder. Extract n paths from the lattice,
+              the path with the highest score is the decoding result.
+            - (4) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        conformer_ctc/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_phone",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The LM dir.
+        It should contain either G_3_gram.pt or G_3_gram.fst.txt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "nhead": 4,
+            "attention_dim": 512,
+            "num_encoder_layers": 12,
+            "num_decoder_layers": 6,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            # parameters for decoder
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    batch: dict,
+    lexicon: Lexicon,
+    sos_id: int,
+    eos_id: int,
+) -> Dict[str, List[List[int]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if decoding method is 1best, the key is the string `no_rescore`.
+               If attention rescoring is used, the key is the string
+               `ngram_lm_scale_xxx_attention_scale_xxx`, where `xxx` is the
+               value of `lm_scale` and `attention_scale`. An example key is
+               `ngram_lm_scale_0.7_attention_scale_0.5`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "attention-decoder", it uses attention rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains the token symbol table and the word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        key = "ctc-decoding"
+        hyps = [[lexicon.token_table[i] for i in ids] for ids in token_ids]
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=lexicon.word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method == "attention-decoder"
+
+    best_path_dict = rescore_with_attention_decoder(
+        lattice=lattice,
+        num_paths=params.num_paths,
+        model=model,
+        memory=memory,
+        memory_key_padding_mask=memory_key_padding_mask,
+        sos_id=sos_id,
+        eos_id=eos_id,
+        nbest_scale=params.nbest_scale,
+    )
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    lexicon: Lexicon,
+    sos_id: int,
+    eos_id: int,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      lexicon:
+        It contains the token symbol table and the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if the decoding method is
+      1best or it may be "ngram_lm_scale_0.7_attention_scale_0.5" if attention
+      rescoring is used. Its value is a list of tuples. Each tuple contains two
+      elements: The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            batch=batch,
+            lexicon=lexicon,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"cer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = MmiTrainingGraphCompiler(
+        args.lang_dir,
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+    else:
+        H = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dls = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            lexicon=lexicon,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/conformer_mmi/label_smoothing.py b/egs/aishell/ASR/conformer_mmi/label_smoothing.py
new file mode 120000
index 000000000..08734abd7
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/label_smoothing.py
@@ -0,0 +1 @@
+../conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_mmi/subsampling.py b/egs/aishell/ASR/conformer_mmi/subsampling.py
new file mode 100644
index 000000000..398837a46
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/subsampling.py
@@ -0,0 +1,153 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape [N, T, idim] to an output
+    with shape [N, T', odim], where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is [N, T, idim].
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim]
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is [N, T, idim].
+
+        Returns:
+          Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
+        """
+        # On entry, x is [N, T, idim]
+        x = x.unsqueeze(1)  # [N, T, idim] -> [N, 1, T, idim] i.e., [N, C, H, W]
+        x = self.conv(x)
+        # Now x is of shape [N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2]
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape [N, ((T-1)//2 - 1))//2, odim]
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape [N, T, idim] to an output
+    with shape [N, T', odim], where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is [N, T, idim].
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim]
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is [N, T, idim].
+
+        Returns:
+          Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x
diff --git a/egs/aishell/ASR/conformer_mmi/train.py b/egs/aishell/ASR/conformer_mmi/train.py
new file mode 100755
index 000000000..09cd6e60c
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/train.py
@@ -0,0 +1,678 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+# Copyright    2021                       Pingfeng Luo
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+from lhotse.utils import fix_random_seed
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.mmi import LFMMILoss
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=90,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_mmi/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_phone",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.7,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * mmi_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - attention_dim: Attention dimension.
+
+        - nhead: Number of heads in multi-head attention.
+                 Must satisfy attention_dim // nhead == 0.
+
+        - num_encoder_layers: Number of attention encoder layers.
+
+        - num_decoder_layers: Number of attention decoder layers.
+
+        - use_feat_batchnorm: Whether to do normalization in the input layer.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - lr_factor: The lr_factor for the optimizer.
+
+        - warm_step: The warm_step for the optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for k2.ctc_loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "attention_dim": 512,
+            "nhead": 4,
+            "num_encoder_layers": 12,
+            "num_decoder_layers": 6,
+            "use_feat_batchnorm": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "lr_factor": 5.0,
+            "warm_step": 80000,
+            "use_pruned_intersect": False,
+            "den_scale": 1.0,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: MmiTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[torch.Tensor, MetricsTracker]:
+    """
+    Compute LF-MMI loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+        # NOTE: We need `encode_supervisions` to sort sequences with
+        # different duration in decreasing order, required by
+        # `k2.intersect_dense` called in `LFMMILoss.forward()`
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    loss_fn = LFMMILoss(
+        graph_compiler=graph_compiler,
+        den_scale=params.den_scale,
+        use_pruned_intersect=params.use_pruned_intersect,
+    )
+
+    mmi_loss = loss_fn(dense_fsa_vec=dense_fsa_vec, texts=texts)
+
+    if params.att_rate != 0.0:
+        token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * mmi_loss + params.att_rate * att_loss
+    else:
+        loss = mmi_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["mmi_loss"] = mmi_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: MmiTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: MmiTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+
+    logging.info("About to create model")
+    if params.att_rate == 0:
+        assert params.num_decoder_layers == 0, f"{params.num_decoder_layers}"
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints and checkpoints["optimizer"]:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AishellAsrDataModule(args)
+    train_cuts = aishell.train_cuts()
+    train_dl = aishell.train_dataloaders(train_cuts)
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/conformer_mmi/transformer.py b/egs/aishell/ASR/conformer_mmi/transformer.py
new file mode 100644
index 000000000..a3e50e385
--- /dev/null
+++ b/egs/aishell/ASR/conformer_mmi/transformer.py
@@ -0,0 +1,924 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from label_smoothing import LabelSmoothingLoss
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        if use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision)
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/aishell/ASR/local/__init__.py b/egs/aishell/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/local/compile_hlg.py b/egs/aishell/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/aishell/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/compile_lg.py b/egs/aishell/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/aishell/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/compute_fbank_aidatatang_200zh.py b/egs/aishell/ASR/local/compute_fbank_aidatatang_200zh.py
new file mode 100755
index 000000000..6a9bb4f42
--- /dev/null
+++ b/egs/aishell/ASR/local/compute_fbank_aidatatang_200zh.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the aidatatang_200zh dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_aidatatang_200zh(num_mel_bins: int = 80, perturb_speed: bool = False):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "test",
+        "dev",
+    )
+    prefix = "aidatatang"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+
+            for sup in m["supervisions"]:
+                sup.custom = {"origin": "aidatatang_200zh"}
+
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and perturb_speed:
+                logging.info(f"Doing speed perturb")
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_aidatatang_200zh(
+        num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed
+    )
diff --git a/egs/aishell/ASR/local/compute_fbank_aishell.py b/egs/aishell/ASR/local/compute_fbank_aishell.py
new file mode 100755
index 000000000..c7000da1c
--- /dev/null
+++ b/egs/aishell/ASR/local/compute_fbank_aishell.py
@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the aishell dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_aishell(num_mel_bins: int = 80, perturb_speed: bool = False):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+    prefix = "aishell"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and perturb_speed:
+                logging.info(f"Doing speed perturb")
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_aishell(
+        num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed
+    )
diff --git a/egs/aishell/ASR/local/compute_fbank_musan.py b/egs/aishell/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/aishell/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/convert_transcript_words_to_tokens.py b/egs/aishell/ASR/local/convert_transcript_words_to_tokens.py
new file mode 120000
index 000000000..2ce13fd69
--- /dev/null
+++ b/egs/aishell/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/convert_transcript_words_to_tokens.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/display_manifest_statistics.py b/egs/aishell/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..c478f7331
--- /dev/null
+++ b/egs/aishell/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in transducer_stateless/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    #  path = "./data/fbank/aishell_cuts_train.jsonl.gz"
+    #  path = "./data/fbank/aishell_cuts_test.jsonl.gz"
+    path = "./data/fbank/aishell_cuts_dev.jsonl.gz"
+    #  path = "./data/fbank/aidatatang_cuts_train.jsonl.gz"
+    #  path = "./data/fbank/aidatatang_cuts_test.jsonl.gz"
+    #  path = "./data/fbank/aidatatang_cuts_dev.jsonl.gz"
+
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+## train (after speed perturb)
+Cuts count: 360294
+Total duration (hours): 455.6
+Speech duration (hours): 455.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    4.6
+std     1.4
+min     1.1
+0.1%    1.8
+0.5%    2.2
+1%      2.3
+5%      2.7
+10%     3.0
+10%     3.0
+25%     3.5
+50%     4.3
+75%     5.4
+90%     6.5
+95%     7.2
+99%     8.8
+99.5%   9.4
+99.9%   10.9
+max     16.1
+
+## test
+Cuts count: 7176
+Total duration (hours): 10.0
+Speech duration (hours): 10.0 (100.0%)
+***
+Duration statistics (seconds):
+mean    5.0
+std     1.6
+min     1.9
+0.1%    2.2
+0.5%    2.4
+1%      2.6
+5%      3.0
+10%     3.2
+10%     3.2
+25%     3.8
+50%     4.7
+75%     5.9
+90%     7.3
+95%     8.2
+99%     9.9
+99.5%   10.7
+99.9%   11.9
+max     14.7
+
+## dev
+Cuts count: 14326
+Total duration (hours): 18.1
+Speech duration (hours): 18.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    4.5
+std     1.3
+min     1.6
+0.1%    2.1
+0.5%    2.3
+1%      2.4
+5%      2.9
+10%     3.1
+10%     3.1
+25%     3.5
+50%     4.3
+75%     5.4
+90%     6.4
+95%     7.0
+99%     8.4
+99.5%   8.9
+99.9%   10.3
+max     12.5
+
+## aidatatang_200zh (train)
+Cuts count: 164905
+Total duration (hours): 139.9
+Speech duration (hours): 139.9 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.1
+std     1.1
+min     1.1
+0.1%    1.5
+0.5%    1.7
+1%      1.8
+5%      2.0
+10%     2.1
+10%     2.1
+25%     2.3
+50%     2.7
+75%     3.4
+90%     4.6
+95%     5.4
+99%     7.1
+99.5%   7.8
+99.9%   9.1
+max     16.3
+
+## aidatatang_200zh (test)
+Cuts count: 48144
+Total duration (hours): 40.2
+Speech duration (hours): 40.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.0
+std     1.1
+min     0.9
+0.1%    1.5
+0.5%    1.8
+1%      1.8
+5%      2.0
+10%     2.1
+10%     2.1
+25%     2.3
+50%     2.6
+75%     3.4
+90%     4.4
+95%     5.2
+99%     6.9
+99.5%   7.5
+99.9%   9.0
+max     21.8
+
+## aidatatang_200zh (dev)
+Cuts count: 24216
+Total duration (hours): 20.2
+Speech duration (hours): 20.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.0
+std     1.0
+min     1.2
+0.1%    1.6
+0.5%    1.7
+1%      1.8
+5%      2.0
+10%     2.1
+10%     2.1
+25%     2.3
+50%     2.7
+75%     3.4
+90%     4.4
+95%     5.1
+99%     6.7
+99.5%   7.3
+99.9%   8.8
+max     11.3
+"""
diff --git a/egs/aishell/ASR/local/generate_unique_lexicon.py b/egs/aishell/ASR/local/generate_unique_lexicon.py
new file mode 120000
index 000000000..c0aea1403
--- /dev/null
+++ b/egs/aishell/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/generate_unique_lexicon.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/prepare_char.py b/egs/aishell/ASR/local/prepare_char.py
new file mode 100755
index 000000000..8cc0502c2
--- /dev/null
+++ b/egs/aishell/ASR/local/prepare_char.py
@@ -0,0 +1,259 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = list(line)
+            for char in chars:
+                if char not in tokens:
+                    tokens[char] = len(tokens)
+    return tokens
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/local/prepare_char_lm_training_data.py b/egs/aishell/ASR/local/prepare_char_lm_training_data.py
new file mode 100755
index 000000000..e7995680b
--- /dev/null
+++ b/egs/aishell/ASR/local/prepare_char_lm_training_data.py
@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Daniel Povey
+#                                                   Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script takes a `tokens.txt` and a text file such as
+./download/lm/aishell-transcript.txt
+and outputs the LM training data to a supplied directory such
+as data/lm_training_char.  The format is as follows:
+It creates a PyTorch archive (.pt file), say data/lm_training.pt, which is a
+representation of a dict with the same format with librispeech receipe
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-char",
+        type=str,
+        help="""Lang dir of asr model, e.g. data/lang_char""",
+    )
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        help="""Input LM training data as text, e.g.
+        download/lm/aishell-train-word.txt""",
+    )
+    parser.add_argument(
+        "--lm-archive",
+        type=str,
+        help="""Path to output archive, e.g. data/lm_training_char/lm_data.pt;
+        look at the source of this script to see the format.""",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    if Path(args.lm_archive).exists():
+        logging.warning(f"{args.lm_archive} exists - skipping")
+        return
+
+    # make token_dict from tokens.txt in order to map characters to tokens.
+    token_dict = {}
+    token_file = args.lang_char + "/tokens.txt"
+
+    with open(token_file, "r") as f:
+        for line in f.readlines():
+            line_list = line.split()
+            token_dict[line_list[0]] = int(line_list[1])
+
+    # word2index is a dictionary from words to integer ids.  No need to reserve
+    # space for epsilon, etc.; the words are just used as a convenient way to
+    # compress the sequences of tokens.
+    word2index = dict()
+
+    word2token = []  # Will be a list-of-list-of-int, representing tokens.
+    sentences = []  # Will be a list-of-list-of-int, representing word-ids.
+
+    if "aishell-lm" in args.lm_data:
+        num_lines_in_total = 120098.0
+        step = 50000
+    elif "valid" in args.lm_data:
+        num_lines_in_total = 14326.0
+        step = 3000
+    elif "test" in args.lm_data:
+        num_lines_in_total = 7176.0
+        step = 3000
+    else:
+        num_lines_in_total = None
+        step = None
+
+    processed = 0
+
+    with open(args.lm_data) as f:
+        while True:
+            line = f.readline()
+            if line == "":
+                break
+
+            if step and processed % step == 0:
+                logging.info(
+                    f"Processed number of lines: {processed} "
+                    f"({processed / num_lines_in_total * 100: .3f}%)"
+                )
+            processed += 1
+
+            line_words = line.split()
+            for w in line_words:
+                if w not in word2index:
+                    w_token = []
+                    for t in w:
+                        if t in token_dict:
+                            w_token.append(token_dict[t])
+                        else:
+                            w_token.append(token_dict["<unk>"])
+                    word2index[w] = len(word2token)
+                    word2token.append(w_token)
+            sentences.append([word2index[w] for w in line_words])
+
+    logging.info("Constructing ragged tensors")
+    words = k2.ragged.RaggedTensor(word2token)
+    sentences = k2.ragged.RaggedTensor(sentences)
+
+    output = dict(words=words, sentences=sentences)
+
+    num_sentences = sentences.dim0
+    logging.info(f"Computing sentence lengths, num_sentences: {num_sentences}")
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        if step and i % step == 0:
+            logging.info(
+                f"Processed number of lines: {i} ({i / num_sentences * 100: .3f}%)"
+            )
+
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    output["sentence_lengths"] = torch.tensor(sentence_lengths, dtype=torch.int32)
+
+    torch.save(output, args.lm_archive)
+    logging.info(f"Saved to {args.lm_archive}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/aishell/ASR/local/prepare_lang.py b/egs/aishell/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..c8cf9b881
--- /dev/null
+++ b/egs/aishell/ASR/local/prepare_lang.py
@@ -0,0 +1,388 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone")
+    return parser.parse_args()
+
+
+def main():
+    out_dir = Path(get_args().lang_dir)
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/local/prepare_lang_bbpe.py b/egs/aishell/ASR/local/prepare_lang_bbpe.py
new file mode 100755
index 000000000..ddd90622e
--- /dev/null
+++ b/egs/aishell/ASR/local/prepare_lang_bbpe.py
@@ -0,0 +1,267 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bbpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.byte_utils import byte_encode
+from icefall.utils import str2bool, tokenize_by_CJK_char
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_lexicon(
+    model_file: str, words: List[str], oov: str
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    encode_words = [byte_encode(tokenize_by_CJK_char(w)) for w in words]
+    words_pieces_ids: List[List[int]] = sp.encode(encode_words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
+
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bbpe.model"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", args.oov, "#0", "<s>", "</s>"]
+
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_lexicon(model_file, words, args.oov)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/local/prepare_lang_fst.py b/egs/aishell/ASR/local/prepare_lang_fst.py
new file mode 120000
index 000000000..c5787c534
--- /dev/null
+++ b/egs/aishell/ASR/local/prepare_lang_fst.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_fst.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/sort_lm_training_data.py b/egs/aishell/ASR/local/sort_lm_training_data.py
new file mode 120000
index 000000000..1d6ccbe33
--- /dev/null
+++ b/egs/aishell/ASR/local/sort_lm_training_data.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/sort_lm_training_data.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/test_prepare_lang.py b/egs/aishell/ASR/local/test_prepare_lang.py
new file mode 100755
index 000000000..74e025ad7
--- /dev/null
+++ b/egs/aishell/ASR/local/test_prepare_lang.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/local/train_bbpe_model.py b/egs/aishell/ASR/local/train_bbpe_model.py
new file mode 100755
index 000000000..48160897d
--- /dev/null
+++ b/egs/aishell/ASR/local/train_bbpe_model.py
@@ -0,0 +1,112 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+import tempfile
+from pathlib import Path
+
+import sentencepiece as spm
+
+from icefall import byte_encode, tokenize_by_CJK_char
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def _convert_to_bchar(in_path: str, out_path: str):
+    with open(out_path, "w") as f:
+        for line in open(in_path, "r").readlines():
+            f.write(byte_encode(tokenize_by_CJK_char(line)) + "\n")
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    model_file = Path(model_prefix + ".model")
+    if model_file.is_file():
+        print(f"{model_file} exists - skipping")
+        return
+
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    temp = tempfile.NamedTemporaryFile()
+    train_text = temp.name
+
+    _convert_to_bchar(args.transcript, train_text)
+
+    spm.SentencePieceTrainer.train(
+        input=train_text,
+        vocab_size=vocab_size,
+        model_type=model_type,
+        model_prefix=model_prefix,
+        input_sentence_size=input_sentence_size,
+        character_coverage=character_coverage,
+        user_defined_symbols=user_defined_symbols,
+        unk_id=unk_id,
+        bos_id=-1,
+        eos_id=-1,
+    )
+
+    shutil.copyfile(model_file, f"{lang_dir}/bbpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/prepare.sh b/egs/aishell/ASR/prepare.sh
new file mode 100755
index 000000000..9f73a2073
--- /dev/null
+++ b/egs/aishell/ASR/prepare.sh
@@ -0,0 +1,378 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=11
+perturb_speed=true
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/aishell
+#      You can find data_aishell, resource_aishell inside it.
+#      You can download them from https://www.openslr.org/33
+#
+#  - $dl_dir/lm
+#      This directory contains the language model downloaded from
+#        https://huggingface.co/pkufool/aishell_lm
+#
+#        - 3-gram.unpruned.arpa
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bbpe_xxx,
+# data/lang_bbpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/aishell,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/aishell $dl_dir/aishell
+  #
+  # The directory structure is
+  # aishell/
+  # |-- data_aishell
+  # |   |-- transcript
+  # |   `-- wav
+  # `-- resource_aishell
+  #     |-- lexicon.txt
+  #     `-- speaker.info
+
+  if [ ! -d $dl_dir/aishell/data_aishell/wav/train ]; then
+    lhotse download aishell $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/musan
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare aishell manifest"
+  # We assume that you have downloaded the aishell corpus
+  # to $dl_dir/aishell
+  if [ ! -f data/manifests/.aishell_manifests.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare aishell $dl_dir/aishell data/manifests
+    touch data/manifests/.aishell_manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for aishell"
+  if [ ! -f data/fbank/.aishell.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_aishell.py --perturb-speed ${perturb_speed}
+    touch data/fbank/.aishell.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+lang_phone_dir=data/lang_phone
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  mkdir -p $lang_phone_dir
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/aishell/resource_aishell/lexicon.txt |
+    sort | uniq > $lang_phone_dir/lexicon.txt
+
+  ./local/generate_unique_lexicon.py --lang-dir $lang_phone_dir
+
+  if [ ! -f $lang_phone_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_phone_dir
+  fi
+
+
+  # Train a bigram P for MMI training
+  if [ ! -f $lang_phone_dir/transcript_words.txt ]; then
+    log "Generate data to train phone based bigram P"
+    aishell_text=$dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt
+    aishell_train_uid=$dl_dir/aishell/data_aishell/transcript/aishell_train_uid
+    find $dl_dir/aishell/data_aishell/wav/train -name "*.wav" | sed 's/\.wav//g' | awk -F '/' '{print $NF}' > $aishell_train_uid
+    awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $aishell_train_uid $aishell_text |
+	    cut -d " " -f 2- > $lang_phone_dir/transcript_words.txt
+  fi
+
+  if [ ! -f $lang_phone_dir/transcript_tokens.txt ]; then
+    ./local/convert_transcript_words_to_tokens.py \
+      --lexicon $lang_phone_dir/uniq_lexicon.txt \
+      --transcript $lang_phone_dir/transcript_words.txt \
+      --oov "<UNK>" \
+      > $lang_phone_dir/transcript_tokens.txt
+  fi
+
+  if [ ! -f $lang_phone_dir/P.arpa ]; then
+    ./shared/make_kn_lm.py \
+      -ngram-order 2 \
+      -text $lang_phone_dir/transcript_tokens.txt \
+      -lm $lang_phone_dir/P.arpa
+  fi
+
+  if [ ! -f $lang_phone_dir/P.fst.txt ]; then
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_phone_dir/tokens.txt" \
+      --disambig-symbol='#0' \
+      --max-order=2 \
+      $lang_phone_dir/P.arpa > $lang_phone_dir/P.fst.txt
+  fi
+fi
+
+lang_char_dir=data/lang_char
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare char based lang"
+  mkdir -p $lang_char_dir
+  # We reuse words.txt from phone based lexicon
+  # so that the two can share G.pt later.
+
+  # The transcripts in training set, generated in stage 5
+  cp $lang_phone_dir/transcript_words.txt $lang_char_dir/transcript_words.txt
+
+  cat $dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt |
+  cut -d " " -f 2- > $lang_char_dir/text
+
+  (echo '<eps> 0'; echo '!SIL 1'; echo '<SPOKEN_NOISE> 2'; echo '<UNK> 3';) \
+    > $lang_char_dir/words.txt
+
+  cat $lang_char_dir/text | sed 's/ /\n/g' | sort -u | sed '/^$/d' \
+     | awk '{print $1" "NR+3}' >> $lang_char_dir/words.txt
+
+  num_lines=$(< $lang_char_dir/words.txt wc -l)
+  (echo "#0 $num_lines"; echo "<s> $(($num_lines + 1))"; echo "</s> $(($num_lines + 2))";) \
+    >> $lang_char_dir/words.txt
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    ./local/prepare_char.py --lang-dir $lang_char_dir
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare Byte BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bbpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    cp $lang_char_dir/words.txt $lang_dir
+    cp $lang_char_dir/text $lang_dir
+
+    if [ ! -f $lang_dir/bbpe.model ]; then
+      ./local/train_bbpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/text
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bbpe.py --lang-dir $lang_dir
+    fi
+  done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare G"
+
+  mkdir -p data/lm
+
+  # Train LM on transcripts
+  if [ ! -f data/lm/3-gram.unpruned.arpa ]; then
+    python3 ./shared/make_kn_lm.py \
+      -ngram-order 3 \
+      -text $lang_char_dir/transcript_words.txt \
+      -lm data/lm/3-gram.unpruned.arpa
+  fi
+
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+  if [ ! -f data/lm/G_3_gram_char.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_phone_dir/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram_phone.fst.txt
+
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_char_dir/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram_char.fst.txt
+  fi
+
+  if [ ! -f $lang_char_dir/HLG.fst ]; then
+    ./local/prepare_lang_fst.py  \
+      --lang-dir $lang_char_dir \
+      --ngram-G ./data/lm/G_3_gram_char.fst.txt
+  fi
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compile LG & HLG"
+  ./local/compile_hlg.py --lang-dir $lang_phone_dir --lm G_3_gram_phone
+  ./local/compile_hlg.py --lang-dir $lang_char_dir --lm G_3_gram_char
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bbpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir --lm G_3_gram_char
+  done
+
+  ./local/compile_lg.py --lang-dir $lang_phone_dir --lm G_3_gram_phone
+  ./local/compile_lg.py --lang-dir $lang_char_dir --lm G_3_gram_char
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bbpe_${vocab_size}
+    ./local/compile_lg.py --lang-dir $lang_dir --lm G_3_gram_char
+  done
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Generate LM training data"
+
+  log "Processing char based data"
+  out_dir=data/lm_training_char
+  mkdir -p $out_dir $dl_dir/lm
+
+  if [ ! -f $dl_dir/lm/aishell-train-word.txt ]; then
+    cp $lang_phone_dir/transcript_words.txt $dl_dir/lm/aishell-train-word.txt
+  fi
+
+  # training words
+  ./local/prepare_char_lm_training_data.py \
+    --lang-char data/lang_char \
+    --lm-data $dl_dir/lm/aishell-train-word.txt \
+    --lm-archive $out_dir/lm_data.pt
+
+  # valid words
+  if [ ! -f $dl_dir/lm/aishell-valid-word.txt ]; then
+    aishell_text=$dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt
+    aishell_valid_uid=$dl_dir/aishell/data_aishell/transcript/aishell_valid_uid
+    find $dl_dir/aishell/data_aishell/wav/dev -name "*.wav" | sed 's/\.wav//g' | awk -F '/' '{print $NF}' > $aishell_valid_uid
+    awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $aishell_valid_uid $aishell_text |
+	    cut -d " " -f 2- > $dl_dir/lm/aishell-valid-word.txt
+  fi
+
+  ./local/prepare_char_lm_training_data.py \
+    --lang-char data/lang_char \
+    --lm-data $dl_dir/lm/aishell-valid-word.txt \
+    --lm-archive $out_dir/lm_data_valid.pt
+
+  # test words
+  if [ ! -f $dl_dir/lm/aishell-test-word.txt ]; then
+    aishell_text=$dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt
+    aishell_test_uid=$dl_dir/aishell/data_aishell/transcript/aishell_test_uid
+    find $dl_dir/aishell/data_aishell/wav/test -name "*.wav" | sed 's/\.wav//g' | awk -F '/' '{print $NF}' > $aishell_test_uid
+    awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $aishell_test_uid $aishell_text |
+	    cut -d " " -f 2- > $dl_dir/lm/aishell-test-word.txt
+  fi
+
+  ./local/prepare_char_lm_training_data.py \
+    --lang-char data/lang_char \
+    --lm-data $dl_dir/lm/aishell-test-word.txt \
+    --lm-archive $out_dir/lm_data_test.pt
+fi
+
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Sort LM training data"
+  # Sort LM training data by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of tokens
+  # in a sentence.
+
+  out_dir=data/lm_training_char
+  mkdir -p $out_dir
+  ln -snf ../../../librispeech/ASR/local/sort_lm_training_data.py local/
+
+  ./local/sort_lm_training_data.py \
+    --in-lm-data $out_dir/lm_data.pt \
+    --out-lm-data $out_dir/sorted_lm_data.pt \
+    --out-statistics $out_dir/statistics.txt
+
+  ./local/sort_lm_training_data.py \
+    --in-lm-data $out_dir/lm_data_valid.pt \
+    --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+    --out-statistics $out_dir/statistics-valid.txt
+
+  ./local/sort_lm_training_data.py \
+    --in-lm-data $out_dir/lm_data_test.pt \
+    --out-lm-data $out_dir/sorted_lm_data-test.pt \
+    --out-statistics $out_dir/statistics-test.txt
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 11: Train RNN LM model"
+  python ../../../icefall/rnn_lm/train.py \
+    --start-epoch 0 \
+    --world-size 1 \
+    --num-epochs 20 \
+    --use-fp16 0 \
+    --embedding-dim 512 \
+    --hidden-dim 512 \
+    --num-layers 2 \
+    --batch-size 400 \
+    --exp-dir rnnlm_char/exp \
+    --lm-data $out_dir/sorted_lm_data.pt \
+    --lm-data-valid $out_dir/sorted_lm_data-valid.pt \
+    --vocab-size 4336 \
+    --master-port 12345
+fi
diff --git a/egs/aishell/ASR/prepare_aidatatang_200zh.sh b/egs/aishell/ASR/prepare_aidatatang_200zh.sh
new file mode 100755
index 000000000..ec89450df
--- /dev/null
+++ b/egs/aishell/ASR/prepare_aidatatang_200zh.sh
@@ -0,0 +1,59 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/aidatatang_200zh
+#      You can find "corpus" and "transcript" inside it.
+#      You can download it at
+#       https://openslr.org/62/
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  if [ ! -f $dl_dir/aidatatang_200zh/transcript/aidatatang_200_zh_transcript.txt ]; then
+    lhotse download aidatatang-200zh $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare manifest"
+  # We assume that you have downloaded the aidatatang_200zh corpus
+  # to $dl_dir/aidatatang_200zh
+  if [ ! -f data/manifests/.aidatatang_200zh_manifests.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare aidatatang-200zh $dl_dir data/manifests
+    touch data/manifests/.aidatatang_200zh_manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Process aidatatang_200zh"
+  if [ ! -f data/fbank/.aidatatang_200zh_fbank.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_aidatatang_200zh.py --perturb-speed True
+    touch data/fbank/.aidatatang_200zh_fbank.done
+  fi
+fi
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/conformer.py b/egs/aishell/ASR/pruned_transducer_stateless2/conformer.py
new file mode 120000
index 000000000..a65957180
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/decode.py b/egs/aishell/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..fb6c7c481
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,549 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if i >= 1:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+    dev_dl = aishell.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/export.py b/egs/aishell/ASR/pruned_transducer_stateless2/export.py
new file mode 100755
index 000000000..2ce5cfe69
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,206 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --jit 0 \
+  --epoch 29 \
+  --avg 5
+
+It will generate a file exp_dir/pretrained-epoch-29-avg-5.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do::
+
+    cd /path/to/exp_dir
+    ln -s pretrained-epoch-29-avg-5.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default=Path("pruned_transducer_stateless2/exp"),
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if i >= 1:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / f"cpu_jit-epoch-{params.epoch}-avg-{params.avg}.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = (
+            params.exp_dir / f"pretrained-epoch-{params.epoch}-avg-{params.avg}.pt"
+        )
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/model.py b/egs/aishell/ASR/pruned_transducer_stateless2/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/optim.py b/egs/aishell/ASR/pruned_transducer_stateless2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless2/pretrained.py
new file mode 100755
index 000000000..82c10f129
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/pretrained.py
@@ -0,0 +1,327 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method modified_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method fast_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="Maximum number of symbols per frame. "
+        "Use only when --method is greedy_search",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lens = [f.size(0) for f in features]
+    feature_lens = torch.tensor(feature_lens, device=device)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lens)
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    logging.info(f"Using {params.method}")
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported decoding method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = []
+    for hyp in hyp_list:
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/train.py b/egs/aishell/ASR/pruned_transducer_stateless2/train.py
new file mode 100755
index 000000000..60f014c48
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,1036 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 90 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --max-duration 200 \
+
+"""
+
+
+import argparse
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse import CutSet
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], "
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer,
+                    "train/current_",
+                    params.batch_idx_train,
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(cuts: CutSet) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 12.0
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aishell = AishellAsrDataModule(args)
+    train_dl = aishell.train_dataloaders(aishell.train_cuts())
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    logging.info(f"start training from epoch {params.start_epoch}")
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0 if params.start_epoch == 1 else 1.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py b/egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py
new file mode 120000
index 000000000..9a799406b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/aidatatang_200zh.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/aishell.py b/egs/aishell/ASR/pruned_transducer_stateless3/aishell.py
new file mode 120000
index 000000000..1b5f38a54
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/aishell.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/aishell.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py
new file mode 120000
index 000000000..ae3bdd1e0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/conformer.py b/egs/aishell/ASR/pruned_transducer_stateless3/conformer.py
new file mode 120000
index 000000000..c7c1a4b6e
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/decode.py b/egs/aishell/ASR/pruned_transducer_stateless3/decode.py
new file mode 100755
index 000000000..27c64efaa
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/decode.py
@@ -0,0 +1,791 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+
+(5) modified beam search (with LM shallow fusion)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+(6) modified beam search with LM shallow fusion + LODR
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --max-duration 600 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --decoding-method modified_beam_search_LODR \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.48 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.28 \
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=2,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+    params.datatang_prob = 0
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = params.lang_dir / f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            lm_filename,
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+
+    else:
+        LM = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+    test_dl = asr_datamodule.test_dataloaders(test_cuts)
+    dev_dl = asr_datamodule.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless3/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/export.py b/egs/aishell/ASR/pruned_transducer_stateless3/export.py
new file mode 100755
index 000000000..723414167
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/export.py
@@ -0,0 +1,274 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --jit 0 \
+  --epoch 29 \
+  --avg 5
+
+It will generate a file exp_dir/pretrained-epoch-29-avg-5.pt
+
+To use the generated file with `pruned_transducer_stateless3/decode.py`,
+you can do::
+
+    cd /path/to/exp_dir
+    ln -s pretrained-epoch-29-avg-5.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./pruned_transducer_stateless3/decode.py \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default=Path("pruned_transducer_stateless3/exp"),
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+    params.datatang_prob = 0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / f"cpu_jit-epoch-{params.epoch}-avg-{params.avg}.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = (
+            params.exp_dir / f"pretrained-epoch-{params.epoch}-avg-{params.avg}.pt"
+        )
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless3/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/lstmp.py b/egs/aishell/ASR/pruned_transducer_stateless3/lstmp.py
new file mode 120000
index 000000000..557e18aa1
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/lstmp.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless3/lstmp.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/model.py b/egs/aishell/ASR/pruned_transducer_stateless3/model.py
new file mode 100644
index 000000000..a4dda0d6d
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/model.py
@@ -0,0 +1,232 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        decoder_datatang: Optional[nn.Module] = None,
+        joiner_datatang: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim). Its output shape is (N, T, U, vocab_size).
+            Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          encoder_dim:
+            Output dimension of the encoder network.
+          decoder_dim:
+            Output dimension of the decoder network.
+          joiner_dim:
+            Input dimension of the joiner network.
+          vocab_size:
+            Output dimension of the joiner network.
+          decoder_datatang:
+            Optional. The decoder network for the aidatatang_200zh dataset.
+          joiner_datatang:
+            Optional. The joiner network for the aidatatang_200zh dataset.
+        """
+        super().__init__()
+
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_datatang = decoder_datatang
+        self.joiner_datatang = joiner_datatang
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+        if decoder_datatang is not None:
+            self.simple_am_proj_datatang = ScaledLinear(
+                encoder_dim, vocab_size, initial_speed=0.5
+            )
+            self.simple_lm_proj_datatang = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        aishell: bool = True,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          aishell:
+            True to use the decoder and joiner for the aishell dataset.
+            False to use the decoder and joiner for the aidatatang_200zh
+            dataset.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(encoder_out_lens > 0)
+
+        if aishell:
+            decoder = self.decoder
+            simple_lm_proj = self.simple_lm_proj
+            simple_am_proj = self.simple_am_proj
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_datatang
+            simple_lm_proj = self.simple_lm_proj_datatang
+            simple_am_proj = self.simple_am_proj_datatang
+            joiner = self.joiner_datatang
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = encoder_out_lens
+
+        lm = simple_lm_proj(decoder_out)
+        am = simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=joiner.encoder_proj(encoder_out),
+            lm=joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/optim.py b/egs/aishell/ASR/pruned_transducer_stateless3/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py
new file mode 100755
index 000000000..ead393e6e
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py
@@ -0,0 +1,328 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method modified_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method fast_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="Maximum number of symbols per frame. "
+        "Use only when --method is greedy_search",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+    params.datatang_prob = 0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lens = [f.size(0) for f in features]
+    feature_lens = torch.tensor(feature_lens, device=device)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lens)
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    logging.info(f"Using {params.method}")
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported decoding method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = []
+    for hyp in hyp_list:
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless3/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/scaling_converter.py b/egs/aishell/ASR/pruned_transducer_stateless3/scaling_converter.py
new file mode 120000
index 000000000..db93d155b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/train.py b/egs/aishell/ASR/pruned_transducer_stateless3/train.py
new file mode 100755
index 000000000..7c23041ca
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/train.py
@@ -0,0 +1,1251 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./prepare.sh
+
+# If you use a non-zero value for --datatang-prob, you also need to run
+./prepare_aidatatang_200zh.sh
+
+If you use --datatang-prob=0, then you don't need to run the above script.
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 0 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --max-duration 300 \
+  --datatang-prob 0.2
+
+# For mix precision training:
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from aidatatang_200zh import AIDatatang200zh
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--datatang-prob",
+        type=float,
+        default=0.0,
+        help="""The probability to select a batch from the
+        aidatatang_200zh dataset.
+        If it is set to 0, you don't need to download the data
+        for aidatatang_200zh.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    if params.datatang_prob > 0:
+        decoder_datatang = get_decoder_model(params)
+        joiner_datatang = get_joiner_model(params)
+    else:
+        decoder_datatang = None
+        joiner_datatang = None
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        decoder_datatang=decoder_datatang,
+        joiner_datatang=joiner_datatang,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_aishell(c: Cut) -> bool:
+    """Return True if this cut is from the AIShell dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of aidatatang_200zh to
+      dict(origin='aidatatang_200zh')
+      See ../local/process_aidatatang_200zh.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    aishell = is_aishell(supervisions["cut"][0])
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            aishell=aishell,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    datatang_train_dl: Optional[torch.utils.data.DataLoader],
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    aishell_tot_loss = MetricsTracker()
+    datatang_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.datatang_prob, params.datatang_prob]
+
+    iter_aishell = iter(train_dl)
+    if datatang_train_dl is not None:
+        iter_datatang = iter(datatang_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        if datatang_train_dl is not None:
+            idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+            dl = iter_aishell if idx == 0 else iter_datatang
+        else:
+            dl = iter_aishell
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            break
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        aishell = is_aishell(batch["supervisions"]["cut"][0])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            if datatang_train_dl is not None:
+                tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            if aishell:
+                aishell_tot_loss = (
+                    aishell_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "aishell"  # for logging only
+            else:
+                datatang_tot_loss = (
+                    datatang_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "datatang"
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            if datatang_train_dl is not None:
+                datatang_str = f"datatang_tot_loss[{datatang_tot_loss}], "
+                tot_loss_str = f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+            else:
+                tot_loss_str = ""
+                datatang_str = ""
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"{tot_loss_str}"
+                f"aishell_tot_loss[{aishell_tot_loss}], "
+                f"{datatang_str}"
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                if datatang_train_dl is not None:
+                    # If it is None, tot_loss is the same as aishell_tot_loss.
+                    tot_loss.write_summary(
+                        tb_writer, "train/tot_", params.batch_idx_train
+                    )
+                aishell_tot_loss.write_summary(
+                    tb_writer, "train/aishell_tot_", params.batch_idx_train
+                )
+                if datatang_train_dl is not None:
+                    datatang_tot_loss.write_summary(
+                        tb_writer, "train/datatang_tot_", params.batch_idx_train
+                    )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    if datatang_train_dl is not None:
+        loss_value = tot_loss["loss"] / tot_loss["frames"]
+    else:
+        loss_value = aishell_tot_loss["loss"] / aishell_tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(cuts: CutSet) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 12.0
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        if params.datatang_prob > 0:
+            find_unused_parameters = True
+        else:
+            find_unused_parameters = False
+
+        model = DDP(
+            model,
+            device_ids=[rank],
+            find_unused_parameters=find_unused_parameters,
+        )
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    train_cuts = aishell.train_cuts()
+    train_cuts = filter_short_and_long_utterances(train_cuts)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    if params.datatang_prob > 0:
+        datatang = AIDatatang200zh(manifest_dir=args.manifest_dir)
+        train_datatang_cuts = datatang.train_cuts()
+        train_datatang_cuts = filter_short_and_long_utterances(train_datatang_cuts)
+        train_datatang_cuts = train_datatang_cuts.repeat(times=None)
+        datatang_train_dl = asr_datamodule.train_dataloaders(
+            train_datatang_cuts,
+            on_the_fly_feats=False,
+            cuts_musan=cuts_musan,
+        )
+    else:
+        datatang_train_dl = None
+        logging.info("Not using aidatatang_200zh for training")
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    for dl in [
+        train_dl,
+        # datatang_train_dl
+    ]:
+        if dl is not None:
+            scan_pessimistic_batches_for_oom(
+                model=model,
+                train_dl=dl,
+                optimizer=optimizer,
+                graph_compiler=graph_compiler,
+                params=params,
+            )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    logging.info(f"start training from epoch {params.start_epoch}")
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+        if datatang_train_dl is not None:
+            datatang_train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            datatang_train_dl=datatang_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0 if params.start_epoch == 1 else 1.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    assert 0 <= args.datatang_prob < 1, args.datatang_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/aishell.py b/egs/aishell/ASR/pruned_transducer_stateless7/aishell.py
new file mode 120000
index 000000000..ce581b950
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/aishell.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/aishell.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless7/asr_datamodule.py
new file mode 120000
index 000000000..ae3bdd1e0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless7/beam_search.py
new file mode 120000
index 000000000..e9bbcf2a9
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7/decode.py
new file mode 100755
index 000000000..696eea906
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/decode.py
@@ -0,0 +1,685 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import ContextGraph, LmScorer, NgramLm
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        key = f"beam_size_{params.beam_size}"
+        if params.has_contexts:
+            key += f"-context-score-{params.context_score}"
+        else:
+            key += "-no-context-words"
+        return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+
+        store_transcripts(filename=recog_path, texts=results_char)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.has_contexts:
+            params.suffix += f"-context-score-{params.context_score}"
+        else:
+            params.suffix += "-no-contexts-words"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build([(c, 0.0) for c in contexts])
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+    test_dl = asr_datamodule.test_dataloaders(test_cuts)
+    dev_dl = asr_datamodule.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless7/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/decoder2.py b/egs/aishell/ASR/pruned_transducer_stateless7/decoder2.py
new file mode 100644
index 000000000..0345db0a8
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/decoder2.py
@@ -0,0 +1,87 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+        )
+        self.blank_id = blank_id
+
+        assert context_size == 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        if torch.jit.is_tracing():
+            # This is for exporting to PNNX via ONNX
+            embedding_out = self.embedding(y)
+        else:
+            embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+
+        embedding_out = F.relu(embedding_out)
+        return embedding_out
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py b/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py
new file mode 100755
index 000000000..6027273b2
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py
@@ -0,0 +1,1256 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./prepare.sh
+
+If you use --datatang-prob=0, then you don't need to run the above script.
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from decoder2 import Decoder
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 12.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        # T = ((c.num_frames - 7) // 2 + 1) // 2
+        # tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        # if T < len(tokens):
+        #     logging.warning(
+        #         f"Exclude cut with ID {c.id} from training. "
+        #         f"Number of frames (before subsampling): {c.num_frames}. "
+        #         f"Number of frames (after subsampling): {T}. "
+        #         f"Text: {c.supervisions[0].text}. "
+        #         f"Tokens: {tokens}. "
+        #         f"Number of tokens: {len(tokens)}"
+        #     )
+        #     return False
+
+        return True
+
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    train_cuts = aishell.train_cuts()
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+        sampler_state_dict=sampler_state_dict,
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         graph_compiler=graph_compiler,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    logging.info(f"start training from epoch {params.start_epoch}")
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless7/encoder_interface.py
new file mode 120000
index 000000000..0c2673d46
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py b/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py
new file mode 100755
index 000000000..39d988cd0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py
@@ -0,0 +1,590 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Xiaoyu Yang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/marcoyang/icefall-asr-aishell-zipformer-pruned-transducer-stateless7-2023-03-21
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/marcoyang/icefall-asr-aishell-zipformer-pruned-transducer-stateless7-2023-03-21/
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from decoder2 import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/export.py b/egs/aishell/ASR/pruned_transducer_stateless7/export.py
new file mode 120000
index 000000000..2713792e6
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/export.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/jit_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7/jit_pretrained.py
new file mode 100755
index 000000000..e61190649
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/jit_pretrained.py
@@ -0,0 +1,278 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --lang-dir ./data/lang_char \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7/exp/cpu_jit.pt \
+  --lang-dir ./data/lang_char \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    lexicon = Lexicon(args.lang_dir)
+    token_table = lexicon.token_table
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    hyps = [[token_table[t] for t in tokens] for tokens in hyps]
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless7/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/model.py b/egs/aishell/ASR/pruned_transducer_stateless7/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/onnx_check.py b/egs/aishell/ASR/pruned_transducer_stateless7/onnx_check.py
new file mode 120000
index 000000000..e97d1c0aa
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/onnx_check.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/onnx_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7/onnx_pretrained.py
new file mode 100755
index 000000000..a92182e8d
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/onnx_pretrained.py
@@ -0,0 +1,422 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-1224000-avg-14.pt"
+
+cd exp
+ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless3/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+3. Run this file
+
+./pruned_transducer_stateless3/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 2-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, its shape is (N, T', joiner_dim)
+            - encoder_out_lens, its shape is (N,)
+        """
+        out = self.encoder.run(
+            [
+                self.encoder.get_outputs()[0].name,
+                self.encoder.get_outputs()[1].name,
+            ],
+            {
+                self.encoder.get_inputs()[0].name: x.numpy(),
+                self.encoder.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, joiner_dim)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.run_decoder(decoder_input)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        # current_encoder_out's shape: (batch_size, joiner_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+        logits = model.run_joiner(current_encoder_out, decoder_out)
+
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                dtype=torch.int64,
+            )
+            decoder_out = model.run_decoder(decoder_input)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += symbol_table[i]
+        return text.replace("▁", " ").strip()
+
+    context_size = model.context_size
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = token_ids_to_words(hyp[context_size:])
+        s += f"{filename}:\n{words}\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/optim.py b/egs/aishell/ASR/pruned_transducer_stateless7/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7/pretrained.py
new file mode 120000
index 000000000..068f0f57f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless7/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/scaling_converter.py b/egs/aishell/ASR/pruned_transducer_stateless7/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train.py b/egs/aishell/ASR/pruned_transducer_stateless7/train.py
new file mode 100755
index 000000000..9d9dd4288
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/train.py
@@ -0,0 +1,1255 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./prepare.sh
+
+If you use --datatang-prob=0, then you don't need to run the above script.
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 12.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        # T = ((c.num_frames - 7) // 2 + 1) // 2
+        # tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        # if T < len(tokens):
+        #     logging.warning(
+        #         f"Exclude cut with ID {c.id} from training. "
+        #         f"Number of frames (before subsampling): {c.num_frames}. "
+        #         f"Number of frames (after subsampling): {T}. "
+        #         f"Text: {c.supervisions[0].text}. "
+        #         f"Tokens: {tokens}. "
+        #         f"Number of tokens: {len(tokens)}"
+        #     )
+        #     return False
+
+        return True
+
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    train_cuts = aishell.train_cuts()
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+        sampler_state_dict=sampler_state_dict,
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         graph_compiler=graph_compiler,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    logging.info(f"start training from epoch {params.start_epoch}")
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/zipformer.py b/egs/aishell/ASR/pruned_transducer_stateless7/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/__init__.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py
new file mode 100755
index 000000000..da9000164
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decode.py
@@ -0,0 +1,818 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import (
+    LmScorer,
+    NgramLm,
+    byte_encode,
+    smart_byte_decode,
+    tokenize_by_CJK_char,
+)
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_500/bbpe.model",
+        help="Path to the byte BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bbpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+        If you use fast_beam_search_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.25,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--ilme-scale",
+        type=float,
+        default=0.2,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for the internal language model estimation.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "fast_beam_search_LG":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            ilme_scale=params.ilme_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        ref_texts = []
+        for tx in supervisions["text"]:
+            ref_texts.append(byte_encode(tokenize_by_CJK_char(tx)))
+
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(ref_texts),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(smart_byte_decode(sp.decode(hyp)).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+            key += f"_ilme_scale_{params.ilme_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+            params.suffix += f"-ilme-scale-{params.ilme_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bbpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+
+    test_dl = aishell.test_dataloaders(test_cuts)
+    dev_dl = aishell.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/export.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/export.py
new file mode 100755
index 000000000..4e82b45d3
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/export.py
@@ -0,0 +1,320 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7_bbpe/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./pruned_transducer_stateless7_bbpe/decode.py \
+        --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bbpe_500/bbpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/pkufool/icefall_asr_aishell_pruned_transducer_stateless7_bbpe
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/pkufool/icefall_asr_aishell_pruned_transducer_stateless7_bbpe
+    # You will find the pre-trained model in icefall_asr_aishell_pruned_transducer_stateless7_bbpe/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_bbpe/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_500/bbpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bbpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py
new file mode 100755
index 000000000..0c43bf74b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py
@@ -0,0 +1,274 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 49 \
+  --avg 28 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7_bbpe/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7_bbpe/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bbpe_500/bbpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall import smart_byte_decode
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = smart_byte_decode(sp.decode(hyp))
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/model.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/optim.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/pretrained.py
new file mode 100755
index 000000000..ea5bda4db
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/pretrained.py
@@ -0,0 +1,345 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 48 \
+  --avg 29
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) modified beam search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) fast beam search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Note: ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_bbpe/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import smart_byte_decode
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(smart_byte_decode(sp.decode(hyp)).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/scaling_converter.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/test_model.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/test_model.py
new file mode 120000
index 000000000..7ceac5d10
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/test_model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py
new file mode 100755
index 000000000..3858bafd7
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py
@@ -0,0 +1,1250 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_bbpe/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_bbpe/exp \
+  --max-duration 400
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_bbpe/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_bbpe/exp \
+  --max-duration 800
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut, CutSet
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import byte_encode, diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+    tokenize_by_CJK_char,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_bbpe/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_500/bbpe.model",
+        help="Path to the Byte BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 2000,
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bbpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    aishell = AishellAsrDataModule(args)
+
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 12.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    def tokenize_and_encode_text(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = byte_encode(tokenize_by_CJK_char(text))
+        c.supervisions[0].text = text
+        return c
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_cuts = train_cuts.map(tokenize_and_encode_text)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_cuts = valid_cuts.map(tokenize_and_encode_text)
+
+    valid_dl = aishell.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/zipformer.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md
new file mode 120000
index 000000000..a784292cd
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/README.md
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py
new file mode 100755
index 000000000..99110d6b6
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py
@@ -0,0 +1,735 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += 30
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, 30),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        key = f"beam_size_{params.beam_size}"
+        if params.has_contexts:
+            key += f"-context-score-{params.context_score}"
+        else:
+            key += "-no-context-words"
+        return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.has_contexts:
+            params.suffix += f"-context-score-{params.context_score}"
+        else:
+            params.suffix += "-no-contexts-words"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build([(c, 0.0) for c in contexts])
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+
+    test_dl = aishell.test_dataloaders(test_cuts)
+    dev_dl = aishell.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    import time
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        start = time.time()
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+        logging.info(f"Elasped time for {test_set}: {time.time() - start}")
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
new file mode 120000
index 000000000..ca8fed319
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
new file mode 100755
index 000000000..0fba3b58f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
@@ -0,0 +1,1254 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer_for_ncnn_export_only import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+        is_pnnx=True,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    aishell = AishellAsrDataModule(args)
+
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        # T = ((c.num_frames - 7) // 2 + 1) // 2
+        # tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        # if T < len(tokens):
+        #     logging.warning(
+        #         f"Exclude cut with ID {c.id} from training. "
+        #         f"Number of frames (before subsampling): {c.num_frames}. "
+        #         f"Number of frames (after subsampling): {T}. "
+        #         f"Text: {c.supervisions[0].text}. "
+        #         f"Tokens: {tokens}. "
+        #         f"Number of tokens: {len(tokens)}"
+        #     )
+        #     return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = aishell.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
new file mode 120000
index 000000000..72e43c297
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
new file mode 120000
index 000000000..3b36924ef
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
new file mode 120000
index 000000000..eca5e2956
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
new file mode 120000
index 000000000..57a0cd0a0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py
new file mode 120000
index 000000000..2acafdc61
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
new file mode 120000
index 000000000..5d9c6ba00
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
new file mode 120000
index 000000000..457131699
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
new file mode 120000
index 000000000..2b8fa3cbb
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py
new file mode 120000
index 000000000..e17d4f734
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
new file mode 120000
index 000000000..8eea90e04
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
new file mode 120000
index 000000000..28bf7bb82
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
new file mode 120000
index 000000000..c8548d459
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
new file mode 120000
index 000000000..ae4d9bb04
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py
new file mode 120000
index 000000000..9510b8fde
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
new file mode 120000
index 000000000..92c3904af
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
new file mode 120000
index 000000000..1199a61d6
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
new file mode 100755
index 000000000..6b4f183cf
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
@@ -0,0 +1,627 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+import os
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        [
+                            token_table[result]
+                            for result in decode_streams[i].decoding_result()
+                        ],
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    [
+                        token_table[result]
+                        for result in decode_streams[i].decoding_result()
+                    ],
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    valid_cuts = aishell.valid_cuts()
+
+    test_sets = ["test", "valid"]
+    cuts = [test_cuts, valid_cuts]
+
+    for test_set, test_cut in zip(test_sets, cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py
new file mode 120000
index 000000000..1259849e0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py
new file mode 100755
index 000000000..2e1044658
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -0,0 +1,1251 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    aishell = AishellAsrDataModule(args)
+
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+            # In pruned RNN-T, we require that T >= S
+            # where T is the number of feature frames after subsampling
+            # and S is the number of tokens in the utterance
+
+            # In ./zipformer.py, the conv module uses the following expression
+            # for subsampling
+            # T = ((c.num_frames - 7) // 2 + 1) // 2
+            # tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+            # if T < len(tokens):
+            #     logging.warning(
+            #         f"Exclude cut with ID {c.id} from training. "
+            #         f"Number of frames (before subsampling): {c.num_frames}. "
+            #         f"Number of frames (after subsampling): {T}. "
+            #         f"Text: {c.supervisions[0].text}. "
+            #         f"Tokens: {tokens}. "
+            #         f"Number of tokens: {len(tokens)}"
+            #     )
+            return False
+
+        return True
+
+    # train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = aishell.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         graph_compiler=graph_compiler,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py
new file mode 120000
index 000000000..ec183baa7
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
new file mode 120000
index 000000000..d301e1f9b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/shared b/egs/aishell/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/aishell/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/README.md b/egs/aishell/ASR/tdnn_lstm_ctc/README.md
new file mode 100644
index 000000000..a2d80a785
--- /dev/null
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/README.md
@@ -0,0 +1,4 @@
+
+Please visit
+<https://icefall.readthedocs.io/en/latest/recipes/aishell/tdnn_lstm_ctc.html>
+for how to run this recipe.
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/__init__.py b/egs/aishell/ASR/tdnn_lstm_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
new file mode 100644
index 000000000..6abe6c084
--- /dev/null
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
@@ -0,0 +1,377 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class AishellAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.info("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest_lazy(
+            self.args.manifest_dir / "aishell_cuts_train.jsonl.gz"
+        )
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "aishell_cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "aishell_cuts_test.jsonl.gz")
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/decode.py b/egs/aishell/ASR/tdnn_lstm_ctc/decode.py
new file mode 100755
index 000000000..824ca2a92
--- /dev/null
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/decode.py
@@ -0,0 +1,393 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from model import TdnnLstm
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import get_lattice, nbest_decoding, one_best_decoding
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=19,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Supported values are:
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+        """,
+    )
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=30,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is nbest.
+        """,
+    )
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp/"),
+            "lang_dir": Path("data/lang_phone"),
+            "lm_dir": Path("data/lm"),
+            # parameters for tdnn_lstm_ctc
+            "subsampling_factor": 3,
+            "feature_dim": 80,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    lexicon: Lexicon,
+) -> Dict[str, List[List[int]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if the decoding method is 1best, the key is the string
+               `no_rescore`. If the decoding method is nbest, the key is the
+               string `no_rescore-xxx`, xxx is the num_paths.
+
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains word symbol table.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is [N, T, C]
+
+    feature = feature.permute(0, 2, 1)  # now feature is [N, C, T]
+
+    nnet_output = model(feature)
+    # nnet_output is [N, T, C]
+
+    supervisions = batch["supervisions"]
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    assert params.method in ["1best", "nbest"]
+    if params.method == "1best":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        key = "no_rescore"
+    else:
+        best_path = nbest_decoding(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            use_double_scores=params.use_double_scores,
+        )
+        key = f"no_rescore-{params.num_paths}"
+    hyps = get_texts(best_path)
+    hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+    return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    lexicon: Lexicon,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      lexicon:
+        It contains word symbol table.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if decoding method is 1best,
+      or it may be "no-rescoer-100" if decoding method is nbest.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            lexicon=lexicon,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        # We compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results_char)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"cer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu"))
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+
+    model.to(device)
+    model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+    # CAUTION: `test_sets` is for displaying only.
+    # If you want to skip test-clean, you have to skip
+    # it inside the for loop. That is, use
+    #
+    #   if test_set == 'test-clean': continue
+    #
+    test_sets = ["test"]
+    test_dls = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            lexicon=lexicon,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/model.py b/egs/aishell/ASR/tdnn_lstm_ctc/model.py
new file mode 100644
index 000000000..1731e1ebe
--- /dev/null
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/model.py
@@ -0,0 +1,100 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class TdnnLstm(nn.Module):
+    def __init__(
+        self, num_features: int, num_classes: int, subsampling_factor: int = 3
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            It reduces the number of output frames by this factor.
+        """
+        super().__init__()
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=500,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+            nn.Conv1d(
+                in_channels=500,
+                out_channels=500,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+            nn.Conv1d(
+                in_channels=500,
+                out_channels=500,
+                kernel_size=3,
+                stride=self.subsampling_factor,  # stride: subsampling_factor!
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+        )
+        self.lstms = nn.ModuleList(
+            [nn.LSTM(input_size=500, hidden_size=500, num_layers=1) for _ in range(5)]
+        )
+        self.lstm_bnorms = nn.ModuleList(
+            [nn.BatchNorm1d(num_features=500, affine=False) for _ in range(5)]
+        )
+        self.dropout = nn.Dropout(0.2)
+        self.linear = nn.Linear(in_features=500, out_features=self.num_classes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Its shape is [N, C, T]
+
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x = self.tdnn(x)
+        x = x.permute(2, 0, 1)  # (N, C, T) -> (T, N, C) -> how LSTM expects it
+        for lstm, bnorm in zip(self.lstms, self.lstm_bnorms):
+            x_new, _ = lstm(x)
+            x_new = bnorm(x_new.permute(1, 2, 0)).permute(
+                2, 0, 1
+            )  # (T, N, C) -> (N, C, T) -> (T, N, C)
+            x_new = self.dropout(x_new)
+            x = x_new + x  # skip connections
+        x = x.transpose(
+            1, 0
+        )  # (T, N, C) -> (N, T, C) -> linear expects "features" in the last dim
+        x = self.linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/pretrained.py b/egs/aishell/ASR/tdnn_lstm_ctc/pretrained.py
new file mode 100644
index 000000000..7e7213501
--- /dev/null
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/pretrained.py
@@ -0,0 +1,220 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from model import TdnnLstm
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Use the best path as decoding output. Only the transformer encoder
+        output is used for decoding. We call it HLG decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "num_classes": 220,
+            "sample_rate": 16000,
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+    if not hasattr(HLG, "lm_scores"):
+        # For whole-lattice-rescoring and attention-decoder
+        HLG.lm_scores = HLG.scores.clone()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    features = features.permute(0, 2, 1)  # now features is [N, C, T]
+
+    with torch.no_grad():
+        nnet_output = model(features)
+        # nnet_output is [N, T, C]
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    assert params.method == "1best"
+    logging.info("Use HLG decoding")
+    best_path = one_best_decoding(
+        lattice=lattice, use_double_scores=params.use_double_scores
+    )
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/train.py b/egs/aishell/ASR/tdnn_lstm_ctc/train.py
new file mode 100755
index 000000000..e574cf89b
--- /dev/null
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/train.py
@@ -0,0 +1,627 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./tdnn_lstm_ctc/train.py \
+    --world-size 4 \
+    --num-epochs 20 \
+    --max-duration 300
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional
+
+import k2
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import AishellAsrDataModule
+from lhotse.utils import fix_random_seed
+from model import TdnnLstm
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, encode_supervisions, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn_lstm_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - exp_dir: It specifies the directory where all training related
+                   files, e.g., checkpoints, log, etc, are saved
+
+        - lang_dir: It contains language related input files such as
+                    "lexicon.txt"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp_lr1e-4"),
+            "lang_dir": Path("data/lang_phone"),
+            "lr": 1e-4,
+            "feature_dim": 80,
+            "weight_decay": 5e-4,
+            "subsampling_factor": 3,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+):
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of TdnnLstm in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is [N, T, C]
+    feature = feature.permute(0, 2, 1)  # now feature is [N, C, T]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(feature)
+        # nnet_output is [N, T, C]
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervisions = batch["supervisions"]
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+    decoding_graph = graph_compiler.compile(texts)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert loss.requires_grad == is_training
+
+    # train_frames and valid_frames are used for printing.
+    if is_training:
+        params.train_frames = supervision_segments[:, 2].sum().item()
+    else:
+        params.valid_frames = supervision_segments[:, 2].sum().item()
+
+    return loss
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> None:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = 0.0
+    tot_frames = 0.0
+    for batch_idx, batch in enumerate(valid_dl):
+        loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        loss_cpu = loss.detach().cpu().item()
+        tot_loss += loss_cpu
+        tot_frames += params.valid_frames
+
+    if world_size > 1:
+        s = torch.tensor([tot_loss, tot_frames], device=loss.device)
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        s = s.cpu().tolist()
+        tot_loss = s[0]
+        tot_frames = s[1]
+
+    params.valid_loss = tot_loss / tot_frames
+
+    if params.valid_loss < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = params.valid_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = 0.0  # reset after params.reset_interval of batches
+    tot_frames = 0.0  # reset after params.reset_interval of batches
+
+    params.tot_loss = 0.0
+    params.tot_frames = 0.0
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        loss_cpu = loss.detach().cpu().item()
+
+        tot_frames += params.train_frames
+        tot_loss += loss_cpu
+        tot_avg_loss = tot_loss / tot_frames
+
+        params.tot_frames += params.train_frames
+        params.tot_loss += loss_cpu
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"batch avg loss {loss_cpu/params.train_frames:.4f}, "
+                f"total avg loss: {tot_avg_loss:.4f}, "
+                f"batch size: {batch_size}"
+            )
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/current_loss",
+                    loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+
+                tb_writer.add_scalar(
+                    "train/tot_avg_loss",
+                    tot_avg_loss,
+                    params.batch_idx_train,
+                )
+
+        if batch_idx > 0 and batch_idx % params.reset_interval == 0:
+            tot_loss = 0
+            tot_frames = 0
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(
+                f"Epoch {params.cur_epoch}, valid loss {params.valid_loss:.4f},"
+                f" best valid loss: {params.best_valid_loss:.4f} "
+                f"best valid epoch: {params.best_valid_epoch}"
+            )
+
+    params.train_loss = params.tot_loss / params.tot_frames
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.AdamW(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    scheduler = StepLR(optimizer, step_size=8, gamma=0.1)
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    aishell = AishellAsrDataModule(args)
+    train_dl = aishell.train_dataloaders(aishell.train_cuts())
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if epoch > params.start_epoch:
+            logging.info(f"epoch {epoch}, lr: {scheduler.get_last_lr()[0]}")
+
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/lr",
+                scheduler.get_last_lr()[0],
+                params.batch_idx_train,
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        scheduler.step()
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless/README.md b/egs/aishell/ASR/transducer_stateless/README.md
new file mode 100644
index 000000000..622cb837c
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/README.md
@@ -0,0 +1,21 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+You can use the following command to start the training:
+
+```bash
+cd egs/aishell/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./transducer_stateless/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 250 \
+  --lr-factor 2.5
+```
diff --git a/egs/aishell/ASR/transducer_stateless/__init__.py b/egs/aishell/ASR/transducer_stateless/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/transducer_stateless/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless/asr_datamodule.py
new file mode 120000
index 000000000..a73848de9
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/asr_datamodule.py
@@ -0,0 +1 @@
+../conformer_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless/beam_search.py b/egs/aishell/ASR/transducer_stateless/beam_search.py
new file mode 100644
index 000000000..de0a8d0f5
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/beam_search.py
@@ -0,0 +1,337 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import numpy as np
+import torch
+from model import Transducer
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(current_encoder_out, decoder_out)
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys
+    log_prob: float
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None):
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self):
+        return self._data
+
+    #  def add(self, ys: List[int], log_prob: float):
+    def add(self, hyp: Hypothesis):
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]
+            old_hyp.log_prob = np.logaddexp(old_hyp.log_prob, hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: float) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          that have `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for key, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0))
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        # TODO(fangjun): Implement prefix search to update the `log_prob`
+        # of hypotheses in A
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            cached_key = y_star.key
+
+            if cached_key not in decoder_cache:
+                decoder_input = torch.tensor(
+                    [y_star.ys[-context_size:]], device=device
+                ).reshape(1, context_size)
+
+                decoder_out = model.decoder(decoder_input, need_pad=False)
+                decoder_cache[cached_key] = decoder_out
+            else:
+                decoder_out = decoder_cache[cached_key]
+
+            cached_key += f"-t-{t}"
+            if cached_key not in joint_cache:
+                logits = model.joiner(current_encoder_out, decoder_out)
+
+                # TODO(fangjun): Ccale the blank posterior
+
+                log_prob = logits.log_softmax(dim=-1)
+                # log_prob is (1, 1, 1, vocab_size)
+                log_prob = log_prob.squeeze()
+                # Now log_prob is (vocab_size,)
+                joint_cache[cached_key] = log_prob
+            else:
+                log_prob = joint_cache[cached_key]
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob.item()
+
+            # ys[:] returns a copy of ys
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for i, v in zip(indices.tolist(), values.tolist()):
+                if i == blank_id:
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys
diff --git a/egs/aishell/ASR/transducer_stateless/conformer.py b/egs/aishell/ASR/transducer_stateless/conformer.py
new file mode 100644
index 000000000..78424aea2
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/conformer.py
@@ -0,0 +1,898 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Transformer
+
+from icefall.utils import make_pad_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        output_dim (int): Number of output dimension
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            output_dim=output_dim,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+        assert x.size(0) == lengths.max().item()
+        mask = make_pad_mask(lengths)
+
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, N, C)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.LayerNorm(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        # x is (batch, channels, time)
+        x = x.permute(0, 2, 1)
+        x = self.norm(x)
+        x = x.permute(0, 2, 1)
+
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/aishell/ASR/transducer_stateless/decode.py b/egs/aishell/ASR/transducer_stateless/decode.py
new file mode 100755
index 000000000..d23f4f883
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/decode.py
@@ -0,0 +1,457 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import beam_search, greedy_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --decoding-method is beam_search",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=3,
+        help="Maximum number of symbols per frame",
+    )
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        transducer_stateless/exp/pretrained.pt. Note: only model.state_dict()
+        is saved. pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains the token symbol table and the word symbol table.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.decoding_method == "greedy_search":
+            hyp = greedy_search(
+                model=model,
+                encoder_out=encoder_out_i,
+                max_sym_per_frame=params.max_sym_per_frame,
+            )
+        elif params.decoding_method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in ("greedy_search", "beam_search")
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if params.decoding_method == "beam_search":
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    # params.blank_id = graph_compiler.texts_to_ids("<blk>")[0][0]
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dls = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless/decoder.py b/egs/aishell/ASR/transducer_stateless/decoder.py
new file mode 100644
index 000000000..130f080ec
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/decoder.py
@@ -0,0 +1,100 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+        else:
+            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
+            # when inference with torch.jit.script and context_size == 1
+            self.conv = nn.Identity()
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U) with blank prepended.
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, embedding_dim).
+        """
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        return embedding_out
diff --git a/egs/aishell/ASR/transducer_stateless/encoder_interface.py b/egs/aishell/ASR/transducer_stateless/encoder_interface.py
new file mode 100644
index 000000000..257facce4
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/encoder_interface.py
@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")
diff --git a/egs/aishell/ASR/transducer_stateless/export.py b/egs/aishell/ASR/transducer_stateless/export.py
new file mode 100755
index 000000000..01de5d772
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/export.py
@@ -0,0 +1,248 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#           2022 Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless/export.py \
+  --exp-dir ./transducer_stateless/exp \
+  --lang-dir data/lang_char \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless/decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./transducer_stateless/decode.py \
+        --exp-dir ./transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless/joiner.py b/egs/aishell/ASR/transducer_stateless/joiner.py
new file mode 100644
index 000000000..2ef3f1de6
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/joiner.py
@@ -0,0 +1,54 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, C).
+        Returns:
+          Return a tensor of shape (N, T, U, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == decoder_out.size(2)
+
+        encoder_out = encoder_out.unsqueeze(2)
+        # Now encoder_out is (N, T, 1, C)
+
+        decoder_out = decoder_out.unsqueeze(1)
+        # Now decoder_out is (N, 1, U, C)
+
+        logit = encoder_out + decoder_out
+        logit = torch.tanh(logit)
+
+        output = self.output_linear(logit)
+
+        return output
diff --git a/egs/aishell/ASR/transducer_stateless/model.py b/egs/aishell/ASR/transducer_stateless/model.py
new file mode 100644
index 000000000..591bbe44f
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/model.py
@@ -0,0 +1,112 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        decoder_out = self.decoder(sos_y_padded)
+
+        logits = self.joiner(encoder_out, decoder_out)
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        loss = k2.rnnt_loss(logits, y_padded, blank_id, boundary)
+
+        return loss
diff --git a/egs/aishell/ASR/transducer_stateless/pretrained.py b/egs/aishell/ASR/transducer_stateless/pretrained.py
new file mode 100755
index 000000000..40f430e13
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/pretrained.py
@@ -0,0 +1,321 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method greedy_search \
+        /path/to/foo.wav \
+        /path/to/bar.wav \
+
+(1) beam search
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav \
+
+You can also use `./transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./transducer_stateless/exp/pretrained.pt is generated by
+./transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import kaldifeat
+import torch
+import torch.nn as nn
+import torchaudio
+from beam_search import beam_search, greedy_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=3,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = graph_compiler.texts_to_ids("<blk>")[0][0]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+    for i in range(num_waves):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.method == "greedy_search":
+            hyp = greedy_search(
+                model=model,
+                encoder_out=encoder_out_i,
+                max_sym_per_frame=params.max_sym_per_frame,
+            )
+        elif params.method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported method: {params.method}")
+
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless/subsampling.py b/egs/aishell/ASR/transducer_stateless/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless/test_decoder.py b/egs/aishell/ASR/transducer_stateless/test_decoder.py
new file mode 100755
index 000000000..fe0bdee70
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/test_decoder.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/aishell/ASR
+    python ./transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, embedding_dim)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, embedding_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless/train.py b/egs/aishell/ASR/transducer_stateless/train.py
new file mode 100755
index 000000000..62ffff473
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/train.py
@@ -0,0 +1,675 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(x=feature, x_lens=feature_lens, y=y)
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AishellAsrDataModule(args)
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 12.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = aishell.train_dataloaders(train_cuts)
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless/transformer.py b/egs/aishell/ASR/transducer_stateless/transformer.py
new file mode 100644
index 000000000..b3ff153c1
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless/transformer.py
@@ -0,0 +1,416 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+from icefall.utils import make_pad_mask
+
+
+class Transformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          dropout:
+            Dropout in encoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.output_dim = output_dim
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(0) == lengths.max().item()
+
+        mask = make_pad_mask(lengths)
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/README.md b/egs/aishell/ASR/transducer_stateless_modified-2/README.md
new file mode 100644
index 000000000..b3c539670
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/README.md
@@ -0,0 +1,59 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+Different from `../transducer_stateless_modified`, this folder
+uses extra data, i.e., http://www.openslr.org/62/, during training.
+
+You can use the following command to start the training:
+
+```bash
+cd egs/aishell/ASR
+./prepare.sh --stop-stage 6
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_stateless_modified-2/train.py \
+  --world-size 3 \
+  --num-epochs 90 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_modified-2/exp-2 \
+  --max-duration 250 \
+  --lr-factor 2.0 \
+  --context-size 2 \
+  --modified-transducer-prob 0.25 \
+  --datatang-prob 0.2
+```
+
+To decode, you can use
+
+```bash
+for epoch in 89; do
+  for avg in 30 38; do
+  ./transducer_stateless_modified-2/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless_modified-2/exp-2 \
+    --max-duration 100 \
+    --context-size 2 \
+    --decoding-method greedy_search \
+    --max-sym-per-frame 1
+  done
+done
+
+for epoch in 89; do
+  for avg in 38; do
+    ./transducer_stateless_modified-2/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless_modified-2/exp-2 \
+    --max-duration 100 \
+    --context-size 2 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+  done
+done
+```
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/__init__.py b/egs/aishell/ASR/transducer_stateless_modified-2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/aidatatang_200zh.py b/egs/aishell/ASR/transducer_stateless_modified-2/aidatatang_200zh.py
new file mode 100644
index 000000000..26d4ee111
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/aidatatang_200zh.py
@@ -0,0 +1,53 @@
+# Copyright      2021  Piotr Żelasko
+#                2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+
+
+class AIDatatang200zh:
+    def __init__(self, manifest_dir: str):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files::
+
+                - aidatatang_cuts_dev.jsonl.gz
+                - aidatatang_cuts_train.jsonl.gz
+                - aidatatang_cuts_test.jsonl.gz
+        """
+        self.manifest_dir = Path(manifest_dir)
+
+    def train_cuts(self) -> CutSet:
+        f = self.manifest_dir / "aidatatang_cuts_train.jsonl.gz"
+        logging.info(f"About to get train cuts from {f}")
+        cuts_train = load_manifest_lazy(f)
+        return cuts_train
+
+    def valid_cuts(self) -> CutSet:
+        f = self.manifest_dir / "aidatatang_cuts_valid.jsonl.gz"
+        logging.info(f"About to get valid cuts from {f}")
+        cuts_valid = load_manifest_lazy(f)
+        return cuts_valid
+
+    def test_cuts(self) -> CutSet:
+        f = self.manifest_dir / "aidatatang_cuts_test.jsonl.gz"
+        logging.info(f"About to get test cuts from {f}")
+        cuts_test = load_manifest_lazy(f)
+        return cuts_test
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/aishell.py b/egs/aishell/ASR/transducer_stateless_modified-2/aishell.py
new file mode 100644
index 000000000..ddeca4d88
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/aishell.py
@@ -0,0 +1,53 @@
+# Copyright      2021  Piotr Żelasko
+#                2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+
+
+class AIShell:
+    def __init__(self, manifest_dir: str):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files::
+
+                - aishell_cuts_dev.jsonl.gz
+                - aishell_cuts_train.jsonl.gz
+                - aishell_cuts_test.jsonl.gz
+        """
+        self.manifest_dir = Path(manifest_dir)
+
+    def train_cuts(self) -> CutSet:
+        f = self.manifest_dir / "aishell_cuts_train.jsonl.gz"
+        logging.info(f"About to get train cuts from {f}")
+        cuts_train = load_manifest_lazy(f)
+        return cuts_train
+
+    def valid_cuts(self) -> CutSet:
+        f = self.manifest_dir / "aishell_cuts_dev.jsonl.gz"
+        logging.info(f"About to get valid cuts from {f}")
+        cuts_valid = load_manifest_lazy(f)
+        return cuts_valid
+
+    def test_cuts(self) -> CutSet:
+        f = self.manifest_dir / "aishell_cuts_test.jsonl.gz"
+        logging.info(f"About to get test cuts from {f}")
+        cuts_test = load_manifest_lazy(f)
+        return cuts_test
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py
new file mode 100644
index 000000000..cd8dd821c
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py
@@ -0,0 +1,297 @@
+# Copyright      2021  Piotr Żelasko
+#                2022  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import inspect
+import logging
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig
+from lhotse.dataset import (
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures, PrecomputedFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class AsrDataModule:
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler "
+            "(you might want to increase it for larger datasets).",
+        )
+
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available. Used only in dev/test CutSet",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        on_the_fly_feats: bool,
+        cuts_musan: Optional[CutSet] = None,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            Cuts for training.
+          cuts_musan:
+            If not None, it is the cuts for mixing.
+          on_the_fly_feats:
+            True to use OnTheFlyFeatures;
+            False to use PrecomputedFeatures.
+        """
+        transforms = []
+        if cuts_musan is not None:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        input_transforms = []
+
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        # NOTE: the PerturbSpeed transform should be added only if we
+        # remove it from data prep stage.
+        # Add on-the-fly speed perturbation; since originally it would
+        # have increased epoch size by 3, we will apply prob 2/3 and use
+        # 3x more epochs.
+        # Speed perturbation probably should come first before
+        # concatenation, but in principle the transforms order doesn't have
+        # to be strict (e.g. could be randomized)
+        # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+        # Drop feats to be on the safe side.
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_strategy=(
+                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+                if on_the_fly_feats
+                else PrecomputedFeatures()
+            ),
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        logging.info("Using DynamicBucketingSampler.")
+        train_sampler = DynamicBucketingSampler(
+            cuts_train,
+            max_duration=self.args.max_duration,
+            shuffle=self.args.shuffle,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        logging.info("About to create train dataloader")
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/beam_search.py b/egs/aishell/ASR/transducer_stateless_modified-2/beam_search.py
new file mode 120000
index 000000000..e188617a8
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/conformer.py b/egs/aishell/ASR/transducer_stateless_modified-2/conformer.py
new file mode 120000
index 000000000..88975988f
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/conformer.py
@@ -0,0 +1 @@
+../transducer_stateless_modified/conformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/decode.py b/egs/aishell/ASR/transducer_stateless_modified-2/decode.py
new file mode 100755
index 000000000..d164b6890
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/decode.py
@@ -0,0 +1,513 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_stateless_modified-2/decode.py \
+    --epoch 89 \
+    --avg 38 \
+    --exp-dir ./transducer_stateless_modified-2/exp \
+    --max-duration 100 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./transducer_stateless_modified-2/decode.py \
+    --epoch 89 \
+    --avg 38 \
+    --exp-dir ./transducer_stateless_modified-2/exp \
+    --max-duration 100 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./transducer_stateless_modified-2/decode.py \
+    --epoch 89 \
+    --avg 38 \
+    --exp-dir ./transducer_stateless_modified-2/exp \
+    --max-duration 100 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+(4) fast beam search
+./transducer_stateless_modified-2/decode.py \
+    --epoch 89 \
+    --avg 38 \
+    --exp-dir ./transducer_stateless_modified-2/exp \
+    --max-duration 100 \
+    --decoding-method fast_beam_search \
+    --beam-size 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_modified-2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      token_table:
+        It maps token ID to a string.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    test_cuts = aishell.test_cuts()
+    test_dl = asr_datamodule.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dls = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/decoder.py b/egs/aishell/ASR/transducer_stateless_modified-2/decoder.py
new file mode 120000
index 000000000..bdfcea5c2
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/decoder.py
@@ -0,0 +1 @@
+../transducer_stateless_modified/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/encoder_interface.py b/egs/aishell/ASR/transducer_stateless_modified-2/encoder_interface.py
new file mode 120000
index 000000000..a2a5f22cf
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless_modified/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/export.py b/egs/aishell/ASR/transducer_stateless_modified-2/export.py
new file mode 100755
index 000000000..c1081c32b
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/export.py
@@ -0,0 +1,246 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless_modified-2/export.py \
+  --exp-dir ./transducer_stateless_modified-2/exp \
+  --epoch 89 \
+  --avg 38
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless_modified-2/decode.py`,
+you can do::
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./transducer_stateless_modified-2/decode.py \
+        --exp-dir ./transducer_stateless_modified-2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default=Path("transducer_stateless_modified-2/exp"),
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/joiner.py b/egs/aishell/ASR/transducer_stateless_modified-2/joiner.py
new file mode 120000
index 000000000..e9e435ecd
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/joiner.py
@@ -0,0 +1 @@
+../transducer_stateless_modified/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/model.py b/egs/aishell/ASR/transducer_stateless_modified-2/model.py
new file mode 100644
index 000000000..086957d0b
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/model.py
@@ -0,0 +1,163 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+from typing import Optional
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        decoder_datatang: Optional[nn.Module] = None,
+        joiner_datatang: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          decoder_datatang:
+            The decoder for the aidatatang_200zh dataset.
+          joiner_datatang:
+            The joiner for the aidatatang_200zh dataset.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+        if decoder_datatang is not None:
+            assert hasattr(decoder_datatang, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_datatang = decoder_datatang
+        self.joiner_datatang = joiner_datatang
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        aishell: bool = True,
+        modified_transducer_prob: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          modified_transducer_prob:
+            The probability to use modified transducer loss.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        if aishell:
+            decoder = self.decoder
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_datatang
+            joiner = self.joiner_datatang
+
+        decoder_out = decoder(sos_y_padded)
+
+        # +1 here since a blank is prepended to each utterance.
+        logits = joiner(
+            encoder_out=encoder_out,
+            decoder_out=decoder_out,
+            encoder_out_len=x_lens,
+            decoder_out_len=y_lens + 1,
+        )
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        # We don't put this `import` at the beginning of the file
+        # as it is required only in the training, not during the
+        # reference stage
+        import optimized_transducer
+
+        assert 0 <= modified_transducer_prob <= 1
+
+        if modified_transducer_prob == 0:
+            one_sym_per_frame = False
+        elif random.random() < modified_transducer_prob:
+            # random.random() returns a float in the range [0, 1)
+            one_sym_per_frame = True
+        else:
+            one_sym_per_frame = False
+
+        loss = optimized_transducer.transducer_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+            one_sym_per_frame=one_sym_per_frame,
+            from_log_softmax=False,
+        )
+
+        return loss
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/pretrained.py b/egs/aishell/ASR/transducer_stateless_modified-2/pretrained.py
new file mode 100755
index 000000000..5d8ca2e11
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/pretrained.py
@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless_modified-2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless_modified-2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) modified beam search
+./transducer_stateless_modified-2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method modified_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) fast beam search
+./transducer_stateless_modified-2/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method fast_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="Maximum number of symbols per frame. "
+        "Use only when --method is greedy_search",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lens = [f.size(0) for f in features]
+    feature_lens = torch.tensor(feature_lens, device=device)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lens)
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    logging.info(f"Using {params.method}")
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported decoding method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = []
+    for hyp in hyp_list:
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/subsampling.py b/egs/aishell/ASR/transducer_stateless_modified-2/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/test_decoder.py b/egs/aishell/ASR/transducer_stateless_modified-2/test_decoder.py
new file mode 120000
index 000000000..fbe1679ea
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/test_decoder.py
@@ -0,0 +1 @@
+../transducer_stateless_modified/test_decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/train.py b/egs/aishell/ASR/transducer_stateless_modified-2/train.py
new file mode 100755
index 000000000..8fb7d1e49
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/train.py
@@ -0,0 +1,866 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./prepare.sh
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_stateless_modified-2/train.py \
+  --world-size 3 \
+  --num-epochs 90 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_modified-2/exp-2 \
+  --max-duration 250 \
+  --lr-factor 2.0 \
+  --context-size 2 \
+  --modified-transducer-prob 0.25 \
+  --datatang-prob 0.2
+"""
+
+
+import argparse
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from aidatatang_200zh import AIDatatang200zh
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_modified-2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--modified-transducer-prob",
+        type=float,
+        default=0.25,
+        help="""The probability to use modified transducer loss.
+        In modified transduer, it limits the maximum number of symbols
+        per frame to 1. See also the option --max-sym-per-frame in
+        transducer_stateless/decode.py
+        """,
+    )
+
+    parser.add_argument(
+        "--datatang-prob",
+        type=float,
+        default=0.2,
+        help="The probability to select a batch from the aidatatang_200zh dataset",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 800,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    decoder_datatang = get_decoder_model(params)
+    joiner_datatang = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        decoder_datatang=decoder_datatang,
+        joiner_datatang=joiner_datatang,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_aishell(c: Cut) -> bool:
+    """Return True if this cut is from the AIShell dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of aidatatang_200zh to
+      dict(origin='aidatatang_200zh')
+      See ../local/process_aidatatang_200zh.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    aishell = is_aishell(supervisions["cut"][0])
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            aishell=aishell,
+            modified_transducer_prob=params.modified_transducer_prob,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    datatang_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      datatang_train_dl:
+        Dataloader for the aidatatang_200zh training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    aishell_tot_loss = MetricsTracker()
+    datatang_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.datatang_prob, params.datatang_prob]
+
+    iter_aishell = iter(train_dl)
+    iter_datatang = iter(datatang_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_aishell if idx == 0 else iter_datatang
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            break
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        aishell = is_aishell(batch["supervisions"]["cut"][0])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+        if aishell:
+            aishell_tot_loss = (
+                aishell_tot_loss * (1 - 1 / params.reset_interval)
+            ) + loss_info
+            prefix = "aishell"  # for logging only
+        else:
+            datatang_tot_loss = (
+                datatang_tot_loss * (1 - 1 / params.reset_interval)
+            ) + loss_info
+            prefix = "datatang"
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"aishell_tot_loss[{aishell_tot_loss}], "
+                f"datatang_tot_loss[{datatang_tot_loss}], "
+                f"batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                aishell_tot_loss.write_summary(
+                    tb_writer, "train/aishell_tot_", params.batch_idx_train
+                )
+                datatang_tot_loss.write_summary(
+                    tb_writer, "train/datatang_tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(cuts: CutSet) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 12.0
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    seed = 42
+    fix_random_seed(seed)
+    rng = random.Random(seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+
+    train_cuts = aishell.train_cuts()
+    train_cuts = filter_short_and_long_utterances(train_cuts)
+
+    datatang = AIDatatang200zh(manifest_dir=args.manifest_dir)
+    train_datatang_cuts = datatang.train_cuts()
+    train_datatang_cuts = filter_short_and_long_utterances(train_datatang_cuts)
+    train_datatang_cuts = train_datatang_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    datatang_train_dl = asr_datamodule.train_dataloaders(
+        train_datatang_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    for dl in [
+        train_dl,
+        # datatang_train_dl
+    ]:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+        datatang_train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            datatang_train_dl=datatang_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    assert 0 <= args.datatang_prob < 1, args.datatang_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/transformer.py b/egs/aishell/ASR/transducer_stateless_modified-2/transformer.py
new file mode 120000
index 000000000..4320d1105
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/transformer.py
@@ -0,0 +1 @@
+../transducer_stateless_modified/transformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/README.md b/egs/aishell/ASR/transducer_stateless_modified/README.md
new file mode 100644
index 000000000..9709eb9a0
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/README.md
@@ -0,0 +1,21 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+You can use the following command to start the training:
+
+```bash
+cd egs/aishell/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./transducer_stateless_modified/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_modified/exp \
+  --max-duration 250 \
+  --lr-factor 2.5
+```
diff --git a/egs/aishell/ASR/transducer_stateless_modified/__init__.py b/egs/aishell/ASR/transducer_stateless_modified/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/transducer_stateless_modified/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless_modified/asr_datamodule.py
new file mode 120000
index 000000000..a73848de9
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/asr_datamodule.py
@@ -0,0 +1 @@
+../conformer_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/beam_search.py b/egs/aishell/ASR/transducer_stateless_modified/beam_search.py
new file mode 120000
index 000000000..e188617a8
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/conformer.py b/egs/aishell/ASR/transducer_stateless_modified/conformer.py
new file mode 120000
index 000000000..8be0dc864
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/decode.py b/egs/aishell/ASR/transducer_stateless_modified/decode.py
new file mode 100755
index 000000000..0a7d87fe8
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/decode.py
@@ -0,0 +1,514 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_stateless_modified/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless_modified/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./transducer_stateless_modified/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless_modified/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./transducer_stateless_modified/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless_modified/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./transducer_stateless_modified/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless_modified/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_modified/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      token_table:
+        It maps token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dls = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified/decoder.py b/egs/aishell/ASR/transducer_stateless_modified/decoder.py
new file mode 120000
index 000000000..82337f7ef
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/encoder_interface.py b/egs/aishell/ASR/transducer_stateless_modified/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/export.py b/egs/aishell/ASR/transducer_stateless_modified/export.py
new file mode 100755
index 000000000..3e14ad69c
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/export.py
@@ -0,0 +1,246 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless_modified/export.py \
+  --exp-dir ./transducer_stateless_modified/exp \
+  --epoch 64 \
+  --avg 33
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless_modified/decode.py`,
+you can do::
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./transducer_stateless_modified/decode.py \
+        --exp-dir ./transducer_stateless_modified/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default=Path("transducer_stateless_modified/exp"),
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified/joiner.py b/egs/aishell/ASR/transducer_stateless_modified/joiner.py
new file mode 120000
index 000000000..1aec6bfaf
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/model.py b/egs/aishell/ASR/transducer_stateless_modified/model.py
new file mode 120000
index 000000000..16ddd93f0
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/pretrained.py b/egs/aishell/ASR/transducer_stateless_modified/pretrained.py
new file mode 100755
index 000000000..9e4459247
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/pretrained.py
@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless_modified/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless_modified/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) modified beam search
+./transducer_stateless_modified/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method modified_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) fast beam search
+./transducer_stateless_modified/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method fast_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="Maximum number of symbols per frame. "
+        "Use only when --method is greedy_search",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lens = [f.size(0) for f in features]
+    feature_lens = torch.tensor(feature_lens, device=device)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lens)
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    logging.info(f"Using {params.method}")
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported decoding method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = []
+    for hyp in hyp_list:
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified/subsampling.py b/egs/aishell/ASR/transducer_stateless_modified/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/transducer_stateless_modified/test_decoder.py b/egs/aishell/ASR/transducer_stateless_modified/test_decoder.py
new file mode 100755
index 000000000..fe0bdee70
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/test_decoder.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/aishell/ASR
+    python ./transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, embedding_dim)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, embedding_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified/train.py b/egs/aishell/ASR/transducer_stateless_modified/train.py
new file mode 100755
index 000000000..5f116f2bd
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/train.py
@@ -0,0 +1,746 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_stateless_modified/train.py \
+  --world-size 3 \
+  --num-epochs 65 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_modified/exp \
+  --max-duration 250 \
+  --lr-factor 2.0 \
+  --context-size 2 \
+  --modified-transducer-prob 0.25
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_modified/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--modified-transducer-prob",
+        type=float,
+        default=0.25,
+        help="""The probability to use modified transducer loss.
+        In modified transduer, it limits the maximum number of symbols
+        per frame to 1. See also the option --max-sym-per-frame in
+        transducer_stateless/decode.py
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 800,
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            modified_transducer_prob=params.modified_transducer_prob,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(42)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AishellAsrDataModule(args)
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 12.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = aishell.train_dataloaders(train_cuts)
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified/transformer.py b/egs/aishell/ASR/transducer_stateless_modified/transformer.py
new file mode 120000
index 000000000..214afed39
--- /dev/null
+++ b/egs/aishell/ASR/transducer_stateless_modified/transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/__init__.py b/egs/aishell/ASR/zipformer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell/ASR/zipformer/asr_datamodule.py b/egs/aishell/ASR/zipformer/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/beam_search.py b/egs/aishell/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/decode.py b/egs/aishell/ASR/zipformer/decode.py
new file mode 100755
index 000000000..1968904ae
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/decode.py
@@ -0,0 +1,814 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao
+#                                                 Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(3) fast beam search (trivial_graph)
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(4) fast beam search (LG)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from train import add_model_arguments, get_model, get_params
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest_oracle
+        If you use fast_beam_search_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--ilme-scale",
+        type=float,
+        default=0.2,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for the internal language model estimation.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    x, x_lens = model.encoder_embed(feature, feature_lens)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+    encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_LG":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            blank_penalty=params.blank_penalty,
+            ilme_scale=params.ilme_scale,
+        )
+        for hyp in hyp_tokens:
+            sentence = "".join([lexicon.word_table[i] for i in hyp])
+            hyps.append(list(sentence))
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=graph_compiler.texts_to_ids(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            blank_penalty=params.blank_penalty,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                    blank_penalty=params.blank_penalty,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                    blank_penalty=params.blank_penalty,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    key = f"blank_penalty_{params.blank_penalty}"
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search_" + key: hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key += f"_beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ilme_scale_{params.ilme_scale}"
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}_" + key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list("".join(text.split())) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "modified_beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest_oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"_ilme_scale_{params.ilme_scale}"
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+    params.suffix += f"-blank-penalty-{params.blank_penalty}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
+            )
+        return T > 0
+
+    dev_cuts = aishell.valid_cuts()
+    dev_cuts = dev_cuts.filter(remove_short_utt)
+    dev_dl = aishell.valid_dataloaders(dev_cuts)
+
+    test_cuts = aishell.test_cuts()
+    test_cuts = test_cuts.filter(remove_short_utt)
+    test_dl = aishell.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/zipformer/decode_stream.py b/egs/aishell/ASR/zipformer/decode_stream.py
new file mode 120000
index 000000000..b8d8ddfc4
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decode_stream.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/decoder.py b/egs/aishell/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/encoder_interface.py b/egs/aishell/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/export-onnx-streaming.py b/egs/aishell/ASR/zipformer/export-onnx-streaming.py
new file mode 120000
index 000000000..2962eb784
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/export-onnx.py b/egs/aishell/ASR/zipformer/export-onnx.py
new file mode 120000
index 000000000..70a15683c
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/export.py b/egs/aishell/ASR/zipformer/export.py
new file mode 120000
index 000000000..dfc1bec08
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/jit_pretrained.py b/egs/aishell/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/jit_pretrained_streaming.py b/egs/aishell/ASR/zipformer/jit_pretrained_streaming.py
new file mode 120000
index 000000000..1962351e9
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/joiner.py b/egs/aishell/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/model.py b/egs/aishell/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/onnx_check.py b/egs/aishell/ASR/zipformer/onnx_check.py
new file mode 120000
index 000000000..f3dd42004
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_check.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/onnx_decode.py b/egs/aishell/ASR/zipformer/onnx_decode.py
new file mode 100755
index 000000000..17c6eceb4
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/onnx_decode.py
@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+"""
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from lhotse.cut import Cut
+from onnx_pretrained import OnnxModel, greedy_search
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_char/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, token_table: k2.SymbolTable, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      token_table:
+        Mapping ids to tokens.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    hyps = [[token_table[h] for h in hyp] for hyp in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      token_table:
+        Mapping ids to tokens.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, token_table=token_table, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = list(ref_text)
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+    assert token_table[0] == "<blk>"
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+
+    aishell = AishellAsrDataModule(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
+            )
+        return T > 0
+
+    dev_cuts = aishell.valid_cuts()
+    dev_cuts = dev_cuts.filter(remove_short_utt)
+    dev_dl = aishell.valid_dataloaders(dev_cuts)
+
+    test_cuts = aishell.test_net_cuts()
+    test_cuts = test_cuts.filter(remove_short_utt)
+    test_dl = aishell.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(
+            dl=test_dl, model=model, token_table=token_table
+        )
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/zipformer/onnx_pretrained-streaming.py b/egs/aishell/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 120000
index 000000000..cfea104c2
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/onnx_pretrained.py b/egs/aishell/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/optim.py b/egs/aishell/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/pretrained.py b/egs/aishell/ASR/zipformer/pretrained.py
new file mode 120000
index 000000000..0bd71dde4
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/scaling.py b/egs/aishell/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/scaling_converter.py b/egs/aishell/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/streaming_beam_search.py b/egs/aishell/ASR/zipformer/streaming_beam_search.py
new file mode 120000
index 000000000..b1ed54557
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/streaming_decode.py b/egs/aishell/ASR/zipformer/streaming_decode.py
new file mode 100755
index 000000000..c3820447a
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/streaming_decode.py
@@ -0,0 +1,880 @@
+#!/usr/bin/env python3
+# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
+#                                                  Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+from asr_datamodule import AishellAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch import Tensor, nn
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="Path to the lang dir(containing lexicon, tokens, etc.)",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_init_states(
+    model: nn.Module,
+    batch_size: int = 1,
+    device: torch.device = torch.device("cpu"),
+) -> List[torch.Tensor]:
+    """
+    Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+    is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+    states[-2] is the cached left padding for ConvNeXt module,
+    of shape (batch_size, num_channels, left_pad, num_freqs)
+    states[-1] is processed_lens of shape (batch,), which records the number
+    of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+    """
+    states = model.encoder.get_init_states(batch_size, device)
+
+    embed_states = model.encoder_embed.get_init_states(batch_size, device)
+    states.append(embed_states)
+
+    processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+    states.append(processed_lens)
+
+    return states
+
+
+def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance. For element-n,
+        state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
+        state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
+        cached_val2, cached_conv1, cached_conv2).
+        state_list[n][-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+        state_list[n][-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+    """
+    batch_size = len(state_list)
+    assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
+    tot_num_layers = (len(state_list[0]) - 2) // 6
+
+    batch_states = []
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key = torch.cat(
+            [state_list[i][layer_offset] for i in range(batch_size)], dim=1
+        )
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn = torch.cat(
+            [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1 = torch.cat(
+            [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2 = torch.cat(
+            [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1 = torch.cat(
+            [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2 = torch.cat(
+            [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
+        )
+        batch_states += [
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ]
+
+    cached_embed_left_pad = torch.cat(
+        [state_list[i][-2] for i in range(batch_size)], dim=0
+    )
+    batch_states.append(cached_embed_left_pad)
+
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
+    batch_states.append(processed_lens)
+
+    return batch_states
+
+
+def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+        batch_states: A list of cached tensors of all encoder layers. For layer-i,
+          states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+          cached_conv1, cached_conv2).
+          state_list[-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+          states[-1] is processed_lens of shape (batch,), which records the number
+          of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Returns:
+        state_list: A list of list. Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+    """
+    assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
+    tot_num_layers = (len(batch_states) - 2) // 6
+
+    processed_lens = batch_states[-1]
+    batch_size = processed_lens.shape[0]
+
+    state_list = [[] for _ in range(batch_size)]
+
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1_list = batch_states[layer_offset + 2].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2_list = batch_states[layer_offset + 3].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1_list = batch_states[layer_offset + 4].chunk(
+            chunks=batch_size, dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2_list = batch_states[layer_offset + 5].chunk(
+            chunks=batch_size, dim=0
+        )
+        for i in range(batch_size):
+            state_list[i] += [
+                cached_key_list[i],
+                cached_nonlin_attn_list[i],
+                cached_val1_list[i],
+                cached_val2_list[i],
+                cached_conv1_list[i],
+                cached_conv2_list[i],
+            ]
+
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(cached_embed_left_pad_list[i])
+
+    processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(processed_lens_list[i])
+
+    return state_list
+
+
+def streaming_forward(
+    features: Tensor,
+    feature_lens: Tensor,
+    model: nn.Module,
+    states: List[Tensor],
+    chunk_size: int,
+    left_context_len: int,
+) -> Tuple[Tensor, Tensor, List[Tensor]]:
+    """
+    Returns encoder outputs, output lengths, and updated states.
+    """
+    cached_embed_left_pad = states[-2]
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        cached_left_pad=cached_embed_left_pad,
+    )
+    assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    # processed_mask is used to mask out initial states
+    processed_mask = torch.arange(left_context_len, device=x.device).expand(
+        x.size(0), left_context_len
+    )
+    processed_lens = states[-1]  # (batch,)
+    # (batch, left_context_size)
+    processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+    # Update processed lengths
+    new_processed_lens = processed_lens + x_lens
+
+    # (batch, left_context_size + chunk_size)
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    encoder_states = states[:-2]
+    (
+        encoder_out,
+        encoder_out_lens,
+        new_encoder_states,
+    ) = model.encoder.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=encoder_states,
+        src_key_padding_mask=src_key_padding_mask,
+    )
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    new_states = new_encoder_states + [
+        new_cached_embed_left_pad,
+        new_processed_lens,
+    ]
+    return encoder_out, encoder_out_lens, new_states
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+    chunk_size = int(params.chunk_size)
+    left_context_len = int(params.left_context_frames)
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []  # Used in fast-beam-search
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(chunk_size * 2)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # Make sure the length after encoder_embed is at least 1.
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    tail_length = chunk_size * 2 + 7 + 2 * 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+
+    encoder_out, encoder_out_lens, new_states = streaming_forward(
+        features=features,
+        feature_lens=feature_lens,
+        model=model,
+        states=states,
+        chunk_size=chunk_size,
+        left_context_len=left_context_len,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(
+            model=model,
+            encoder_out=encoder_out,
+            streams=decode_streams,
+            blank_penalty=params.blank_penalty,
+        )
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = torch.tensor(processed_lens, device=device)
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+            blank_penalty=params.blank_penalty,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+            blank_penalty=params.blank_penalty,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      lexicon:
+        The Lexicon.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = get_init_states(model=model, batch_size=1, device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        if audio.max() > 1:
+            logging.warning(
+                f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
+                f"Clipping to [-1, 1]."
+            )
+            audio = np.clip(audio, -1, 1)
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=30)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        list(decode_streams[i].ground_truth.strip()),
+                        [
+                            lexicon.token_table[idx]
+                            for idx in decode_streams[i].decoding_result()
+                        ],
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    [
+                        lexicon.token_table[idx]
+                        for idx in decode_streams[i].decoding_result()
+                    ],
+                )
+            )
+            del decode_streams[i]
+
+    key = f"blank_penalty_{params.blank_penalty}"
+    if params.decoding_method == "greedy_search":
+        key = f"greedy_search_{key}"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}_{key}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}_{key}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    assert params.causal, params.causal
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
+    assert (
+        "," not in params.left_context_frames
+    ), "left_context_frames should be one value in decoding."
+    params.suffix += f"-chunk-{params.chunk_size}"
+    params.suffix += f"-left-context-{params.left_context_frames}"
+    params.suffix += f"-blank-penalty-{params.blank_penalty}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+
+    dev_cuts = aishell.valid_cuts()
+    test_cuts = aishell.test_cuts()
+
+    test_sets = ["dev", "test"]
+    test_cuts = [dev_cuts, test_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/zipformer/subsampling.py b/egs/aishell/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/zipformer/train.py b/egs/aishell/ASR/zipformer/train.py
new file mode 100755
index 000000000..d381649e4
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/train.py
@@ -0,0 +1,1350 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 12 \
+  --start-epoch 1 \
+  --exp-dir zipformer/exp \
+  --training-subset L
+  --lr-epochs 1.5 \
+  --max-duration 350
+
+# For mix precision training:
+
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 12 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --training-subset L \
+  --lr-epochs 1.5 \
+  --max-duration 750
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="""Feedforward dimension of the zipformer encoder layers, per stack, comma separated.""",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="""Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="""Embedding dimension in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="""Query/key dimension per head in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="""Value dimension per head in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="""Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="""Unmasked dimensions in the encoders, relates to augmentation during training. A single int or comma-separated list.  Must be <= each corresponding encoder_dim.""",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="""Sizes of convolutional kernels in convolution modules in each encoder stack: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="""Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. Must be just -1 if --causal=False""",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="""Maximum left-contexts for causal training, measured in frames which will
+        be converted to a number of chunks.  If splitting into chunks,
+        chunk left-context frames will be chosen randomly from this list; else not relevant.""",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="""Reference batch duration for purposes of adjusting batch counts for setting various schedules inside the model""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="""The context size in the decoder. 1 means bigram; 2 means tri-gram""",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="""The prune range for rnnt loss, it means how many symbols(context)
+        we are using to compute the loss""",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="""The scale to smooth the loss with lm
+        (output of prediction network) part.""",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="""The scale to smooth the loss with am (output of encoder network) part.""",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="""To get pruning ranges, we will calculate a simple version
+        loss(joiner is just addition), this simple loss also uses for
+        training (as a regularization item). We will scale the simple loss
+        with this parameter before adding to the final loss.""",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, _ = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    aishell = AishellAsrDataModule(args)
+
+    train_cuts = aishell.train_cuts()
+    valid_cuts = aishell.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 15 seconds
+        #
+        # Caution: There is a reason to select 15.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 12.0:
+            # logging.warning(
+            #    f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0]
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_dl = aishell.valid_dataloaders(valid_cuts)
+
+    if False and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      graph_compiler:
+        The compiler to encode texts to ids.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    texts = supervisions["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/zipformer/zipformer.py b/egs/aishell/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/aishell/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/README.md b/egs/aishell2/ASR/README.md
new file mode 100644
index 000000000..4e786af11
--- /dev/null
+++ b/egs/aishell2/ASR/README.md
@@ -0,0 +1,23 @@
+
+# Introduction
+
+This recipe contains various different ASR models trained with Aishell2.
+
+In AISHELL-2, 1000 hours of clean read-speech data from iOS is published, which is free for academic usage. On top of AISHELL-2 corpus, an improved recipe is developed and released, containing key components for industrial applications, such as Chinese word segmentation, flexible vocabulary expension and phone set transformation etc. Pipelines support various state-of-the-art techniques, such as time-delayed neural networks and Lattic-Free MMI objective funciton. In addition, we also release dev and test data from other channels (Android and Mic).
+
+(From [AISHELL-2: Transforming Mandarin ASR Research Into Industrial Scale](https://arxiv.org/abs/1808.10583))
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless5 in librispeech recipe  |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/aishell2/ASR/RESULTS.md b/egs/aishell2/ASR/RESULTS.md
new file mode 100644
index 000000000..32ad74b50
--- /dev/null
+++ b/egs/aishell2/ASR/RESULTS.md
@@ -0,0 +1,87 @@
+## Results
+
+### Aishell2 char-based training results 
+
+#### Pruned transducer stateless 5
+
+Using the codes from this commit https://github.com/k2-fsa/icefall/pull/465.
+
+When training with context size equals to 1, the WERs are
+
+|                                    |  dev-ios  | test-ios | comment                      |
+|------------------------------------|-------|----------|----------------------------------|
+|          greedy search             | 5.57  | 5.89     | --epoch 25, --avg 5, --max-duration 600  |
+| modified beam search (beam size 4) | 5.32  | 5.56     | --epoch 25, --avg 5, --max-duration 600  |
+| fast beam search (set as default)  | 5.5   |  5.78    | --epoch 25, --avg 5, --max-duration 600 |
+| fast beam search nbest             | 5.46  |  5.74    | --epoch 25, --avg 5, --max-duration 600 |
+| fast beam search oracle            | 1.92  |  2.2     | --epoch 25, --avg 5, --max-duration 600 |
+| fast beam search nbest LG          | 5.59  |  5.93    | --epoch 25, --avg 5, --max-duration 600 |
+
+The training command for reproducing is given below:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --lang-dir data/lang_char \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir /result \
+  --max-duration 300 \
+  --use-fp16 0 \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --context-size 1
+```
+
+The decoding command is:
+```bash
+for method in greedy_search modified_beam_search fast_beam_search fast_beam_search_nbest  fast_beam_search_nbest_oracle fast_beam_search_nbest_LG; do
+  ./pruned_transducer_stateless5/decode.py \
+    --epoch 25 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method $method \
+    --max-sym-per-frame 1 \
+    --num-encoder-layers 24 \
+    --dim-feedforward 1536 \
+    --nhead 8 \
+    --encoder-dim 384 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    --context-size 1 \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5 \
+    --context-size 1 \
+    --use-averaged-model True
+done
+```
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/RXyX4QjQQVKjBS2eQ2Qajg/#scalars
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/yuekai/icefall-asr-aishell2-pruned-transducer-stateless5-B-2022-07-12>
+
+When training with context size equals to 2, the WERs are
+
+|                                    |  dev-ios  | test-ios | comment                      |
+|------------------------------------|-------|----------|----------------------------------|
+|          greedy search             | 5.47  |  5.81    | --epoch 25, --avg 5, --max-duration 600  |
+| modified beam search (beam size 4) | 5.38  |  5.61    | --epoch 25, --avg 5, --max-duration 600  |
+| fast beam search (set as default)  | 5.36  |  5.61    | --epoch 25, --avg 5, --max-duration 600  |
+| fast beam search nbest             | 5.37  |  5.6     | --epoch 25, --avg 5, --max-duration 600 |
+| fast beam search oracle            | 2.04  |  2.2     | --epoch 25, --avg 5, --max-duration 600 |
+| fast beam search nbest LG          | 5.59  |  5.82     | --epoch 25, --avg 5, --max-duration 600 |
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/5AxJ8LHoSre8kDAuLp4L7Q/#scalars
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/yuekai/icefall-asr-aishell2-pruned-transducer-stateless5-A-2022-07-12>
diff --git a/egs/aishell2/ASR/local/__init__.py b/egs/aishell2/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell2/ASR/local/compile_lg.py b/egs/aishell2/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/aishell2/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/local/compute_fbank_aishell2.py b/egs/aishell2/ASR/local/compute_fbank_aishell2.py
new file mode 100755
index 000000000..1fb1621ff
--- /dev/null
+++ b/egs/aishell2/ASR/local/compute_fbank_aishell2.py
@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the aishell2 dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_aishell2(num_mel_bins: int = 80, perturb_speed: bool = False):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+    prefix = "aishell2"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and perturb_speed:
+                logging.info(f"Doing speed perturb")
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_aishell2(
+        num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed
+    )
diff --git a/egs/aishell2/ASR/local/compute_fbank_musan.py b/egs/aishell2/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/aishell2/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/local/display_manifest_statistics.py b/egs/aishell2/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..14844cbf3
--- /dev/null
+++ b/egs/aishell2/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in transducer_stateless/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    paths = [
+        "./data/fbank/aishell2_cuts_train.jsonl.gz",
+        "./data/fbank/aishell2_cuts_dev.jsonl.gz",
+        "./data/fbank/aishell2_cuts_test.jsonl.gz",
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest_lazy(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/aishell2_cuts_train.jsonl.gz
+Cuts count: 3026106
+Total duration (hours): 3021.2
+Speech duration (hours): 3021.2 (100.0%)
+***
+Duration statistics (seconds):
+mean	3.6
+std	1.5
+min	0.3
+25%	2.4
+50%	3.3
+75%	4.4
+99%	8.2
+99.5%	8.9
+99.9%	10.6
+max	21.5
+Starting display the statistics for ./data/fbank/aishell2_cuts_dev.jsonl.gz
+Cuts count: 2500
+Total duration (hours): 2.0
+Speech duration (hours): 2.0 (100.0%)
+***
+Duration statistics (seconds):
+mean	2.9
+std	1.0
+min	1.1
+25%	2.2
+50%	2.7
+75%	3.4
+99%	6.3
+99.5%	6.7
+99.9%	7.8
+max	9.4
+Starting display the statistics for ./data/fbank/aishell2_cuts_test.jsonl.gz
+Cuts count: 5000
+Total duration (hours): 4.0
+Speech duration (hours): 4.0 (100.0%)
+***
+Duration statistics (seconds):
+mean	2.9
+std	1.0
+min	1.1
+25%	2.2
+50%	2.7
+75%	3.3
+99%	6.2
+99.5%	6.6
+99.9%	7.7
+max	8.5
+"""
diff --git a/egs/aishell2/ASR/local/prepare_char.py b/egs/aishell2/ASR/local/prepare_char.py
new file mode 120000
index 000000000..8779181e5
--- /dev/null
+++ b/egs/aishell2/ASR/local/prepare_char.py
@@ -0,0 +1 @@
+../../../aidatatang_200zh/ASR/local/prepare_char.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/local/prepare_lang.py b/egs/aishell2/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..5d88dc1c8
--- /dev/null
+++ b/egs/aishell2/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/local/prepare_words.py b/egs/aishell2/ASR/local/prepare_words.py
new file mode 120000
index 000000000..e58fabb8f
--- /dev/null
+++ b/egs/aishell2/ASR/local/prepare_words.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/prepare_words.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/local/text2segments.py b/egs/aishell2/ASR/local/text2segments.py
new file mode 120000
index 000000000..7d68a39c3
--- /dev/null
+++ b/egs/aishell2/ASR/local/text2segments.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/text2segments.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/local/text2token.py b/egs/aishell2/ASR/local/text2token.py
new file mode 120000
index 000000000..81e459d69
--- /dev/null
+++ b/egs/aishell2/ASR/local/text2token.py
@@ -0,0 +1 @@
+../../../aidatatang_200zh/ASR/local/text2token.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/prepare.sh b/egs/aishell2/ASR/prepare.sh
new file mode 100755
index 000000000..a5eb9bd13
--- /dev/null
+++ b/egs/aishell2/ASR/prepare.sh
@@ -0,0 +1,186 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=30
+stage=0
+stop_stage=7
+perturb_speed=true
+
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, you need to apply aishell2 through
+# their official website.
+# https://www.aishelltech.com/aishell_2
+#
+#  - $dl_dir/aishell2
+#
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/aishell2,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/aishell2 $dl_dir/aishell2
+  #
+  # The directory structure is
+  # aishell2/
+  # |-- AISHELL-2
+  # |   |-- iOS
+  #         |-- data
+  #             |-- wav
+  #             |-- trans.txt
+  #         |-- dev
+  #             |-- wav
+  #             |-- trans.txt
+  #         |-- test
+  #             |-- wav
+  #             |-- trans.txt
+
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/musan
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare aishell2 manifest"
+  # We assume that you have downloaded and unzip the aishell2 corpus
+  # to $dl_dir/aishell2
+  if [ ! -f data/manifests/.aishell2_manifests.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare aishell2 $dl_dir/aishell2 data/manifests -j $nj
+    touch data/manifests/.aishell2_manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for aishell2"
+  if [ ! -f data/fbank/.aishell2.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_aishell2.py --perturb-speed ${perturb_speed}
+    touch data/fbank/.aishell2.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+lang_char_dir=data/lang_char
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare char based lang"
+  mkdir -p $lang_char_dir
+
+  # Prepare text.
+  # Note: in Linux, you can install jq with the following command:
+  # 1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+  # 2. chmod +x ./jq
+  # 3. cp jq /usr/bin
+  if [ ! -f $lang_char_dir/text ]; then
+    gunzip -c data/manifests/aishell2_supervisions_train.jsonl.gz \
+      | jq '.text' | sed 's/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text
+  fi
+
+  # The implementation of chinese word segmentation for text,
+  # and it will take about 15 minutes.
+  # If you can't install paddle-tiny with python 3.8, please refer to
+  # https://github.com/fxsjy/jieba/issues/920
+  if [ ! -f $lang_char_dir/text_words_segmentation ]; then
+    python3 ./local/text2segments.py \
+      --input-file $lang_char_dir/text \
+      --output-file $lang_char_dir/text_words_segmentation
+  fi
+
+  cat $lang_char_dir/text_words_segmentation | sed 's/ /\n/g' \
+    | sort -u | sed '/^$/d' | uniq > $lang_char_dir/words_no_ids.txt
+
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    python3 ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    python3 ./local/prepare_char.py
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  if [ ! -f ${lang_char_dir}/3-gram.unpruned.arpa ]; then
+    ./shared/make_kn_lm.py \
+      -ngram-order 3 \
+      -text $lang_char_dir/text_words_segmentation \
+      -lm $lang_char_dir/3-gram.unpruned.arpa
+  fi
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building LG
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_char_dir/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $lang_char_dir/3-gram.unpruned.arpa > data/lm/G_3_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compile LG"
+  ./local/compile_lg.py --lang-dir $lang_char_dir
+fi
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/__init__.py b/egs/aishell2/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 100644
index 000000000..8f6a88f59
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1,406 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AiShell2AsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. ios, android, mic).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to gen cuts from aishell2_cuts_train.jsonl.gz")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell2_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to gen cuts from aishell2_cuts_dev.jsonl.gz")
+        return load_manifest_lazy(self.args.manifest_dir / "aishell2_cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to gen cuts from aishell2_cuts_test.jsonl.gz")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell2_cuts_test.jsonl.gz"
+        )
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/beam_search.py b/egs/aishell2/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/conformer.py b/egs/aishell2/ASR/pruned_transducer_stateless5/conformer.py
new file mode 120000
index 000000000..c7c1a4b6e
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/decode.py b/egs/aishell2/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..9e44b4e34
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,776 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 25 \
+        --avg 5 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --lang-dir data/lang_char \
+        --max-duration 600 \
+        --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 25 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 25 \
+        --avg 5 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --lang-dir data/lang_char \
+        --max-duration 600 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless5/decode.py \
+        --epoch 25 \
+        --avg 5 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --lang-dir data/lang_char \
+        --max-duration 600 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 25 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 25 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 25 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AiShell2AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            sentence = "".join([lexicon.word_table[i] for i in hyp])
+            hyps.append(list(sentence))
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=graph_compiler.texts_to_ids(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AiShell2AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.unk_id = lexicon.token_table["<unk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell2 = AiShell2AsrDataModule(args)
+
+    valid_cuts = aishell2.valid_cuts()
+    test_cuts = aishell2.test_cuts()
+
+    # use ios sets for dev and test
+    dev_dl = aishell2.valid_dataloaders(valid_cuts)
+    test_dl = aishell2.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/decoder.py b/egs/aishell2/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/aishell2/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/export.py b/egs/aishell2/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..8a5be94d0
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --lang-dir data/lang_char
+  --epoch 25 \
+  --avg 5
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell2/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.unk_id = lexicon.token_table["<unk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/joiner.py b/egs/aishell2/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/model.py b/egs/aishell2/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/optim.py b/egs/aishell2/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/pretrained.py b/egs/aishell2/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100755
index 000000000..bc3ae7abf
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,334 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir ./data/lang_char \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir ./data/lang_char \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir ./data/lang_char \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.unk_id = lexicon.token_table["<unk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = "".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/scaling.py b/egs/aishell2/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/train.py b/egs/aishell2/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..8c7448d4c
--- /dev/null
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1108 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+# Copyright    2022  Nvidia                   (authors: Yuekai Zhang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --lang-dir data/lang_char \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --max-duration 300 \
+  --use-fp16 0 \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+# For mix precision training:
+
+./pruned_transducer_stateless5/train.py \
+  --lang-dir data/lang_char \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AiShell2AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    assert type(y) == list
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aishell2 = AiShell2AsrDataModule(args)
+
+    train_cuts = aishell2.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 8 seconds
+        #
+        # Caution: There is a reason to select 8.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 8.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell2.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell2.valid_cuts()
+    valid_dl = aishell2.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+
+def main():
+    parser = get_parser()
+    AiShell2AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell2/ASR/shared b/egs/aishell2/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/aishell2/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/aishell4/ASR/README.md b/egs/aishell4/ASR/README.md
new file mode 100644
index 000000000..67fa17790
--- /dev/null
+++ b/egs/aishell4/ASR/README.md
@@ -0,0 +1,23 @@
+
+# Introduction
+
+This recipe contains some various ASR models trained with Aishell4 (including S, M and L three subsets).
+
+The AISHELL-4 is a sizable real-recorded Mandarin speech dataset collected by 8-channel circular microphone array for speech processing in conference scenarios. The dataset consists of 211 recorded meeting sessions, each containing 4 to 8 speakers, with a total length of 120 hours. This dataset aims to bridge the advanced research on multi-speaker processing and the practical application scenario in three aspects. With real recorded meetings, AISHELL-4 provides realistic acoustics and rich natural speech characteristics in conversation such as short pause, speech overlap, quick speaker turn, noise, etc. Meanwhile, the accurate transcription and speaker voice activity are provided for each meeting in AISHELL-4. This allows the researchers to explore different aspects in meeting processing, ranging from individual tasks such as speech front-end processing, speech recognition and speaker diarization, to multi-modality modeling and joint optimization of relevant tasks. 
+
+(From [Open Speech and Language Resources](https://www.openslr.org/111/))
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/aishell4/ASR/RESULTS.md b/egs/aishell4/ASR/RESULTS.md
new file mode 100644
index 000000000..9bd062f1d
--- /dev/null
+++ b/egs/aishell4/ASR/RESULTS.md
@@ -0,0 +1,117 @@
+## Results
+
+### Aishell4 Char training results (Pruned Transducer Stateless5)
+
+#### 2022-06-13
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/399.
+
+When use-averaged-model=False, the CERs are
+|                                    |    test    | comment                                  |
+|------------------------------------|------------|------------------------------------------|
+|          greedy search             | 30.05      | --epoch 30, --avg 25, --max-duration 800 |
+| modified beam search (beam size 4) | 29.16      | --epoch 30, --avg 25, --max-duration 800 |
+| fast beam search (set as default)  | 29.20      | --epoch 30, --avg 25, --max-duration 1500|
+
+When use-averaged-model=True, the CERs are
+|                                    |    test    | comment                                                              |
+|------------------------------------|------------|----------------------------------------------------------------------|
+|          greedy search             | 29.89      | --iter 36000, --avg 8, --max-duration 800  --use-averaged-model=True |
+| modified beam search (beam size 4) | 28.91      | --iter 36000, --avg 8, --max-duration 800  --use-averaged-model=True |
+| fast beam search (set as default)  | 29.08      | --iter 36000, --avg 8, --max-duration 1500 --use-averaged-model=True |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 4000
+
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/tjaVRKERS8C10SzhpBcxSQ/#scalars
+
+When use-averaged-model=False, the decoding command is:
+```
+epoch=30
+avg=25
+
+## greedy search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 800
+
+## modified beam search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 800 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --lang-dir ./data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+When use-averaged-model=True, the decoding command is:
+```
+iter=36000
+avg=8
+
+## greedy search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 800 \
+  --use-averaged-model True
+
+## modified beam search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 800 \
+  --decoding-method modified_beam_search \
+  --beam-size 4 \
+  --use-averaged-model True
+
+## fast beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --lang-dir ./data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        --use-averaged-model True
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_aishell4_pruned_transducer_stateless5>
diff --git a/egs/aishell4/ASR/local/__init__.py b/egs/aishell4/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell4/ASR/local/compute_fbank_aishell4.py b/egs/aishell4/ASR/local/compute_fbank_aishell4.py
new file mode 100755
index 000000000..f19163988
--- /dev/null
+++ b/egs/aishell4/ASR/local/compute_fbank_aishell4.py
@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the aidatatang_200zh dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import ChunkedLilcomHdf5Writer, CutSet, Fbank, FbankConfig
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_aishell4(num_mel_bins: int = 80, perturb_speed: bool = False):
+    src_dir = Path("data/manifests/aishell4")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train_S",
+        "train_M",
+        "train_L",
+        "test",
+    )
+    prefix = "aishell4"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and perturb_speed:
+                logging.info(f"Doing speed perturb")
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=ChunkedLilcomHdf5Writer,
+            )
+
+            logging.info("About splitting cuts into smaller chunks")
+            cut_set = cut_set.trim_to_supervisions(
+                keep_overlapping=False,
+                min_duration=None,
+            )
+
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_aishell4(
+        num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed
+    )
diff --git a/egs/aishell4/ASR/local/compute_fbank_musan.py b/egs/aishell4/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/aishell4/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/local/display_manifest_statistics.py b/egs/aishell4/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..b79e55eef
--- /dev/null
+++ b/egs/aishell4/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,113 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest
+
+
+def main():
+    paths = [
+        "./data/fbank/cuts_train_S.json.gz",
+        "./data/fbank/cuts_train_M.json.gz",
+        "./data/fbank/cuts_train_L.json.gz",
+        "./data/fbank/cuts_test.json.gz",
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/cuts_train_S.json.gz
+Cuts count: 91995
+Total duration (hours): 95.8
+Speech duration (hours): 95.8 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.7
+std     7.1
+min     0.1
+25%     0.9
+50%     2.5
+75%     5.4
+99%     15.3
+99.5%   17.5
+99.9%   23.3
+max     1021.7
+Starting display the statistics for ./data/fbank/cuts_train_M.json.gz
+Cuts count: 177195
+Total duration (hours): 179.5
+Speech duration (hours): 179.5 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.6
+std     6.4
+min     0.0
+25%     0.9
+50%     2.4
+75%     5.2
+99%     14.9
+99.5%   17.0
+99.9%   23.5
+max     990.4
+Starting display the statistics for ./data/fbank/cuts_train_L.json.gz
+Cuts count: 37572
+Total duration (hours): 49.1
+Speech duration (hours): 49.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    4.7
+std     4.0
+min     0.2
+25%     1.6
+50%     3.7
+75%     6.7
+99%     17.5
+99.5%   19.8
+99.9%   26.2
+max     87.4
+Starting display the statistics for ./data/fbank/cuts_test.json.gz
+Cuts count: 10574
+Total duration (hours): 12.1
+Speech duration (hours): 12.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    4.1
+std     3.4
+min     0.2
+25%     1.4
+50%     3.2
+75%     5.8
+99%     14.4
+99.5%   14.9
+99.9%   16.5
+max     17.9
+"""
diff --git a/egs/aishell4/ASR/local/prepare_char.py b/egs/aishell4/ASR/local/prepare_char.py
new file mode 100755
index 000000000..6b440dfb3
--- /dev/null
+++ b/egs/aishell4/ASR/local/prepare_char.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = list(line)
+            for char in chars:
+                if char not in tokens:
+                    tokens[char] = len(tokens)
+    return tokens
+
+
+def main():
+    lang_dir = Path("data/lang_char")
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/local/prepare_lang.py b/egs/aishell4/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..c8cf9b881
--- /dev/null
+++ b/egs/aishell4/ASR/local/prepare_lang.py
@@ -0,0 +1,388 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone")
+    return parser.parse_args()
+
+
+def main():
+    out_dir = Path(get_args().lang_dir)
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/local/prepare_words.py b/egs/aishell4/ASR/local/prepare_words.py
new file mode 100755
index 000000000..65aca2983
--- /dev/null
+++ b/egs/aishell4/ASR/local/prepare_words.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input words.txt without ids:
+    - words_no_ids.txt
+and generates the new words.txt with related ids.
+    - words.txt
+"""
+
+
+import argparse
+import logging
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare words.txt",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/words_no_ids.txt",
+        type=str,
+        help="the words file without ids for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/words.txt",
+        type=str,
+        help="the words file with ids for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    add_words = ["<eps> 0", "!SIL 1", "<SPOKEN_NOISE> 2", "<UNK> 3"]
+    new_lines.extend(add_words)
+
+    logging.info("Starting reading the input file")
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        idx = 4 + i
+        new_line = str(x.strip("\n")) + " " + str(idx)
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the words.txt")
+    f_out = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/local/test_prepare_lang.py b/egs/aishell4/ASR/local/test_prepare_lang.py
new file mode 100755
index 000000000..74e025ad7
--- /dev/null
+++ b/egs/aishell4/ASR/local/test_prepare_lang.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/local/text2segments.py b/egs/aishell4/ASR/local/text2segments.py
new file mode 100644
index 000000000..3df727c67
--- /dev/null
+++ b/egs/aishell4/ASR/local/text2segments.py
@@ -0,0 +1,83 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input "text", which refers to the transcript file for
+WenetSpeech:
+    - text
+and generates the output file text_word_segmentation which is implemented
+with word segmenting:
+    - text_words_segmentation
+"""
+
+
+import argparse
+
+import jieba
+from tqdm import tqdm
+
+jieba.enable_paddle()
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Chinese Word Segmentation for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/text",
+        type=str,
+        help="the input text file for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/text_words_segmentation",
+        type=str,
+        help="the text implemented with words segmenting for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        x = lines[i].rstrip()
+        seg_list = jieba.cut(x, use_paddle=True)
+        new_line = " ".join(seg_list)
+        new_lines.append(new_line)
+
+    f_new = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_new.write(line)
+        f_new.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/local/text2token.py b/egs/aishell4/ASR/local/text2token.py
new file mode 100755
index 000000000..85047c367
--- /dev/null
+++ b/egs/aishell4/ASR/local/text2token.py
@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument("--space", default="<space>", type=str, help="space symbol")
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument("text", type=str, default=False, nargs="?", help="input text")
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer)
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = "".join(a_flat)
+        print(a_chars)
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/local/text_normalize.py b/egs/aishell4/ASR/local/text_normalize.py
new file mode 100755
index 000000000..5650be502
--- /dev/null
+++ b/egs/aishell4/ASR/local/text_normalize.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Copyright    2022  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input "text_full", which includes three transcript files
+(train_S, train_M and train_L) for AISHELL4:
+    - text_full
+and generates the output file text_normalize which is implemented
+to normalize text:
+    - text
+"""
+
+
+import argparse
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Normalizing for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input",
+        default="data/lang_char/text_full",
+        type=str,
+        help="the input text files for AISHELL4",
+    )
+    parser.add_argument(
+        "--output",
+        default="data/lang_char/text",
+        type=str,
+        help="the text implemented with normalizer for AISHELL4",
+    )
+
+    return parser
+
+
+def text_normalize(str_line: str):
+    line = str_line.strip().rstrip("\n")
+    line = line.replace(" ", "")
+    line = line.replace("<sil>", "")
+    line = line.replace("<%>", "")
+    line = line.replace("<->", "")
+    line = line.replace("<$>", "")
+    line = line.replace("<#>", "")
+    line = line.replace("<_>", "")
+    line = line.replace("<space>", "")
+    line = line.replace("`", "")
+    line = line.replace("&", "")
+    line = line.replace(",", "")
+    line = line.replace("Ａ", "")
+    line = line.replace("ａ", "A")
+    line = line.replace("ｂ", "B")
+    line = line.replace("ｃ", "C")
+    line = line.replace("ｋ", "K")
+    line = line.replace("ｔ", "T")
+    line = line.replace("，", "")
+    line = line.replace("丶", "")
+    line = line.replace("。", "")
+    line = line.replace("、", "")
+    line = line.replace("？", "")
+    line = line.replace("·", "")
+    line = line.replace("*", "")
+    line = line.replace("！", "")
+    line = line.replace("$", "")
+    line = line.replace("+", "")
+    line = line.replace("-", "")
+    line = line.replace("\\", "")
+    line = line.replace("?", "")
+    line = line.replace("￥", "")
+    line = line.replace("%", "")
+    line = line.replace(".", "")
+    line = line.replace("<", "")
+    line = line.replace("＆", "")
+    line = line.upper()
+
+    return line
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input
+    output_file = args.output
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        new_line = text_normalize(lines[i])
+        new_lines.append(new_line)
+
+    f_new = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_new.write(line)
+        f_new.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/prepare.sh b/egs/aishell4/ASR/prepare.sh
new file mode 100755
index 000000000..e8d9eb7b9
--- /dev/null
+++ b/egs/aishell4/ASR/prepare.sh
@@ -0,0 +1,165 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+perturb_speed=true
+
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/aishell4
+#      You can find four directories:train_S, train_M, train_L and test.
+#      You can download it from https://openslr.org/111/
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/aishell4,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/aishell4 $dl_dir/aishell4
+  #
+  if [ ! -f $dl_dir/aishell4/train_L ]; then
+    lhotse download aishell4  $dl_dir/aishell4
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/musan
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare aishell4 manifest"
+  # We assume that you have downloaded the aishell4 corpus
+  # to $dl_dir/aishell4
+  if [ ! -f data/manifests/aishell4/.manifests.done ]; then
+    mkdir -p data/manifests/aishell4
+    lhotse prepare aishell4 $dl_dir/aishell4 data/manifests/aishell4
+    touch data/manifests/aishell4/.manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Process aishell4"
+  if [ ! -f data/fbank/aishell4/.fbank.done ]; then
+    mkdir -p data/fbank/aishell4
+    ./local/compute_fbank_aishell4.py --perturb-speed ${perturb_speed}
+    touch data/fbank/aishell4/.fbank.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for aishell4"
+  if [ ! -f data/fbank/.aishell4.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_aishell4.py --perturb-speed ${perturb_speed}
+    touch data/fbank/.aishell4.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+
+  # Prepare text.
+  # Note: in Linux, you can install jq with the  following command:
+  # wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+  gunzip -c data/manifests/aishell4/aishell4_supervisions_train_S.jsonl.gz \
+    | jq ".text" | sed 's/"//g' \
+    | ./local/text2token.py -t "char" > $lang_char_dir/text_S
+
+  gunzip -c data/manifests/aishell4/aishell4_supervisions_train_M.jsonl.gz \
+    | jq ".text" | sed 's/"//g' \
+    | ./local/text2token.py -t "char" > $lang_char_dir/text_M
+
+  gunzip -c data/manifests/aishell4/aishell4_supervisions_train_L.jsonl.gz \
+    | jq ".text" | sed 's/"//g' \
+    | ./local/text2token.py -t "char" > $lang_char_dir/text_L
+
+  for r in text_S text_M text_L ; do
+    cat $lang_char_dir/$r >> $lang_char_dir/text_full
+  done
+
+  # Prepare text normalize
+  python ./local/text_normalize.py \
+    --input $lang_char_dir/text_full \
+    --output $lang_char_dir/text
+
+  # Prepare words segments
+  python ./local/text2segments.py \
+    --input $lang_char_dir/text \
+    --output $lang_char_dir/text_words_segmentation
+
+  cat $lang_char_dir/text_words_segmentation | sed "s/ /\n/g" \
+    | sort -u | sed "/^$/d" \
+    | uniq > $lang_char_dir/words_no_ids.txt
+
+  # Prepare words.txt
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    ./local/prepare_char.py
+  fi
+fi
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/__init__.py b/egs/aishell4/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 100644
index 000000000..e6db2651f
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1,438 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 for AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class Aishell4AsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=300,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                buffer_size=100000,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            rank=0,
+            world_size=1,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            rank=0,
+            world_size=1,
+            shuffle=False,
+        )
+        logging.info("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_S_cuts(self) -> CutSet:
+        logging.info("About to get S train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell4_cuts_train_S.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_M_cuts(self) -> CutSet:
+        logging.info("About to get M train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell4_cuts_train_M.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_L_cuts(self) -> CutSet:
+        logging.info("About to get L train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell4_cuts_train_L.jsonl.gz"
+        )
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        # Aishell4 doesn't have dev data, here use test to replace dev.
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell4_cuts_test.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "aishell4_cuts_test.jsonl.gz"
+        )
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/beam_search.py b/egs/aishell4/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..ed78bd4bb
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/conformer.py b/egs/aishell4/ASR/pruned_transducer_stateless5/conformer.py
new file mode 120000
index 000000000..c7c1a4b6e
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/decode.py b/egs/aishell4/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..068e2749a
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,623 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When use-averaged-model=True, usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+    --iter 36000 \
+    --avg 8 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 800 \
+    --decoding-method greedy_search \
+    --use-averaged-model True
+
+(2) modified beam search
+./pruned_transducer_stateless5/decode.py \
+    --iter 36000 \
+    --avg 8 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 800 \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    --use-averaged-model True
+
+(3) fast beam search
+./pruned_transducer_stateless5/decode.py \
+    --iter 36000 \
+    --avg 8 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 800 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8 \
+    --use-averaged-model True
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import Aishell4AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from local.text_normalize import text_normalize
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text).replace(" ", "")) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Aishell4AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    def text_normalize_for_cut(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = text.strip("\n").strip("\t")
+        c.supervisions[0].text = text_normalize(text)
+        return c
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell4 = Aishell4AsrDataModule(args)
+    test_cuts = aishell4.test_cuts()
+    test_cuts = test_cuts.map(text_normalize_for_cut)
+    test_dl = aishell4.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dl = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/decoder.py b/egs/aishell4/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..8a5e07bd5
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/aishell4/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..2fc10439b
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/export.py b/egs/aishell4/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..bf9856c60
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,281 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --lang-dir data/lang_char \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell4/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/joiner.py b/egs/aishell4/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..f31b5fd9b
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/local b/egs/aishell4/ASR/pruned_transducer_stateless5/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/model.py b/egs/aishell4/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..be059ba7c
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/optim.py b/egs/aishell4/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..661206562
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/pretrained.py b/egs/aishell4/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100755
index 000000000..ee898c303
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When use-averaged-model=True, usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir data/lang_char \
+    --decoding-method greedy_search \
+    --use-averaged-model True \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir data/lang_char \
+    --use-averaged-model True \
+    --decoding-method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search (not suggest)
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir data/lang_char \
+    --use-averaged-model True \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir data/lang_char \
+    --use-averaged-model True \
+    --decoding-method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --decoding-method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    if params.decoding_method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding-method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/scaling.py b/egs/aishell4/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..be7b111c6
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/test_model.py b/egs/aishell4/ASR/pruned_transducer_stateless5/test_model.py
new file mode 100755
index 000000000..d42c3b4f4
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/test_model.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/aishell4/ASR
+    python ./pruned_transducer_stateless5/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 18
+    params.dim_feedforward = 1024
+    params.encoder_dim = 256
+    params.nhead = 4
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/train.py b/egs/aishell4/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..a354f761e
--- /dev/null
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1079 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import Aishell4AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from local.text_normalize import text_normalize
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 100,
+            "valid_interval": 200,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 400,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+        # print(batch["supervisions"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aishell4 = Aishell4AsrDataModule(args)
+    # Combine all of the training data
+    train_cuts = aishell4.train_S_cuts()
+    train_cuts += aishell4.train_M_cuts()
+    train_cuts += aishell4.train_L_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    def text_normalize_for_cut(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = text.strip("\n").strip("\t")
+        c.supervisions[0].text = text_normalize(text)
+        return c
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.map(text_normalize_for_cut)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell4.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell4.valid_cuts()
+    valid_cuts = valid_cuts.map(text_normalize_for_cut)
+    valid_dl = aishell4.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    Aishell4AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell4/ASR/shared b/egs/aishell4/ASR/shared
new file mode 120000
index 000000000..3a3b28f96
--- /dev/null
+++ b/egs/aishell4/ASR/shared
@@ -0,0 +1 @@
+../../../egs/aishell/ASR/shared
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/README.md b/egs/alimeeting/ASR/README.md
new file mode 100644
index 000000000..257fe38d5
--- /dev/null
+++ b/egs/alimeeting/ASR/README.md
@@ -0,0 +1,19 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with Alimeeting (far).
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless2`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/alimeeting/ASR/RESULTS.md b/egs/alimeeting/ASR/RESULTS.md
new file mode 100644
index 000000000..745795a20
--- /dev/null
+++ b/egs/alimeeting/ASR/RESULTS.md
@@ -0,0 +1,71 @@
+## Results
+
+### Alimeeting Char training results (Pruned Transducer Stateless2)
+
+#### 2022-06-01
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/378.
+
+The WERs are
+|                                    |     eval   |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 31.77      | 34.66      | --epoch 29, --avg 18, --max-duration 100 |
+| modified beam search (beam size 4) | 30.38      | 33.02      | --epoch 29, --avg 18, --max-duration 100 |
+| fast beam search (set as default)  | 31.39      | 34.25      | --epoch 29, --avg 18, --max-duration 1500|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 1000
+
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/AoqgSvZKTZCJhJbOuG3W6g/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=18
+
+## greedy search
+./pruned_transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 100
+
+## modified beam search
+./pruned_transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir ./data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_alimeeting_pruned_transducer_stateless2>
diff --git a/egs/alimeeting/ASR/local/__init__.py b/egs/alimeeting/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py b/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py
new file mode 100755
index 000000000..f8c10648a
--- /dev/null
+++ b/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py
@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the aishell dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_alimeeting(num_mel_bins: int = 80, perturb_speed: bool = False):
+    src_dir = Path("data/manifests/alimeeting")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "eval",
+        "test",
+    )
+
+    prefix = "alimeeting"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and perturb_speed:
+                logging.info(f"Doing speed perturb")
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cur_num_jobs = num_jobs if ex is None else 80
+            cur_num_jobs = min(cur_num_jobs, len(cut_set))
+
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=cur_num_jobs,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+
+            logging.info("About splitting cuts into smaller chunks")
+            cut_set = cut_set.trim_to_supervisions(
+                keep_overlapping=False,
+                min_duration=None,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_alimeeting(
+        num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed
+    )
diff --git a/egs/alimeeting/ASR/local/compute_fbank_musan.py b/egs/alimeeting/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/alimeeting/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/local/display_manifest_statistics.py b/egs/alimeeting/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..16cdecc91
--- /dev/null
+++ b/egs/alimeeting/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,96 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    paths = [
+        "./data/fbank/alimeeting_cuts_train.jsonl.gz",
+        "./data/fbank/alimeeting_cuts_eval.jsonl.gz",
+        "./data/fbank/alimeeting_cuts_test.jsonl.gz",
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest_lazy(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/alimeeting_cuts_train.jsonl.gz
+Cuts count: 559092
+Total duration (hours): 424.6
+Speech duration (hours): 424.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.7
+std     3.0
+min     0.0
+25%     0.7
+50%     1.7
+75%     3.6
+99%     13.6
+99.5%   14.7
+99.9%   16.2
+max     284.3
+Starting display the statistics for ./data/fbank/alimeeting_cuts_eval.jsonl.gz
+Cuts count: 6457
+Total duration (hours): 4.9
+Speech duration (hours): 4.9 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.7
+std     3.1
+min     0.1
+25%     0.6
+50%     1.6
+75%     3.5
+99%     13.6
+99.5%   14.1
+99.9%   14.7
+max     15.8
+Starting display the statistics for ./data/fbank/alimeeting_cuts_test.jsonl.gz
+Cuts count: 16358
+Total duration (hours): 12.5
+Speech duration (hours): 12.5 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.7
+std     2.9
+min     0.1
+25%     0.7
+50%     1.7
+75%     3.5
+99%     13.7
+99.5%   14.2
+99.9%   14.8
+max     15.7
+"""
diff --git a/egs/alimeeting/ASR/local/prepare_char.py b/egs/alimeeting/ASR/local/prepare_char.py
new file mode 100755
index 000000000..6b440dfb3
--- /dev/null
+++ b/egs/alimeeting/ASR/local/prepare_char.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = list(line)
+            for char in chars:
+                if char not in tokens:
+                    tokens[char] = len(tokens)
+    return tokens
+
+
+def main():
+    lang_dir = Path("data/lang_char")
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/local/prepare_lang.py b/egs/alimeeting/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..c8cf9b881
--- /dev/null
+++ b/egs/alimeeting/ASR/local/prepare_lang.py
@@ -0,0 +1,388 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone")
+    return parser.parse_args()
+
+
+def main():
+    out_dir = Path(get_args().lang_dir)
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/local/prepare_words.py b/egs/alimeeting/ASR/local/prepare_words.py
new file mode 100755
index 000000000..65aca2983
--- /dev/null
+++ b/egs/alimeeting/ASR/local/prepare_words.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input words.txt without ids:
+    - words_no_ids.txt
+and generates the new words.txt with related ids.
+    - words.txt
+"""
+
+
+import argparse
+import logging
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare words.txt",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/words_no_ids.txt",
+        type=str,
+        help="the words file without ids for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/words.txt",
+        type=str,
+        help="the words file with ids for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    add_words = ["<eps> 0", "!SIL 1", "<SPOKEN_NOISE> 2", "<UNK> 3"]
+    new_lines.extend(add_words)
+
+    logging.info("Starting reading the input file")
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        idx = 4 + i
+        new_line = str(x.strip("\n")) + " " + str(idx)
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the words.txt")
+    f_out = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/local/test_prepare_lang.py b/egs/alimeeting/ASR/local/test_prepare_lang.py
new file mode 100755
index 000000000..74e025ad7
--- /dev/null
+++ b/egs/alimeeting/ASR/local/test_prepare_lang.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/local/text2segments.py b/egs/alimeeting/ASR/local/text2segments.py
new file mode 100644
index 000000000..27b904fc8
--- /dev/null
+++ b/egs/alimeeting/ASR/local/text2segments.py
@@ -0,0 +1,85 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input "text", which refers to the transcript file for
+WenetSpeech:
+    - text
+and generates the output file text_word_segmentation which is implemented
+with word segmenting:
+    - text_words_segmentation
+"""
+
+
+import argparse
+
+import jieba
+import paddle
+from tqdm import tqdm
+
+paddle.enable_static()
+jieba.enable_paddle()
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Chinese Word Segmentation for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/text",
+        type=str,
+        help="the input text file for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/text_words_segmentation",
+        type=str,
+        help="the text implemented with words segmenting for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        x = lines[i].rstrip()
+        seg_list = jieba.cut(x, use_paddle=True)
+        new_line = " ".join(seg_list)
+        new_lines.append(new_line)
+
+    f_new = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_new.write(line)
+        f_new.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/local/text2token.py b/egs/alimeeting/ASR/local/text2token.py
new file mode 100755
index 000000000..85047c367
--- /dev/null
+++ b/egs/alimeeting/ASR/local/text2token.py
@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument("--space", default="<space>", type=str, help="space symbol")
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument("text", type=str, default=False, nargs="?", help="input text")
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer)
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = "".join(a_flat)
+        print(a_chars)
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/prepare.sh b/egs/alimeeting/ASR/prepare.sh
new file mode 100755
index 000000000..c8fed658d
--- /dev/null
+++ b/egs/alimeeting/ASR/prepare.sh
@@ -0,0 +1,137 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+perturb_speed=true
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/alimeeting
+#     This directory contains the following files downloaded from
+#       https://openslr.org/62/
+#
+#     - Train_Ali_far.tar.gz
+#     - Train_Ali_near.tar.gz
+#     - Test_Ali.tar.gz
+#     - Eval_Ali.tar.gz
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  if [ ! -f $dl_dir/alimeeting/Train_Ali_far.tar.gz ]; then
+    lhotse download ali-meeting $dl_dir/alimeeting
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare alimeeting manifest"
+  # We assume that you have downloaded the alimeeting corpus
+  # to $dl_dir/alimeeting
+  if [ ! -f data/manifests/alimeeting/.manifests.done ]; then
+    mkdir -p data/manifests/alimeeting
+    lhotse prepare ali-meeting $dl_dir/alimeeting data/manifests/alimeeting
+    touch data/manifests/alimeeting/.manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Process alimeeting"
+  if [ ! -f data/fbank/alimeeting/.fbank.done ]; then
+    mkdir -p data/fbank/alimeeting
+    ./local/compute_fbank_alimeeting.py --perturb-speed ${perturb_speed}
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for alimeeting"
+  if [ ! -f data/fbank/.alimeeting.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_alimeeting.py --perturb-speed True
+    touch data/fbank/.alimeeting.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+
+  # Prepare text.
+  # Note: in Linux, you can install jq with the  following command:
+  # wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+  gunzip -c data/manifests/alimeeting/alimeeting_supervisions_train.jsonl.gz \
+    | jq ".text" | sed 's/"//g' \
+    | ./local/text2token.py -t "char" > $lang_char_dir/text
+
+  # Prepare words segments
+  python ./local/text2segments.py \
+    --input $lang_char_dir/text \
+    --output $lang_char_dir/text_words_segmentation
+
+  cat $lang_char_dir/text_words_segmentation | sed "s/ /\n/g" \
+    | sort -u | sed "/^$/d" \
+    | uniq > $lang_char_dir/words_no_ids.txt
+
+  # Prepare words.txt
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    ./local/prepare_char.py
+  fi
+fi
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/__init__.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 100644
index 000000000..5ad80817a
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1,411 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    load_manifest,
+    load_manifest_lazy,
+    set_caching_enabled,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+set_caching_enabled(False)
+torch.set_num_threads(1)
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AlimeetingAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/dev/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=300,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                buffer_size=30000,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        dev_iter_dataset = IterableDatasetWrapper(
+            dataset=validate,
+            sampler=valid_sampler,
+        )
+        valid_dl = DataLoader(
+            dev_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        test_iter_dataset = IterableDatasetWrapper(
+            dataset=test,
+            sampler=sampler,
+        )
+        test_dl = DataLoader(
+            test_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "alimeeting_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "alimeeting_cuts_eval.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "alimeeting_cuts_test.jsonl.gz"
+        )
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/beam_search.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/conformer.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/conformer.py
new file mode 120000
index 000000000..a65957180
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/decode.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..6c170c392
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,596 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When training with the far data, usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 29 \
+        --avg 18 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 29 \
+        --avg 18 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 29 \
+        --avg 18 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AlimeetingAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--batch",
+        type=int,
+        default=None,
+        help="It specifies the batch checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 50
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text).replace(" ", "")) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AlimeetingAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    elif params.batch is not None:
+        filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints([filenames], device=device))
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+        average = average_checkpoints(filenames, device=device)
+        checkpoint = {"model": average}
+        torch.save(
+            checkpoint,
+            "pruned_transducer_stateless2/exp/pretrained_epoch_29_avg_18.pt",
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # Note: Please use "pip install webdataset==0.1.103"
+    # for installing the webdataset.
+    import glob
+    import os
+
+    from lhotse import CutSet
+    from lhotse.dataset.webdataset import export_to_webdataset
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    alimeeting = AlimeetingAsrDataModule(args)
+
+    dev = "eval"
+    test = "test"
+
+    if not os.path.exists(f"{dev}/shared-0.tar"):
+        os.makedirs(dev)
+        dev_cuts = alimeeting.valid_cuts()
+        export_to_webdataset(
+            dev_cuts,
+            output_path=f"{dev}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    if not os.path.exists(f"{test}/shared-0.tar"):
+        os.makedirs(test)
+        test_cuts = alimeeting.test_cuts()
+        export_to_webdataset(
+            test_cuts,
+            output_path=f"{test}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    dev_shards = [
+        str(path) for path in sorted(glob.glob(os.path.join(dev, "shared-*.tar")))
+    ]
+    cuts_dev_webdataset = CutSet.from_webdataset(
+        dev_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    test_shards = [
+        str(path) for path in sorted(glob.glob(os.path.join(test, "shared-*.tar")))
+    ]
+    cuts_test_webdataset = CutSet.from_webdataset(
+        test_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    def remove_short_and_long_utt(c: Cut):
+        return 1.0 <= c.duration
+
+    cuts_dev_webdataset = cuts_dev_webdataset.filter(remove_short_and_long_utt)
+    cuts_test_webdataset = cuts_test_webdataset.filter(remove_short_and_long_utt)
+
+    dev_dl = alimeeting.valid_dataloaders(cuts_dev_webdataset)
+    test_dl = alimeeting.test_dataloaders(cuts_test_webdataset)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/decoder.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py
new file mode 100644
index 000000000..8e5cc6075
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,178 @@
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 29 \
+  --avg 18
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/alimeeting/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/joiner.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/model.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/optim.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/pretrained.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/pretrained.py
new file mode 100644
index 000000000..f5a0dd8c8
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/pretrained.py
@@ -0,0 +1,338 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		         2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Here, the far data is used for training, usage:
+
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(3) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
+./pruned_transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and modified_beam_search ",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/scaling.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py
new file mode 100644
index 000000000..30154291d
--- /dev/null
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,958 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 1000
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 1000
+  --use-fp16 True
+
+"""
+
+import argparse
+import logging
+import os
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AlimeetingAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12359,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+    Explanation of options saved in `params`:
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+        - best_train_epoch: It is the epoch that has the best training loss.
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+        - subsampling_factor:  The subsampling factor for the model.
+        - encoder_dim: Hidden dim for multi-head attention model.
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 10,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 400,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 200,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    alimeeting = AlimeetingAsrDataModule(args)
+
+    train_cuts = alimeeting.train_cuts()
+    valid_cuts = alimeeting.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 15.0 seconds
+        #
+        # Caution: There is a reason to select 10.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 15.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    valid_dl = alimeeting.valid_dataloaders(valid_cuts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = alimeeting.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if not params.print_diagnostics and params.start_batch == 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AlimeetingAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR/shared b/egs/alimeeting/ASR/shared
new file mode 120000
index 000000000..3a3b28f96
--- /dev/null
+++ b/egs/alimeeting/ASR/shared
@@ -0,0 +1 @@
+../../../egs/aishell/ASR/shared
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/README.md b/egs/alimeeting/ASR_v2/README.md
new file mode 100644
index 000000000..f70327501
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/README.md
@@ -0,0 +1,38 @@
+
+# Introduction
+
+This recipe trains multi-domain ASR models for AliMeeting. By multi-domain, we mean that
+we train a single model on close-talk and far-field conditions. This recipe optionally
+uses [GSS]-based enhancement for far-field array microphone.
+We pool data in the following 4 ways and train a single model on the pooled data:
+
+(i) individual headset microphone (IHM)
+(ii) IHM with simulated reverb
+(iii) Single distant microphone (SDM)
+(iv) GSS-enhanced array microphones
+
+This is different from `alimeeting/ASR` since that recipe trains a model only on the
+far-field audio. Additionally, we use text normalization here similar to the original
+M2MeT challenge, so the results should be more comparable to those from Table 4 of
+the [paper](https://arxiv.org/abs/2110.07393).
+
+The following additional packages need to be installed to run this recipe:
+* `pip install jieba`
+* `pip install paddlepaddle`
+* `pip install git+https://github.com/desh2608/gss.git`
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+## Performance Record
+
+### pruned_transducer_stateless7
+
+The following are decoded using `modified_beam_search`:
+
+| Evaluation set           | eval WER    | test WER |
+|--------------------------|------------|---------|
+| IHM                      |  9.58  | 11.53 |
+| SDM                      |  23.37  | 25.85 |
+| MDM (GSS-enhanced)       |  11.82  | 14.22 |
+
+See [RESULTS](/egs/alimeeting/ASR_v2/RESULTS.md) for details.
diff --git a/egs/alimeeting/ASR_v2/RESULTS.md b/egs/alimeeting/ASR_v2/RESULTS.md
new file mode 100644
index 000000000..15b24250d
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/RESULTS.md
@@ -0,0 +1,90 @@
+## Results (CER)
+
+#### 2022-12-09
+
+#### Zipformer (pruned_transducer_stateless7)
+
+Zipformer encoder + non-current decoder. The decoder
+contains only an embedding layer, a Conv1d (with kernel size 2) and a linear
+layer (to transform tensor dim).
+
+All the results below are using a single model that is trained by combining the following
+data: IHM, IHM+reverb, SDM, and GSS-enhanced MDM. Speed perturbation and MUSAN noise
+augmentation are applied on top of the pooled data.
+
+**WERs for IHM:**
+
+|                           | eval | test | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             |  10.13  |  12.21  | --epoch 15 --avg 8 --max-duration 500 |
+| modified beam search      |  9.58  |  11.53  | --epoch 15 --avg 8 --max-duration 500 --beam-size 4 |
+| fast beam search          |  9.92  |  12.07  | --epoch 15 --avg 8 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+**WERs for SDM:**
+
+|                           | eval | test | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             |  23.70  |  26.41  | --epoch 15 --avg 8 --max-duration 500 |
+| modified beam search      |  23.37  |  25.85  | --epoch 15 --avg 8 --max-duration 500 --beam-size 4 |
+| fast beam search          |  23.60  |  26.38  | --epoch 15 --avg 8 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+**WERs for GSS-enhanced MDM:**
+
+|                           | eval | test | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             |  12.24  |  14.99  | --epoch 15 --avg 8 --max-duration 500 |
+| modified beam search      |  11.82  |  14.22  | --epoch 15 --avg 8 --max-duration 500 --beam-size 4 |
+| fast beam search          |  12.30  |  14.98  | --epoch 15 --avg 8 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 15 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 300 \
+  --max-cuts 100 \
+  --prune-range 5 \
+  --lr-factor 5 \
+  --lm-scale 0.25 \
+  --use-fp16 True
+```
+
+The decoding command is:
+```
+# greedy search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 15 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method greedy_search
+
+# modified beam search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 15 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+# fast beam search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 15 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+Pretrained model is available at <https://huggingface.co/desh2608/icefall-asr-alimeeting-pruned-transducer-stateless7>
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/EzmVahMMTb2YfKWXwQ2dyQ/#scalars>
diff --git a/egs/alimeeting/ASR_v2/local/__init__.py b/egs/alimeeting/ASR_v2/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/alimeeting/ASR_v2/local/compute_fbank_alimeeting.py b/egs/alimeeting/ASR_v2/local/compute_fbank_alimeeting.py
new file mode 100755
index 000000000..833d11c72
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/compute_fbank_alimeeting.py
@@ -0,0 +1,217 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the AliMeeting dataset.
+For the training data, we prepare IHM, reverberated IHM, SDM, and GSS-enhanced
+audios. For the test data, we separately prepare IHM, SDM, and GSS-enhanced
+parts (which are the 3 evaluation settings).
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+from lhotse import CutSet, LilcomChunkyWriter
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_ami(perturb_speed: bool = False):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    manifests_ihm = read_manifests_if_cached(
+        dataset_parts=["train", "eval", "test"],
+        output_dir=src_dir,
+        prefix="alimeeting-ihm",
+        suffix="jsonl.gz",
+    )
+    manifests_sdm = read_manifests_if_cached(
+        dataset_parts=["train", "eval", "test"],
+        output_dir=src_dir,
+        prefix="alimeeting-sdm",
+        suffix="jsonl.gz",
+    )
+    # For GSS we already have cuts so we read them directly.
+    manifests_gss = read_manifests_if_cached(
+        dataset_parts=["train", "eval", "test"],
+        output_dir=src_dir,
+        prefix="alimeeting-gss",
+        suffix="jsonl.gz",
+    )
+
+    def _extract_feats(
+        cuts: CutSet, storage_path: Path, manifest_path: Path, speed_perturb: bool
+    ) -> None:
+        if speed_perturb:
+            logging.info(f"Doing speed perturb")
+            cuts = cuts + cuts.perturb_speed(0.9) + cuts.perturb_speed(1.1)
+        _ = cuts.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=storage_path,
+            manifest_path=manifest_path,
+            batch_duration=5000,
+            num_workers=8,
+            storage_type=LilcomChunkyWriter,
+        )
+
+    logging.info(
+        "Preparing training cuts: IHM + reverberated IHM + SDM + GSS (optional)"
+    )
+
+    logging.info("Processing train split IHM")
+    cuts_ihm = (
+        CutSet.from_manifests(**manifests_ihm["train"])
+        .trim_to_supervisions(keep_overlapping=False, keep_all_channels=False)
+        .modify_ids(lambda x: x + "-ihm")
+    )
+    _extract_feats(
+        cuts_ihm,
+        output_dir / "feats_train_ihm",
+        src_dir / "cuts_train_ihm.jsonl.gz",
+        perturb_speed,
+    )
+
+    logging.info("Processing train split IHM + reverberated IHM")
+    cuts_ihm_rvb = cuts_ihm.reverb_rir()
+    _extract_feats(
+        cuts_ihm_rvb,
+        output_dir / "feats_train_ihm_rvb",
+        src_dir / "cuts_train_ihm_rvb.jsonl.gz",
+        perturb_speed,
+    )
+
+    logging.info("Processing train split SDM")
+    cuts_sdm = (
+        CutSet.from_manifests(**manifests_sdm["train"])
+        .trim_to_supervisions(keep_overlapping=False)
+        .modify_ids(lambda x: x + "-sdm")
+    )
+    _extract_feats(
+        cuts_sdm,
+        output_dir / "feats_train_sdm",
+        src_dir / "cuts_train_sdm.jsonl.gz",
+        perturb_speed,
+    )
+
+    logging.info("Processing train split GSS")
+    cuts_gss = (
+        CutSet.from_manifests(**manifests_gss["train"])
+        .trim_to_supervisions(keep_overlapping=False)
+        .modify_ids(lambda x: x + "-gss")
+    )
+    _extract_feats(
+        cuts_gss,
+        output_dir / "feats_train_gss",
+        src_dir / "cuts_train_gss.jsonl.gz",
+        perturb_speed,
+    )
+
+    logging.info("Preparing test cuts: IHM, SDM, GSS (optional)")
+    for split in ["eval", "test"]:
+        logging.info(f"Processing {split} IHM")
+        cuts_ihm = (
+            CutSet.from_manifests(**manifests_ihm[split])
+            .trim_to_supervisions(keep_overlapping=False, keep_all_channels=False)
+            .compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{split}_ihm",
+                manifest_path=src_dir / f"cuts_{split}_ihm.jsonl.gz",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        logging.info(f"Processing {split} SDM")
+        cuts_sdm = (
+            CutSet.from_manifests(**manifests_sdm[split])
+            .trim_to_supervisions(keep_overlapping=False)
+            .compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{split}_sdm",
+                manifest_path=src_dir / f"cuts_{split}_sdm.jsonl.gz",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        logging.info(f"Processing {split} GSS")
+        cuts_gss = (
+            CutSet.from_manifests(**manifests_gss[split])
+            .trim_to_supervisions(keep_overlapping=False)
+            .compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{split}_gss",
+                manifest_path=src_dir / f"cuts_{split}_gss.jsonl.gz",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+
+    compute_fbank_ami(perturb_speed=args.perturb_speed)
diff --git a/egs/alimeeting/ASR_v2/local/compute_fbank_musan.py b/egs/alimeeting/ASR_v2/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/local/prepare_alimeeting_enhanced.py b/egs/alimeeting/ASR_v2/local/prepare_alimeeting_enhanced.py
new file mode 100644
index 000000000..f1512efa5
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/prepare_alimeeting_enhanced.py
@@ -0,0 +1,158 @@
+#!/usr/local/bin/python
+# -*- coding: utf-8 -*-
+# Data preparation for AliMeeting GSS-enhanced dataset.
+
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+from lhotse import Recording, RecordingSet, SupervisionSet
+from lhotse.qa import fix_manifests
+from lhotse.recipes.utils import read_manifests_if_cached
+from lhotse.utils import fastcopy
+from tqdm import tqdm
+
+logging.basicConfig(
+    format="%(asctime)s %(levelname)-8s %(message)s",
+    level=logging.INFO,
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+
+
+def get_args():
+    import argparse
+
+    parser = argparse.ArgumentParser(description="AMI enhanced dataset preparation.")
+    parser.add_argument(
+        "manifests_dir",
+        type=Path,
+        help="Path to directory containing AliMeeting manifests.",
+    )
+    parser.add_argument(
+        "enhanced_dir",
+        type=Path,
+        help="Path to enhanced data directory.",
+    )
+    parser.add_argument(
+        "--num-jobs",
+        "-j",
+        type=int,
+        default=1,
+        help="Number of parallel jobs to run.",
+    )
+    parser.add_argument(
+        "--min-segment-duration",
+        "-d",
+        type=float,
+        default=0.0,
+        help="Minimum duration of a segment in seconds.",
+    )
+    return parser.parse_args()
+
+
+def find_recording_and_create_new_supervision(enhanced_dir, supervision):
+    """
+    Given a supervision (corresponding to original AMI recording), this function finds the
+    enhanced recording correspoding to the supervision, and returns this recording and
+    a new supervision whose start and end times are adjusted to match the enhanced recording.
+    """
+    file_name = Path(
+        f"{supervision.recording_id}-{supervision.speaker}-{int(100*supervision.start):06d}_{int(100*supervision.end):06d}.flac"
+    )
+    save_path = enhanced_dir / f"{supervision.recording_id}" / file_name
+    if save_path.exists():
+        recording = Recording.from_file(save_path)
+        if recording.duration == 0:
+            logging.warning(f"Skipping {save_path} which has duration 0 seconds.")
+            return None
+
+        # Old supervision is wrt to the original recording, we create new supervision
+        # wrt to the enhanced segment
+        new_supervision = fastcopy(
+            supervision,
+            recording_id=recording.id,
+            start=0,
+            duration=recording.duration,
+        )
+        return recording, new_supervision
+    else:
+        logging.warning(f"{save_path} does not exist.")
+        return None
+
+
+def main(args):
+    # Get arguments
+    manifests_dir = args.manifests_dir
+    enhanced_dir = args.enhanced_dir
+
+    # Load manifests from cache if they exist (saves time)
+    manifests = read_manifests_if_cached(
+        dataset_parts=["train", "eval", "test"],
+        output_dir=manifests_dir,
+        prefix="alimeeting-sdm",
+        suffix="jsonl.gz",
+    )
+    if not manifests:
+        raise ValueError(
+            "AliMeeting SDM manifests not found in {}".format(manifests_dir)
+        )
+
+    with ThreadPoolExecutor(args.num_jobs) as ex:
+        for part in ["train", "eval", "test"]:
+            logging.info(f"Processing {part}...")
+            supervisions_orig = manifests[part]["supervisions"].filter(
+                lambda s: s.duration >= args.min_segment_duration
+            )
+            futures = []
+
+            for supervision in tqdm(
+                supervisions_orig,
+                desc="Distributing tasks",
+            ):
+                futures.append(
+                    ex.submit(
+                        find_recording_and_create_new_supervision,
+                        enhanced_dir,
+                        supervision,
+                    )
+                )
+
+            recordings = []
+            supervisions = []
+            for future in tqdm(
+                futures,
+                total=len(futures),
+                desc="Processing tasks",
+            ):
+                result = future.result()
+                if result is not None:
+                    recording, new_supervision = result
+                    recordings.append(recording)
+                    supervisions.append(new_supervision)
+
+            # Remove duplicates from the recordings
+            recordings_nodup = {}
+            for recording in recordings:
+                if recording.id not in recordings_nodup:
+                    recordings_nodup[recording.id] = recording
+                else:
+                    logging.warning("Recording {} is duplicated.".format(recording.id))
+            recordings = RecordingSet.from_recordings(recordings_nodup.values())
+            supervisions = SupervisionSet.from_segments(supervisions)
+
+            recordings, supervisions = fix_manifests(
+                recordings=recordings, supervisions=supervisions
+            )
+
+            logging.info(f"Writing {part} enhanced manifests")
+            recordings.to_file(
+                manifests_dir / f"alimeeting-gss_recordings_{part}.jsonl.gz"
+            )
+            supervisions.to_file(
+                manifests_dir / f"alimeeting-gss_supervisions_{part}.jsonl.gz"
+            )
+
+
+if __name__ == "__main__":
+    args = get_args()
+    main(args)
diff --git a/egs/alimeeting/ASR_v2/local/prepare_alimeeting_gss.sh b/egs/alimeeting/ASR_v2/local/prepare_alimeeting_gss.sh
new file mode 100755
index 000000000..76db19832
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/prepare_alimeeting_gss.sh
@@ -0,0 +1,98 @@
+#!/bin/bash
+# This script is used to run GSS-based enhancement on AMI data.
+set -euo pipefail
+nj=4
+stage=0
+
+. shared/parse_options.sh || exit 1
+
+if [ $# != 2 ]; then
+   echo "Wrong #arguments ($#, expected 2)"
+   echo "Usage: local/prepare_alimeeting_gss.sh [options] <data-dir> <exp-dir>"
+   echo "e.g. local/prepare_alimeeting_gss.sh data/manifests exp/ami_gss"
+   echo "main options (for others, see top of script file)"
+   echo "  --nj <nj>                                # number of parallel jobs"
+   echo "  --stage <stage>                          # stage to start running from"
+   exit 1;
+fi
+
+DATA_DIR=$1
+EXP_DIR=$2
+
+mkdir -p $EXP_DIR
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 1 ]; then
+  log "Stage 1: Prepare cut sets"
+  for part in train eval test; do
+    lhotse cut simple \
+      -r $DATA_DIR/alimeeting-mdm_recordings_${part}.jsonl.gz \
+      -s $DATA_DIR/alimeeting-mdm_supervisions_${part}.jsonl.gz \
+      $EXP_DIR/cuts_${part}.jsonl.gz
+  done
+fi
+
+if [ $stage -le 2 ]; then
+  log "Stage 2: Trim cuts to supervisions (1 cut per supervision segment)"
+  for part in train eval test; do
+    lhotse cut trim-to-supervisions --discard-overlapping \
+        $EXP_DIR/cuts_${part}.jsonl.gz $EXP_DIR/cuts_per_segment_${part}.jsonl.gz
+  done
+fi
+
+if [ $stage -le 3 ]; then
+  log "Stage 3: Split manifests for multi-GPU processing (optional)"
+  for part in train eval test; do
+    gss utils split $nj $EXP_DIR/cuts_per_segment_${part}.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_${part}_split$nj
+  done
+fi
+
+if [ $stage -le 4 ]; then
+  log "Stage 4: Enhance train segments using GSS (requires GPU)"
+  # for train, we use smaller context and larger batches to speed-up processing
+  for JOB in $(seq $nj); do
+    gss enhance cuts $EXP_DIR/cuts_train.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.JOB.jsonl.gz $EXP_DIR/enhanced \
+      --bss-iterations 10 \
+      --context-duration 5.0 \
+      --use-garbage-class \
+      --channels 0,1,2,3,4,5,6,7 \
+      --min-segment-length 0.05 \
+      --max-segment-length 25.0 \
+      --max-batch-duration 60.0 \
+      --num-buckets 4 \
+      --num-workers 4
+  done
+fi
+
+if [ $stage -le 5 ]; then
+  log "Stage 5: Enhance eval/test segments using GSS (using GPU)"
+  # for eval/test, we use larger context and smaller batches to get better quality
+  for part in eval test; do
+    for JOB in $(seq $nj); do
+      gss enhance cuts $EXP_DIR/cuts_${part}.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.JOB.jsonl.gz \
+      $EXP_DIR/enhanced \
+      --bss-iterations 10 \
+      --context-duration 15.0 \
+      --use-garbage-class \
+      --channels 0,1,2,3,4,5,6,7 \
+      --min-segment-length 0.05 \
+      --max-segment-length 16.0 \
+      --max-batch-duration 45.0 \
+      --num-buckets 4 \
+      --num-workers 4
+    done
+  done
+fi
+
+if [ $stage -le 6 ]; then
+  log "Stage 6: Prepare manifests for GSS-enhanced data"
+  python local/prepare_alimeeting_enhanced.py $DATA_DIR $EXP_DIR/enhanced -j $nj --min-segment-duration 0.05
+fi
diff --git a/egs/alimeeting/ASR_v2/local/prepare_char.py b/egs/alimeeting/ASR_v2/local/prepare_char.py
new file mode 120000
index 000000000..ee5dd34f1
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/prepare_char.py
@@ -0,0 +1 @@
+../../ASR/local/prepare_char.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/local/prepare_words.py b/egs/alimeeting/ASR_v2/local/prepare_words.py
new file mode 120000
index 000000000..970bfd60c
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/prepare_words.py
@@ -0,0 +1 @@
+../../ASR/local/prepare_words.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/local/text2segments.py b/egs/alimeeting/ASR_v2/local/text2segments.py
new file mode 120000
index 000000000..bf4547794
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/text2segments.py
@@ -0,0 +1 @@
+../../ASR/local/text2segments.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/local/text2token.py b/egs/alimeeting/ASR_v2/local/text2token.py
new file mode 120000
index 000000000..f6b8531b6
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/local/text2token.py
@@ -0,0 +1 @@
+../../ASR/local/text2token.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/prepare.sh b/egs/alimeeting/ASR_v2/prepare.sh
new file mode 100755
index 000000000..1098840f8
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/prepare.sh
@@ -0,0 +1,125 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+use_gss=true  # Use GSS-based enhancement with MDM setting
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/alimeeting
+#     This directory contains the following files downloaded from
+#       https://openslr.org/62/
+#
+#     - Train_Ali_far.tar.gz
+#     - Train_Ali_near.tar.gz
+#     - Test_Ali.tar.gz
+#     - Eval_Ali.tar.gz
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  if [ ! -f $dl_dir/alimeeting/Train_Ali_far.tar.gz ]; then
+    lhotse download ali-meeting $dl_dir/alimeeting
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare alimeeting manifest"
+  # We assume that you have downloaded the alimeeting corpus
+  # to $dl_dir/alimeeting
+  for part in ihm sdm mdm; do
+    mkdir -p data/manifests/alimeeting
+    lhotse prepare ali-meeting --mic $part --save-mono --normalize-text m2met \
+      $dl_dir/alimeeting data/manifests
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ] && [ $use_gss = true ]; then
+  log "Stage 3: Apply GSS enhancement on MDM data (this stage requires a GPU)"
+  # We assume that you have installed the GSS package: https://github.com/desh2608/gss
+  local/prepare_alimeeting_gss.sh data/manifests exp/alimeeting_gss
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  python local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for alimeeting"
+  mkdir -p data/fbank
+  python local/compute_fbank_alimeeting.py --perturb-speed True
+  log "Combine features from train splits"
+  lhotse combine data/manifests/cuts_train_{ihm,ihm_rvb,sdm,gss}.jsonl.gz - | shuf |\
+    gzip -c > data/manifests/cuts_train_all.jsonl.gz
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+
+  # Prepare text.
+  # Note: in Linux, you can install jq with the  following command:
+  # wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+  gunzip -c data/manifests/alimeeting-sdm_supervisions_train.jsonl.gz \
+    | jq ".text" | sed 's/"//g' \
+    | ./local/text2token.py -t "char" > $lang_char_dir/text
+
+  # Prepare words segments
+  python ./local/text2segments.py \
+    --input $lang_char_dir/text \
+    --output $lang_char_dir/text_words_segmentation
+
+  cat $lang_char_dir/text_words_segmentation | sed "s/ /\n/g" \
+    | sort -u | sed "/^$/d" \
+    | uniq > $lang_char_dir/words_no_ids.txt
+
+  # Prepare words.txt
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    ./local/prepare_char.py
+  fi
+fi
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/__init__.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py
new file mode 100644
index 000000000..9d288218a
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py
@@ -0,0 +1,418 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import re
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.cut import Cut
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AlimeetingAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description=(
+                "These options are used for the preparation of "
+                "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+                "effective batch sizes, sampling strategies, applied data "
+                "augmentations, etc."
+            ),
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help=(
+                "When enabled, select noise from MUSAN and mix it "
+                "with training dataset. "
+            ),
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help=(
+                "When enabled, utterances (cuts) will be concatenated "
+                "to minimize the amount of padding."
+            ),
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help=(
+                "Determines the maximum duration of a concatenated cut "
+                "relative to the duration of the longest cut in a batch."
+            ),
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help=(
+                "The amount of padding (in seconds) inserted between "
+                "concatenated cuts. This padding is filled with noise when "
+                "noise augmentation is used."
+            ),
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=100.0,
+            help=(
+                "Maximum pooled recordings duration (seconds) in a "
+                "single batch. You can reduce it if it causes CUDA OOM."
+            ),
+        )
+        group.add_argument(
+            "--max-cuts", type=int, default=None, help="Maximum cuts in a single batch."
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=50,
+            help=(
+                "The number of buckets for the BucketingSampler"
+                "(you might want to increase it for larger datasets)."
+            ),
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help=(
+                "When enabled, use on-the-fly cut mixing and feature "
+                "extraction. Will drop existing precomputed feature manifests "
+                "if available."
+            ),
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help=(
+                "When enabled (=default), the examples will be "
+                "shuffled for each epoch."
+            ),
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=8,
+            help=(
+                "The number of training dataloader workers that " "collect the batches."
+            ),
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help=(
+                "Used only when --enable-spec-aug is True. "
+                "It specifies the factor for time warping in SpecAugment. "
+                "Larger values mean more warping. "
+                "A value less than 1 means to disable time warp."
+            ),
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                "Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=2,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+            )
+
+        logging.info("Using DynamicBucketingSampler.")
+        train_sampler = DynamicBucketingSampler(
+            cuts_train,
+            max_duration=self.args.max_duration,
+            max_cuts=self.args.max_cuts,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
+        )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=True,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts, max_duration=self.args.max_duration, shuffle=False
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    def remove_short_cuts(self, cut: Cut) -> bool:
+        """
+        See: https://github.com/k2-fsa/icefall/issues/500
+        Basically, the zipformer model subsamples the input using the following formula:
+        num_out_frames = ((num_in_frames - 7)//2 + 1)//2
+        For num_out_frames to be at least 1, num_in_frames must be at least 9.
+        """
+        return cut.duration >= 0.09
+
+    @lru_cache()
+    def train_cuts(self, sp: Optional[Any] = None) -> CutSet:
+        logging.info("About to get AMI train cuts")
+
+        def _remove_short_and_long_utt(c: Cut):
+            if c.duration < 0.1 or c.duration > 25.0:
+                return False
+
+            # In pruned RNN-T, we require that T >= S
+            # where T is the number of feature frames after subsampling
+            # and S is the number of tokens in the utterance
+
+            # In ./zipformer.py, the conv module uses the following expression
+            # for subsampling
+            T = ((c.num_frames - 7) // 2 + 1) // 2
+            tokens = c.supervisions[0].text
+            return T >= len(tokens)
+
+        cuts_train = load_manifest_lazy(
+            self.args.manifest_dir / "cuts_train_all.jsonl.gz"
+        )
+
+        return cuts_train.filter(_remove_short_and_long_utt)
+
+    @lru_cache()
+    def eval_ihm_cuts(self) -> CutSet:
+        logging.info("About to get AliMeeting IHM eval cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_eval_ihm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def eval_sdm_cuts(self) -> CutSet:
+        logging.info("About to get AliMeeting SDM eval cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_eval_sdm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def eval_gss_cuts(self) -> CutSet:
+        if not (self.args.manifest_dir / "cuts_eval_gss.jsonl.gz").exists():
+            logging.info("No GSS dev cuts found")
+            return None
+        logging.info("About to get AliMeeting GSS-enhanced eval cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_eval_gss.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def test_ihm_cuts(self) -> CutSet:
+        logging.info("About to get AliMeeting IHM test cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_ihm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def test_sdm_cuts(self) -> CutSet:
+        logging.info("About to get AliMeeting SDM test cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_sdm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def test_gss_cuts(self) -> CutSet:
+        if not (self.args.manifest_dir / "cuts_test_gss.jsonl.gz").exists():
+            logging.info("No GSS test cuts found")
+            return None
+        logging.info("About to get AliMeeting GSS-enhanced test cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_gss.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/beam_search.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/beam_search.py
new file mode 120000
index 000000000..37516affc
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/beam_search.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/decode.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/decode.py
new file mode 100755
index 000000000..2741e0eeb
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/decode.py
@@ -0,0 +1,692 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 15 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 15 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 15 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AlimeetingAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text).replace(" ", "")) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AlimeetingAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest_LG",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if "fast_beam_search" in params.decoding_method:
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    alimeeting = AlimeetingAsrDataModule(args)
+
+    eval_ihm_cuts = alimeeting.eval_ihm_cuts()
+    test_ihm_cuts = alimeeting.test_ihm_cuts()
+    eval_sdm_cuts = alimeeting.eval_sdm_cuts()
+    test_sdm_cuts = alimeeting.test_sdm_cuts()
+    eval_gss_cuts = alimeeting.eval_gss_cuts()
+    test_gss_cuts = alimeeting.test_gss_cuts()
+
+    eval_ihm_dl = alimeeting.test_dataloaders(eval_ihm_cuts)
+    test_ihm_dl = alimeeting.test_dataloaders(test_ihm_cuts)
+    eval_sdm_dl = alimeeting.test_dataloaders(eval_sdm_cuts)
+    test_sdm_dl = alimeeting.test_dataloaders(test_sdm_cuts)
+    if eval_gss_cuts is not None:
+        eval_gss_dl = alimeeting.test_dataloaders(eval_gss_cuts)
+    if test_gss_cuts is not None:
+        test_gss_dl = alimeeting.test_dataloaders(test_gss_cuts)
+
+    test_sets = {
+        "eval_ihm": (eval_ihm_dl, eval_ihm_cuts),
+        "test_ihm": (test_ihm_dl, test_ihm_cuts),
+        "eval_sdm": (eval_sdm_dl, eval_sdm_cuts),
+        "test_sdm": (test_sdm_dl, test_sdm_cuts),
+    }
+    if eval_gss_cuts is not None:
+        test_sets["eval_gss"] = (eval_gss_dl, eval_gss_cuts)
+    if test_gss_cuts is not None:
+        test_sets["test_gss"] = (test_gss_dl, test_gss_cuts)
+
+    for test_set in test_sets:
+        logging.info(f"Decoding {test_set}")
+        dl, cuts = test_sets[test_set]
+        results_dict = decode_dataset(
+            dl=dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/decoder.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/encoder_interface.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/encoder_interface.py
new file mode 120000
index 000000000..0c2673d46
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/export.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/export.py
new file mode 100755
index 000000000..23a88dd29
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/export.py
@@ -0,0 +1,320 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless7/decode.py \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=15,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=8,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/jit_pretrained.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/jit_pretrained.py
new file mode 120000
index 000000000..a44034e34
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/joiner.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/model.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/optim.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/pretrained.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/pretrained.py
new file mode 120000
index 000000000..068f0f57f
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/pretrained.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/scaling.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/scaling_converter.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/test_model.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/test_model.py
new file mode 120000
index 000000000..7ceac5d10
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/test_model.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py
new file mode 100755
index 000000000..8f09f1aa5
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py
@@ -0,0 +1,1175 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 15 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 150 \
+    --use-fp16 True
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AlimeetingAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=15,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=5000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=10,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 100,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = ((feature_lens - 7) // 2).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    alimeeting = AlimeetingAsrDataModule(args)
+
+    train_cuts = alimeeting.train_cuts()
+    train_dl = alimeeting.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = alimeeting.eval_ihm_cuts()
+    valid_dl = alimeeting.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         graph_compiler=graph_compiler,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AlimeetingAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/zipformer.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/alimeeting/ASR_v2/shared b/egs/alimeeting/ASR_v2/shared
new file mode 120000
index 000000000..3a3b28f96
--- /dev/null
+++ b/egs/alimeeting/ASR_v2/shared
@@ -0,0 +1 @@
+../../../egs/aishell/ASR/shared
\ No newline at end of file
diff --git a/egs/ami/ASR/README.md b/egs/ami/ASR/README.md
new file mode 100644
index 000000000..1c9714bd4
--- /dev/null
+++ b/egs/ami/ASR/README.md
@@ -0,0 +1,48 @@
+# AMI
+
+This is an ASR recipe for the AMI corpus. AMI provides recordings from the speaker's
+headset and lapel microphones, and also 2 array microphones containing 8 channels each.
+We pool data in the following 4 ways and train a single model on the pooled data:
+
+(i) individual headset microphone (IHM)
+(ii) IHM with simulated reverb
+(iii) Single distant microphone (SDM)
+(iv) GSS-enhanced array microphones
+
+Speed perturbation and MUSAN noise augmentation are additionally performed on the pooled
+data. Here are the statistics of the combined training data:
+
+```python
+>>> cuts_train.describe()
+Cuts count: 1222053
+Total duration (hh:mm:ss): 905:00:28
+Speech duration (hh:mm:ss): 905:00:28 (99.9%)
+Duration statistics (seconds):
+mean    2.7
+std     2.8
+min     0.0
+25%     0.6
+50%     1.6
+75%     3.8
+99%     12.3
+99.5%   13.9
+99.9%   18.4
+max     36.8
+```
+
+**Note:** This recipe additionally uses [GSS](https://github.com/desh2608/gss) for enhancement
+of far-field array microphones, but this is optional (see `prepare.sh` for details).
+
+## Performance Record
+
+### pruned_transducer_stateless7
+
+The following are decoded using `modified_beam_search`:
+
+| Evaluation set           | dev WER    | test WER |
+|--------------------------|------------|---------|
+| IHM                      |  18.92  | 17.40 |
+| SDM                      |  31.25  | 32.21 |
+| MDM (GSS-enhanced)       |  21.67  | 22.43 |
+
+See [RESULTS](/egs/ami/ASR/RESULTS.md) for details.
diff --git a/egs/ami/ASR/RESULTS.md b/egs/ami/ASR/RESULTS.md
new file mode 100644
index 000000000..163986021
--- /dev/null
+++ b/egs/ami/ASR/RESULTS.md
@@ -0,0 +1,92 @@
+## Results
+
+### AMI training results (Pruned Transducer)
+
+#### 2022-11-20
+
+#### Zipformer (pruned_transducer_stateless7)
+
+Zipformer encoder + non-current decoder. The decoder
+contains only an embedding layer, a Conv1d (with kernel size 2) and a linear
+layer (to transform tensor dim).
+
+All the results below are using a single model that is trained by combining the following
+data: IHM, IHM+reverb, SDM, and GSS-enhanced MDM. Speed perturbation and MUSAN noise
+augmentation are applied on top of the pooled data.
+
+**WERs for IHM:**
+
+|                           | dev | test | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             |  19.25  |  17.83  | --epoch 14 --avg 8 --max-duration 500 |
+| modified beam search      |  18.92  |  17.40  | --epoch 14 --avg 8 --max-duration 500 --beam-size 4 |
+| fast beam search          |  19.44  |  18.04  | --epoch 14 --avg 8 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+**WERs for SDM:**
+
+|                           | dev | test | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             |  31.32  |  32.38  | --epoch 14 --avg 8 --max-duration 500 |
+| modified beam search      |  31.25  |  32.21  | --epoch 14 --avg 8 --max-duration 500 --beam-size 4 |
+| fast beam search          |  31.11  |  32.10  | --epoch 14 --avg 8 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+**WERs for GSS-enhanced MDM:**
+
+|                           | dev | test | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             |  22.05  |  22.93  | --epoch 14 --avg 8 --max-duration 500 |
+| modified beam search      |  21.67  |  22.43  | --epoch 14 --avg 8 --max-duration 500 --beam-size 4 |
+| fast beam search          |  22.21  |  22.83  | --epoch 14 --avg 8 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 15 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 150 \
+  --max-cuts 150 \
+  --prune-range 5 \
+  --lr-factor 5 \
+  --lm-scale 0.25 \
+  --use-fp16 True
+```
+
+The decoding command is:
+```
+# greedy search
+./pruned_transducer_stateless7/decode.py \
+        --epoch 14 \
+        --avg 8 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method greedy_search
+
+# modified beam search
+./pruned_transducer_stateless7/decode.py \
+        --iter 105000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+# fast beam search
+./pruned_transducer_stateless7/decode.py \
+        --iter 105000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+Pretrained model is available at <https://huggingface.co/desh2608/icefall-asr-ami-pruned-transducer-stateless7>
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/VH10QOTBTbuYpWx994Onrg/#scalars>
diff --git a/egs/ami/ASR/local/__init__.py b/egs/ami/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/ami/ASR/local/compute_fbank_ami.py b/egs/ami/ASR/local/compute_fbank_ami.py
new file mode 100755
index 000000000..4892b40e3
--- /dev/null
+++ b/egs/ami/ASR/local/compute_fbank_ami.py
@@ -0,0 +1,194 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the AMI dataset.
+For the training data, we pool together IHM, reverberated IHM, and GSS-enhanced
+audios. For the test data, we separately prepare IHM, SDM, and GSS-enhanced
+parts (which are the 3 evaluation settings).
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+import math
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+from lhotse import CutSet, LilcomChunkyWriter
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_ami():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    manifests_ihm = read_manifests_if_cached(
+        dataset_parts=["train", "dev", "test"],
+        output_dir=src_dir,
+        prefix="ami-ihm",
+        suffix="jsonl.gz",
+    )
+    manifests_sdm = read_manifests_if_cached(
+        dataset_parts=["train", "dev", "test"],
+        output_dir=src_dir,
+        prefix="ami-sdm",
+        suffix="jsonl.gz",
+    )
+    # For GSS we already have cuts so we read them directly.
+    manifests_gss = read_manifests_if_cached(
+        dataset_parts=["train", "dev", "test"],
+        output_dir=src_dir,
+        prefix="ami-gss",
+        suffix="jsonl.gz",
+    )
+
+    def _extract_feats(cuts: CutSet, storage_path: Path, manifest_path: Path) -> None:
+        cuts = cuts + cuts.perturb_speed(0.9) + cuts.perturb_speed(1.1)
+        _ = cuts.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=storage_path,
+            manifest_path=manifest_path,
+            batch_duration=5000,
+            num_workers=8,
+            storage_type=LilcomChunkyWriter,
+        )
+
+    logging.info(
+        "Preparing training cuts: IHM + reverberated IHM + SDM + GSS (optional)"
+    )
+
+    logging.info("Processing train split IHM")
+    cuts_ihm = (
+        CutSet.from_manifests(**manifests_ihm["train"])
+        .trim_to_supervisions(keep_overlapping=False, keep_all_channels=False)
+        .modify_ids(lambda x: x + "-ihm")
+    )
+    _extract_feats(
+        cuts_ihm,
+        output_dir / "feats_train_ihm",
+        src_dir / "cuts_train_ihm.jsonl.gz",
+    )
+
+    logging.info("Processing train split IHM + reverberated IHM")
+    cuts_ihm_rvb = cuts_ihm.reverb_rir()
+    _extract_feats(
+        cuts_ihm_rvb,
+        output_dir / "feats_train_ihm_rvb",
+        src_dir / "cuts_train_ihm_rvb.jsonl.gz",
+    )
+
+    logging.info("Processing train split SDM")
+    cuts_sdm = (
+        CutSet.from_manifests(**manifests_sdm["train"])
+        .trim_to_supervisions(keep_overlapping=False)
+        .modify_ids(lambda x: x + "-sdm")
+    )
+    _extract_feats(
+        cuts_sdm,
+        output_dir / "feats_train_sdm",
+        src_dir / "cuts_train_sdm.jsonl.gz",
+    )
+
+    logging.info("Processing train split GSS")
+    cuts_gss = (
+        CutSet.from_manifests(**manifests_gss["train"])
+        .trim_to_supervisions(keep_overlapping=False)
+        .modify_ids(lambda x: x + "-gss")
+    )
+    _extract_feats(
+        cuts_gss,
+        output_dir / "feats_train_gss",
+        src_dir / "cuts_train_gss.jsonl.gz",
+    )
+
+    logging.info("Preparing test cuts: IHM, SDM, GSS (optional)")
+    for split in ["dev", "test"]:
+        logging.info(f"Processing {split} IHM")
+        cuts_ihm = (
+            CutSet.from_manifests(**manifests_ihm[split])
+            .trim_to_supervisions(keep_overlapping=False, keep_all_channels=False)
+            .compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{split}_ihm",
+                manifest_path=src_dir / f"cuts_{split}_ihm.jsonl.gz",
+                batch_duration=5000,
+                num_workers=8,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        logging.info(f"Processing {split} SDM")
+        cuts_sdm = (
+            CutSet.from_manifests(**manifests_sdm[split])
+            .trim_to_supervisions(keep_overlapping=False)
+            .compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{split}_sdm",
+                manifest_path=src_dir / f"cuts_{split}_sdm.jsonl.gz",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        logging.info(f"Processing {split} GSS")
+        cuts_gss = (
+            CutSet.from_manifests(**manifests_gss[split])
+            .trim_to_supervisions(keep_overlapping=False)
+            .compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{split}_gss",
+                manifest_path=src_dir / f"cuts_{split}_gss.jsonl.gz",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_ami()
diff --git a/egs/ami/ASR/local/compute_fbank_musan.py b/egs/ami/ASR/local/compute_fbank_musan.py
new file mode 100755
index 000000000..1fcf951f9
--- /dev/null
+++ b/egs/ami/ASR/local/compute_fbank_musan.py
@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, LilcomChunkyWriter, combine
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    musan_cuts_path = src_dir / "musan_cuts.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    # create chunks of Musan with duration 5 - 10 seconds
+    _ = (
+        CutSet.from_manifests(
+            recordings=combine(part["recordings"] for part in manifests.values())
+        )
+        .cut_into_windows(10.0)
+        .filter(lambda c: c.duration > 5)
+        .compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / "musan_feats",
+            manifest_path=musan_cuts_path,
+            batch_duration=500,
+            num_workers=4,
+            storage_type=LilcomChunkyWriter,
+        )
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()
diff --git a/egs/ami/ASR/local/prepare_ami_enhanced.py b/egs/ami/ASR/local/prepare_ami_enhanced.py
new file mode 100644
index 000000000..bed220eb3
--- /dev/null
+++ b/egs/ami/ASR/local/prepare_ami_enhanced.py
@@ -0,0 +1,158 @@
+#!/usr/local/bin/python
+# -*- coding: utf-8 -*-
+# Data preparation for AMI GSS-enhanced dataset.
+
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+from lhotse import Recording, RecordingSet, SupervisionSet
+from lhotse.qa import fix_manifests
+from lhotse.recipes.utils import read_manifests_if_cached
+from lhotse.utils import fastcopy
+from tqdm import tqdm
+
+logging.basicConfig(
+    format="%(asctime)s %(levelname)-8s %(message)s",
+    level=logging.INFO,
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+
+
+def get_args():
+    import argparse
+
+    parser = argparse.ArgumentParser(description="AMI enhanced dataset preparation.")
+    parser.add_argument(
+        "manifests_dir",
+        type=Path,
+        help="Path to directory containing AMI manifests.",
+    )
+    parser.add_argument(
+        "enhanced_dir",
+        type=Path,
+        help="Path to enhanced data directory.",
+    )
+    parser.add_argument(
+        "--num-jobs",
+        "-j",
+        type=int,
+        default=1,
+        help="Number of parallel jobs to run.",
+    )
+    parser.add_argument(
+        "--min-segment-duration",
+        "-d",
+        type=float,
+        default=0.0,
+        help="Minimum duration of a segment in seconds.",
+    )
+    return parser.parse_args()
+
+
+def find_recording_and_create_new_supervision(enhanced_dir, supervision):
+    """
+    Given a supervision (corresponding to original AMI recording), this function finds the
+    enhanced recording correspoding to the supervision, and returns this recording and
+    a new supervision whose start and end times are adjusted to match the enhanced recording.
+    """
+    file_name = Path(
+        f"{supervision.recording_id}-{supervision.speaker}-{int(100*supervision.start):06d}_{int(100*supervision.end):06d}.flac"
+    )
+    save_path = enhanced_dir / f"{supervision.recording_id}" / file_name
+    if save_path.exists():
+        recording = Recording.from_file(save_path)
+        if recording.duration == 0:
+            logging.warning(f"Skipping {save_path} which has duration 0 seconds.")
+            return None
+
+        # Old supervision is wrt to the original recording, we create new supervision
+        # wrt to the enhanced segment
+        new_supervision = fastcopy(
+            supervision,
+            recording_id=recording.id,
+            start=0,
+            duration=recording.duration,
+        )
+        return recording, new_supervision
+    else:
+        logging.warning(f"{save_path} does not exist.")
+        return None
+
+
+def main(args):
+    # Get arguments
+    manifests_dir = args.manifests_dir
+    enhanced_dir = args.enhanced_dir
+
+    # Load manifests from cache if they exist (saves time)
+    manifests = read_manifests_if_cached(
+        dataset_parts=["train", "dev", "test"],
+        output_dir=manifests_dir,
+        prefix="ami-sdm",
+        suffix="jsonl.gz",
+    )
+    if not manifests:
+        raise ValueError("AMI SDM manifests not found in {}".format(manifests_dir))
+
+    with ThreadPoolExecutor(args.num_jobs) as ex:
+        for part in ["train", "dev", "test"]:
+            logging.info(f"Processing {part}...")
+            supervisions_orig = manifests[part]["supervisions"].filter(
+                lambda s: s.duration >= args.min_segment_duration
+            )
+            # Remove TS3009d supervisions since they are not present in the enhanced data
+            supervisions_orig = supervisions_orig.filter(
+                lambda s: s.recording_id != "TS3009d"
+            )
+            futures = []
+
+            for supervision in tqdm(
+                supervisions_orig,
+                desc="Distributing tasks",
+            ):
+                futures.append(
+                    ex.submit(
+                        find_recording_and_create_new_supervision,
+                        enhanced_dir,
+                        supervision,
+                    )
+                )
+
+            recordings = []
+            supervisions = []
+            for future in tqdm(
+                futures,
+                total=len(futures),
+                desc="Processing tasks",
+            ):
+                result = future.result()
+                if result is not None:
+                    recording, new_supervision = result
+                    recordings.append(recording)
+                    supervisions.append(new_supervision)
+
+            # Remove duplicates from the recordings
+            recordings_nodup = {}
+            for recording in recordings:
+                if recording.id not in recordings_nodup:
+                    recordings_nodup[recording.id] = recording
+                else:
+                    logging.warning("Recording {} is duplicated.".format(recording.id))
+            recordings = RecordingSet.from_recordings(recordings_nodup.values())
+            supervisions = SupervisionSet.from_segments(supervisions)
+
+            recordings, supervisions = fix_manifests(
+                recordings=recordings, supervisions=supervisions
+            )
+
+            logging.info(f"Writing {part} enhanced manifests")
+            recordings.to_file(manifests_dir / f"ami-gss_recordings_{part}.jsonl.gz")
+            supervisions.to_file(
+                manifests_dir / f"ami-gss_supervisions_{part}.jsonl.gz"
+            )
+
+
+if __name__ == "__main__":
+    args = get_args()
+    main(args)
diff --git a/egs/ami/ASR/local/prepare_ami_gss.sh b/egs/ami/ASR/local/prepare_ami_gss.sh
new file mode 100755
index 000000000..d5422458b
--- /dev/null
+++ b/egs/ami/ASR/local/prepare_ami_gss.sh
@@ -0,0 +1,98 @@
+#!/bin/bash
+# This script is used to run GSS-based enhancement on AMI data.
+set -euo pipefail
+nj=4
+stage=0
+
+. shared/parse_options.sh || exit 1
+
+if [ $# != 2 ]; then
+   echo "Wrong #arguments ($#, expected 2)"
+   echo "Usage: local/prepare_ami_gss.sh [options] <data-dir> <exp-dir>"
+   echo "e.g. local/prepare_ami_gss.sh data/manifests exp/ami_gss"
+   echo "main options (for others, see top of script file)"
+   echo "  --nj <nj>                                # number of parallel jobs"
+   echo "  --stage <stage>                          # stage to start running from"
+   exit 1;
+fi
+
+DATA_DIR=$1
+EXP_DIR=$2
+
+mkdir -p $EXP_DIR
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 1 ]; then
+  log "Stage 1: Prepare cut sets"
+  for part in train dev test; do
+    lhotse cut simple \
+      -r $DATA_DIR/ami-mdm_recordings_${part}.jsonl.gz \
+      -s $DATA_DIR/ami-mdm_supervisions_${part}.jsonl.gz \
+      $EXP_DIR/cuts_${part}.jsonl.gz
+  done
+fi
+
+if [ $stage -le 2 ]; then
+  log "Stage 2: Trim cuts to supervisions (1 cut per supervision segment)"
+  for part in train dev test; do
+    lhotse cut trim-to-supervisions --discard-overlapping \
+        $EXP_DIR/cuts_${part}.jsonl.gz $EXP_DIR/cuts_per_segment_${part}.jsonl.gz
+  done
+fi
+
+if [ $stage -le 3 ]; then
+  log "Stage 3: Split manifests for multi-GPU processing (optional)"
+  for part in train; do
+    gss utils split $nj $EXP_DIR/cuts_per_segment_${part}.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_${part}_split$nj
+  done
+fi
+
+if [ $stage -le 4 ]; then
+  log "Stage 4: Enhance train segments using GSS (requires GPU)"
+  # for train, we use smaller context and larger batches to speed-up processing
+  for JOB in $(seq $nj); do
+    gss enhance cuts $EXP_DIR/cuts_train.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.JOB.jsonl.gz $EXP_DIR/enhanced \
+      --bss-iterations 10 \
+      --context-duration 5.0 \
+      --use-garbage-class \
+      --channels 0,1,2,3,4,5,6,7 \
+      --min-segment-length 0.05 \
+      --max-segment-length 35.0 \
+      --max-batch-duration 60.0 \
+      --num-buckets 3 \
+      --num-workers 2
+  done
+fi
+
+if [ $stage -le 5 ]; then
+  log "Stage 5: Enhance dev/test segments using GSS (using GPU)"
+  # for dev/test, we use larger context and smaller batches to get better quality
+  for part in dev test; do
+    for JOB in $(seq $nj); do
+      gss enhance cuts $EXP_DIR/cuts_${part}.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.JOB.jsonl.gz \
+      $EXP_DIR/enhanced \
+      --bss-iterations 10 \
+      --context-duration 15.0 \
+      --use-garbage-class \
+      --channels 0,1,2,3,4,5,6,7 \
+      --min-segment-length 0.05 \
+      --max-segment-length 30.0 \
+      --max-batch-duration 45.0 \
+      --num-buckets 3 \
+      --num-workers 2
+    done
+  done
+fi
+
+if [ $stage -le 6 ]; then
+  log "Stage 6: Prepare manifests for GSS-enhanced data"
+  python local/prepare_ami_enhanced.py $DATA_DIR $EXP_DIR/enhanced -j $nj --min-segment-duration 0.05
+fi
diff --git a/egs/ami/ASR/local/prepare_lang_bpe.py b/egs/ami/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/ami/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/ami/ASR/local/train_bpe_model.py b/egs/ami/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/ami/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/ami/ASR/prepare.sh b/egs/ami/ASR/prepare.sh
new file mode 100755
index 000000000..fb21a8ec6
--- /dev/null
+++ b/egs/ami/ASR/prepare.sh
@@ -0,0 +1,144 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+use_gss=true  # Use GSS-based enhancement with MDM setting
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/amicorpus
+#      You can find audio and transcripts in this path.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+#
+#  - $dl_dir/{LDC2004S13,LDC2005S13,LDC2004T19,LDC2005T19}
+#      These contain the Fisher English audio and transcripts. We will
+#      only use the transcripts as extra LM training data (similar to Kaldi).
+#
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+vocab_size=500
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/amicorpus,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/amicorpus $dl_dir/amicorpus
+  #
+  if [ ! -d $dl_dir/amicorpus ]; then
+    lhotse download ami --mic ihm $dl_dir/amicorpus
+    lhotse download ami --mic mdm $dl_dir/amicorpus
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare AMI manifests"
+  # We assume that you have downloaded the AMI corpus
+  # to $dl_dir/amicorpus. We perform text normalization for the transcripts.
+  mkdir -p data/manifests
+  for mic in ihm sdm mdm; do
+    lhotse prepare ami --mic $mic --partition full-corpus-asr --normalize-text kaldi \
+      --max-words-per-segment 30 $dl_dir/amicorpus data/manifests/
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ] && [ $use_gss = true ]; then
+  log "Stage 3: Apply GSS enhancement on MDM data (this stage requires a GPU)"
+  # We assume that you have installed the GSS package: https://github.com/desh2608/gss
+  local/prepare_ami_gss.sh data/manifests exp/ami_gss
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank features for AMI"
+  mkdir -p data/fbank
+  python local/compute_fbank_ami.py
+  log "Combine features from train splits"
+  lhotse combine data/manifests/cuts_train_{ihm,ihm_rvb,sdm,gss}.jsonl.gz - | shuf |\
+    gzip -c > data/manifests/cuts_train_all.jsonl.gz
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank features for musan"
+  mkdir -p data/fbank
+  python local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Dump transcripts for BPE model training."
+  mkdir -p data/lm
+  cat <(gunzip -c data/manifests/ami-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g')> data/lm/transcript_words.txt
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare BPE based lang"
+
+  lang_dir=data/lang_bpe_${vocab_size}
+  mkdir -p $lang_dir
+
+  # Add special words to words.txt
+  echo "<eps> 0" > $lang_dir/words.txt
+  echo "!SIL 1" >> $lang_dir/words.txt
+  echo "<UNK> 2" >> $lang_dir/words.txt
+
+  # Add regular words to words.txt
+  cat data/lm/transcript_words.txt | grep -o -E '\w+' | sort -u | awk '{print $0,NR+2}' >> $lang_dir/words.txt
+
+  # Add remaining special word symbols expected by LM scripts.
+  num_words=$(cat $lang_dir/words.txt | wc -l)
+  echo "<s> ${num_words}" >> $lang_dir/words.txt
+  num_words=$(cat $lang_dir/words.txt | wc -l)
+  echo "</s> ${num_words}" >> $lang_dir/words.txt
+  num_words=$(cat $lang_dir/words.txt | wc -l)
+  echo "#0 ${num_words}" >> $lang_dir/words.txt
+
+  ./local/train_bpe_model.py \
+    --lang-dir $lang_dir \
+    --vocab-size $vocab_size \
+    --transcript data/lm/transcript_words.txt
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+  fi
+fi
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/__init__.py b/egs/ami/ASR/pruned_transducer_stateless7/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py
new file mode 100644
index 000000000..79474f1d8
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -0,0 +1,430 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import re
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.cut import Cut
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AmiAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description=(
+                "These options are used for the preparation of "
+                "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+                "effective batch sizes, sampling strategies, applied data "
+                "augmentations, etc."
+            ),
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help=(
+                "When enabled, select noise from MUSAN and mix it "
+                "with training dataset. "
+            ),
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help=(
+                "When enabled, utterances (cuts) will be concatenated "
+                "to minimize the amount of padding."
+            ),
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help=(
+                "Determines the maximum duration of a concatenated cut "
+                "relative to the duration of the longest cut in a batch."
+            ),
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help=(
+                "The amount of padding (in seconds) inserted between "
+                "concatenated cuts. This padding is filled with noise when "
+                "noise augmentation is used."
+            ),
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=100.0,
+            help=(
+                "Maximum pooled recordings duration (seconds) in a "
+                "single batch. You can reduce it if it causes CUDA OOM."
+            ),
+        )
+        group.add_argument(
+            "--max-cuts", type=int, default=None, help="Maximum cuts in a single batch."
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=50,
+            help=(
+                "The number of buckets for the BucketingSampler"
+                "(you might want to increase it for larger datasets)."
+            ),
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help=(
+                "When enabled, use on-the-fly cut mixing and feature "
+                "extraction. Will drop existing precomputed feature manifests "
+                "if available."
+            ),
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help=(
+                "When enabled (=default), the examples will be "
+                "shuffled for each epoch."
+            ),
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=8,
+            help=(
+                "The number of training dataloader workers that " "collect the batches."
+            ),
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help=(
+                "Used only when --enable-spec-aug is True. "
+                "It specifies the factor for time warping in SpecAugment. "
+                "Larger values mean more warping. "
+                "A value less than 1 means to disable time warp."
+            ),
+        )
+        group.add_argument(
+            "--ihm-only",
+            type=str2bool,
+            default=False,
+            help="When enabled, only use IHM data for training.",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                "Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=2,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+            )
+
+        logging.info("Using DynamicBucketingSampler.")
+        train_sampler = DynamicBucketingSampler(
+            cuts_train,
+            max_duration=self.args.max_duration,
+            max_cuts=self.args.max_cuts,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
+        )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=True,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts, max_duration=self.args.max_duration, shuffle=False
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    def remove_short_cuts(self, cut: Cut) -> bool:
+        """
+        See: https://github.com/k2-fsa/icefall/issues/500
+        Basically, the zipformer model subsamples the input using the following formula:
+        num_out_frames = (num_in_frames - 7)//2
+        For num_out_frames to be at least 1, num_in_frames must be at least 9.
+        """
+        return cut.duration >= 0.09
+
+    @lru_cache()
+    def train_cuts(self, sp: Optional[Any] = None) -> CutSet:
+        logging.info("About to get AMI train cuts")
+
+        def _remove_short_and_long_utt(c: Cut):
+            if c.duration < 0.2 or c.duration > 25.0:
+                return False
+
+            # In pruned RNN-T, we require that T >= S
+            # where T is the number of feature frames after subsampling
+            # and S is the number of tokens in the utterance
+
+            # In ./zipformer.py, the conv module uses the following expression
+            # for subsampling
+            T = ((c.num_frames - 7) // 2 + 1) // 2
+            tokens = sp.encode(c.supervisions[0].text, out_type=str)
+            return T >= len(tokens)
+
+        if self.args.ihm_only:
+            cuts_train = load_manifest_lazy(
+                self.args.manifest_dir / "cuts_train_ihm.jsonl.gz"
+            )
+        else:
+            cuts_train = load_manifest_lazy(
+                self.args.manifest_dir / "cuts_train_all.jsonl.gz"
+            )
+
+        return cuts_train.filter(_remove_short_and_long_utt)
+
+    @lru_cache()
+    def dev_ihm_cuts(self) -> CutSet:
+        logging.info("About to get AMI IHM dev cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_dev_ihm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def dev_sdm_cuts(self) -> CutSet:
+        logging.info("About to get AMI SDM dev cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_dev_sdm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def dev_gss_cuts(self) -> CutSet:
+        if not (self.args.manifest_dir / "cuts_dev_gss.jsonl.gz").exists():
+            logging.info("No GSS dev cuts found")
+            return None
+        logging.info("About to get AMI GSS-enhanced dev cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_dev_gss.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def test_ihm_cuts(self) -> CutSet:
+        logging.info("About to get AMI IHM test cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_ihm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def test_sdm_cuts(self) -> CutSet:
+        logging.info("About to get AMI SDM test cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_sdm.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
+
+    @lru_cache()
+    def test_gss_cuts(self) -> CutSet:
+        if not (self.args.manifest_dir / "cuts_test_gss.jsonl.gz").exists():
+            logging.info("No GSS test cuts found")
+            return None
+        logging.info("About to get AMI GSS-enhanced test cuts")
+        cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_gss.jsonl.gz")
+        return cs.filter(self.remove_short_cuts)
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/beam_search.py b/egs/ami/ASR/pruned_transducer_stateless7/beam_search.py
new file mode 120000
index 000000000..37516affc
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/beam_search.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/decode.py b/egs/ami/ASR/pruned_transducer_stateless7/decode.py
new file mode 100755
index 000000000..9999894d1
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/decode.py
@@ -0,0 +1,739 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+        --iter 105000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless7/decode.py \
+        --iter 105000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode.py \
+        --iter 105000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --max-duration 500 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless7/decode.py \
+        --iter 105000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AmiAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    word_table: Optional[k2.SymbolTable] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+      word_table:
+        The word symbol table.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+    word_table: Optional[k2.SymbolTable] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        wers_filename = params.res_dir / f"wers-{test_set_name}-{params.suffix}.txt"
+        with open(wers_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        # we also compute CER for AMI dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        cers_filename = params.res_dir / f"cers-{test_set_name}-{params.suffix}.txt"
+        with open(cers_filename, "w") as f:
+            cer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_cers[key] = cer
+
+        logging.info("Wrote detailed error stats to {}".format(wers_filename))
+
+    test_set_wers = {k: v for k, v in sorted(test_set_wers.items(), key=lambda x: x[1])}
+    test_set_cers = {k: v for k, v in sorted(test_set_cers.items(), key=lambda x: x[1])}
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER\tCER", file=f)
+        for key in test_set_wers:
+            print(
+                "{}\t{}\t{}".format(key, test_set_wers[key], test_set_cers[key]),
+                file=f,
+            )
+
+    s = "\nFor {}, WER/CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key in test_set_wers:
+        s += "{}\t{}\t{}{}\n".format(key, test_set_wers[key], test_set_cers[key], note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AmiAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest_LG",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(f"{params.lang_dir}/bpe.model")
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    ami = AmiAsrDataModule(args)
+
+    dev_ihm_cuts = ami.dev_ihm_cuts()
+    test_ihm_cuts = ami.test_ihm_cuts()
+    dev_sdm_cuts = ami.dev_sdm_cuts()
+    test_sdm_cuts = ami.test_sdm_cuts()
+    dev_gss_cuts = ami.dev_gss_cuts()
+    test_gss_cuts = ami.test_gss_cuts()
+
+    dev_ihm_dl = ami.test_dataloaders(dev_ihm_cuts)
+    test_ihm_dl = ami.test_dataloaders(test_ihm_cuts)
+    dev_sdm_dl = ami.test_dataloaders(dev_sdm_cuts)
+    test_sdm_dl = ami.test_dataloaders(test_sdm_cuts)
+    if dev_gss_cuts is not None:
+        dev_gss_dl = ami.test_dataloaders(dev_gss_cuts)
+    if test_gss_cuts is not None:
+        test_gss_dl = ami.test_dataloaders(test_gss_cuts)
+
+    test_sets = {
+        "dev_ihm": (dev_ihm_dl, dev_ihm_cuts),
+        "test_ihm": (test_ihm_dl, test_ihm_cuts),
+        "dev_sdm": (dev_sdm_dl, dev_sdm_cuts),
+        "test_sdm": (test_sdm_dl, test_sdm_cuts),
+    }
+    if dev_gss_cuts is not None:
+        test_sets["dev_gss"] = (dev_gss_dl, dev_gss_cuts)
+    if test_gss_cuts is not None:
+        test_sets["test_gss"] = (test_gss_dl, test_gss_cuts)
+
+    for test_set in test_sets:
+        logging.info(f"Decoding {test_set}")
+        dl, cuts = test_sets[test_set]
+        results_dict = decode_dataset(
+            dl=dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/decoder.py b/egs/ami/ASR/pruned_transducer_stateless7/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/encoder_interface.py b/egs/ami/ASR/pruned_transducer_stateless7/encoder_interface.py
new file mode 120000
index 000000000..0c2673d46
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/export.py b/egs/ami/ASR/pruned_transducer_stateless7/export.py
new file mode 120000
index 000000000..2713792e6
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/export.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/joiner.py b/egs/ami/ASR/pruned_transducer_stateless7/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/model.py b/egs/ami/ASR/pruned_transducer_stateless7/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/optim.py b/egs/ami/ASR/pruned_transducer_stateless7/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/scaling.py b/egs/ami/ASR/pruned_transducer_stateless7/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/scaling_converter.py b/egs/ami/ASR/pruned_transducer_stateless7/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/train.py b/egs/ami/ASR/pruned_transducer_stateless7/train.py
new file mode 100755
index 000000000..9b67141c0
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/train.py
@@ -0,0 +1,1182 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 15 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 150 \
+    --use-fp16 True
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AmiAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=11,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=5000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=10,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 100,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"]
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = supervisions["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = ((feature_lens - 7) // 2).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    ami = AmiAsrDataModule(args)
+
+    # Here is the duration statistics of the training set.
+    # Cuts count: 1230033
+    # Total duration (hh:mm:ss): 904:25:34
+    # Speech duration (hh:mm:ss): 904:25:34 (100.0%)
+    # Duration statistics (seconds):
+    # mean	2.6
+    # std	2.8
+    # min	0.0
+    # 25%	0.6
+    # 50%	1.6
+    # 75%	3.8
+    # 99%	12.3
+    # 99.5%	13.9
+    # 99.9%	18.3
+    # max	36.8
+
+    train_cuts = ami.train_cuts(sp=sp)
+    train_dl = ami.train_dataloaders(train_cuts, sampler_state_dict=sampler_state_dict)
+
+    valid_cuts = ami.dev_ihm_cuts()
+    valid_dl = ami.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AmiAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/zipformer.py b/egs/ami/ASR/pruned_transducer_stateless7/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/ami/ASR/pruned_transducer_stateless7/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/ami/ASR/shared b/egs/ami/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/ami/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/ami/SURT/README.md b/egs/ami/SURT/README.md
new file mode 100644
index 000000000..74a8ba014
--- /dev/null
+++ b/egs/ami/SURT/README.md
@@ -0,0 +1,156 @@
+# Introduction
+
+This is a multi-talker ASR recipe for the AMI and ICSI datasets. We train a Streaming
+Unmixing and Recognition Transducer (SURT) model for the task.
+
+Please refer to the `egs/libricss/SURT` recipe README for details about the task and the
+model.
+
+## Description of the recipe
+
+### Pre-requisites
+
+The recipes in this directory need the following packages to be installed:
+
+- [meeteval](https://github.com/fgnt/meeteval)
+- [einops](https://github.com/arogozhnikov/einops)
+
+Additionally, we initialize the model with the pre-trained model from the LibriCSS recipe.
+Please download this checkpoint (see below) or train the LibriCSS recipe first.
+
+### Training
+
+To train the model, run the following from within `egs/ami/SURT`:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+python dprnn_zipformer/train.py \
+    --use-fp16 True \
+    --exp-dir dprnn_zipformer/exp/surt_base \
+    --world-size 4 \
+    --max-duration 500 \
+    --max-duration-valid 250 \
+    --max-cuts 200 \
+    --num-buckets 50 \
+    --num-epochs 30 \
+    --enable-spec-aug True \
+    --enable-musan False \
+    --ctc-loss-scale 0.2 \
+    --heat-loss-scale 0.2 \
+    --base-lr 0.004 \
+    --model-init-ckpt exp/libricss_base.pt \
+    --chunk-width-randomization True \
+    --num-mask-encoder-layers 4 \
+    --num-encoder-layers 2,2,2,2,2
+```
+
+The above is for SURT-base (~26M). For SURT-large (~38M), use:
+
+```bash
+    --model-init-ckpt exp/libricss_large.pt \
+    --num-mask-encoder-layers 6 \
+    --num-encoder-layers 2,4,3,2,4 \
+    --model-init-ckpt exp/zipformer_large.pt \
+```
+
+**NOTE:** You may need to decrease the `--max-duration` for SURT-large to avoid OOM.
+
+### Adaptation
+
+The training step above only trains on simulated mixtures. For best results, we also
+adapt the final model on the AMI+ICSI train set. For this, run the following from within
+`egs/ami/SURT`:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+
+python dprnn_zipformer/train_adapt.py \
+    --use-fp16 True \
+    --exp-dir dprnn_zipformer/exp/surt_base_adapt \
+    --world-size 4 \
+    --max-duration 500 \
+    --max-duration-valid 250 \
+    --max-cuts 200 \
+    --num-buckets 50 \
+    --num-epochs 8 \
+    --lr-epochs 2 \
+    --enable-spec-aug True \
+    --enable-musan False \
+    --ctc-loss-scale 0.2 \
+    --base-lr 0.0004 \
+    --model-init-ckpt dprnn_zipformer/exp/surt_base/epoch-30.pt \
+    --chunk-width-randomization True \
+    --num-mask-encoder-layers 4 \
+    --num-encoder-layers 2,2,2,2,2
+```
+
+For SURT-large, use the following config:
+
+```bash
+    --num-mask-encoder-layers 6 \
+    --num-encoder-layers 2,4,3,2,4 \
+    --model-init-ckpt dprnn_zipformer/exp/surt_large/epoch-30.pt \
+    --num-epochs 15 \
+    --lr-epochs 4 \
+```
+
+
+### Decoding
+
+To decode the model, run the following from within `egs/ami/SURT`:
+
+#### Greedy search
+
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+
+python dprnn_zipformer/decode.py \
+    --epoch 20 --avg 1 --use-averaged-model False \
+    --exp-dir dprnn_zipformer/exp/surt_base_adapt \
+    --max-duration 250 \
+    --decoding-method greedy_search
+```
+
+#### Beam search
+
+```bash
+python dprnn_zipformer/decode.py \
+    --epoch 20 --avg 1 --use-averaged-model False \
+    --exp-dir dprnn_zipformer/exp/surt_base_adapt \
+    --max-duration 250 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+```
+
+## Results (using beam search)
+
+**AMI**
+
+| Model      | IHM-Mix |  SDM |  MDM |
+|------------|:-------:|:----:|:----:|
+| SURT-base  |   39.8  | 65.4 | 46.6 |
+|   + adapt  |   37.4  | 46.9 | 43.7 |
+| SURT-large |   36.8  | 62.5 | 44.4 |
+|    + adapt |   **35.1**  | **44.6** | **41.4** |
+
+**ICSI**
+
+| Model      | IHM-Mix |  SDM |
+|------------|:-------:|:----:|
+| SURT-base  |   28.3  | 60.0 |
+|   + adapt  |   26.3  | 33.9 |
+| SURT-large |   27.8  | 59.7 |
+|    + adapt |   **24.4**  | **32.3** |
+
+## Pre-trained models and logs
+
+* LibriCSS pre-trained model (for initialization): [base](https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer/tree/main/exp/surt_base) [large](https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer/tree/main/exp/surt_large)
+
+* Pre-trained models: <https://huggingface.co/desh2608/icefall-surt-ami-dprnn-zipformer>
+
+* Training logs:
+    - surt_base: <https://tensorboard.dev/experiment/8awy98VZSWegLmH4l2JWSA/>
+    - surt_base_adapt: <https://tensorboard.dev/experiment/aGVgXVzYRDKbGUbPekcNjg/>
+    - surt_large: <https://tensorboard.dev/experiment/ZXMkez0VSYKbPLqRk4clOQ/>
+    - surt_large_adapt: <https://tensorboard.dev/experiment/WLKL1e7bTVyEjSonYSNYwg/>
diff --git a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py
new file mode 100644
index 000000000..1549c1631
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py
@@ -0,0 +1,399 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutMix,
+    DynamicBucketingSampler,
+    K2SurtDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AmiAsrDataModule:
+    """
+    DataModule for k2 SURT experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--max-duration-valid",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--max-cuts",
+            type=int,
+            default=100,
+            help="Maximum number of cuts in a single batch. You can "
+            "reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help=(
+                "When enabled, use on-the-fly cut mixing and feature "
+                "extraction. Will drop existing precomputed feature manifests "
+                "if available."
+            ),
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        sources: bool = False,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SurtDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+            return_sources=sources,
+            strict=False,
+        )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                quadratic_duration=30.0,
+                max_cuts=self.args.max_cuts,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                max_cuts=self.args.max_cuts,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+
+        logging.info("About to create dev dataset")
+        validate = K2SurtDataset(
+            input_strategy=OnTheFlyFeatures(
+                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            )
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+            return_sources=False,
+            strict=False,
+        )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration_valid,
+            quadratic_duration=30.0,
+            max_cuts=self.args.max_cuts,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SurtDataset(
+            input_strategy=OnTheFlyFeatures(
+                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            )
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+            return_sources=False,
+            strict=False,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration_valid,
+            max_cuts=self.args.max_cuts,
+            shuffle=False,
+        )
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+        return test_dl
+
+    @lru_cache()
+    def aimix_train_cuts(
+        self,
+        rvb_affix: str = "clean",
+        sources: bool = True,
+    ) -> CutSet:
+        logging.info("About to get train cuts")
+        source_affix = "_sources" if sources else ""
+        cs = load_manifest_lazy(
+            self.args.manifest_dir / f"cuts_train_{rvb_affix}{source_affix}.jsonl.gz"
+        )
+        cs = cs.filter(lambda c: c.duration >= 1.0 and c.duration <= 30.0)
+        return cs
+
+    @lru_cache()
+    def train_cuts(
+        self,
+    ) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "cuts_train_ami_icsi.jsonl.gz"
+        )
+
+    @lru_cache()
+    def ami_cuts(self, split: str = "dev", type: str = "sdm") -> CutSet:
+        logging.info(f"About to get AMI {split} {type} cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / f"cuts_ami-{type}_{split}.jsonl.gz"
+        )
+
+    @lru_cache()
+    def icsi_cuts(self, split: str = "dev", type: str = "sdm") -> CutSet:
+        logging.info(f"About to get ICSI {split} {type} cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / f"cuts_icsi-{type}_{split}.jsonl.gz"
+        )
diff --git a/egs/ami/SURT/dprnn_zipformer/beam_search.py b/egs/ami/SURT/dprnn_zipformer/beam_search.py
new file mode 120000
index 000000000..581b29833
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/beam_search.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/decode.py b/egs/ami/SURT/dprnn_zipformer/decode.py
new file mode 100755
index 000000000..d1a1eddc9
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/decode.py
@@ -0,0 +1,622 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./dprnn_zipformer/decode.py \
+    --epoch 20 \
+    --avg 1 \
+    --use-averaged-model false \
+    --exp-dir ./dprnn_zipformer/exp_adapt \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./dprnn_zipformer/decode.py \
+    --epoch 20 \
+    --avg 1 \
+    --use-averaged-model false \
+    --exp-dir ./dprnn_zipformer/exp_adapt \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./dprnn_zipformer/decode.py \
+    --epoch 20 \
+    --avg 1 \
+    --use-averaged-model false \
+    --exp-dir ./dprnn_zipformer/exp_adapt \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AmiAsrDataModule
+from beam_search import (
+    beam_search,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.utils import EPSILON
+from train import add_model_arguments, get_params, get_surt_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_surt_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="dprnn_zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    feature_lens = batch["input_lens"].to(device)
+
+    # Apply the mask encoder
+    B, T, F = feature.shape
+    processed = model.mask_encoder(feature)  # B,T,F*num_channels
+    masks = processed.view(B, T, F, params.num_channels).unbind(dim=-1)
+    x_masked = [feature * m for m in masks]
+
+    # Recognition
+    # Stack the inputs along the batch axis
+    h = torch.cat(x_masked, dim=0)
+    h_lens = torch.cat([feature_lens for _ in range(params.num_channels)], dim=0)
+    encoder_out, encoder_out_lens = model.encoder(x=h, x_lens=h_lens)
+
+    if model.joint_encoder_layer is not None:
+        encoder_out = model.joint_encoder_layer(encoder_out)
+
+    def _group_channels(hyps: List[str]) -> List[List[str]]:
+        """
+        Currently we have a batch of size M*B, where M is the number of
+        channels and B is the batch size. We need to group the hypotheses
+        into B groups, each of which contains M hypotheses.
+
+        Example:
+            hyps = ['a1', 'b1', 'c1', 'a2', 'b2', 'c2']
+            _group_channels(hyps) = [['a1', 'a2'], ['b1', 'b2'], ['c1', 'c2']]
+        """
+        assert len(hyps) == B * params.num_channels
+        out_hyps = []
+        for i in range(B):
+            out_hyps.append(hyps[i::B])
+        return out_hyps
+
+    hyps = []
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp)
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp)
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp))
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": _group_channels(hyps)}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: _group_channels(hyps)}
+    else:
+        return {f"beam_size_{params.beam_size}": _group_channels(hyps)}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        cut_ids = [cut.id for cut in batch["cuts"]]
+        cuts_batch = batch["cuts"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            for cut_id, hyp_words in zip(cut_ids, hyps):
+                # Reference is a list of supervision texts sorted by start time.
+                ref_words = [
+                    s.text.strip()
+                    for s in sorted(
+                        cuts_batch[cut_id].supervisions, key=lambda s: s.start
+                    )
+                ]
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(cut_ids)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_surt_error_stats(
+                f,
+                f"{test_set_name}-{key}",
+                results,
+                enable_log=True,
+                num_channels=params.num_channels,
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LmScorer.add_arguments(parser)
+    AmiAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "modified_beam_search",
+    ), f"Decoding method {params.decoding_method} is not supported."
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    ami = AmiAsrDataModule(args)
+
+    # NOTE(@desh2608): we filter segments longer than 120s to avoid OOM errors in decoding.
+    # However, 99.9% of the segments are shorter than 120s, so this should not
+    # substantially affect the results. In future, we will implement an overlapped
+    # inference method to avoid OOM errors.
+
+    test_sets = {}
+    for split in ["dev", "test"]:
+        for type in ["ihm-mix", "sdm", "mdm8-bf"]:
+            test_sets[f"ami-{split}_{type}"] = (
+                ami.ami_cuts(split=split, type=type)
+                .trim_to_supervision_groups(max_pause=0.0)
+                .filter(lambda c: 0.1 < c.duration < 120.0)
+                .to_eager()
+            )
+
+    for split in ["dev", "test"]:
+        for type in ["ihm-mix", "sdm"]:
+            test_sets[f"icsi-{split}_{type}"] = (
+                ami.icsi_cuts(split=split, type=type)
+                .trim_to_supervision_groups(max_pause=0.0)
+                .filter(lambda c: 0.1 < c.duration < 120.0)
+                .to_eager()
+            )
+
+    for test_set, test_cuts in test_sets.items():
+        test_dl = ami.test_dataloaders(test_cuts)
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ami/SURT/dprnn_zipformer/decoder.py b/egs/ami/SURT/dprnn_zipformer/decoder.py
new file mode 120000
index 000000000..c34865c25
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/decoder.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/dprnn.py b/egs/ami/SURT/dprnn_zipformer/dprnn.py
new file mode 120000
index 000000000..8918beb32
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/dprnn.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/dprnn.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/encoder_interface.py b/egs/ami/SURT/dprnn_zipformer/encoder_interface.py
new file mode 120000
index 000000000..0ba945d0f
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/encoder_interface.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/export.py b/egs/ami/SURT/dprnn_zipformer/export.py
new file mode 120000
index 000000000..3deae4471
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/export.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/export.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/joiner.py b/egs/ami/SURT/dprnn_zipformer/joiner.py
new file mode 120000
index 000000000..79fbe8769
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/joiner.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/model.py b/egs/ami/SURT/dprnn_zipformer/model.py
new file mode 120000
index 000000000..ae8c65c99
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/model.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/model.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/optim.py b/egs/ami/SURT/dprnn_zipformer/optim.py
new file mode 120000
index 000000000..366d0f7a2
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/optim.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/optim.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/scaling.py b/egs/ami/SURT/dprnn_zipformer/scaling.py
new file mode 120000
index 000000000..f11d49d77
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/scaling.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/scaling_converter.py b/egs/ami/SURT/dprnn_zipformer/scaling_converter.py
new file mode 120000
index 000000000..1533cbe0e
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/test_model.py b/egs/ami/SURT/dprnn_zipformer/test_model.py
new file mode 120000
index 000000000..1259849e0
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py
\ No newline at end of file
diff --git a/egs/ami/SURT/dprnn_zipformer/train.py b/egs/ami/SURT/dprnn_zipformer/train.py
new file mode 100755
index 000000000..cd5fafc34
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/train.py
@@ -0,0 +1,1420 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+cd egs/ami/SURT/
+./prepare.sh
+
+./dprnn_zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir dprnn_zipformer/exp \
+  --max-duration 650
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from itertools import chain
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AmiAsrDataModule
+from decoder import Decoder
+from dprnn import DPRNN
+from einops.layers.torch import Rearrange
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import LOG_EPSILON, fix_random_seed
+from model import SURT
+from optim import Eden, ScaledAdam
+from scaling import ScaledLinear, ScaledLSTM
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-mask-encoder-layers",
+        type=int,
+        default=4,
+        help="Number of layers in the DPRNN based mask encoder.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-dim",
+        type=int,
+        default=256,
+        help="Hidden dimension of the LSTM blocks in DPRNN.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-segment-size",
+        type=int,
+        default=32,
+        help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the "
+        "decode-chunk-length of the zipformer encoder.",
+    )
+
+    parser.add_argument(
+        "--chunk-width-randomization",
+        type=bool,
+        default=False,
+        help="Whether to randomize the chunk width in DPRNN.",
+    )
+
+    # Zipformer config is based on:
+    # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,2,2,2",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="768,768,768,768,768",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--use-joint-encoder-layer",
+        type=str,
+        default="lstm",
+        choices=["linear", "lstm", "none"],
+        help="Whether to use a joint layer to combine all branches.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conv_lstm_transducer_stateless_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--model-init-ckpt",
+        type=str,
+        default=None,
+        help="""The model checkpoint to initialize the model (either full or part).
+        If not specified, the model is randomly initialized.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.004, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--heat-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for HEAT loss on separated sources.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=1,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 2000,
+            # parameters for SURT
+            "num_channels": 2,
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed
+            # parameters for Noam
+            "model_warm_step": 5000,  # arg given to model, not for lrate
+            # parameters for ctc loss
+            "beam_size": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_mask_encoder_model(params: AttributeDict) -> nn.Module:
+    mask_encoder = DPRNN(
+        feature_dim=params.feature_dim,
+        input_size=params.mask_encoder_dim,
+        hidden_size=params.mask_encoder_dim,
+        output_size=params.feature_dim * params.num_channels,
+        segment_size=params.mask_encoder_segment_size,
+        num_blocks=params.num_mask_encoder_layers,
+        chunk_width_randomization=params.chunk_width_randomization,
+    )
+    return mask_encoder
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_joint_encoder_layer(params: AttributeDict) -> nn.Module:
+    class TakeFirst(nn.Module):
+        def forward(self, x):
+            return x[0]
+
+    if params.use_joint_encoder_layer == "linear":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            nn.Linear(
+                params.num_channels * encoder_dim, params.num_channels * encoder_dim
+            ),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "lstm":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            ScaledLSTM(
+                input_size=params.num_channels * encoder_dim,
+                hidden_size=params.num_channels * encoder_dim,
+                num_layers=1,
+                bias=True,
+                batch_first=True,
+                dropout=0.0,
+                bidirectional=False,
+            ),
+            TakeFirst(),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "none":
+        joint_layer = None
+    else:
+        raise ValueError(
+            f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}"
+        )
+    return joint_layer
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_surt_model(
+    params: AttributeDict,
+) -> nn.Module:
+    mask_encoder = get_mask_encoder_model(params)
+    encoder = get_encoder_model(params)
+    joint_layer = get_joint_encoder_layer(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = SURT(
+        mask_encoder=mask_encoder,
+        encoder=encoder,
+        joint_encoder_layer=joint_layer,
+        decoder=decoder,
+        joiner=joiner,
+        num_channels=params.num_channels,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor:
+    """
+    Compute HEAT loss for separated sources using the output of mask encoder.
+    Args:
+      x_masked:
+        The output of mask encoder. It is a tensor of shape (B, T, C).
+      batch:
+        A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()`
+        for the content in it.
+      num_channels:
+        The number of output branches in the SURT model.
+    """
+    B, T, D = x_masked[0].shape
+    device = x_masked[0].device
+
+    # Create training targets for each channel.
+    targets = []
+    for i in range(num_channels):
+        target = torch.ones_like(x_masked[i]) * LOG_EPSILON
+        targets.append(target)
+
+    source_feats = batch["source_feats"]
+    source_boundaries = batch["source_boundaries"]
+    input_lens = batch["input_lens"].to(device)
+    # Assign sources to channels based on the HEAT criteria
+    for b in range(B):
+        cut_source_feats = source_feats[b]
+        cut_source_boundaries = source_boundaries[b]
+        last_seg_end = [0 for _ in range(num_channels)]
+        for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries):
+            assigned = False
+            end = min(end, T)
+            source_feat = source_feat[: end - start, :]
+            for i in range(num_channels):
+                if start >= last_seg_end[i]:
+                    targets[i][b, start:end, :] += source_feat.to(device)
+                    last_seg_end[i] = max(end, last_seg_end[i])
+                    assigned = True
+                    break
+            if not assigned:
+                min_end_channel = last_seg_end.index(min(last_seg_end))
+                targets[min_end_channel][b, start:end, :] += source_feat.to(device)
+                last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel])
+
+    # Get padding mask based on input lengths
+    pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1)
+    pad_mask = pad_mask.unsqueeze(-1)
+
+    # Compute masked loss for each channel
+    losses = torch.zeros((num_channels, B, T, D), device=device)
+    for i in range(num_channels):
+        loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none")
+        # Apply padding mask to loss
+        loss.masked_fill_(pad_mask, 0)
+        losses[i] = loss
+
+    # loss: C x B x T x D. pad_mask: B x T x 1
+    # We want to compute loss for each item in the batch. Each item has loss given
+    # by the sum over C, and average over T and D. For T, we need to use the padding.
+    loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device)
+    return loss
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"].to(device)
+    feature_lens = batch["input_lens"].to(device)
+
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+
+    # The dataloader returns text as a list of cuts, each of which is a list of channel
+    # text. We flatten this to a list where all channels are together, i.e., it looks like
+    # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2].
+    text = [val for tup in zip(*batch["text"]) for val in tup]
+    assert len(text) == len(feature) * params.num_channels
+
+    # Convert all channel texts to token IDs and create a ragged tensor.
+    y = sp.encode(text, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.model_warm_step
+
+    with torch.set_grad_enabled(is_training):
+        (simple_loss, pruned_loss, ctc_loss, x_masked) = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            reduction="none",
+            subsampling_factor=params.subsampling_factor,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        ctc_loss_is_finite = torch.isfinite(ctc_loss)
+
+        # Compute HEAT loss
+        if is_training and params.heat_loss_scale > 0.0:
+            heat_loss = compute_heat_loss(
+                x_masked, batch, num_channels=params.num_channels
+            )
+        else:
+            heat_loss = torch.tensor(0.0, device=device)
+
+        heat_loss_is_finite = torch.isfinite(heat_loss)
+        is_finite = (
+            simple_loss_is_finite
+            & pruned_loss_is_finite
+            & ctc_loss_is_finite
+            & heat_loss_is_finite
+        )
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_losses: {simple_loss}\n"
+                f"pruned_losses: {pruned_loss}\n"
+                f"ctc_losses: {ctc_loss}\n"
+                f"heat_losses: {heat_loss}\n"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+            ctc_loss = ctc_loss[ctc_loss_is_finite]
+            heat_loss = heat_loss[heat_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if (
+                torch.all(~simple_loss_is_finite)
+                or torch.all(~pruned_loss_is_finite)
+                or torch.all(~ctc_loss_is_finite)
+                or torch.all(~heat_loss_is_finite)
+            ):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss_sum = simple_loss.sum()
+        pruned_loss_sum = pruned_loss.sum()
+        ctc_loss_sum = ctc_loss.sum()
+        heat_loss_sum = heat_loss.sum()
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+        loss = (
+            simple_loss_scale * simple_loss_sum
+            + pruned_loss_scale * pruned_loss_sum
+            + params.ctc_loss_scale * ctc_loss_sum
+            + params.heat_loss_scale * heat_loss_sum
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss_sum.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss_sum.detach().cpu().item()
+    if params.ctc_loss_scale > 0.0:
+        info["ctc_loss"] = ctc_loss_sum.detach().cpu().item()
+    if params.heat_loss_scale > 0.0:
+        info["heat_loss"] = heat_loss_sum.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    torch.cuda.empty_cache()
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = batch["inputs"].shape[0]
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+
+    if checkpoints is None and params.model_init_ckpt is not None:
+        logging.info(
+            f"Initializing model with checkpoint from {params.model_init_ckpt}"
+        )
+        init_ckpt = torch.load(params.model_init_ckpt, map_location=device)
+        model.load_state_dict(init_ckpt["model"], strict=False)
+
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    ami = AmiAsrDataModule(args)
+
+    train_cuts = ami.aimix_train_cuts(rvb_affix="comb", sources=True)
+    dev_cuts = ami.ami_cuts(split="dev", type="ihm-mix")
+    dev_cuts = dev_cuts.trim_to_supervision_groups(max_pause=0.0).filter(
+        lambda c: 0.2 <= c.duration <= 60.0
+    )
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = ami.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        sources=True,
+    )
+    valid_dl = ami.valid_dataloaders(dev_cuts)
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = [sp.encode(text_ch) for text_ch in batch["text"]]
+    num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y]
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def main():
+    parser = get_parser()
+    AmiAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ami/SURT/dprnn_zipformer/train_adapt.py b/egs/ami/SURT/dprnn_zipformer/train_adapt.py
new file mode 100755
index 000000000..9f3b4425f
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/train_adapt.py
@@ -0,0 +1,1411 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+# ./dprnn_zipformer/train.py should be run before this script.
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./dprnn_zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir dprnn_zipformer/exp_adapt \
+  --model-init-ckpt dprnn_zipformer/exp/epoch-30.pt \
+  --max-duration 550
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from itertools import chain
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AmiAsrDataModule
+from decoder import Decoder
+from dprnn import DPRNN
+from einops.layers.torch import Rearrange
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import LOG_EPSILON, fix_random_seed
+from model import SURT
+from optim import Eden, ScaledAdam
+from scaling import ScaledLinear, ScaledLSTM
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-mask-encoder-layers",
+        type=int,
+        default=4,
+        help="Number of layers in the DPRNN based mask encoder.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-dim",
+        type=int,
+        default=256,
+        help="Hidden dimension of the LSTM blocks in DPRNN.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-segment-size",
+        type=int,
+        default=32,
+        help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the "
+        "decode-chunk-length of the zipformer encoder.",
+    )
+
+    parser.add_argument(
+        "--chunk-width-randomization",
+        type=bool,
+        default=False,
+        help="Whether to randomize the chunk width in DPRNN.",
+    )
+
+    # Zipformer config is based on:
+    # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,2,2,2",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="768,768,768,768,768",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--use-joint-encoder-layer",
+        type=str,
+        default="linear",
+        choices=["linear", "lstm", "none"],
+        help="Whether to use a joint layer to combine all branches.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conv_lstm_transducer_stateless_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--model-init-ckpt",
+        type=str,
+        default=None,
+        help="""The model checkpoint to initialize the model (either full or part).
+        If not specified, the model is randomly initialized.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.0001, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=2,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=1,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 2000,
+            # parameters for SURT
+            "num_channels": 2,
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed
+            # parameters for Noam
+            "model_warm_step": 5000,  # arg given to model, not for lrate
+            # parameters for ctc loss
+            "beam_size": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_mask_encoder_model(params: AttributeDict) -> nn.Module:
+    mask_encoder = DPRNN(
+        feature_dim=params.feature_dim,
+        input_size=params.mask_encoder_dim,
+        hidden_size=params.mask_encoder_dim,
+        output_size=params.feature_dim * params.num_channels,
+        segment_size=params.mask_encoder_segment_size,
+        num_blocks=params.num_mask_encoder_layers,
+        chunk_width_randomization=params.chunk_width_randomization,
+    )
+    return mask_encoder
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_joint_encoder_layer(params: AttributeDict) -> nn.Module:
+    class TakeFirst(nn.Module):
+        def forward(self, x):
+            return x[0]
+
+    if params.use_joint_encoder_layer == "linear":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            nn.Linear(
+                params.num_channels * encoder_dim, params.num_channels * encoder_dim
+            ),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "lstm":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            ScaledLSTM(
+                input_size=params.num_channels * encoder_dim,
+                hidden_size=params.num_channels * encoder_dim,
+                num_layers=1,
+                bias=True,
+                batch_first=True,
+                dropout=0.0,
+                bidirectional=False,
+            ),
+            TakeFirst(),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "none":
+        joint_layer = None
+    else:
+        raise ValueError(
+            f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}"
+        )
+    return joint_layer
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_surt_model(
+    params: AttributeDict,
+) -> nn.Module:
+    mask_encoder = get_mask_encoder_model(params)
+    encoder = get_encoder_model(params)
+    joint_layer = get_joint_encoder_layer(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = SURT(
+        mask_encoder=mask_encoder,
+        encoder=encoder,
+        joint_encoder_layer=joint_layer,
+        decoder=decoder,
+        joiner=joiner,
+        num_channels=params.num_channels,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor:
+    """
+    Compute HEAT loss for separated sources using the output of mask encoder.
+    Args:
+      x_masked:
+        The output of mask encoder. It is a tensor of shape (B, T, C).
+      batch:
+        A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()`
+        for the content in it.
+      num_channels:
+        The number of output branches in the SURT model.
+    """
+    B, T, D = x_masked[0].shape
+    device = x_masked[0].device
+
+    # Create training targets for each channel.
+    targets = []
+    for i in range(num_channels):
+        target = torch.ones_like(x_masked[i]) * LOG_EPSILON
+        targets.append(target)
+
+    source_feats = batch["source_feats"]
+    source_boundaries = batch["source_boundaries"]
+    input_lens = batch["input_lens"].to(device)
+    # Assign sources to channels based on the HEAT criteria
+    for b in range(B):
+        cut_source_feats = source_feats[b]
+        cut_source_boundaries = source_boundaries[b]
+        last_seg_end = [0 for _ in range(num_channels)]
+        for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries):
+            assigned = False
+            for i in range(num_channels):
+                if start >= last_seg_end[i]:
+                    targets[i][b, start:end, :] += source_feat.to(device)
+                    last_seg_end[i] = max(end, last_seg_end[i])
+                    assigned = True
+                    break
+            if not assigned:
+                min_end_channel = last_seg_end.index(min(last_seg_end))
+                targets[min_end_channel][b, start:end, :] += source_feat
+                last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel])
+
+    # Get padding mask based on input lengths
+    pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1)
+    pad_mask = pad_mask.unsqueeze(-1)
+
+    # Compute masked loss for each channel
+    losses = torch.zeros((num_channels, B, T, D), device=device)
+    for i in range(num_channels):
+        loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none")
+        # Apply padding mask to loss
+        loss.masked_fill_(pad_mask, 0)
+        losses[i] = loss
+
+    # loss: C x B x T x D. pad_mask: B x T x 1
+    # We want to compute loss for each item in the batch. Each item has loss given
+    # by the sum over C, and average over T and D. For T, we need to use the padding.
+    loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device)
+    return loss
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"].to(device)
+    feature_lens = batch["input_lens"].to(device)
+
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+
+    # The dataloader returns text as a list of cuts, each of which is a list of channel
+    # text. We flatten this to a list where all channels are together, i.e., it looks like
+    # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2].
+    text = [val for tup in zip(*batch["text"]) for val in tup]
+    assert len(text) == len(feature) * params.num_channels
+
+    # Convert all channel texts to token IDs and create a ragged tensor.
+    y = sp.encode(text, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.model_warm_step
+
+    with torch.set_grad_enabled(is_training):
+        (simple_loss, pruned_loss, ctc_loss, x_masked) = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            reduction="none",
+            subsampling_factor=params.subsampling_factor,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        ctc_loss_is_finite = torch.isfinite(ctc_loss)
+
+        # Compute HEAT loss
+        if is_training and params.heat_loss_scale > 0.0:
+            heat_loss = compute_heat_loss(
+                x_masked, batch, num_channels=params.num_channels
+            )
+        else:
+            heat_loss = torch.tensor(0.0, device=device)
+
+        heat_loss_is_finite = torch.isfinite(heat_loss)
+        is_finite = (
+            simple_loss_is_finite
+            & pruned_loss_is_finite
+            & ctc_loss_is_finite
+            & heat_loss_is_finite
+        )
+        if not torch.all(is_finite):
+            # logging.info(
+            #     "Not all losses are finite!\n"
+            #     f"simple_losses: {simple_loss}\n"
+            #     f"pruned_losses: {pruned_loss}\n"
+            #     f"ctc_losses: {ctc_loss}\n"
+            #     f"heat_losses: {heat_loss}\n"
+            # )
+            # display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+            ctc_loss = ctc_loss[ctc_loss_is_finite]
+            heat_loss = heat_loss[heat_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if (
+                torch.all(~simple_loss_is_finite)
+                or torch.all(~pruned_loss_is_finite)
+                or torch.all(~ctc_loss_is_finite)
+                or torch.all(~heat_loss_is_finite)
+            ):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss_sum = simple_loss.sum()
+        pruned_loss_sum = pruned_loss.sum()
+        ctc_loss_sum = ctc_loss.sum()
+        heat_loss_sum = heat_loss.sum()
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+        loss = (
+            simple_loss_scale * simple_loss_sum
+            + pruned_loss_scale * pruned_loss_sum
+            + params.ctc_loss_scale * ctc_loss_sum
+            + params.heat_loss_scale * heat_loss_sum
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss_sum.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss_sum.detach().cpu().item()
+    if params.ctc_loss_scale > 0.0:
+        info["ctc_loss"] = ctc_loss_sum.detach().cpu().item()
+    if params.heat_loss_scale > 0.0:
+        info["heat_loss"] = heat_loss_sum.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    torch.cuda.empty_cache()
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = batch["inputs"].shape[0]
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+
+    if checkpoints is None and params.model_init_ckpt is not None:
+        logging.info(
+            f"Initializing model with checkpoint from {params.model_init_ckpt}"
+        )
+        init_ckpt = torch.load(params.model_init_ckpt, map_location=device)
+        model.load_state_dict(init_ckpt["model"], strict=False)
+
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    ami = AmiAsrDataModule(args)
+
+    train_cuts = ami.train_cuts()
+    train_cuts = train_cuts.filter(lambda c: 0.5 <= c.duration <= 35.0)
+    dev_cuts = ami.ami_cuts(split="dev", type="ihm-mix")
+    dev_cuts = dev_cuts.trim_to_supervision_groups(max_pause=0.0).filter(
+        lambda c: 0.2 <= c.duration <= 60.0
+    )
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = ami.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+    )
+    valid_dl = ami.valid_dataloaders(dev_cuts)
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = [sp.encode(text_ch) for text_ch in batch["text"]]
+    num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y]
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def main():
+    parser = get_parser()
+    AmiAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ami/SURT/dprnn_zipformer/zipformer.py b/egs/ami/SURT/dprnn_zipformer/zipformer.py
new file mode 120000
index 000000000..59b772024
--- /dev/null
+++ b/egs/ami/SURT/dprnn_zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../libricss/SURT/dprnn_zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/ami/SURT/local/add_source_feats.py b/egs/ami/SURT/local/add_source_feats.py
new file mode 100755
index 000000000..0917b88a6
--- /dev/null
+++ b/egs/ami/SURT/local/add_source_feats.py
@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file adds source features as temporal arrays to the mixture manifests.
+It looks for manifests in the directory data/manifests.
+"""
+import logging
+from pathlib import Path
+
+import numpy as np
+from lhotse import CutSet, LilcomChunkyWriter, load_manifest, load_manifest_lazy
+from tqdm import tqdm
+
+
+def add_source_feats():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    logging.info("Reading mixed cuts")
+    mixed_cuts_clean = load_manifest_lazy(src_dir / "cuts_train_clean.jsonl.gz")
+    mixed_cuts_reverb = load_manifest_lazy(src_dir / "cuts_train_reverb.jsonl.gz")
+
+    logging.info("Reading source cuts")
+    source_cuts = load_manifest(src_dir / "ihm_cuts_train_trimmed.jsonl.gz")
+
+    logging.info("Adding source features to the mixed cuts")
+    pbar = tqdm(total=len(mixed_cuts_clean), desc="Adding source features")
+    with CutSet.open_writer(
+        src_dir / "cuts_train_clean_sources.jsonl.gz"
+    ) as cut_writer_clean, CutSet.open_writer(
+        src_dir / "cuts_train_reverb_sources.jsonl.gz"
+    ) as cut_writer_reverb, LilcomChunkyWriter(
+        output_dir / "feats_train_clean_sources"
+    ) as source_feat_writer:
+        for cut_clean, cut_reverb in zip(mixed_cuts_clean, mixed_cuts_reverb):
+            assert cut_reverb.id == cut_clean.id + "_rvb"
+            source_feats = []
+            source_feat_offsets = []
+            cur_offset = 0
+            for sup in sorted(
+                cut_clean.supervisions, key=lambda s: (s.start, s.speaker)
+            ):
+                source_cut = source_cuts[sup.id]
+                source_feats.append(source_cut.load_features())
+                source_feat_offsets.append(cur_offset)
+                cur_offset += source_cut.num_frames
+            cut_clean.source_feats = source_feat_writer.store_array(
+                cut_clean.id, np.concatenate(source_feats, axis=0)
+            )
+            cut_clean.source_feat_offsets = source_feat_offsets
+            cut_writer_clean.write(cut_clean)
+            # Also write the reverb cut
+            cut_reverb.source_feats = cut_clean.source_feats
+            cut_reverb.source_feat_offsets = cut_clean.source_feat_offsets
+            cut_writer_reverb.write(cut_reverb)
+            pbar.update(1)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    add_source_feats()
diff --git a/egs/ami/SURT/local/compute_fbank_aimix.py b/egs/ami/SURT/local/compute_fbank_aimix.py
new file mode 100755
index 000000000..91b3a060b
--- /dev/null
+++ b/egs/ami/SURT/local/compute_fbank_aimix.py
@@ -0,0 +1,185 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the synthetically mixed AMI and ICSI
+train set.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+import random
+import warnings
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+import torchaudio
+from lhotse import (
+    AudioSource,
+    LilcomChunkyWriter,
+    Recording,
+    load_manifest,
+    load_manifest_lazy,
+)
+from lhotse.audio import set_ffmpeg_torchaudio_info_enabled
+from lhotse.cut import MixedCut, MixTrack, MultiCut
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.utils import fix_random_seed, uuid4
+from tqdm import tqdm
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+torchaudio.set_audio_backend("soundfile")
+set_ffmpeg_torchaudio_info_enabled(False)
+
+
+def compute_fbank_aimix():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    train_cuts = load_manifest_lazy(src_dir / "ai-mix_cuts_clean_full.jsonl.gz")
+
+    # only uses RIRs and noises from REVERB challenge
+    real_rirs = load_manifest(src_dir / "real-rir_recordings_all.jsonl.gz").filter(
+        lambda r: "RVB2014" in r.id
+    )
+    noises = load_manifest(src_dir / "iso-noise_recordings_all.jsonl.gz").filter(
+        lambda r: "RVB2014" in r.id
+    )
+
+    # Apply perturbation to the training cuts
+    logging.info("Applying perturbation to the training cuts")
+    train_cuts_rvb = train_cuts.map(
+        lambda c: augment(
+            c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True
+        )
+    )
+
+    logging.info("Extracting fbank features for training cuts")
+    _ = train_cuts.compute_and_store_features_batch(
+        extractor=extractor,
+        storage_path=output_dir / "ai-mix_feats_clean",
+        manifest_path=src_dir / "cuts_train_clean.jsonl.gz",
+        batch_duration=5000,
+        num_workers=4,
+        storage_type=LilcomChunkyWriter,
+        overwrite=True,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        _ = train_cuts_rvb.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / "ai-mix_feats_reverb",
+            manifest_path=src_dir / "cuts_train_reverb.jsonl.gz",
+            batch_duration=5000,
+            num_workers=4,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+
+def augment(cut, perturb_snr=False, rirs=None, noises=None, perturb_loudness=False):
+    """
+    Given a mixed cut, this function optionally applies the following augmentations:
+    - Perturbing the SNRs of the tracks (in range [-5, 5] dB)
+    - Reverberation using a randomly selected RIR
+    - Adding noise
+    - Perturbing the loudness (in range [-20, -25] dB)
+    """
+    out_cut = cut.drop_features()
+
+    # Perturb the SNRs (optional)
+    if perturb_snr:
+        snrs = [random.uniform(-5, 5) for _ in range(len(cut.tracks))]
+        for i, (track, snr) in enumerate(zip(out_cut.tracks, snrs)):
+            if i == 0:
+                # Skip the first track since it is the reference
+                continue
+            track.snr = snr
+
+    # Reverberate the cut (optional)
+    if rirs is not None:
+        # Select an RIR at random
+        rir = random.choice(rirs)
+        # Select a channel at random
+        rir_channel = random.choice(list(range(rir.num_channels)))
+        # Reverberate the cut
+        out_cut = out_cut.reverb_rir(rir_recording=rir, rir_channels=[rir_channel])
+
+    # Add noise (optional)
+    if noises is not None:
+        # Select a noise recording at random
+        noise = random.choice(noises).to_cut()
+        if isinstance(noise, MultiCut):
+            noise = noise.to_mono()[0]
+        # Select an SNR at random
+        snr = random.uniform(10, 30)
+        # Repeat the noise to match the duration of the cut
+        noise = repeat_cut(noise, out_cut.duration)
+        out_cut = MixedCut(
+            id=out_cut.id,
+            tracks=[
+                MixTrack(cut=out_cut, type="MixedCut"),
+                MixTrack(cut=noise, type="DataCut", snr=snr),
+            ],
+        )
+
+    # Perturb the loudness (optional)
+    if perturb_loudness:
+        target_loudness = random.uniform(-20, -25)
+        out_cut = out_cut.normalize_loudness(target_loudness, mix_first=True)
+    return out_cut
+
+
+def repeat_cut(cut, duration):
+    while cut.duration < duration:
+        cut = cut.mix(cut, offset_other_by=cut.duration)
+    return cut.truncate(duration=duration)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    fix_random_seed(42)
+    compute_fbank_aimix()
diff --git a/egs/ami/SURT/local/compute_fbank_ami.py b/egs/ami/SURT/local/compute_fbank_ami.py
new file mode 100755
index 000000000..351b41765
--- /dev/null
+++ b/egs/ami/SURT/local/compute_fbank_ami.py
@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the AMI dataset.
+We compute features for full recordings (i.e., without trimming to supervisions).
+This way we can create arbitrary segmentations later.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+import math
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+from lhotse import CutSet, LilcomChunkyWriter
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_ami():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    manifests = {}
+    for part in ["ihm-mix", "sdm", "mdm8-bf"]:
+        manifests[part] = read_manifests_if_cached(
+            dataset_parts=["train", "dev", "test"],
+            output_dir=src_dir,
+            prefix=f"ami-{part}",
+            suffix="jsonl.gz",
+        )
+
+    for part in ["ihm-mix", "sdm", "mdm8-bf"]:
+        for split in ["train", "dev", "test"]:
+            logging.info(f"Processing {part} {split}")
+            cuts = CutSet.from_manifests(
+                **manifests[part][split]
+            ).compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"ami-{part}_{split}_feats",
+                manifest_path=src_dir / f"cuts_ami-{part}_{split}.jsonl.gz",
+                batch_duration=5000,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+            )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_ami()
diff --git a/egs/ami/SURT/local/compute_fbank_icsi.py b/egs/ami/SURT/local/compute_fbank_icsi.py
new file mode 100755
index 000000000..4e2ff3f3b
--- /dev/null
+++ b/egs/ami/SURT/local/compute_fbank_icsi.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the ICSI dataset.
+We compute features for full recordings (i.e., without trimming to supervisions).
+This way we can create arbitrary segmentations later.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+import math
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+from lhotse import CutSet, LilcomChunkyWriter
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_icsi():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    manifests = {}
+    for part in ["ihm-mix", "sdm"]:
+        manifests[part] = read_manifests_if_cached(
+            dataset_parts=["train"],
+            output_dir=src_dir,
+            prefix=f"icsi-{part}",
+            suffix="jsonl.gz",
+        )
+
+    for part in ["ihm-mix", "sdm"]:
+        for split in ["train"]:
+            logging.info(f"Processing {part} {split}")
+            cuts = CutSet.from_manifests(
+                **manifests[part][split]
+            ).compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"icsi-{part}_{split}_feats",
+                manifest_path=src_dir / f"cuts_icsi-{part}_{split}.jsonl.gz",
+                batch_duration=5000,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+                overwrite=True,
+            )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_icsi()
diff --git a/egs/ami/SURT/local/compute_fbank_ihm.py b/egs/ami/SURT/local/compute_fbank_ihm.py
new file mode 100755
index 000000000..56f54aa21
--- /dev/null
+++ b/egs/ami/SURT/local/compute_fbank_ihm.py
@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the trimmed sub-segments which will be
+used for simulating the training mixtures.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+import math
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+import torchaudio
+from lhotse import CutSet, LilcomChunkyWriter, load_manifest
+from lhotse.audio import set_ffmpeg_torchaudio_info_enabled
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+from tqdm import tqdm
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+torchaudio.set_audio_backend("soundfile")
+set_ffmpeg_torchaudio_info_enabled(False)
+
+
+def compute_fbank_ihm():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    manifests = {}
+    for data in ["ami", "icsi"]:
+        manifests[data] = read_manifests_if_cached(
+            dataset_parts=["train"],
+            output_dir=src_dir,
+            types=["recordings", "supervisions"],
+            prefix=f"{data}-ihm",
+            suffix="jsonl.gz",
+        )
+
+    logging.info("Computing features")
+    for data in ["ami", "icsi"]:
+        cs = CutSet.from_manifests(**manifests[data]["train"])
+        cs = cs.trim_to_supervisions(keep_overlapping=False)
+        cs = cs.normalize_loudness(target=-23.0, affix_id=False)
+        cs = cs + cs.perturb_speed(0.9) + cs.perturb_speed(1.1)
+        _ = cs.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / f"{data}-ihm_train_feats",
+            manifest_path=src_dir / f"{data}-ihm_cuts_train.jsonl.gz",
+            batch_duration=5000,
+            num_workers=4,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_ihm()
diff --git a/egs/ami/SURT/local/prepare_ami_train_cuts.py b/egs/ami/SURT/local/prepare_ami_train_cuts.py
new file mode 100755
index 000000000..72fced70d
--- /dev/null
+++ b/egs/ami/SURT/local/prepare_ami_train_cuts.py
@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file creates AMI train segments.
+"""
+import logging
+import math
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+from lhotse import LilcomChunkyWriter, load_manifest_lazy
+from lhotse.cut import Cut, CutSet
+from lhotse.utils import EPSILON, add_durations
+from tqdm import tqdm
+
+
+def cut_into_windows(cuts: CutSet, duration: float):
+    """
+    This function takes a CutSet and cuts each cut into windows of roughly
+    `duration` seconds. By roughly, we mean that we try to adjust for the last supervision
+    that exceeds the duration, or is shorter than the duration.
+    """
+    res = []
+    with tqdm() as pbar:
+        for cut in cuts:
+            pbar.update(1)
+            sups = cut.index_supervisions()[cut.id]
+            sr = cut.sampling_rate
+            start = 0.0
+            end = duration
+            num_tries = 0
+            while start < cut.duration and num_tries < 2:
+                # Find the supervision that are cut by the window endpoint
+                hitlist = [iv for iv in sups.at(end) if iv.begin < end]
+                # If there are no supervisions, we are done
+                if not hitlist:
+                    res.append(
+                        cut.truncate(
+                            offset=start,
+                            duration=add_durations(end, -start, sampling_rate=sr),
+                            keep_excessive_supervisions=False,
+                        )
+                    )
+                    # Update the start and end for the next window
+                    start = end
+                    end = add_durations(end, duration, sampling_rate=sr)
+                else:
+                    # find ratio of durations cut by the window endpoint
+                    ratios = [
+                        add_durations(end, -iv.end, sampling_rate=sr) / iv.length()
+                        for iv in hitlist
+                    ]
+                    # we retain the supervisions that have >50% of their duration
+                    # in the window, and discard the others
+                    retained = []
+                    discarded = []
+                    for iv, ratio in zip(hitlist, ratios):
+                        if ratio > 0.5:
+                            retained.append(iv)
+                        else:
+                            discarded.append(iv)
+                    cur_end = max(iv.end for iv in retained) if retained else end
+                    res.append(
+                        cut.truncate(
+                            offset=start,
+                            duration=add_durations(cur_end, -start, sampling_rate=sr),
+                            keep_excessive_supervisions=False,
+                        )
+                    )
+                    # For the next window, we start at the earliest discarded supervision
+                    next_start = min(iv.begin for iv in discarded) if discarded else end
+                    next_end = add_durations(next_start, duration, sampling_rate=sr)
+                    # It may happen that next_start is the same as start, in which case
+                    # we will advance the window anyway
+                    if next_start == start:
+                        logging.warning(
+                            f"Next start is the same as start: {next_start} == {start} for cut {cut.id}"
+                        )
+                        start = end + EPSILON
+                        end = add_durations(start, duration, sampling_rate=sr)
+                        num_tries += 1
+                    else:
+                        start = next_start
+                        end = next_end
+    return CutSet.from_cuts(res)
+
+
+def prepare_train_cuts():
+    src_dir = Path("data/manifests")
+
+    logging.info("Loading the manifests")
+    train_cuts_ihm = load_manifest_lazy(
+        src_dir / "cuts_ami-ihm-mix_train.jsonl.gz"
+    ).map(lambda c: c.with_id(f"{c.id}_ihm-mix"))
+    train_cuts_sdm = load_manifest_lazy(src_dir / "cuts_ami-sdm_train.jsonl.gz").map(
+        lambda c: c.with_id(f"{c.id}_sdm")
+    )
+    train_cuts_mdm = load_manifest_lazy(
+        src_dir / "cuts_ami-mdm8-bf_train.jsonl.gz"
+    ).map(lambda c: c.with_id(f"{c.id}_mdm8-bf"))
+
+    # Combine all cuts into one CutSet
+    train_cuts = train_cuts_ihm + train_cuts_sdm + train_cuts_mdm
+
+    train_cuts_1 = train_cuts.trim_to_supervision_groups(max_pause=0.5)
+    train_cuts_2 = train_cuts.trim_to_supervision_groups(max_pause=0.0)
+
+    # Combine the two segmentations
+    train_all = train_cuts_1 + train_cuts_2
+
+    # At this point, some of the cuts may be very long. We will cut them into windows of
+    # roughly 30 seconds.
+    logging.info("Cutting the segments into windows of 30 seconds")
+    train_all_30 = cut_into_windows(train_all, duration=30.0)
+    logging.info(f"Number of cuts after cutting into windows: {len(train_all_30)}")
+
+    # Show statistics
+    train_all.describe(full=True)
+
+    # Save the cuts
+    logging.info("Saving the cuts")
+    train_all.to_file(src_dir / "cuts_train_ami.jsonl.gz")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    prepare_train_cuts()
diff --git a/egs/ami/SURT/local/prepare_icsi_train_cuts.py b/egs/ami/SURT/local/prepare_icsi_train_cuts.py
new file mode 100755
index 000000000..818e26bfb
--- /dev/null
+++ b/egs/ami/SURT/local/prepare_icsi_train_cuts.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file creates ICSI train segments.
+"""
+import logging
+from pathlib import Path
+
+from lhotse import load_manifest_lazy
+from prepare_ami_train_cuts import cut_into_windows
+
+
+def prepare_train_cuts():
+    src_dir = Path("data/manifests")
+
+    logging.info("Loading the manifests")
+    train_cuts_ihm = load_manifest_lazy(
+        src_dir / "cuts_icsi-ihm-mix_train.jsonl.gz"
+    ).map(lambda c: c.with_id(f"{c.id}_ihm-mix"))
+    train_cuts_sdm = load_manifest_lazy(src_dir / "cuts_icsi-sdm_train.jsonl.gz").map(
+        lambda c: c.with_id(f"{c.id}_sdm")
+    )
+
+    # Combine all cuts into one CutSet
+    train_cuts = train_cuts_ihm + train_cuts_sdm
+
+    train_cuts_1 = train_cuts.trim_to_supervision_groups(max_pause=0.5)
+    train_cuts_2 = train_cuts.trim_to_supervision_groups(max_pause=0.0)
+
+    # Combine the two segmentations
+    train_all = train_cuts_1 + train_cuts_2
+
+    # At this point, some of the cuts may be very long. We will cut them into windows of
+    # roughly 30 seconds.
+    logging.info("Cutting the segments into windows of 30 seconds")
+    train_all_30 = cut_into_windows(train_all, duration=30.0)
+    logging.info(f"Number of cuts after cutting into windows: {len(train_all_30)}")
+
+    # Show statistics
+    train_all.describe(full=True)
+
+    # Save the cuts
+    logging.info("Saving the cuts")
+    train_all.to_file(src_dir / "cuts_train_icsi.jsonl.gz")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    prepare_train_cuts()
diff --git a/egs/ami/SURT/local/prepare_lang_bpe.py b/egs/ami/SURT/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/ami/SURT/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/ami/SURT/local/train_bpe_model.py b/egs/ami/SURT/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/ami/SURT/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/ami/SURT/prepare.sh b/egs/ami/SURT/prepare.sh
new file mode 100755
index 000000000..ea4e5baf2
--- /dev/null
+++ b/egs/ami/SURT/prepare.sh
@@ -0,0 +1,195 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/ami
+#      You can find audio and transcripts for AMI in this path.
+#
+#  - $dl_dir/icsi
+#      You can find audio and transcripts for ICSI in this path.
+#
+#  - $dl_dir/rirs_noises
+#      This directory contains the RIRS_NOISES corpus downloaded from https://openslr.org/28/.
+#
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+vocab_size=500
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/amicorpus,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/amicorpus $dl_dir/amicorpus
+  #
+  if [ ! -d $dl_dir/amicorpus ]; then
+    for mic in ihm ihm-mix sdm mdm8-bf; do
+      lhotse download ami --mic $mic $dl_dir/amicorpus
+    done
+  fi
+
+  # If you have pre-downloaded it to /path/to/icsi,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/icsi $dl_dir/icsi
+  #
+  if [ ! -d $dl_dir/icsi ]; then
+    lhotse download icsi $dl_dir/icsi
+  fi
+
+  # If you have pre-downloaded it to /path/to/rirs_noises,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/rirs_noises $dl_dir/
+  #
+  if [ ! -d $dl_dir/rirs_noises ]; then
+    lhotse download rirs_noises $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare AMI manifests"
+  # We assume that you have downloaded the AMI corpus
+  # to $dl_dir/amicorpus. We perform text normalization for the transcripts.
+  mkdir -p data/manifests
+  for mic in ihm ihm-mix sdm mdm8-bf; do
+    log "Preparing AMI manifest for $mic"
+    lhotse prepare ami --mic $mic --max-words-per-segment 30 --merge-consecutive $dl_dir/amicorpus data/manifests/
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare ICSI manifests"
+  # We assume that you have downloaded the ICSI corpus
+  # to $dl_dir/icsi. We perform text normalization for the transcripts.
+  mkdir -p data/manifests
+  log "Preparing ICSI manifest"
+  for mic in ihm ihm-mix sdm; do
+    lhotse prepare icsi --mic $mic $dl_dir/icsi data/manifests/
+  done
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare RIRs"
+  # We assume that you have downloaded the RIRS_NOISES corpus
+  # to $dl_dir/rirs_noises
+  lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises data/manifests
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 3: Extract features for AMI and ICSI recordings"
+  python local/compute_fbank_ami.py
+  python local/compute_fbank_icsi.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Create sources for simulating mixtures"
+  # In the following script, we speed-perturb the IHM recordings and extract features.
+  python local/compute_fbank_ihm.py
+  lhotse combine data/manifests/ami-ihm_cuts_train.jsonl.gz \
+    data/manifests/icsi-ihm_cuts_train.jsonl.gz - |\
+    lhotse cut trim-to-alignments --type word --max-pause 0.5 - - |\
+    lhotse filter 'duration<=12.0' - - |\
+    shuf | gzip -c > data/manifests/ihm_cuts_train_trimmed.jsonl.gz
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Create training mixtures"
+  lhotse workflows simulate-meetings \
+    --method conversational \
+    --same-spk-pause 0.5 \
+    --diff-spk-pause 0.5 \
+    --diff-spk-overlap 1.0 \
+    --prob-diff-spk-overlap 0.8 \
+    --num-meetings 200000 \
+    --num-speakers-per-meeting 2,3 \
+    --max-duration-per-speaker 15.0 \
+    --max-utterances-per-speaker 3 \
+    --seed 1234 \
+    --num-jobs 2 \
+    data/manifests/ihm_cuts_train_trimmed.jsonl.gz \
+    data/manifests/ai-mix_cuts_clean.jsonl.gz
+
+  python local/compute_fbank_aimix.py
+
+  # Add source features to the manifest (will be used for masking loss)
+  # This may take ~2 hours.
+  python local/add_source_feats.py
+
+  # Combine clean and reverb
+  cat <(gunzip -c data/manifests/cuts_train_clean_sources.jsonl.gz) \
+    <(gunzip -c data/manifests/cuts_train_reverb_sources.jsonl.gz) |\
+    shuf | gzip -c > data/manifests/cuts_train_comb_sources.jsonl.gz
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Create training mixtures from real sessions"
+  python local/prepare_ami_train_cuts.py
+  python local/prepare_icsi_train_cuts.py
+
+  # Combine AMI and ICSI
+  cat <(gunzip -c data/manifests/cuts_train_ami.jsonl.gz) \
+    <(gunzip -c data/manifests/cuts_train_icsi.jsonl.gz) |\
+    shuf | gzip -c > data/manifests/cuts_train_ami_icsi.jsonl.gz
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Dump transcripts for BPE model training (using AMI and ICSI)."
+  mkdir -p data/lm
+  cat <(gunzip -c data/manifests/ami-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \
+      <(gunzip -c data/manifests/icsi-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \
+  > data/lm/transcript_words.txt
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Prepare BPE based lang (combining AMI and ICSI)"
+
+  lang_dir=data/lang_bpe_${vocab_size}
+  mkdir -p $lang_dir
+
+  # Add special words to words.txt
+  echo "<eps> 0" > $lang_dir/words.txt
+  echo "!SIL 1" >> $lang_dir/words.txt
+  echo "<UNK> 2" >> $lang_dir/words.txt
+
+  # Add regular words to words.txt
+  cat data/lm/transcript_words.txt | grep -o -E '\w+' | sort -u | awk '{print $0,NR+2}' >> $lang_dir/words.txt
+
+  # Add remaining special word symbols expected by LM scripts.
+  num_words=$(cat $lang_dir/words.txt | wc -l)
+  echo "<s> ${num_words}" >> $lang_dir/words.txt
+  num_words=$(cat $lang_dir/words.txt | wc -l)
+  echo "</s> ${num_words}" >> $lang_dir/words.txt
+  num_words=$(cat $lang_dir/words.txt | wc -l)
+  echo "#0 ${num_words}" >> $lang_dir/words.txt
+
+  ./local/train_bpe_model.py \
+    --lang-dir $lang_dir \
+    --vocab-size $vocab_size \
+    --transcript data/lm/transcript_words.txt
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+  fi
+fi
diff --git a/egs/ami/SURT/shared b/egs/ami/SURT/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/ami/SURT/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/README.md b/egs/commonvoice/ASR/README.md
new file mode 100644
index 000000000..a4582499b
--- /dev/null
+++ b/egs/commonvoice/ASR/README.md
@@ -0,0 +1,18 @@
+# Introduction
+
+This recipe includes some different ASR models trained with Common Voice
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                                           |
+|---------------------------------------|---------------------|--------------------|---------------------------------------------------|
+| `pruned_transducer_stateless7`        | Zipformer           | Embedding + Conv1d | First experiment with Zipformer from Dan          |
+
+The decoder in `transducer_stateless` is modified from the paper
+[RNN-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/commonvoice/ASR/RESULTS.md b/egs/commonvoice/ASR/RESULTS.md
new file mode 100644
index 000000000..2c158d91d
--- /dev/null
+++ b/egs/commonvoice/ASR/RESULTS.md
@@ -0,0 +1,84 @@
+## Results
+### GigaSpeech BPE training results (Pruned Stateless Transducer 7)
+
+#### [pruned_transducer_stateless7](./pruned_transducer_stateless7)
+
+See #997  for more details.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+The best WER, as of 2023-04-17, for Common Voice English 13.0 (cv-corpus-13.0-2023-03-09/en) is below:
+
+Results are:
+
+|                      |  Dev  | Test  |
+|----------------------|-------|-------|
+|    greedy search     | 9.96  | 12.54 |
+| modified beam search | 9.86  | 12.48 |
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \ 
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 550
+```
+
+and the following commands for decoding:
+
+```bash
+# greedy search
+./pruned_transducer_stateless7/decode.py \
+  --epoch 30 \
+  --avg 5 \
+  --decoding-method greedy_search \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --bpe-model data/en/lang_bpe_500/bpe.model \
+  --max-duration 600
+
+# modified beam search
+./pruned_transducer_stateless7/decode.py \
+  --epoch 30 \
+  --avg 5 \
+  --decoding-method modified_beam_search \
+  --beam-size 4 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --bpe-model data/en/lang_bpe_500/bpe.model \
+  --max-duration 600
+```
+
+Pretrained model is available at
+<https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17>
+
+The tensorboard log for training is available at
+<https://tensorboard.dev/experiment/j4pJQty6RMOkMJtRySREKw/>
+
+
+### Commonvoice (fr) BPE training results (Pruned Stateless Transducer 7_streaming)
+
+#### [pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
+
+See #1018  for more details.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+The best WER for Common Voice French 12.0 (cv-corpus-12.0-2022-12-07/fr) is below:
+
+Results are:
+
+|    decoding method   | Test  |
+|----------------------|-------|
+|    greedy search     | 9.95  | 
+| modified beam search | 9.57  |
+|   fast beam search   | 9.67  |
+
+Note: This best result is trained on the full librispeech and gigaspeech, and then fine-tuned on the full commonvoice.
+
+Detailed experimental results and Pretrained model are available at
+<https://huggingface.co/shaojieli/icefall-asr-commonvoice-fr-pruned-transducer-stateless7-streaming-2023-04-02>
+
diff --git a/egs/commonvoice/ASR/local/compile_hlg.py b/egs/commonvoice/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/commonvoice/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/local/compile_lg.py b/egs/commonvoice/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/commonvoice/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/local/compute_fbank_commonvoice_dev_test.py b/egs/commonvoice/ASR/local/compute_fbank_commonvoice_dev_test.py
new file mode 100755
index 000000000..a0b4d224c
--- /dev/null
+++ b/egs/commonvoice/ASR/local/compute_fbank_commonvoice_dev_test.py
@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the CommonVoice dataset.
+It looks for manifests in the directory data/${lang}/manifests.
+
+The generated fbank features are saved in data/${lang}/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--language",
+        type=str,
+        help="""Language of Common Voice""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_commonvoice_dev_test(language: str):
+    src_dir = Path(f"data/{language}/manifests")
+    output_dir = Path(f"data/{language}/fbank")
+    num_workers = 16
+    batch_duration = 200
+
+    subsets = ("dev", "test")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    for partition in subsets:
+        cuts_path = output_dir / f"cv-{language}_cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping.")
+            continue
+
+        raw_cuts_path = output_dir / f"cv-{language}_cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/cv-{language}_feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_commonvoice_dev_test(language=args.language)
diff --git a/egs/commonvoice/ASR/local/compute_fbank_commonvoice_splits.py b/egs/commonvoice/ASR/local/compute_fbank_commonvoice_splits.py
new file mode 100755
index 000000000..0564f6ec6
--- /dev/null
+++ b/egs/commonvoice/ASR/local/compute_fbank_commonvoice_splits.py
@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.             (Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+    set_audio_duration_mismatch_tolerance,
+    set_caching_enabled,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--language",
+        type=str,
+        help="""Language of Common Voice""",
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the train subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_commonvoice_splits(args):
+    subset = "train"
+    num_splits = args.num_splits
+    language = args.language
+    output_dir = f"data/{language}/fbank/cv-{language}_{subset}_split_{num_splits}"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    set_audio_duration_mismatch_tolerance(0.01)  # 10ms tolerance
+    set_caching_enabled(False)
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"cv-{language}_cuts_{subset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"cv-{language}_cuts_{subset}_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/cv-{language}_feats_{subset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_commonvoice_splits(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/local/compute_fbank_musan.py b/egs/commonvoice/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/commonvoice/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/local/filter_cuts.py b/egs/commonvoice/ASR/local/filter_cuts.py
new file mode 120000
index 000000000..27aca1729
--- /dev/null
+++ b/egs/commonvoice/ASR/local/filter_cuts.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/filter_cuts.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/local/prepare_lang_bpe.py b/egs/commonvoice/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/commonvoice/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/local/preprocess_commonvoice.py b/egs/commonvoice/ASR/local/preprocess_commonvoice.py
new file mode 100755
index 000000000..5f6aa3ec0
--- /dev/null
+++ b/egs/commonvoice/ASR/local/preprocess_commonvoice.py
@@ -0,0 +1,134 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import re
+from pathlib import Path
+from typing import Optional
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--language",
+        type=str,
+        help="""Language of Common Voice""",
+    )
+
+    return parser.parse_args()
+
+
+def normalize_text(utt: str, language: str) -> str:
+    utt = re.sub(r"[{0}]+".format("-"), " ", utt)
+    utt = re.sub("’", "'", utt)
+    if language == "en":
+        return re.sub(r"[^a-zA-Z\s]", "", utt).upper()
+    if language == "fr":
+        return re.sub(r"[^A-ZÀÂÆÇÉÈÊËÎÏÔŒÙÛÜ' ]", "", utt).upper()
+
+
+def preprocess_commonvoice(
+    language: str,
+    dataset: Optional[str] = None,
+):
+    src_dir = Path(f"data/{language}/manifests")
+    output_dir = Path(f"data/{language}/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    if dataset is None:
+        dataset_parts = (
+            "dev",
+            "test",
+            "train",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    logging.info("Loading manifest")
+    prefix = f"cv-{language}"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix=suffix,
+        prefix=prefix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"{prefix}_cuts_{partition}_raw.{suffix}"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        logging.info(f"Normalizing text in {partition}")
+        for sup in m["supervisions"]:
+            text = str(sup.text)
+            orig_text = text
+            sup.text = normalize_text(sup.text, language)
+            text = str(sup.text)
+            if len(orig_text) != len(text):
+                logging.info(
+                    f"\nOriginal text vs normalized text:\n{orig_text}\n{text}"
+                )
+
+        # Create long-recording cut manifests.
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        ).resample(16000)
+
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    preprocess_commonvoice(
+        language=args.language,
+        dataset=args.dataset,
+    )
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/local/train_bpe_model.py b/egs/commonvoice/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/commonvoice/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/local/validate_bpe_lexicon.py b/egs/commonvoice/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/commonvoice/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/prepare.sh b/egs/commonvoice/ASR/prepare.sh
new file mode 100755
index 000000000..edac0e8e6
--- /dev/null
+++ b/egs/commonvoice/ASR/prepare.sh
@@ -0,0 +1,304 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=16
+stage=-1
+stop_stage=100
+
+# Split data/${lang}set to this number of pieces
+# This is to avoid OOM during feature extraction.
+num_splits=1000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/$release/$lang
+#      This directory contains the following files downloaded from
+#       https://mozilla-common-voice-datasets.s3.dualstack.us-west-2.amazonaws.com/${release}/${release}-${lang}.tar.gz
+#
+#     - clips
+#     - dev.tsv
+#     - invalidated.tsv
+#     - other.tsv
+#     - reported.tsv
+#     - test.tsv
+#     - train.tsv
+#     - validated.tsv
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+release=cv-corpus-12.0-2022-12-07
+lang=fr
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/${lang}/lang_bpe_xxx,
+# data/${lang}/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data/${lang}".
+# You can safely remove "data/${lang}" and rerun this script to regenerate it.
+mkdir -p data/${lang}
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if ! command -v ffmpeg &> /dev/null; then
+  echo "This dataset requires ffmpeg"
+  echo "Please install ffmpeg first"
+  echo ""
+  echo "  sudo apt-get install ffmpeg"
+  exit 1
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/$release,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/$release $dl_dir/$release
+  #
+  if [ ! -d $dl_dir/$release/$lang/clips ]; then
+    lhotse download commonvoice --languages $lang --release $release $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare CommonVoice manifest"
+  # We assume that you have downloaded the CommonVoice corpus
+  # to $dl_dir/$release
+  mkdir -p data/${lang}/manifests
+  if [ ! -e data/${lang}/manifests/.cv-${lang}.done ]; then
+    lhotse prepare commonvoice --language $lang -j $nj $dl_dir/$release data/${lang}/manifests
+    touch data/${lang}/manifests/.cv-${lang}.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Preprocess CommonVoice manifest"
+  if [ ! -e data/${lang}/fbank/.preprocess_complete ]; then
+    ./local/preprocess_commonvoice.py  --language $lang
+    touch data/${lang}/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for dev and test subsets of CommonVoice"
+  mkdir -p data/${lang}/fbank
+  if [ ! -e data/${lang}/fbank/.cv-${lang}_dev_test.done ]; then
+    ./local/compute_fbank_commonvoice_dev_test.py --language $lang
+    touch data/${lang}/fbank/.cv-${lang}_dev_test.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Split train subset into ${num_splits} pieces"
+  split_dir=data/${lang}/fbank/cv-${lang}_train_split_${num_splits}
+  if [ ! -e $split_dir/.cv-${lang}_train_split.done ]; then
+    lhotse split $num_splits ./data/${lang}/fbank/cv-${lang}_cuts_train_raw.jsonl.gz $split_dir
+    touch $split_dir/.cv-${lang}_train_split.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compute features for train subset of CommonVoice"
+  if [ ! -e data/${lang}/fbank/.cv-${lang}_train.done ]; then
+    ./local/compute_fbank_commonvoice_splits.py \
+      --num-workers $nj \
+      --batch-duration 200 \
+      --start 0 \
+      --num-splits $num_splits \
+      --language $lang
+    touch data/${lang}/fbank/.cv-${lang}_train.done
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Combine features for train"
+  if [ ! -f data/${lang}/fbank/cv-${lang}_cuts_train.jsonl.gz ]; then
+    pieces=$(find data/${lang}/fbank/cv-${lang}_train_split_${num_splits} -name "cv-${lang}_cuts_train.*.jsonl.gz")
+    lhotse combine $pieces data/${lang}/fbank/cv-${lang}_cuts_train.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/${lang}/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      file=$(
+        find "data/${lang}/fbank/cv-${lang}_cuts_train.jsonl.gz"
+      )
+      gunzip -c ${file} | awk -F '"' '{print $30}' > $lang_dir/transcript_words.txt
+
+      # Ensure space only appears once
+      sed -i 's/\t/ /g' $lang_dir/transcript_words.txt
+      sed -i 's/[ ][ ]*/ /g' $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/words.txt ]; then
+      cat $lang_dir/transcript_words.txt | sed 's/ /\n/g' \
+        | sort -u | sed '/^$/d' > $lang_dir/words.txt
+      (echo '!SIL'; echo '<SPOKEN_NOISE>'; echo '<UNK>'; ) |
+        cat - $lang_dir/words.txt | sort | uniq | awk '
+        BEGIN {
+          print "<eps> 0";
+        }
+        {
+          if ($1 == "<s>") {
+            print "<s> is in the vocabulary!" | "cat 1>&2"
+            exit 1;
+          }
+          if ($1 == "</s>") {
+            print "</s> is in the vocabulary!" | "cat 1>&2"
+            exit 1;
+          }
+          printf("%s %d\n", $1, NR);
+        }
+        END {
+          printf("#0 %d\n", NR+1);
+          printf("<s> %d\n", NR+2);
+          printf("</s> %d\n", NR+3);
+        }' > $lang_dir/words || exit 1;
+      mv $lang_dir/words $lang_dir/words.txt
+    fi
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+      log "Converting L.pt to L.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L.pt \
+        $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+      log "Converting L_disambig.pt to L_disambig.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L_disambig.pt \
+        $lang_dir/L_disambig.fst
+    fi
+  done
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Prepare G"
+  # We assume you have install kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/${lang}/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir/lm
+    #3-gram used in building HLG, 4-gram used for LM rescoring
+    for ngram in 3 4; do
+      if [ ! -f $lang_dir/lm/${ngram}gram.arpa ]; then
+        ./shared/make_kn_lm.py \
+          -ngram-order ${ngram} \
+          -text $lang_dir/transcript_words.txt \
+          -lm $lang_dir/lm/${ngram}gram.arpa
+      fi
+
+      if [ ! -f $lang_dir/lm/${ngram}gram.fst.txt ]; then
+        python3 -m kaldilm \
+          --read-symbol-table="$lang_dir/words.txt" \
+          --disambig-symbol='#0' \
+          --max-order=${ngram} \
+          $lang_dir/lm/${ngram}gram.arpa > $lang_dir/lm/G_${ngram}_gram.fst.txt
+      fi
+    done
+  done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Compile HLG"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/${lang}/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+
+    # Note If ./local/compile_hlg.py throws OOM,
+    # please switch to the following command
+    #
+    # ./local/compile_hlg_using_openfst.py --lang-dir $lang_dir
+  done
+fi
+
+# Compile LG for RNN-T fast_beam_search decoding
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Compile LG"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/${lang}/lang_bpe_${vocab_size}
+    ./local/compile_lg.py --lang-dir $lang_dir
+  done
+fi
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/__init__.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py
new file mode 100644
index 000000000..546e9f9dd
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -0,0 +1,420 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class CommonVoiceAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. CommonVoice test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--language",
+            type=str,
+            default="en",
+            help="""Language of Common Voice""",
+        )
+        group.add_argument(
+            "--cv-manifest-dir",
+            type=Path,
+            default=Path("data/en/fbank"),
+            help="Path to directory with CommonVoice train/dev/test cuts.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with the other cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_dev.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_test.jsonl.gz"
+        )
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/beam_search.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/decode.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/decode.py
new file mode 100755
index 000000000..52b2fbcab
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/decode.py
@@ -0,0 +1,962 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search with RNNLM shallow fusion
+./pruned_transducer_stateless5/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+(9) modified beam search with LM shallow fusion + LODR
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --max-duration 600 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --decoding-method modified_beam_search_LODR \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.4 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16 \
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import CommonVoiceAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+    modified_beam_search_ngram_rescoring,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/en/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/en/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=3,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring, or LODR
+            """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+
+    else:
+        LM = None
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    commonvoice = CommonVoiceAsrDataModule(args)
+
+    dev_cuts = commonvoice.dev_cuts()
+    test_cuts = commonvoice.test_cuts()
+
+    dev_dl = commonvoice.valid_dataloaders(dev_cuts)
+    test_dl = commonvoice.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/decoder.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/encoder_interface.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/export-onnx.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/export-onnx.py
new file mode 100755
index 000000000..2b9f2293a
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/export-onnx.py
@@ -0,0 +1,601 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Yifan   Yang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --bpe-model $repo/data/en/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/en/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless7",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/export.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/export.py
new file mode 100755
index 000000000..53705321e
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/export.py
@@ -0,0 +1,321 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Yifan   Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/en/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 5 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/en/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 5
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/commonvoice/ASR
+    ./pruned_transducer_stateless7/decode.py \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/en/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17 
+    # You will find the pre-trained model in icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/en/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/joiner.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/model.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py
new file mode 100755
index 000000000..19c518eaf
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py
@@ -0,0 +1,240 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Yifan   Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script checks that exported onnx models produce the same output
+with the given torchscript model for the same input.
+
+We use the pre-trained model from
+https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+2. Export the model via torchscript (torch.jit.script())
+
+./pruned_transducer_stateless7/export.py \
+  --bpe-model $repo/data/en/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --jit 1
+
+It will generate the following file in $repo/exp:
+    - cpu_jit.pt
+
+3. Export the model to ONNX
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --bpe-model $repo/data/en/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+4. Run this file
+
+./pruned_transducer_stateless7/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-9999-avg-1.onnx
+"""
+
+import argparse
+import logging
+
+from icefall import is_module_available
+from onnx_pretrained import OnnxModel
+
+import torch
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript model",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    C = 80
+    for i in range(3):
+        N = torch.randint(low=1, high=20, size=(1,)).item()
+        T = torch.randint(low=30, high=50, size=(1,)).item()
+        logging.info(f"test_encoder: iter {i}, N={N}, T={T}")
+
+        x = torch.rand(N, T, C)
+        x_lens = torch.randint(low=30, high=T + 1, size=(N,))
+        x_lens[0] = T
+
+        torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens)
+        torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out)
+
+        onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens)
+
+        assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), (
+            (torch_encoder_out - onnx_encoder_out).abs().max()
+        )
+
+
+def test_decoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    context_size = onnx_model.context_size
+    vocab_size = onnx_model.vocab_size
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_decoder: iter {i}, N={N}")
+        x = torch.randint(
+            low=1,
+            high=vocab_size,
+            size=(N, context_size),
+            dtype=torch.int64,
+        )
+        torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False]))
+        torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out)
+        torch_decoder_out = torch_decoder_out.squeeze(1)
+
+        onnx_decoder_out = onnx_model.run_decoder(x)
+        assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
+            (torch_decoder_out - onnx_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1]
+    decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1]
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_joiner: iter {i}, N={N}")
+        encoder_out = torch.rand(N, encoder_dim)
+        decoder_out = torch.rand(N, decoder_dim)
+
+        projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out)
+        projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out)
+
+        torch_joiner_out = torch_model.joiner(encoder_out, decoder_out)
+        onnx_joiner_out = onnx_model.run_joiner(
+            projected_encoder_out, projected_decoder_out
+        )
+
+        assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
+            (torch_joiner_out - onnx_joiner_out).abs().max()
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_model = torch.jit.load(args.jit_filename)
+
+    onnx_model = OnnxModel(
+        encoder_model_filename=args.onnx_encoder_filename,
+        decoder_model_filename=args.onnx_decoder_filename,
+        joiner_model_filename=args.onnx_joiner_filename,
+    )
+
+    logging.info("Test encoder")
+    test_encoder(torch_model, onnx_model)
+
+    logging.info("Test decoder")
+    test_decoder(torch_model, onnx_model)
+
+    logging.info("Test joiner")
+    test_joiner(torch_model, onnx_model)
+    logging.info("Finished checking ONNX models")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# See https://github.com/pytorch/pytorch/issues/38342
+# and https://github.com/pytorch/pytorch/issues/33354
+#
+# If we don't do this, the delay increases whenever there is
+# a new request that changes the actual batch size.
+# If you use `py-spy dump --pid <server-pid> --native`, you will
+# see a lot of time is spent in re-compiling the torch script model.
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    torch.manual_seed(20220727)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_pretrained.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_pretrained.py
new file mode 100755
index 000000000..cf6ddfa36
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_pretrained.py
@@ -0,0 +1,422 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Yifan   Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+We use the pre-trained model from
+https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-cv-corpus-13.0-2023-03-09-en-pruned-transducer-stateless7-2023-04-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --bpe-model $repo/data/en/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+3. Run this file
+
+./pruned_transducer_stateless7/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+  --tokens $repo/data/en/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 2-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, its shape is (N, T', joiner_dim)
+            - encoder_out_lens, its shape is (N,)
+        """
+        out = self.encoder.run(
+            [
+                self.encoder.get_outputs()[0].name,
+                self.encoder.get_outputs()[1].name,
+            ],
+            {
+                self.encoder.get_inputs()[0].name: x.numpy(),
+                self.encoder.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, joiner_dim)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.run_decoder(decoder_input)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        # current_encoder_out's shape: (batch_size, joiner_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+        logits = model.run_joiner(current_encoder_out, decoder_out)
+
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                dtype=torch.int64,
+            )
+            decoder_out = model.run_decoder(decoder_input)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += symbol_table[i]
+        return text.replace("▁", " ").strip()
+
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = token_ids_to_words(hyp)
+        s += f"{filename}:\n{words}\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/optim.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/pretrained.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/pretrained.py
new file mode 100755
index 000000000..a22d1b4ba
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/pretrained.py
@@ -0,0 +1,355 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/en/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 5
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/en/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/en/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/en/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/en/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/scaling.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/scaling_converter.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
new file mode 100755
index 000000000..4aedeffe4
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
@@ -0,0 +1,1239 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import CommonVoiceAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/en/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    commonvoice = CommonVoiceAsrDataModule(args)
+
+    train_cuts = commonvoice.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = commonvoice.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = commonvoice.dev_cuts()
+    valid_dl = commonvoice.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/zipformer.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md
new file mode 100644
index 000000000..6c20bab2c
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md
@@ -0,0 +1,9 @@
+This recipe implements Streaming Zipformer-Transducer model.
+
+See https://k2-fsa.github.io/icefall/recipes/Streaming-ASR/librispeech/zipformer_transducer.html for detailed tutorials.
+
+[./emformer.py](./emformer.py) and [./train.py](./train.py)
+are basically the same as
+[./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py).
+The only purpose of [./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py)
+is for exporting to [sherpa-ncnn](https://github.com/k2-fsa/sherpa-ncnn).
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/beam_search.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/beam_search.py
new file mode 120000
index 000000000..d7349b0a3
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py
new file mode 100644
index 000000000..cafa4111d
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py
@@ -0,0 +1,422 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class CommonVoiceAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--language",
+            type=str,
+            default="fr",
+            help="""Language of Common Voice""",
+        )
+        group.add_argument(
+            "--cv-manifest-dir",
+            type=Path,
+            default=Path("data/fr/fbank"),
+            help="Path to directory with CommonVoice train/dev/test cuts.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_dev.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.cv_manifest_dir / f"cv-{self.args.language}_cuts_test.jsonl.gz"
+        )
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decode.py
new file mode 100755
index 000000000..30f7c1e77
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decode.py
@@ -0,0 +1,810 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from commonvoice_fr import CommonVoiceAsrDataModule
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += 30
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, 30),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+
+    # errs_info = (
+    #    params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    # )
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{key}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    commonvoice = CommonVoiceAsrDataModule(args)
+
+    test_cuts = commonvoice.test_cuts()
+
+    test_dl = commonvoice.test_dataloaders(test_cuts)
+
+    test_sets = "test-cv"
+
+    results_dict = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+        sp=sp,
+        word_table=word_table,
+        decoding_graph=decoding_graph,
+    )
+
+    save_results(
+        params=params,
+        test_set_name=test_sets,
+        results_dict=results_dict,
+    )
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decode_stream.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
new file mode 120000
index 000000000..ca8fed319
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decoder.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
new file mode 100755
index 000000000..0426bc9a3
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
@@ -0,0 +1,1258 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from commonvoice_fr import CommonVoiceAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer_for_ncnn_export_only import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+        is_pnnx=True,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    commonvoice = CommonVoiceAsrDataModule(args)
+
+    train_cuts = commonvoice.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = commonvoice.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = commonvoice.dev_cuts()
+    valid_dl = commonvoice.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
new file mode 120000
index 000000000..cb673b3eb
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
new file mode 120000
index 000000000..72e43c297
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
new file mode 120000
index 000000000..3b36924ef
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-onnx.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
new file mode 120000
index 000000000..57a0cd0a0
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export.py
new file mode 120000
index 000000000..2acafdc61
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py
new file mode 100755
index 000000000..3a10c5d81
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py
@@ -0,0 +1,1342 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from commonvoice_fr import CommonVoiceAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_finetune_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument("--do-finetune", type=str2bool, default=False)
+
+    parser.add_argument(
+        "--init-modules",
+        type=str,
+        default=None,
+        help="""
+        Modules to be initialized. It matches all parameters starting with
+        a specific key. The keys are given with Comma seperated. If None,
+        all modules will be initialised. For example, if you only want to
+        initialise all parameters staring with "encoder", use "encoder";
+        if you want to initialise parameters starting with encoder or decoder,
+        use "encoder,joiner".
+        """,
+    )
+
+    parser.add_argument(
+        "--finetune-ckpt",
+        type=str,
+        default=None,
+        help="Fine-tuning from which checkpoint (a path to a .pt file)",
+    )
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="""Embedding dimension in the 2 blocks of zipformer encoder
+        layers, comma separated
+        """,
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers,\
+        comma separated; not the same as embedding dimension.
+        """,
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="""Unmasked dimensions in the encoders, relates to augmentation
+        during training. Must be <= each of encoder_dims. Empirically, less
+        than 256 seems to make performance worse.
+        """,
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="""Path to the BPE model.
+        This should be the bpe model of the original model
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.005, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=100000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. During fine-tuning, we set this very large so that the
+        learning rate slowly decays with number of batches. You may tune
+        its value by yourself.
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=100,
+        help="""Number of epochs that affects how rapidly the learning rate
+        decreases. During fine-tuning, we set this very large so that the
+        learning rate slowly decays with number of batches. You may tune
+        its value by yourself.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+    add_finetune_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def load_model_params(
+    ckpt: str, model: nn.Module, init_modules: List[str] = None, strict: bool = True
+):
+    """Load model params from checkpoint
+
+    Args:
+        ckpt (str): Path to the checkpoint
+        model (nn.Module): model to be loaded
+
+    """
+    logging.info(f"Loading checkpoint from {ckpt}")
+    checkpoint = torch.load(ckpt, map_location="cpu")
+
+    # if module list is empty, load the whole model from ckpt
+    if not init_modules:
+        if next(iter(checkpoint["model"])).startswith("module."):
+            logging.info("Loading checkpoint saved by DDP")
+
+            dst_state_dict = model.state_dict()
+            src_state_dict = checkpoint["model"]
+            for key in dst_state_dict.keys():
+                src_key = "{}.{}".format("module", key)
+                dst_state_dict[key] = src_state_dict.pop(src_key)
+            assert len(src_state_dict) == 0
+            model.load_state_dict(dst_state_dict, strict=strict)
+        else:
+            model.load_state_dict(checkpoint["model"], strict=strict)
+    else:
+        src_state_dict = checkpoint["model"]
+        dst_state_dict = model.state_dict()
+        for module in init_modules:
+            logging.info(f"Loading parameters starting with prefix {module}")
+            src_keys = [k for k in src_state_dict.keys() if k.startswith(module)]
+            dst_keys = [k for k in dst_state_dict.keys() if k.startswith(module)]
+            assert set(src_keys) == set(dst_keys)  # two sets should match exactly
+            for key in src_keys:
+                dst_state_dict[key] = src_state_dict.pop(key)
+
+        model.load_state_dict(dst_state_dict, strict=strict)
+
+    return None
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it. The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have
+            # different behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    # load model parameters for model fine-tuning
+    if params.do_finetune:
+        modules = params.init_modules.split(",") if params.init_modules else None
+        checkpoints = load_model_params(
+            ckpt=params.finetune_ckpt, model=model, init_modules=modules
+        )
+    else:
+        assert params.start_epoch > 0, params.start_epoch
+        checkpoints = load_checkpoint_if_available(
+            params=params, model=model, model_avg=model_avg
+        )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    commonvoice = CommonVoiceAsrDataModule(args)
+
+    train_cuts = commonvoice.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = commonvoice.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = commonvoice.dev_cuts()
+    valid_dl = commonvoice.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(
+        parser
+    )  # you may replace this with your own dataset
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/generate_model_from_checkpoint.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/generate_model_from_checkpoint.py
new file mode 100755
index 000000000..3fd14aa47
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/generate_model_from_checkpoint.py
@@ -0,0 +1,281 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) use the averaged model with checkpoint exp_dir/epoch-xxx.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --epoch 28 \
+    --avg 15 \
+    --use-averaged-model True \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `epoch-28-avg-15-use-averaged-model.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15-use-averaged-model.pt")`.
+
+(2) use the averaged model with checkpoint exp_dir/checkpoint-iter.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --iter 22000 \
+    --avg 5 \
+    --use-averaged-model True \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `iter-22000-avg-5-use-averaged-model.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5-use-averaged-model.pt")`.
+
+(3) use the original model with checkpoint exp_dir/epoch-xxx.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --epoch 28 \
+    --avg 15 \
+    --use-averaged-model False \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `epoch-28-avg-15.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15.pt")`.
+
+(4) use the original model with checkpoint exp_dir/checkpoint-iter.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --iter 22000 \
+    --avg 5 \
+    --use-averaged-model False \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `iter-22000-avg-5.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5.pt")`.
+"""
+
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+import sentencepiece as spm
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model."
+        "If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    print("Script started")
+
+    device = torch.device("cpu")
+    print(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    print("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            print(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+            filename = params.exp_dir / f"iter-{params.iter}-avg-{params.avg}.pt"
+            torch.save({"model": model.state_dict()}, filename)
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+            filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+            torch.save({"model": model.state_dict()}, filename)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            print(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+            filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+            torch.save({"model": model.state_dict()}, filename)
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            print(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+            filename = (
+                params.exp_dir
+                / f"iter-{params.iter}-avg-{params.avg}-use-averaged-model.pt"
+            )
+            torch.save({"model": model.state_dict()}, filename)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            print(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+            filename = (
+                params.exp_dir
+                / f"epoch-{params.epoch}-avg-{params.avg}-use-averaged-model.pt"
+            )
+            torch.save({"model": model.state_dict()}, filename)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    print("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
new file mode 120000
index 000000000..5d9c6ba00
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
new file mode 120000
index 000000000..457131699
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
new file mode 120000
index 000000000..2b8fa3cbb
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/joiner.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/model.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/model.py
new file mode 120000
index 000000000..e17d4f734
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_check.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
new file mode 120000
index 000000000..28bf7bb82
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
new file mode 120000
index 000000000..c8548d459
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
new file mode 120000
index 000000000..ae4d9bb04
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/optim.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/pretrained.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/pretrained.py
new file mode 120000
index 000000000..9510b8fde
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/scaling.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
new file mode 120000
index 000000000..92c3904af
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
new file mode 120000
index 000000000..2adf271c1
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
new file mode 100755
index 000000000..dbe65d0a7
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
@@ -0,0 +1,612 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding_method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from commonvoice_fr import CommonVoiceAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    idx = 0
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+        audio: np.ndarray = cut.load_audio()
+        if audio.max() > 1 or audio.min() < -1:
+            audio = audio / max(abs(audio.max()), abs(audio.min()))
+            print(audio)
+            print(audio.max())
+            print(audio.min())
+            print(cut)
+            idx += 1
+            print(idx)
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    commonvoice = CommonVoiceAsrDataModule(args)
+    test_cuts = commonvoice.test_cuts()
+    test_sets = "test-cv"
+
+    results_dict = decode_dataset(
+        cuts=test_cuts,
+        params=params,
+        model=model,
+        sp=sp,
+        decoding_graph=decoding_graph,
+    )
+
+    save_results(
+        params=params,
+        test_set_name=test_sets,
+        results_dict=results_dict,
+    )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/test_model.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/test_model.py
new file mode 100755
index 000000000..5400df804
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/test_model.py
@@ -0,0 +1,150 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless7_streaming/test_model.py
+"""
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    # Test jit script
+    convert_scaled_to_non_scaled(model, inplace=True)
+    # We won't use the forward() method of the model in C++, so just ignore
+    # it here.
+    # Otherwise, one of its arguments is a ragged tensor and is not
+    # torch scriptabe.
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    print("Using torch.jit.script")
+    model = torch.jit.script(model)
+
+
+def test_model_jit_trace():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    # Test encoder
+    def _test_encoder():
+        encoder = model.encoder
+        assert encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+            encoder.decode_chunk_size,
+            params.decode_chunk_len,
+        )
+        T = params.decode_chunk_len + 7
+
+        x = torch.zeros(1, T, 80, dtype=torch.float32)
+        x_lens = torch.full((1,), T, dtype=torch.int32)
+        states = encoder.get_init_state(device=x.device)
+        encoder.__class__.forward = encoder.__class__.streaming_forward
+        traced_encoder = torch.jit.trace(encoder, (x, x_lens, states))
+
+        states1 = encoder.get_init_state(device=x.device)
+        states2 = traced_encoder.get_init_state(device=x.device)
+        for i in range(5):
+            x = torch.randn(1, T, 80, dtype=torch.float32)
+            x_lens = torch.full((1,), T, dtype=torch.int32)
+            y1, _, states1 = encoder.streaming_forward(x, x_lens, states1)
+            y2, _, states2 = traced_encoder(x, x_lens, states2)
+            assert torch.allclose(y1, y2, atol=1e-6), (i, (y1 - y2).abs().mean())
+
+    # Test decoder
+    def _test_decoder():
+        decoder = model.decoder
+        y = torch.zeros(10, decoder.context_size, dtype=torch.int64)
+        need_pad = torch.tensor([False])
+
+        traced_decoder = torch.jit.trace(decoder, (y, need_pad))
+        d1 = decoder(y, need_pad)
+        d2 = traced_decoder(y, need_pad)
+        assert torch.equal(d1, d2), (d1 - d2).abs().mean()
+
+    # Test joiner
+    def _test_joiner():
+        joiner = model.joiner
+        encoder_out_dim = joiner.encoder_proj.weight.shape[1]
+        decoder_out_dim = joiner.decoder_proj.weight.shape[1]
+        encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+        decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+        traced_joiner = torch.jit.trace(joiner, (encoder_out, decoder_out))
+        j1 = joiner(encoder_out, decoder_out)
+        j2 = traced_joiner(encoder_out, decoder_out)
+        assert torch.equal(j1, j2), (j1 - j2).abs().mean()
+
+    _test_encoder()
+    _test_decoder()
+    _test_joiner()
+
+
+def main():
+    test_model()
+    test_model_jit_trace()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py
new file mode 100755
index 000000000..a9bc9c2a2
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -0,0 +1,1256 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from commonvoice_fr import CommonVoiceAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/fr/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    commonvoice = CommonVoiceAsrDataModule(args)
+
+    train_cuts = commonvoice.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = commonvoice.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = commonvoice.dev_cuts()
+    valid_dl = commonvoice.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    CommonVoiceAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/zipformer.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/zipformer.py
new file mode 120000
index 000000000..ec183baa7
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
new file mode 120000
index 000000000..d301e1f9b
--- /dev/null
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
\ No newline at end of file
diff --git a/egs/commonvoice/ASR/shared b/egs/commonvoice/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/commonvoice/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/csj/ASR/.gitignore b/egs/csj/ASR/.gitignore
new file mode 100644
index 000000000..cd0e20c4c
--- /dev/null
+++ b/egs/csj/ASR/.gitignore
@@ -0,0 +1,8 @@
+librispeech_*
+todelete*
+lang*
+notify_tg.py
+finetune_*
+misc.ini
+.vscode/*
+offline/*
diff --git a/egs/csj/ASR/README.md b/egs/csj/ASR/README.md
new file mode 100644
index 000000000..95c2ec6ac
--- /dev/null
+++ b/egs/csj/ASR/README.md
@@ -0,0 +1,11 @@
+# Introduction
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+These are the types of architectures currently available.
+
+|                                       | Encoder             | Decoder            | Comment                                           |
+|---------------------------------------|---------------------|--------------------|---------------------------------------------------|
+| `pruned_transducer_stateless7_streaming` | Streaming Zipformer | Embedding + Conv1d | Adapted from librispeech pruned_transducer_stateless7_streaming |
diff --git a/egs/csj/ASR/RESULTS.md b/egs/csj/ASR/RESULTS.md
new file mode 100644
index 000000000..56fdb899f
--- /dev/null
+++ b/egs/csj/ASR/RESULTS.md
@@ -0,0 +1,200 @@
+# Results
+
+## Streaming Zipformer-Transducer (Pruned Stateless Transducer + Streaming Zipformer)
+
+### [pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
+
+See <https://github.com/k2-fsa/icefall/pull/892> for more details.
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/TeoWenShen/icefall-asr-csj-pruned-transducer-stateless7-streaming-230208>
+
+Number of model parameters: 75688409, i.e. 75.7M.
+
+#### training on disfluent transcript
+
+The CERs are:
+
+| decoding method | chunk size | eval1 | eval2 | eval3 | excluded | valid | average | decoding mode |
+| --------------- | ---------- | ----- | ----- | ----- | -------- | ----- | ------- | ------------- |
+| fast beam search | 320ms | 5.39 | 4.08 | 4.16 | 5.4 | 5.02 | --epoch 30 --avg 17 | simulated streaming |
+| fast beam search | 320ms | 5.34 | 4.1 | 4.26 | 5.61 | 4.91 | --epoch 30 --avg 17 | chunk-wise |
+| greedy search | 320ms | 5.43 | 4.14 | 4.31 | 5.48 | 4.88 | --epoch 30 --avg 17 | simulated streaming |
+| greedy search | 320ms | 5.44 | 4.14 | 4.39 | 5.7 | 4.98 | --epoch 30 --avg 17 | chunk-wise |
+| modified beam search | 320ms | 5.2 | 3.95 | 4.09 | 5.12 | 4.75 | --epoch 30 --avg 17 | simulated streaming |
+| modified beam search | 320ms | 5.18 | 4.07 | 4.12 | 5.36 | 4.77 | --epoch 30 --avg 17 | chunk-wise |
+| fast beam search | 640ms | 5.01 | 3.78 | 3.96 | 4.85 | 4.6 | --epoch 30 --avg 17 | simulated streaming |
+| fast beam search | 640ms | 4.97 | 3.88 | 3.96 | 4.91 | 4.61 | --epoch 30 --avg 17 | chunk-wise |
+| greedy search | 640ms | 5.02 | 3.84 | 4.14 | 5.02 | 4.59 | --epoch 30 --avg 17 | simulated streaming |
+| greedy search | 640ms | 5.32 | 4.22 | 4.33 | 5.39 | 4.99 | --epoch 30 --avg 17 | chunk-wise |
+| modified beam search | 640ms | 4.78 | 3.66 | 3.85 | 4.72 | 4.42 | --epoch 30 --avg 17 | simulated streaming |
+| modified beam search | 640ms | 5.77 | 4.72 | 4.73 | 5.85 | 5.36 | --epoch 30 --avg 17 | chunk-wise |
+
+Note: `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
+while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
+
+The training command was:
+```bash
+./pruned_transducer_stateless7_streaming/train.py \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30 \
+  --max-duration 375 \
+  --transcript-mode disfluent \
+  --lang data/lang_char \
+  --manifest-dir /mnt/host/corpus/csj/fbank \
+  --pad-feature 30 \
+  --musan-dir /mnt/host/corpus/musan/musan/fbank
+```
+
+The simulated streaming decoding command was:
+```bash
+for chunk in 64 32; do
+    for m in greedy_search fast_beam_search modified_beam_search; do
+        python pruned_transducer_stateless7_streaming/decode.py \
+            --feedforward-dims  "1024,1024,2048,2048,1024" \
+            --exp-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30 \
+            --epoch 30 \
+            --avg 17 \
+            --max-duration 350 \
+            --decoding-method $m \
+            --manifest-dir /mnt/host/corpus/csj/fbank \
+            --lang data/lang_char \
+            --transcript-mode disfluent \
+            --res-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30/github/sim_"$chunk"_"$m" \
+            --decode-chunk-len $chunk \
+            --pad-feature 30 \
+            --gpu 0
+    done
+done
+```
+
+The streaming chunk-wise decoding command was:
+```bash
+for chunk in 64 32; do
+    for m in greedy_search fast_beam_search modified_beam_search; do
+        python pruned_transducer_stateless7_streaming/streaming_decode.py \
+            --feedforward-dims  "1024,1024,2048,2048,1024" \
+            --exp-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30 \
+            --epoch 30 \
+            --avg 17 \
+            --max-duration 350 \
+            --decoding-method $m \
+            --manifest-dir /mnt/host/corpus/csj/fbank \
+            --lang data/lang_char \
+            --transcript-mode disfluent \
+            --res-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30/github/stream_"$chunk"_"$m" \
+            --decode-chunk-len $chunk \
+            --gpu 2 \
+            --num-decode-streams 40
+    done
+done
+```
+
+#### training on fluent transcript
+
+The CERs are:
+
+| decoding method | chunk size | eval1 | eval2 | eval3 | excluded | valid | average | decoding mode |
+| --------------- | ---------- | ----- | ----- | ----- | -------- | ----- | ------- | ------------- |
+| fast beam search | 320ms | 4.19 | 3.63 | 3.77 | 4.43 | 4.09 | --epoch 30 --avg 12 | simulated streaming |
+| fast beam search | 320ms | 4.06 | 3.55 | 3.66 | 4.70 | 4.04 | --epoch 30 --avg 12 | chunk-wise |
+| greedy search | 320ms | 4.22 | 3.62 | 3.82 | 4.45 | 3.98 | --epoch 30 --avg 12 | simulated streaming |
+| greedy search | 320ms | 4.13 | 3.61 | 3.85 | 4.67 | 4.05 | --epoch 30 --avg 12 | chunk-wise |
+| modified beam search | 320ms | 4.02 | 3.43 | 3.62 | 4.43 | 3.81 | --epoch 30 --avg 12 | simulated streaming |
+| modified beam search | 320ms | 3.97 | 3.43 | 3.59 | 4.99 | 3.88 | --epoch 30 --avg 12 | chunk-wise |
+| fast beam search | 640ms | 3.80 | 3.31 | 3.55 | 4.16 | 3.90 | --epoch 30 --avg 12 | simulated streaming |
+| fast beam search | 640ms | 3.81 | 3.34 | 3.46 | 4.58 | 3.85 | --epoch 30 --avg 12 | chunk-wise |
+| greedy search | 640ms | 3.92 | 3.38 | 3.65 | 4.31 | 3.88 | --epoch 30 --avg 12 | simulated streaming |
+| greedy search | 640ms | 3.98 | 3.38 | 3.64 | 4.54 | 4.01 | --epoch 30 --avg 12 | chunk-wise |
+| modified beam search | 640ms | 3.72 | 3.26 | 3.39 | 4.10 | 3.65 | --epoch 30 --avg 12 | simulated streaming |
+| modified beam search | 640ms | 3.78 | 3.32 | 3.45 | 4.81 | 3.81 | --epoch 30 --avg 12 | chunk-wise |
+
+Note: `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
+while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
+
+The training command was:
+```bash
+./pruned_transducer_stateless7_streaming/train.py \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30 \
+  --max-duration 375 \
+  --transcript-mode fluent \
+  --lang data/lang_char \
+  --manifest-dir /mnt/host/corpus/csj/fbank \
+  --pad-feature 30 \
+  --musan-dir /mnt/host/corpus/musan/musan/fbank
+```
+
+The simulated streaming decoding command was:
+```bash
+for chunk in 64 32; do
+    for m in greedy_search fast_beam_search modified_beam_search; do
+        python pruned_transducer_stateless7_streaming/decode.py \
+            --feedforward-dims  "1024,1024,2048,2048,1024" \
+            --exp-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30 \
+            --epoch 30 \
+            --avg 12 \
+            --max-duration 350 \
+            --decoding-method $m \
+            --manifest-dir /mnt/host/corpus/csj/fbank \
+            --lang data/lang_char \
+            --transcript-mode fluent \
+            --res-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30/github/sim_"$chunk"_"$m" \
+            --decode-chunk-len $chunk \
+            --pad-feature 30 \
+            --gpu 1
+    done
+done
+```
+
+The streaming chunk-wise decoding command was:
+```bash
+for chunk in 64 32; do
+    for m in greedy_search fast_beam_search modified_beam_search; do
+        python pruned_transducer_stateless7_streaming/streaming_decode.py \
+            --feedforward-dims  "1024,1024,2048,2048,1024" \
+            --exp-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30 \
+            --epoch 30 \
+            --avg 12 \
+            --max-duration 350 \
+            --decoding-method $m \
+            --manifest-dir /mnt/host/corpus/csj/fbank \
+            --lang data/lang_char \
+            --transcript-mode fluent \
+            --res-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30/github/stream_"$chunk"_"$m" \
+            --decode-chunk-len $chunk \
+            --gpu 3 \
+            --num-decode-streams 40
+    done
+done
+```
+
+#### Comparing disfluent to fluent
+
+$$ \texttt{CER}^{f}_d = \frac{\texttt{sub}_f + \texttt{ins} + \texttt{del}_f}{N_f} $$
+
+This comparison evaluates the disfluent model on the fluent transcript (calculated by `disfluent_recogs_to_fluent.py`), forgiving the disfluent model's mistakes on fillers and partial words. It is meant as an illustrative metric only, so that the disfluent and fluent models can be compared.
+
+| decoding method | chunk size | eval1 (d vs f) | eval2  (d vs f) | eval3  (d vs f) | excluded  (d vs f) | valid  (d vs f) | decoding mode |
+| --------------- | ---------- | -------------- | --------------- | -------------- | -------------------- | --------------- | ----------- |
+| fast beam search | 320ms | 4.54 vs 4.19 | 3.44 vs 3.63 | 3.56 vs 3.77 | 4.22 vs 4.43 | 4.22 vs 4.09 | simulated streaming |
+| fast beam search | 320ms | 4.48 vs 4.06 | 3.41 vs 3.55 | 3.65 vs 3.66 | 4.26 vs 4.7 | 4.08 vs 4.04 | chunk-wise |
+| greedy search | 320ms | 4.53 vs 4.22 | 3.48 vs 3.62 | 3.69 vs 3.82 | 4.38 vs 4.45 | 4.05 vs 3.98 | simulated streaming |
+| greedy search | 320ms | 4.53 vs 4.13 | 3.46 vs 3.61 | 3.71 vs 3.85 | 4.48 vs 4.67 | 4.12 vs 4.05 | chunk-wise |
+| modified beam search | 320ms | 4.45 vs 4.02 | 3.38 vs 3.43 | 3.57 vs 3.62 | 4.19 vs 4.43 | 4.04 vs 3.81 | simulated streaming |
+| modified beam search | 320ms | 4.44 vs 3.97 | 3.47 vs 3.43 | 3.56 vs 3.59 | 4.28 vs 4.99 | 4.04 vs 3.88 | chunk-wise |
+| fast beam search | 640ms | 4.14 vs 3.8 | 3.12 vs 3.31 | 3.38 vs 3.55 | 3.72 vs 4.16 | 3.81 vs 3.9 | simulated streaming |
+| fast beam search | 640ms | 4.05 vs 3.81 | 3.23 vs 3.34 | 3.36 vs 3.46 | 3.65 vs 4.58 | 3.78 vs 3.85 | chunk-wise |
+| greedy search | 640ms | 4.1 vs 3.92 | 3.17 vs 3.38 | 3.5 vs 3.65 | 3.87 vs 4.31 | 3.77 vs 3.88 | simulated streaming |
+| greedy search | 640ms | 4.41 vs 3.98 | 3.56 vs 3.38 | 3.69 vs 3.64 | 4.26 vs 4.54 | 4.16 vs 4.01 | chunk-wise |
+| modified beam search | 640ms | 4 vs 3.72 | 3.08 vs 3.26 | 3.33 vs 3.39 | 3.75 vs 4.1 | 3.71 vs 3.65 | simulated streaming |
+| modified beam search | 640ms | 5.05 vs 3.78 | 4.22 vs 3.32 | 4.26 vs 3.45 | 5.02 vs 4.81 | 4.73 vs 3.81 | chunk-wise |
+| average (d - f) |  | 0.43 | -0.02 | -0.02 | -0.34 | 0.13 |  |
diff --git a/egs/csj/ASR/local/add_transcript_mode.py b/egs/csj/ASR/local/add_transcript_mode.py
new file mode 100644
index 000000000..f6b4b2caf
--- /dev/null
+++ b/egs/csj/ASR/local/add_transcript_mode.py
@@ -0,0 +1,94 @@
+import argparse
+import logging
+from configparser import ConfigParser
+from pathlib import Path
+from typing import List
+
+from lhotse import CutSet, SupervisionSet
+from lhotse.recipes.csj import CSJSDBParser
+
+ARGPARSE_DESCRIPTION = """
+This script adds transcript modes to an existing CutSet or SupervisionSet.
+"""
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+        description=ARGPARSE_DESCRIPTION,
+    )
+    parser.add_argument(
+        "-f",
+        "--fbank-dir",
+        type=Path,
+        help="Path to directory where manifests are stored.",
+    )
+    parser.add_argument(
+        "-c",
+        "--config",
+        type=Path,
+        nargs="+",
+        help="Path to config file for transcript parsing.",
+    )
+    return parser.parse_args()
+
+
+def get_CSJParsers(config_files: List[Path]) -> List[CSJSDBParser]:
+    parsers = []
+    for config_file in config_files:
+        config = ConfigParser()
+        config.optionxform = str
+        assert config.read(config_file), f"{config_file} could not be found."
+        decisions = {}
+        for k, v in config["DECISIONS"].items():
+            try:
+                decisions[k] = int(v)
+            except ValueError:
+                decisions[k] = v
+        parsers.append(
+            (config["CONSTANTS"].get("MODE"), CSJSDBParser(decisions=decisions))
+        )
+    return parsers
+
+
+def main():
+    args = get_args()
+    logging.basicConfig(
+        format=("%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"),
+        level=logging.INFO,
+    )
+    parsers = get_CSJParsers(args.config)
+    config = ConfigParser()
+    config.optionxform = str
+    assert config.read(args.config), args.config
+    decisions = {}
+    for k, v in config["DECISIONS"].items():
+        try:
+            decisions[k] = int(v)
+        except ValueError:
+            decisions[k] = v
+
+    logging.info(f"Adding {', '.join(x[0] for x in parsers)} transcript mode.")
+
+    manifests = args.fbank_dir.glob("csj_cuts_*.jsonl.gz")
+    assert manifests, f"No cuts to be found in {args.fbank_dir}"
+
+    for manifest in manifests:
+        results = []
+        logging.info(f"Adding transcript modes to {manifest.name} now.")
+        cutset = CutSet.from_file(manifest)
+        for cut in cutset:
+            for name, parser in parsers:
+                cut.supervisions[0].custom[name] = parser.parse(
+                    cut.supervisions[0].custom["raw"]
+                )
+            cut.supervisions[0].text = ""
+            results.append(cut)
+        results = CutSet.from_items(results)
+        res_file = manifest.as_posix()
+        manifest.replace(manifest.parent / ("bak." + manifest.name))
+        results.to_file(res_file)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/local/compute_fbank_csj.py b/egs/csj/ASR/local/compute_fbank_csj.py
new file mode 100644
index 000000000..ce560025d
--- /dev/null
+++ b/egs/csj/ASR/local/compute_fbank_csj.py
@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+# Copyright    2023  The University of Electro-Communications  (Author: Teo Wen Shen)  # noqa
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import List, Tuple
+
+import torch
+
+# fmt: off
+from lhotse import (  # See the following for why LilcomChunkyWriter is preferred; https://github.com/k2-fsa/icefall/pull/404; https://github.com/lhotse-speech/lhotse/pull/527
+    CutSet,
+    Fbank,
+    FbankConfig,
+    LilcomChunkyWriter,
+    RecordingSet,
+    SupervisionSet,
+)
+from lhotse.recipes.csj import concat_csj_supervisions
+
+# fmt: on
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+RNG_SEED = 42
+# concat_params_train = [
+#     {"gap": 1.0, "maxlen": 10.0},
+#     {"gap": 1.5, "maxlen": 8.0},
+#     {"gap": 1.0, "maxlen": 18.0},
+# ]
+
+concat_params = {"gap": 1.0, "maxlen": 10.0}
+
+
+def make_cutset_blueprints(
+    manifest_dir: Path,
+) -> List[Tuple[str, CutSet]]:
+
+    cut_sets = []
+    logging.info("Creating non-train cuts.")
+
+    # Create eval datasets
+    for i in range(1, 4):
+        sps = sorted(
+            SupervisionSet.from_file(
+                manifest_dir / f"csj_supervisions_eval{i}.jsonl.gz"
+            ),
+            key=lambda x: x.id,
+        )
+
+        cut_set = CutSet.from_manifests(
+            recordings=RecordingSet.from_file(
+                manifest_dir / f"csj_recordings_eval{i}.jsonl.gz"
+            ),
+            supervisions=concat_csj_supervisions(sps, **concat_params),
+        )
+        cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+        cut_sets.append((f"eval{i}", cut_set))
+
+    # Create excluded dataset
+    sps = sorted(
+        SupervisionSet.from_file(manifest_dir / "csj_supervisions_excluded.jsonl.gz"),
+        key=lambda x: x.id,
+    )
+    cut_set = CutSet.from_manifests(
+        recordings=RecordingSet.from_file(
+            manifest_dir / "csj_recordings_excluded.jsonl.gz"
+        ),
+        supervisions=concat_csj_supervisions(sps, **concat_params),
+    )
+    cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+    cut_sets.append(("excluded", cut_set))
+
+    # Create valid dataset
+    sps = sorted(
+        SupervisionSet.from_file(manifest_dir / "csj_supervisions_valid.jsonl.gz"),
+        key=lambda x: x.id,
+    )
+    cut_set = CutSet.from_manifests(
+        recordings=RecordingSet.from_file(
+            manifest_dir / "csj_recordings_valid.jsonl.gz"
+        ),
+        supervisions=concat_csj_supervisions(sps, **concat_params),
+    )
+    cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+    cut_sets.append(("valid", cut_set))
+
+    logging.info("Creating train cuts.")
+
+    # Create train dataset
+    sps = sorted(
+        SupervisionSet.from_file(manifest_dir / "csj_supervisions_core.jsonl.gz")
+        + SupervisionSet.from_file(manifest_dir / "csj_supervisions_noncore.jsonl.gz"),
+        key=lambda x: x.id,
+    )
+
+    recording = RecordingSet.from_file(
+        manifest_dir / "csj_recordings_core.jsonl.gz"
+    ) + RecordingSet.from_file(manifest_dir / "csj_recordings_noncore.jsonl.gz")
+
+    train_set = CutSet.from_manifests(
+        recordings=recording, supervisions=concat_csj_supervisions(sps, **concat_params)
+    ).trim_to_supervisions(keep_overlapping=False)
+    train_set = train_set + train_set.perturb_speed(0.9) + train_set.perturb_speed(1.1)
+
+    cut_sets.append(("train", train_set))
+
+    return cut_sets
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "-m", "--manifest-dir", type=Path, help="Path to save manifests"
+    )
+    parser.add_argument(
+        "-f", "--fbank-dir", type=Path, help="Path to save fbank features"
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    extractor = Fbank(FbankConfig(num_mel_bins=80))
+    num_jobs = min(16, os.cpu_count())
+
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    if (args.fbank_dir / ".done").exists():
+        logging.info(
+            "Previous fbank computed for CSJ found. "
+            f"Delete {args.fbank_dir / '.done'} to allow recomputing fbank."
+        )
+        return
+    else:
+        cut_sets = make_cutset_blueprints(args.manifest_dir)
+        for part, cut_set in cut_sets:
+            logging.info(f"Processing {part}")
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                num_jobs=num_jobs,
+                storage_path=(args.fbank_dir / f"feats_{part}").as_posix(),
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(args.fbank_dir / f"csj_cuts_{part}.jsonl.gz")
+
+        logging.info("All fbank computed for CSJ.")
+        (args.fbank_dir / ".done").touch()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/local/compute_fbank_musan.py b/egs/csj/ASR/local/compute_fbank_musan.py
new file mode 100644
index 000000000..c942df98e
--- /dev/null
+++ b/egs/csj/ASR/local/compute_fbank_musan.py
@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter, combine
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+ARGPARSE_DESCRIPTION = """
+This file computes fbank features of the musan dataset.
+
+"""
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan(manifest_dir: Path, fbank_dir: Path):
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=manifest_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    musan_cuts_path = fbank_dir / "musan_cuts.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        # create chunks of Musan with duration 5 - 10 seconds
+        musan_cuts = (
+            CutSet.from_manifests(
+                recordings=combine(part["recordings"] for part in manifests.values())
+            )
+            .cut_into_windows(10.0)
+            .filter(lambda c: c.duration > 5)
+            .compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{fbank_dir}/musan_feats",
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        musan_cuts.to_file(musan_cuts_path)
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        description=ARGPARSE_DESCRIPTION,
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+
+    parser.add_argument(
+        "-m", "--manifest-dir", type=Path, help="Path to save manifests"
+    )
+    parser.add_argument(
+        "-f", "--fbank-dir", type=Path, help="Path to save fbank features"
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = get_args()
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan(args.manifest_dir, args.fbank_dir)
diff --git a/egs/csj/ASR/local/conf/disfluent.ini b/egs/csj/ASR/local/conf/disfluent.ini
new file mode 100644
index 000000000..4f0a9ec0e
--- /dev/null
+++ b/egs/csj/ASR/local/conf/disfluent.ini
@@ -0,0 +1,79 @@
+[CONSTANTS]
+; # Name of this mode
+MODE = disfluent
+
+[DECISIONS]
+; # フィラー、感情表出系感動詞
+; # 0 to remain, 1 to delete
+; # Example: '(F ぎょっ)'
+F = 0
+; # 言い直し、いいよどみなどによる語断片
+; # 0 to remain, 1 to delete
+; # Example: '(D だ)(D だいが) 大学の学部の会議'
+D = 0
+; # 助詞、助動詞、接辞の言い直し
+; # 0 to remain, 1 to delete
+; # Example: '西洋 (D2 的)(F えー)(D ふ) 風というか'
+D2 = 0
+; # 聞き取りや語彙の判断に自信がない場合
+; # 0 to remain, 1 to delete
+; # Example: (? 字数) の
+; # If no option: empty string is returned regardless of output
+; # Example: '(?) で'
+? = 0
+; # タグ?で、値は複数の候補が想定される場合
+; # 0 for main guess with matching morph info, 1 for second guess
+; # Example:  '(? 次数, 実数)', '(? これ,ここで)＋(? 説明＋し＋た＋方＋が＋いい＋か＋な)'
+?, = 0
+; # 音や言葉に関するメタ的な引用
+; # 0 to remain, 1 to delete
+; # Example: '助詞の (M は) は (M は) と書くが発音は (M わ)'
+M = 0
+; # 外国語や古語、方言など
+; # 0 to remain, 1 to delete
+; # Example: '(O ザッツファイン)'
+O = 0
+; # 講演者の名前、差別語、誹謗中傷など
+; # 0 to remain, 1 to delete
+; # Example: '国語研の (R ××) です'
+R = 0
+; # 非朗読対象発話（朗読における言い間違い等）
+; # 0 to remain, 1 to delete
+; # Example: '(X 実際は) 実際には'
+X = 0
+; # アルファベットや算用数字、記号の表記
+; # 0 to use Japanese form, 1 to use alphabet form
+; # Example: '(A シーディーアール;ＣＤ－Ｒ)'
+A = 1
+; # タグAで、単語は算用数字の場合
+; # 0 to use Japanese form, 1 to use Arabic numerals
+; # Example: (A 二千;２０００)
+A_num = 0
+; # 何らかの原因で漢字表記できなくなった場合
+; # 0 to use broken form, 1 to use orthodox form
+; # Example: '(K たち (F えー) ばな;橘)'
+K = 1
+; # 転訛、発音の怠けなど、一時的な発音エラー
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(W ギーツ;ギジュツ)'
+W = 1
+; # 語の読みに関する知識レベルのいい間違い
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(B シブタイ;ジュータイ)'
+B = 0
+; # 笑いながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(笑 ナニガ)', '(笑 (F エー)＋ソー＋イッ＋タ＋ヨー＋ナ)'
+笑 = 0
+; # 泣きながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(泣 ドンナニ)'
+泣 = 0
+; # 咳をしながら発話
+; # 0 to remain, 1 to delete
+; # Example: 'シャ(咳 リン) ノ'
+咳 = 0
+; # ささやき声や独り言などの小さな声
+; # 0 to remain, 1 to delete
+; # Example: '(L アレコレナンダッケ)', '(L (W コデ;(? コレ,ココデ))＋(? セツメー＋シ＋タ＋ホー＋ガ＋イー＋カ＋ナ))'
+L = 0
diff --git a/egs/csj/ASR/local/conf/fluent.ini b/egs/csj/ASR/local/conf/fluent.ini
new file mode 100644
index 000000000..5d033ed17
--- /dev/null
+++ b/egs/csj/ASR/local/conf/fluent.ini
@@ -0,0 +1,79 @@
+[CONSTANTS]
+; # Name of this mode
+MODE = fluent
+
+[DECISIONS]
+; # フィラー、感情表出系感動詞
+; # 0 to remain, 1 to delete
+; # Example: '(F ぎょっ)'
+F = 1
+; # 言い直し、いいよどみなどによる語断片
+; # 0 to remain, 1 to delete
+; # Example: '(D だ)(D だいが) 大学の学部の会議'
+D = 1
+; # 助詞、助動詞、接辞の言い直し
+; # 0 to remain, 1 to delete
+; # Example: '西洋 (D2 的)(F えー)(D ふ) 風というか'
+D2 = 1
+; # 聞き取りや語彙の判断に自信がない場合
+; # 0 to remain, 1 to delete
+; # Example: (? 字数) の
+; # If no option: empty string is returned regardless of output
+; # Example: '(?) で'
+? = 0
+; # タグ?で、値は複数の候補が想定される場合
+; # 0 for main guess with matching morph info, 1 for second guess
+; # Example:  '(? 次数, 実数)', '(? これ,ここで)＋(? 説明＋し＋た＋方＋が＋いい＋か＋な)'
+?, = 0
+; # 音や言葉に関するメタ的な引用
+; # 0 to remain, 1 to delete
+; # Example: '助詞の (M は) は (M は) と書くが発音は (M わ)'
+M = 0
+; # 外国語や古語、方言など
+; # 0 to remain, 1 to delete
+; # Example: '(O ザッツファイン)'
+O = 0
+; # 講演者の名前、差別語、誹謗中傷など
+; # 0 to remain, 1 to delete
+; # Example: '国語研の (R ××) です'
+R = 0
+; # 非朗読対象発話（朗読における言い間違い等）
+; # 0 to remain, 1 to delete
+; # Example: '(X 実際は) 実際には'
+X = 0
+; # アルファベットや算用数字、記号の表記
+; # 0 to use Japanese form, 1 to use alphabet form
+; # Example: '(A シーディーアール;ＣＤ－Ｒ)'
+A = 1
+; # タグAで、単語は算用数字の場合
+; # 0 to use Japanese form, 1 to use Arabic numerals
+; # Example: (A 二千;２０００)
+A_num = 0
+; # 何らかの原因で漢字表記できなくなった場合
+; # 0 to use broken form, 1 to use orthodox form
+; # Example: '(K たち (F えー) ばな;橘)'
+K = 1
+; # 転訛、発音の怠けなど、一時的な発音エラー
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(W ギーツ;ギジュツ)'
+W = 1
+; # 語の読みに関する知識レベルのいい間違い
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(B シブタイ;ジュータイ)'
+B = 0
+; # 笑いながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(笑 ナニガ)', '(笑 (F エー)＋ソー＋イッ＋タ＋ヨー＋ナ)'
+笑 = 0
+; # 泣きながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(泣 ドンナニ)'
+泣 = 0
+; # 咳をしながら発話
+; # 0 to remain, 1 to delete
+; # Example: 'シャ(咳 リン) ノ'
+咳 = 0
+; # ささやき声や独り言などの小さな声
+; # 0 to remain, 1 to delete
+; # Example: '(L アレコレナンダッケ)', '(L (W コデ;(? コレ,ココデ))＋(? セツメー＋シ＋タ＋ホー＋ガ＋イー＋カ＋ナ))'
+L = 0
diff --git a/egs/csj/ASR/local/conf/number.ini b/egs/csj/ASR/local/conf/number.ini
new file mode 100644
index 000000000..3ada9aa24
--- /dev/null
+++ b/egs/csj/ASR/local/conf/number.ini
@@ -0,0 +1,79 @@
+[CONSTANTS]
+; # Name of this mode
+MODE = number
+
+[DECISIONS]
+; # フィラー、感情表出系感動詞
+; # 0 to remain, 1 to delete
+; # Example: '(F ぎょっ)'
+F = 1
+; # 言い直し、いいよどみなどによる語断片
+; # 0 to remain, 1 to delete
+; # Example: '(D だ)(D だいが) 大学の学部の会議'
+D = 1
+; # 助詞、助動詞、接辞の言い直し
+; # 0 to remain, 1 to delete
+; # Example: '西洋 (D2 的)(F えー)(D ふ) 風というか'
+D2 = 1
+; # 聞き取りや語彙の判断に自信がない場合
+; # 0 to remain, 1 to delete
+; # Example: (? 字数) の
+; # If no option: empty string is returned regardless of output
+; # Example: '(?) で'
+? = 0
+; # タグ?で、値は複数の候補が想定される場合
+; # 0 for main guess with matching morph info, 1 for second guess
+; # Example:  '(? 次数, 実数)', '(? これ,ここで)＋(? 説明＋し＋た＋方＋が＋いい＋か＋な)'
+?, = 0
+; # 音や言葉に関するメタ的な引用
+; # 0 to remain, 1 to delete
+; # Example: '助詞の (M は) は (M は) と書くが発音は (M わ)'
+M = 0
+; # 外国語や古語、方言など
+; # 0 to remain, 1 to delete
+; # Example: '(O ザッツファイン)'
+O = 0
+; # 講演者の名前、差別語、誹謗中傷など
+; # 0 to remain, 1 to delete
+; # Example: '国語研の (R ××) です'
+R = 0
+; # 非朗読対象発話（朗読における言い間違い等）
+; # 0 to remain, 1 to delete
+; # Example: '(X 実際は) 実際には'
+X = 0
+; # アルファベットや算用数字、記号の表記
+; # 0 to use Japanese form, 1 to use alphabet form
+; # Example: '(A シーディーアール;ＣＤ－Ｒ)'
+A = 1
+; # タグAで、単語は算用数字の場合
+; # 0 to use Japanese form, 1 to use Arabic numerals
+; # Example: (A 二千;２０００)
+A_num = 1
+; # 何らかの原因で漢字表記できなくなった場合
+; # 0 to use broken form, 1 to use orthodox form
+; # Example: '(K たち (F えー) ばな;橘)'
+K = 1
+; # 転訛、発音の怠けなど、一時的な発音エラー
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(W ギーツ;ギジュツ)'
+W = 1
+; # 語の読みに関する知識レベルのいい間違い
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(B シブタイ;ジュータイ)'
+B = 0
+; # 笑いながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(笑 ナニガ)', '(笑 (F エー)＋ソー＋イッ＋タ＋ヨー＋ナ)'
+笑 = 0
+; # 泣きながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(泣 ドンナニ)'
+泣 = 0
+; # 咳をしながら発話
+; # 0 to remain, 1 to delete
+; # Example: 'シャ(咳 リン) ノ'
+咳 = 0
+; # ささやき声や独り言などの小さな声
+; # 0 to remain, 1 to delete
+; # Example: '(L アレコレナンダッケ)', '(L (W コデ;(? コレ,ココデ))＋(? セツメー＋シ＋タ＋ホー＋ガ＋イー＋カ＋ナ))'
+L = 0
diff --git a/egs/csj/ASR/local/conf/symbol.ini b/egs/csj/ASR/local/conf/symbol.ini
new file mode 100644
index 000000000..dafd65c9a
--- /dev/null
+++ b/egs/csj/ASR/local/conf/symbol.ini
@@ -0,0 +1,80 @@
+[CONSTANTS]
+; # Name of this mode
+; # From https://www.isca-speech.org/archive/pdfs/interspeech_2022/horii22_interspeech.pdf
+MODE = symbol
+
+[DECISIONS]
+; # フィラー、感情表出系感動詞
+; # 0 to remain, 1 to delete
+; # Example: '(F ぎょっ)'
+F = "＃", ["F"]
+; # 言い直し、いいよどみなどによる語断片
+; # 0 to remain, 1 to delete
+; # Example: '(D だ)(D だいが) 大学の学部の会議'
+D = "＠", ["D"]
+; # 助詞、助動詞、接辞の言い直し
+; # 0 to remain, 1 to delete
+; # Example: '西洋 (D2 的)(F えー)(D ふ) 風というか'
+D2 = "＠", ["D2"]
+; # 聞き取りや語彙の判断に自信がない場合
+; # 0 to remain, 1 to delete
+; # Example: (? 字数) の
+; # If no option: empty string is returned regardless of output
+; # Example: '(?) で'
+? = 0
+; # タグ?で、値は複数の候補が想定される場合
+; # 0 for main guess with matching morph info, 1 for second guess
+; # Example:  '(? 次数, 実数)', '(? これ,ここで)＋(? 説明＋し＋た＋方＋が＋いい＋か＋な)'
+?, = 0
+; # 音や言葉に関するメタ的な引用
+; # 0 to remain, 1 to delete
+; # Example: '助詞の (M は) は (M は) と書くが発音は (M わ)'
+M = 0
+; # 外国語や古語、方言など
+; # 0 to remain, 1 to delete
+; # Example: '(O ザッツファイン)'
+O = 0
+; # 講演者の名前、差別語、誹謗中傷など
+; # 0 to remain, 1 to delete
+; # Example: '国語研の (R ××) です'
+R = 0
+; # 非朗読対象発話（朗読における言い間違い等）
+; # 0 to remain, 1 to delete
+; # Example: '(X 実際は) 実際には'
+X = 0
+; # アルファベットや算用数字、記号の表記
+; # 0 to use Japanese form, 1 to use alphabet form
+; # Example: '(A シーディーアール;ＣＤ－Ｒ)'
+A = 1
+; # タグAで、単語は算用数字の場合
+; # 0 to use Japanese form, 1 to use Arabic numerals
+; # Example: (A 二千;２０００)
+A_num = 1
+; # 何らかの原因で漢字表記できなくなった場合
+; # 0 to use broken form, 1 to use orthodox form
+; # Example: '(K たち (F えー) ばな;橘)'
+K = 1
+; # 転訛、発音の怠けなど、一時的な発音エラー
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(W ギーツ;ギジュツ)'
+W = 1
+; # 語の読みに関する知識レベルのいい間違い
+; # 0 to use wrong form, 1 to use orthodox form
+; # Example: '(B シブタイ;ジュータイ)'
+B = 0
+; # 笑いながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(笑 ナニガ)', '(笑 (F エー)＋ソー＋イッ＋タ＋ヨー＋ナ)'
+笑 = 0
+; # 泣きながら発話
+; # 0 to remain, 1 to delete
+; # Example: '(泣 ドンナニ)'
+泣 = 0
+; # 咳をしながら発話
+; # 0 to remain, 1 to delete
+; # Example: 'シャ(咳 リン) ノ'
+咳 = 0
+; # ささやき声や独り言などの小さな声
+; # 0 to remain, 1 to delete
+; # Example: '(L アレコレナンダッケ)', '(L (W コデ;(? コレ,ココデ))＋(? セツメー＋シ＋タ＋ホー＋ガ＋イー＋カ＋ナ))'
+L = 0
diff --git a/egs/csj/ASR/local/disfluent_recogs_to_fluent.py b/egs/csj/ASR/local/disfluent_recogs_to_fluent.py
new file mode 100644
index 000000000..45c9c7656
--- /dev/null
+++ b/egs/csj/ASR/local/disfluent_recogs_to_fluent.py
@@ -0,0 +1,202 @@
+import argparse
+from pathlib import Path
+
+import kaldialign
+from lhotse import CutSet
+
+ARGPARSE_DESCRIPTION = """
+This helper code takes in a disfluent recogs file generated from icefall.utils.store_transcript,
+compares it against a fluent transcript, and saves the results in a separate directory.
+This is useful to compare disfluent models with fluent models on the same metric.
+
+"""
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+        description=ARGPARSE_DESCRIPTION,
+    )
+    parser.add_argument(
+        "--recogs",
+        type=Path,
+        required=True,
+        help="Path to the recogs-XXX file generated by icefall.utils.store_transcript.",
+    )
+    parser.add_argument(
+        "--cut",
+        type=Path,
+        required=True,
+        help="Path to the cut manifest to be compared to. Assumes that disfluent_tag exists in the custom dict.",
+    )
+    parser.add_argument(
+        "--res-dir", type=Path, required=True, help="Path to save results"
+    )
+    return parser.parse_args()
+
+
+def d2f(stats):
+    """
+    Compare the outputs of a disfluent model against a fluent reference.
+    Indicates a disfluent model's performance only on the content words
+
+    CER^d_f = (sub_f + ins + del_f) / Nf
+
+    """
+    return stats["base"] / stats["Nf"]
+
+
+def calc_cer(refs, hyps):
+    subs = {
+        "F": 0,
+        "D": 0,
+    }
+    ins = 0
+    dels = {
+        "F": 0,
+        "D": 0,
+    }
+    cors = {
+        "F": 0,
+        "D": 0,
+    }
+    dis_ref_len = 0
+    flu_ref_len = 0
+
+    for ref, hyp in zip(refs, hyps):
+        assert (
+            ref[0] == hyp[0]
+        ), f"Expected ref cut id {ref[0]} to be the same as hyp cut id {hyp[0]}."
+        tag = ref[2].copy()
+        ref = ref[1]
+        dis_ref_len += len(ref)
+        # Remember that the 'D' and 'F' tags here refer to CSJ tags, not disfluent and fluent respectively.
+        flu_ref_len += len([t for t in tag if ("D" not in t and "F" not in t)])
+        hyp = hyp[1]
+        ali = kaldialign.align(ref, hyp, "*")
+        tags = ["*" if r[0] == "*" else tag.pop(0) for r in ali]
+        for tag, (ref_word, hyp_word) in zip(tags, ali):
+            if "D" in tag or "F" in tag:
+                tag = "D"
+            else:
+                tag = "F"
+
+            if ref_word == "*":
+                ins += 1
+            elif hyp_word == "*":
+                dels[tag] += 1
+            elif ref_word != hyp_word:
+                subs[tag] += 1
+            else:
+                cors[tag] += 1
+
+    return {
+        "subs": subs,
+        "ins": ins,
+        "dels": dels,
+        "cors": cors,
+        "dis_ref_len": dis_ref_len,
+        "flu_ref_len": flu_ref_len,
+    }
+
+
+def for_each_recogs(recogs_file: Path, refs, out_dir):
+    hyps = []
+    with recogs_file.open() as fin:
+        for line in fin:
+            if "ref" in line:
+                continue
+            cutid, hyp = line.split(":\thyp=")
+            hyps.append((cutid, eval(hyp)))
+
+    assert len(refs) == len(
+        hyps
+    ), f"Expected refs len {len(refs)} and hyps len {len(hyps)} to be equal."
+    stats = calc_cer(refs, hyps)
+    stat_table = ["tag,yes,no"]
+
+    for cer_type in ["subs", "dels", "cors", "ins"]:
+        ret = f"{cer_type}"
+        for df in ["D", "F"]:
+            try:
+                ret += f",{stats[cer_type][df]}"
+            except TypeError:
+                # insertions do not belong to F or D, and is not subscriptable.
+                ret += f",{stats[cer_type]},"
+                break
+        stat_table.append(ret)
+    stat_table = "\n".join(stat_table)
+
+    stats = {
+        "subd": stats["subs"]["D"],
+        "deld": stats["dels"]["D"],
+        "cord": stats["cors"]["D"],
+        "Nf": stats["flu_ref_len"],
+        "base": stats["subs"]["F"] + stats["ins"] + stats["dels"]["F"],
+    }
+
+    cer = d2f(stats)
+    results = [
+        f"{cer:.2%}",
+        f"Nf,{stats['Nf']}",
+    ]
+    results = "\n".join(results)
+
+    with (out_dir / (recogs_file.stem + ".dfcer")).open("w") as fout:
+        fout.write(results)
+        fout.write("\n\n")
+        fout.write(stat_table)
+
+
+def main():
+    args = get_args()
+    recogs_file: Path = args.recogs
+    assert (
+        recogs_file.is_file() or recogs_file.is_dir()
+    ), f"recogs_file cannot be found at {recogs_file}."
+
+    args.res_dir.mkdir(parents=True, exist_ok=True)
+
+    if recogs_file.is_file() and recogs_file.stem.startswith("recogs-"):
+        assert (
+            "csj_cuts" in args.cut.name
+        ), f"Expected {args.cut} to be a cuts manifest."
+
+        refs: CutSet = CutSet.from_file(args.cut)
+        refs = sorted(
+            [
+                (
+                    e.id,
+                    list(e.supervisions[0].custom["disfluent"]),
+                    e.supervisions[0].custom["disfluent_tag"].split(","),
+                )
+                for e in refs
+            ],
+            key=lambda x: x[0],
+        )
+        for_each_recogs(recogs_file, refs, args.res_dir)
+
+    elif recogs_file.is_dir():
+        recogs_file_path = recogs_file
+        for partname in ["eval1", "eval2", "eval3", "excluded", "valid"]:
+            refs: CutSet = CutSet.from_file(args.cut / f"csj_cuts_{partname}.jsonl.gz")
+            refs = sorted(
+                [
+                    (
+                        r.id,
+                        list(r.supervisions[0].custom["disfluent"]),
+                        r.supervisions[0].custom["disfluent_tag"].split(","),
+                    )
+                    for r in refs
+                ],
+                key=lambda x: x[0],
+            )
+            for recogs_file in recogs_file_path.glob(f"recogs-{partname}-*.txt"):
+                for_each_recogs(recogs_file, refs, args.res_dir)
+
+    else:
+        raise TypeError(f"Unrecognised recogs file provided: {recogs_file}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/local/display_manifest_statistics.py b/egs/csj/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..924474d33
--- /dev/null
+++ b/egs/csj/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,328 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2022  The University of Electro-Communications (author: Teo Wen Shen)  # noqa
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest
+
+ARGPARSE_DESCRIPTION = """
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in
+pruned_transducer_stateless5/train.py for usage.
+"""
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description=ARGPARSE_DESCRIPTION,
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+
+    parser.add_argument("--manifest-dir", type=Path, help="Path to cutset manifests")
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_parser()
+
+    for part in ["eval1", "eval2", "eval3", "valid", "excluded", "train"]:
+        path = args.manifest_dir / f"csj_cuts_{part}.jsonl.gz"
+        cuts: CutSet = load_manifest(path)
+
+        print("\n---------------------------------\n")
+        print(path.name + ":")
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+csj_cuts_eval1.jsonl.gz:
+Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 1023     │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 01:55:40 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 6.8      │
+├───────────────────────────┼──────────┤
+│ std                       │ 2.7      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.2      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 4.9      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 7.7      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 9.0      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ max                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 1023     │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 0        │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 1023     │
+╘═══════════════════════════╧══════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 01:55:40 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 01:55:40 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+
+---------------------------------
+
+csj_cuts_eval2.jsonl.gz:
+Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 1025     │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 02:02:07 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 7.1      │
+├───────────────────────────┼──────────┤
+│ std                       │ 2.5      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.1      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 5.9      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 7.9      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 9.1      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ max                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 1025     │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 0        │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 1025     │
+╘═══════════════════════════╧══════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 02:02:07 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 02:02:07 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+
+---------------------------------
+
+csj_cuts_eval3.jsonl.gz:
+Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 865      │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 01:26:44 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 6.0      │
+├───────────────────────────┼──────────┤
+│ std                       │ 3.0      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.3      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 3.3      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 6.8      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 8.7      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ max                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 865      │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 0        │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 865      │
+╘═══════════════════════════╧══════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 01:26:44 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 01:26:44 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+
+---------------------------------
+
+csj_cuts_valid.jsonl.gz:
+Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 3743     │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 06:40:15 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 6.4      │
+├───────────────────────────┼──────────┤
+│ std                       │ 3.0      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.1      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 3.9      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 7.4      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 9.0      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 10.1     │
+├───────────────────────────┼──────────┤
+│ max                       │ 11.8     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 3743     │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 0        │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 3743     │
+╘═══════════════════════════╧══════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 06:40:15 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 06:40:15 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+
+---------------------------------
+
+csj_cuts_excluded.jsonl.gz:
+Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 980      │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 00:56:06 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 3.4      │
+├───────────────────────────┼──────────┤
+│ std                       │ 3.1      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.1      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 0.8      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 2.2      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 5.8      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 9.9      │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 9.9      │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 10.0     │
+├───────────────────────────┼──────────┤
+│ max                       │ 10.0     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 980      │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 0        │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 980      │
+╘═══════════════════════════╧══════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 00:56:06 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 00:56:06 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+
+---------------------------------
+
+csj_cuts_train.jsonl.gz:
+Cut statistics:
+╒═══════════════════════════╤════════════╕
+│ Cuts count:               │ 914151     │
+├───────────────────────────┼────────────┤
+│ Total duration (hh:mm:ss) │ 1695:29:43 │
+├───────────────────────────┼────────────┤
+│ mean                      │ 6.7        │
+├───────────────────────────┼────────────┤
+│ std                       │ 2.9        │
+├───────────────────────────┼────────────┤
+│ min                       │ 0.1        │
+├───────────────────────────┼────────────┤
+│ 25%                       │ 4.6        │
+├───────────────────────────┼────────────┤
+│ 50%                       │ 7.5        │
+├───────────────────────────┼────────────┤
+│ 75%                       │ 8.9        │
+├───────────────────────────┼────────────┤
+│ 99%                       │ 11.0       │
+├───────────────────────────┼────────────┤
+│ 99.5%                     │ 11.0       │
+├───────────────────────────┼────────────┤
+│ 99.9%                     │ 11.1       │
+├───────────────────────────┼────────────┤
+│ max                       │ 18.0       │
+├───────────────────────────┼────────────┤
+│ Recordings available:     │ 914151     │
+├───────────────────────────┼────────────┤
+│ Features available:       │ 0          │
+├───────────────────────────┼────────────┤
+│ Supervisions available:   │ 914151     │
+╘═══════════════════════════╧════════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤════════════╤══════════════════════╕
+│ Total speech duration        │ 1695:29:43 │ 100.00% of recording │
+├──────────────────────────────┼────────────┼──────────────────────┤
+│ Total speaking time duration │ 1695:29:43 │ 100.00% of recording │
+├──────────────────────────────┼────────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00   │ 0.00% of recording   │
+╘══════════════════════════════╧════════════╧══════════════════════╛
+"""
diff --git a/egs/csj/ASR/local/prepare_lang_char.py b/egs/csj/ASR/local/prepare_lang_char.py
new file mode 100644
index 000000000..58b197922
--- /dev/null
+++ b/egs/csj/ASR/local/prepare_lang_char.py
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2022  The University of Electro-Communications  (Author: Teo Wen Shen)  # noqa
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet
+from lhotse.recipes.csj import CSJSDBParser
+
+ARGPARSE_DESCRIPTION = """
+This script gathers all training transcripts, parses them in disfluent mode, and produces a token list that would be the output set of the ASR system.
+
+It outputs 3 files into the lang directory:
+- tokens.txt: a list of tokens in the output set.
+- lang_type: a file that contains the string "char"
+
+"""
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        description=ARGPARSE_DESCRIPTION,
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+
+    parser.add_argument(
+        "train_cut", metavar="train-cut", type=Path, help="Path to the train cut"
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help=(
+            "Name of lang dir. "
+            "If not set, this will default to lang_char_{trans-mode}"
+        ),
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    logging.basicConfig(
+        format=("%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"),
+        level=logging.INFO,
+    )
+
+    sysdef_string = set(["<blk>", "<unk>", "<sos/eos>"])
+
+    # Using disfluent parsing as fluent is a subset of disfluent
+    parser = CSJSDBParser()
+
+    token_set = set()
+    logging.info(f"Creating vocabulary from {args.train_cut}.")
+    train_cut: CutSet = CutSet.from_file(args.train_cut)
+    for cut in train_cut:
+        if "_sp" in cut.id:
+            continue
+
+        text: str = cut.supervisions[0].custom["raw"]
+        for w in parser.parse(text, sep=" ").split(" "):
+            token_set.update(w)
+
+    token_set = ["<blk>"] + sorted(token_set - sysdef_string) + ["<unk>", "<sos/eos>"]
+    args.lang_dir.mkdir(parents=True, exist_ok=True)
+    (args.lang_dir / "tokens.txt").write_text(
+        "\n".join(f"{t}\t{i}" for i, t in enumerate(token_set))
+    )
+
+    (args.lang_dir / "lang_type").write_text("char")
+    logging.info("Done.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/local/utils/asr_datamodule.py b/egs/csj/ASR/local/utils/asr_datamodule.py
new file mode 100644
index 000000000..042b6ecbf
--- /dev/null
+++ b/egs/csj/ASR/local/utils/asr_datamodule.py
@@ -0,0 +1,462 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Union
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AsrVariableTranscriptDataset(K2SpeechRecognitionDataset):
+    def __init__(
+        self,
+        *args,
+        transcript_mode: str = "",
+        return_cuts: bool = False,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.transcript_mode = transcript_mode
+        self.return_cuts = True
+        self._return_cuts = return_cuts
+
+    def __getitem__(self, cuts: CutSet) -> Dict[str, Union[torch.Tensor, List[str]]]:
+        batch = super().__getitem__(cuts)
+
+        if self.transcript_mode:
+            batch["supervisions"]["text"] = [
+                supervision.custom[self.transcript_mode]
+                for cut in batch["supervisions"]["cut"]
+                for supervision in cut.supervisions
+            ]
+
+        if not self._return_cuts:
+            del batch["supervisions"]["cut"]
+
+        return batch
+
+
+class CSJAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--transcript-mode",
+            type=str,
+            default="",
+            help="Mode of transcript in supervision to use.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--musan-dir", type=Path, help="Path to directory with musan cuts. "
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.musan_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = AsrVariableTranscriptDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+            transcript_mode=self.args.transcript_mode,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = AsrVariableTranscriptDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+                transcript_mode=self.args.transcript_mode,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = AsrVariableTranscriptDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+                transcript_mode=self.args.transcript_mode,
+            )
+        else:
+            validate = AsrVariableTranscriptDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+                transcript_mode=self.args.transcript_mode,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+
+        test = AsrVariableTranscriptDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+            transcript_mode=self.args.transcript_mode,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "csj_cuts_train.jsonl.gz")
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get valid cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "csj_cuts_valid.jsonl.gz")
+
+    @lru_cache()
+    def excluded_cuts(self) -> CutSet:
+        logging.info("About to get excluded cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "csj_cuts_excluded.jsonl.gz")
+
+    @lru_cache()
+    def eval1_cuts(self) -> CutSet:
+        logging.info("About to get eval1 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "csj_cuts_eval1.jsonl.gz")
+
+    @lru_cache()
+    def eval2_cuts(self) -> CutSet:
+        logging.info("About to get eval2 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "csj_cuts_eval2.jsonl.gz")
+
+    @lru_cache()
+    def eval3_cuts(self) -> CutSet:
+        logging.info("About to get eval3 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "csj_cuts_eval3.jsonl.gz")
diff --git a/egs/csj/ASR/local/utils/tokenizer.py b/egs/csj/ASR/local/utils/tokenizer.py
new file mode 100644
index 000000000..c9be72be1
--- /dev/null
+++ b/egs/csj/ASR/local/utils/tokenizer.py
@@ -0,0 +1,253 @@
+import argparse
+from pathlib import Path
+from typing import Callable, List, Union
+
+import sentencepiece as spm
+from k2 import SymbolTable
+
+
+class Tokenizer:
+    text2word: Callable[[str], List[str]]
+
+    @staticmethod
+    def add_arguments(parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(title="Lang related options")
+
+        group.add_argument("--lang", type=Path, help="Path to lang directory.")
+
+        group.add_argument(
+            "--lang-type",
+            type=str,
+            default=None,
+            help=(
+                "Either 'bpe' or 'char'. If not provided, it expects lang_dir/lang_type to exists. "
+                "Note: 'bpe' directly loads sentencepiece.SentencePieceProcessor"
+            ),
+        )
+
+    @staticmethod
+    def Load(lang_dir: Path, lang_type="", oov="<unk>"):
+
+        if not lang_type:
+            assert (lang_dir / "lang_type").exists(), "lang_type not specified."
+            lang_type = (lang_dir / "lang_type").read_text().strip()
+
+        tokenizer = None
+
+        if lang_type == "bpe":
+            assert (
+                lang_dir / "bpe.model"
+            ).exists(), f"No BPE .model could be found in {lang_dir}."
+            tokenizer = spm.SentencePieceProcessor()
+            tokenizer.Load(str(lang_dir / "bpe.model"))
+        elif lang_type == "char":
+            tokenizer = CharTokenizer(lang_dir, oov=oov)
+        else:
+            raise NotImplementedError(f"{lang_type} not supported at the moment.")
+
+        return tokenizer
+
+    load = Load
+
+    def PieceToId(self, piece: str) -> int:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    piece_to_id = PieceToId
+
+    def IdToPiece(self, id: int) -> str:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    id_to_piece = IdToPiece
+
+    def GetPieceSize(self) -> int:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    get_piece_size = GetPieceSize
+
+    def __len__(self) -> int:
+        return self.get_piece_size()
+
+    def EncodeAsIdsBatch(self, input: List[str]) -> List[List[int]]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def EncodeAsPiecesBatch(self, input: List[str]) -> List[List[str]]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def EncodeAsIds(self, input: str) -> List[int]:
+        return self.EncodeAsIdsBatch([input])[0]
+
+    def EncodeAsPieces(self, input: str) -> List[str]:
+        return self.EncodeAsPiecesBatch([input])[0]
+
+    def Encode(
+        self, input: Union[str, List[str]], out_type=int
+    ) -> Union[List, List[List]]:
+        if not input:
+            return []
+
+        if isinstance(input, list):
+            if out_type is int:
+                return self.EncodeAsIdsBatch(input)
+            if out_type is str:
+                return self.EncodeAsPiecesBatch(input)
+
+        if out_type is int:
+            return self.EncodeAsIds(input)
+        if out_type is str:
+            return self.EncodeAsPieces(input)
+
+    encode = Encode
+
+    def DecodeIdsBatch(self, input: List[List[int]]) -> List[str]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def DecodePiecesBatch(self, input: List[List[str]]) -> List[str]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def DecodeIds(self, input: List[int]) -> str:
+        return self.DecodeIdsBatch([input])[0]
+
+    def DecodePieces(self, input: List[str]) -> str:
+        return self.DecodePiecesBatch([input])[0]
+
+    def Decode(
+        self,
+        input: Union[int, List[int], List[str], List[List[int]], List[List[str]]],
+    ) -> Union[List[str], str]:
+
+        if not input:
+            return ""
+
+        if isinstance(input, int):
+            return self.id_to_piece(input)
+        elif isinstance(input, str):
+            raise TypeError(
+                "Unlike spm.SentencePieceProcessor, cannot decode from type str."
+            )
+
+        if isinstance(input[0], list):
+            if not input[0] or isinstance(input[0][0], int):
+                return self.DecodeIdsBatch(input)
+
+            if isinstance(input[0][0], str):
+                return self.DecodePiecesBatch(input)
+
+        if isinstance(input[0], int):
+            return self.DecodeIds(input)
+        if isinstance(input[0], str):
+            return self.DecodePieces(input)
+
+        raise RuntimeError("Unknown input type")
+
+    decode = Decode
+
+    def SplitBatch(self, input: List[str]) -> List[List[str]]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def Split(self, input: Union[List[str], str]) -> Union[List[List[str]], List[str]]:
+        if isinstance(input, list):
+            return self.SplitBatch(input)
+        elif isinstance(input, str):
+            return self.SplitBatch([input])[0]
+        raise RuntimeError("Unknown input type")
+
+    split = Split
+
+
+class CharTokenizer(Tokenizer):
+    def __init__(self, lang_dir: Path, oov="<unk>", sep=""):
+        assert (
+            lang_dir / "tokens.txt"
+        ).exists(), f"tokens.txt could not be found in {lang_dir}."
+        token_table = SymbolTable.from_file(lang_dir / "tokens.txt")
+        assert (
+            "#0" not in token_table
+        ), "This tokenizer does not support disambig symbols."
+        self._id2sym = token_table._id2sym
+        self._sym2id = token_table._sym2id
+        self.oov = oov
+        self.oov_id = self._sym2id[oov]
+        self.sep = sep
+        if self.sep:
+            self.text2word = lambda x: x.split(self.sep)
+        else:
+            self.text2word = lambda x: list(x.replace(" ", ""))
+
+    def piece_to_id(self, piece: str) -> int:
+        try:
+            return self._sym2id[piece]
+        except KeyError:
+            return self.oov_id
+
+    def id_to_piece(self, id: int) -> str:
+        return self._id2sym[id]
+
+    def get_piece_size(self) -> int:
+        return len(self._sym2id)
+
+    def EncodeAsIdsBatch(self, input: List[str]) -> List[List[int]]:
+        return [[self.piece_to_id(i) for i in self.text2word(text)] for text in input]
+
+    def EncodeAsPiecesBatch(self, input: List[str]) -> List[List[str]]:
+        return [
+            [i if i in self._sym2id else self.oov for i in self.text2word(text)]
+            for text in input
+        ]
+
+    def DecodeIdsBatch(self, input: List[List[int]]) -> List[str]:
+        return [self.sep.join(self.id_to_piece(i) for i in text) for text in input]
+
+    def DecodePiecesBatch(self, input: List[List[str]]) -> List[str]:
+        return [self.sep.join(text) for text in input]
+
+    def SplitBatch(self, input: List[str]) -> List[List[str]]:
+        return [self.text2word(text) for text in input]
+
+
+def test_CharTokenizer():
+    test_single_string = "こんにちは"
+    test_multiple_string = [
+        "今日はいい天気ですよね",
+        "諏訪湖は綺麗でしょう",
+        "这在词表外",
+        "分かち 書き に し た 文章 です",
+        "",
+    ]
+    test_empty_string = ""
+    sp = Tokenizer.load(Path("lang_char"), "char", oov="<unk>")
+    splitter = sp.split
+    print(sp.encode(test_single_string, out_type=str))
+    print(sp.encode(test_single_string, out_type=int))
+    print(sp.encode(test_multiple_string, out_type=str))
+    print(sp.encode(test_multiple_string, out_type=int))
+    print(sp.encode(test_empty_string, out_type=str))
+    print(sp.encode(test_empty_string, out_type=int))
+    print(sp.decode(sp.encode(test_single_string, out_type=str)))
+    print(sp.decode(sp.encode(test_single_string, out_type=int)))
+    print(sp.decode(sp.encode(test_multiple_string, out_type=str)))
+    print(sp.decode(sp.encode(test_multiple_string, out_type=int)))
+    print(sp.decode(sp.encode(test_empty_string, out_type=str)))
+    print(sp.decode(sp.encode(test_empty_string, out_type=int)))
+    print(splitter(test_single_string))
+    print(splitter(test_multiple_string))
+    print(splitter(test_empty_string))
+
+
+if __name__ == "__main__":
+    test_CharTokenizer()
diff --git a/egs/csj/ASR/local/validate_manifest.py b/egs/csj/ASR/local/validate_manifest.py
new file mode 100644
index 000000000..7f67c64b6
--- /dev/null
+++ b/egs/csj/ASR/local/validate_manifest.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+- Supervision time bounds are within cut time bounds
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def validate_one_supervision_per_cut(c: Cut):
+    if len(c.supervisions) != 1:
+        raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
+
+
+def validate_supervision_and_cut_time_bounds(c: Cut):
+    s = c.supervisions[0]
+
+    # Removed because when the cuts were trimmed from supervisions,
+    # the start time of the supervision can be lesser than cut start time.
+    # https://github.com/lhotse-speech/lhotse/issues/813
+    # if s.start < c.start:
+    #     raise ValueError(
+    #         f"{c.id}: Supervision start time {s.start} is less "
+    #         f"than cut start time {c.start}"
+    #     )
+
+    if s.end > c.end:
+        raise ValueError(
+            f"{c.id}: Supervision end time {s.end} is larger "
+            f"than cut end time {c.end}"
+        )
+
+
+def main():
+    args = get_args()
+
+    manifest = Path(args.manifest)
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    for c in cut_set:
+        validate_one_supervision_per_cut(c)
+        validate_supervision_and_cut_time_bounds(c)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/csj/ASR/prepare.sh b/egs/csj/ASR/prepare.sh
new file mode 100755
index 000000000..52339bb35
--- /dev/null
+++ b/egs/csj/ASR/prepare.sh
@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# We assume the following directories are downloaded.
+#
+#  - $csj_dir
+#     CSJ is assumed to be the USB-type directory, which should contain the following subdirectories:-
+#     - DATA (not used in this script)
+#     - DOC (not used in this script)
+#     - MODEL (not used in this script)
+#     - MORPH
+#       - LDB (not used in this script)
+#       - SUWDIC (not used in this script)
+#       - SDB
+#         - core
+#           - ...
+#         - noncore
+#           - ...
+#     - PLABEL (not used in this script)
+#     - SUMMARY (not used in this script)
+#     - TOOL (not used in this script)
+#     - WAV
+#       - core
+#         - ...
+#       - noncore
+#         - ...
+#     - XML (not used in this script)
+#
+#  - $musan_dir
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#     - music
+#     - noise
+#     - speech
+#
+# By default, this script produces the original transcript like kaldi and espnet. Optionally, you
+# can add other transcript formats by supplying your own config files. A few examples of these
+# config files can be found in local/conf.
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=8
+stage=-1
+stop_stage=100
+
+csj_dir=/mnt/host/corpus/csj
+musan_dir=/mnt/host/corpus/musan/musan
+trans_dir=$csj_dir/transcript
+csj_fbank_dir=/mnt/host/corpus/csj/fbank
+musan_fbank_dir=$musan_dir/fbank
+csj_manifest_dir=data/manifests
+musan_manifest_dir=$musan_dir/manifests
+
+. shared/parse_options.sh || exit 1
+
+mkdir -p data
+
+log() {
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+    log "Stage 1: Prepare CSJ manifest"
+    if [ ! -e $csj_manifest_dir/.csj.done ]; then
+        lhotse prepare csj $csj_dir $csj_manifest_dir -t $trans_dir -j 16
+        touch $csj_manifest_dir/.csj.done
+    fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+    log "Stage 2: Prepare musan manifest"
+    mkdir -p $musan_manifest_dir
+    if [ ! -e $musan_manifest_dir/.musan.done ]; then
+        lhotse prepare musan $musan_dir $musan_manifest_dir
+        touch $musan_manifest_dir/.musan.done
+    fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3: Compute CSJ fbank"
+    if [ ! -e $csj_fbank_dir/.csj-validated.done ]; then
+        python local/compute_fbank_csj.py --manifest-dir $csj_manifest_dir \
+            --fbank-dir $csj_fbank_dir
+        parts=(
+            eval1
+            eval2
+            eval3
+            valid
+            excluded
+            train
+        )
+        for part in ${parts[@]}; do
+            python local/validate_manifest.py --manifest $csj_fbank_dir/csj_cuts_$part.jsonl.gz
+        done
+        touch $csj_fbank_dir/.csj-validated.done
+    fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+    log "Stage 4: Prepare CSJ lang_char"
+    python local/prepare_lang_char.py $csj_fbank_dir/csj_cuts_train.jsonl.gz
+    python local/add_transcript_mode.py -f $csj_fbank_dir -c local/conf/fluent.ini local/conf/number.ini
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for musan"
+    mkdir -p $musan_fbank_dir
+
+    if [ ! -e $musan_fbank_dir/.musan.done ]; then
+        python local/compute_fbank_musan.py --manifest-dir $musan_manifest_dir --fbank-dir $musan_fbank_dir
+        touch $musan_fbank_dir/.musan.done
+    fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+    log "Stage 6: Show manifest statistics"
+    python local/display_manifest_statistics.py --manifest-dir $csj_fbank_dir > $csj_fbank_dir/manifest_statistics.txt
+    cat $csj_fbank_dir/manifest_statistics.txt
+fi
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/TelegramStreamIO.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/TelegramStreamIO.py
new file mode 100644
index 000000000..f5235207a
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/TelegramStreamIO.py
@@ -0,0 +1,76 @@
+import logging
+from configparser import ConfigParser
+
+import requests
+
+
+def escape_html(text: str):
+    """
+    Escapes all html characters in text
+    :param str text:
+    :rtype: str
+    """
+    return text.replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
+
+
+class TelegramStreamIO(logging.Handler):
+
+    API_ENDPOINT = "https://api.telegram.org"
+    MAX_MESSAGE_LEN = 4096
+    formatter = logging.Formatter(
+        "%(asctime)s - %(levelname)s at %(funcName)s "
+        "(line %(lineno)s):\n\n%(message)s"
+    )
+
+    def __init__(self, tg_configfile: str):
+        super(TelegramStreamIO, self).__init__()
+        config = ConfigParser()
+        if not config.read(tg_configfile):
+            raise FileNotFoundError(
+                f"{tg_configfile} not found. " "Retry without --telegram-cred flag."
+            )
+        config = config["TELEGRAM"]
+        token = config["token"]
+        self.chat_id = config["chat_id"]
+        self.url = f"{self.API_ENDPOINT}/bot{token}/sendMessage"
+
+    @staticmethod
+    def setup_logger(params):
+        if not params.telegram_cred:
+            return
+        formatter = logging.Formatter(
+            f"{params.exp_dir.name} %(asctime)s \n%(message)s"
+        )
+        tg = TelegramStreamIO(params.telegram_cred)
+        tg.setLevel(logging.WARN)
+        tg.setFormatter(formatter)
+        logging.getLogger("").addHandler(tg)
+
+    def emit(self, record: logging.LogRecord):
+        """
+        Emit a record.
+        Send the record to the Web server as a percent-encoded dictionary
+        """
+        data = {
+            "chat_id": self.chat_id,
+            "text": self.format(self.mapLogRecord(record)),
+            "parse_mode": "HTML",
+        }
+        try:
+            requests.get(self.url, json=data)
+            # return response.json()
+        except Exception as e:
+            logging.error(f"Failed to send telegram message: {repr(e)}")
+            pass
+
+    def mapLogRecord(self, record):
+        """
+        Default implementation of mapping the log record into a dict
+        that is sent as the CGI data. Overwrite in your class.
+        Contributed by Franz Glasner.
+        """
+
+        for k, v in record.__dict__.items():
+            if isinstance(v, str):
+                setattr(record, k, escape_html(v))
+        return record
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
new file mode 120000
index 000000000..a48591198
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
@@ -0,0 +1 @@
+../local/utils/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/beam_search.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/beam_search.py
new file mode 120000
index 000000000..d7349b0a3
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/decode.py
new file mode 100755
index 000000000..f5a1d750d
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/decode.py
@@ -0,0 +1,846 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --lang data/lang_char \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --lang data/lang_char \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --lang data/lang_char \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --lang data/lang_char \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --lang data/lang_char \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --lang data/lang_char \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --lang data/lang_char \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import CSJAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from tokenizer import Tokenizer
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--gpu",
+        type=int,
+        default=0,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--res-dir",
+        type=Path,
+        default=None,
+        help="The path to save results.",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        help="The lang dir. It should contain at least a word table.",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-graph",
+        type=str,
+        default="",
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--pad-feature",
+        type=int,
+        default=30,
+        help="""
+        Number of frames to pad at the end.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: Tokenizer,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.pad_feature:
+        feature_lens += params.pad_feature
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.pad_feature),
+            value=LOG_EPS,
+        )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.text2word(sp.decode(hyp)))
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: Tokenizer,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = sp.text2word(ref_text)
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    return test_set_wers
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    CSJAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    if not params.res_dir:
+        params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", params.gpu)
+
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> and <unk> are defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    decoding_graph = None
+    word_table = None
+
+    if params.decoding_graph:
+        decoding_graph = k2.Fsa.from_dict(
+            torch.load(params.decoding_graph, map_location=device)
+        )
+    elif "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    csj_corpus = CSJAsrDataModule(args)
+
+    for subdir in ["eval1", "eval2", "eval3", "excluded", "valid"]:
+        results_dict = decode_dataset(
+            dl=csj_corpus.test_dataloaders(getattr(csj_corpus, f"{subdir}_cuts")()),
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+        tot_err = save_results(
+            params=params,
+            test_set_name=subdir,
+            results_dict=results_dict,
+        )
+        with (
+            params.res_dir
+            / (
+                f"{subdir}-{params.decode_chunk_len}_{params.beam_size}"
+                f"_{params.avg}_{params.epoch}.cer"
+            )
+        ).open("w") as fout:
+            if len(tot_err) == 1:
+                fout.write(f"{tot_err[0][1]}")
+            else:
+                fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/decode_stream.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
new file mode 120000
index 000000000..ca8fed319
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/decoder.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/decoder.py
new file mode 120000
index 000000000..1ce277aa6
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
new file mode 100755
index 000000000..685f6ece6
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
@@ -0,0 +1,1295 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import math
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import CSJAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from tokenizer import Tokenizer
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer_for_ncnn_export_only import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+LOG_EPS = math.log(1e-10)
+
+try:
+    from TelegramStreamIO import TelegramStreamIO
+
+    HAS_TELEGRAM = True
+except ImportError:
+    HAS_TELEGRAM = False
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument("--debug", action="store_true", help="Use hardcoded arguments")
+
+    parser.add_argument(
+        "--telegram-cred",
+        type=Path,
+        default=None,
+        help="Telegram credentials to report progress in telegram",
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--pad-feature",
+        type=int,
+        default=0,
+        help="""
+        Number of frames to pad at the end.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+        is_pnnx=True,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.pad_feature:
+        feature_lens += params.pad_feature
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.pad_feature),
+            value=LOG_EPS,
+        )
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: Tokenizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except Exception as e:  # noqa
+            logging.error(e, exc_info=True)
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise e
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            if HAS_TELEGRAM and batch_idx in [0, 500] and not rank:
+                logging.warning(
+                    f"Epoch {params.cur_epoch}, "
+                    f"batch {batch_idx}, loss[{loss_info}], "
+                    f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                    f"lr: {cur_lr:.2e}, "
+                    + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+                )
+            else:
+                logging.info(
+                    f"Epoch {params.cur_epoch}, "
+                    f"batch {batch_idx}, loss[{loss_info}], "
+                    f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                    f"lr: {cur_lr:.2e}, "
+                    + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+                )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            if (
+                HAS_TELEGRAM
+                and batch_idx % (params.valid_interval * 3) == 0
+                and not rank
+            ):
+                log_mode = logging.warning
+            else:
+                log_mode = logging.info
+            log_mode(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            log_mode(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, master_port=params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    if HAS_TELEGRAM and params.telegram_cred:
+        TelegramStreamIO.setup_logger(params)
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(args.lang, args.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 0.3 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.info(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    csj_corpus = CSJAsrDataModule(args)
+    train_cuts = csj_corpus.train_cuts()
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = csj_corpus.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = csj_corpus.valid_cuts()
+    valid_dl = csj_corpus.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: Tokenizer,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: Tokenizer,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    CSJAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
new file mode 120000
index 000000000..cb673b3eb
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
new file mode 100755
index 000000000..b210430c6
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
@@ -0,0 +1,369 @@
+#!/usr/bin/env python3
+
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for more details about how to use this file.
+
+We use
+https://huggingface.co/TeoWenShen/icefall-asr-csj-pruned-transducer-stateless7-streaming-230208
+
+to demonstrate the usage of this file.
+
+1. Download the pre-trained model
+
+cd egs/csj/ASR
+
+repo_url=https://huggingface.co/TeoWenShen/icefall-asr-csj-pruned-transducer-stateless7-streaming-230208
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp_fluent/pretrained.pt"
+
+cd exp_fluent
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export to ncnn
+
+./pruned_transducer_stateless7_streaming/export-for-ncnn.py \
+  --lang $repo/data/lang_char \
+  --exp-dir $repo/exp_fluent/ \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  \
+  --decode-chunk-len 32 \
+  --num-left-chunks 4 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+cd $repo/exp_fluent
+
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+You can find converted models at
+https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-ja-fluent-2023-02-14
+
+Please also have a look at
+https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-ja-fluent-2023-02-14/blob/main/export-for-ncnn-fluent.sh
+
+See ./streaming-ncnn-decode.py
+and
+https://github.com/k2-fsa/sherpa-ncnn
+for usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from tokenizer import Tokenizer
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
+
+    decode_chunk_len = encoder_model.decode_chunk_size * 2
+    pad_length = 7
+    T = decode_chunk_len + pad_length  # 32 + 7 = 39
+
+    logging.info(f"decode_chunk_len: {decode_chunk_len}")
+    logging.info(f"T: {T}")
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.get_init_state()
+
+    traced_model = torch.jit.trace(encoder_model, (x, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
+
+    logging.info(f"device: {device}")
+
+    sp = Tokenizer.load(args.lang, args.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+    assert params.blank_id == 0, params.blank_id
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_pnnx=True)
+
+    encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export.py
new file mode 100644
index 000000000..2d45ecca3
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export.py
@@ -0,0 +1,313 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7_streaming/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/csj/ASR
+    ./pruned_transducer_stateless7_streaming/decode.py \
+        --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --lang data/lang_char
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/TeoWenShen/icefall-asr-csj-pruned-transducer-stateless7-streaming-230208
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/TeoWenShen/icefall-asr-csj-pruned-transducer-stateless7-streaming-230208
+    # You will find the pre-trained model in icefall-asr-csj-pruned-transducer-stateless7-230208/exp_fluent
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from tokenizer import Tokenizer
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
new file mode 100644
index 000000000..ab7c8748a
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+
+"""
+Usage:
+# use -O to skip assertions and avoid some of the TracerWarnings
+python -O pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --decode-chunk-len 32
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from tokenizer import Tokenizer
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+    params: AttributeDict,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    decode_chunk_len = params.decode_chunk_len  # before subsampling
+    pad_length = 7
+    s = f"decode_chunk_len: {decode_chunk_len}"
+    logging.info(s)
+    assert encoder_model.decode_chunk_size == decode_chunk_len // 2, (
+        encoder_model.decode_chunk_size,
+        decode_chunk_len,
+    )
+
+    T = decode_chunk_len + pad_length
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    x_lens = torch.full((1,), T, dtype=torch.int32)
+    states = encoder_model.get_init_state(device=x.device)
+
+    encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename, params)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
new file mode 100644
index 000000000..d84cf04a3
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+# flake8: noqa
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models exported by `torch.jit.trace()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --decode-chunk-len 32
+
+Usage of this script:
+
+./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless7_streaming/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./pruned_transducer_stateless7_streaming/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./pruned_transducer_stateless7_streaming/exp/joiner_jit_trace.pt \
+  --lang data/lang_char \
+  --decode-chunk-len 32 \
+  /path/to/foo.wav \
+"""
+
+import argparse
+import logging
+from typing import List, Optional
+
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+from tokenizer import Tokenizer
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    assert encoder_out.ndim == 2
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32).unsqueeze(0)
+        # decoder_input.shape (1,, 1 context_size)
+        decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1)
+    else:
+        assert decoder_out.ndim == 2
+        assert hyp is not None, hyp
+
+    T = encoder_out.size(0)
+    for i in range(T):
+        cur_encoder_out = encoder_out[i : i + 1]
+        joiner_out = joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32).unsqueeze(0)
+            decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1)
+
+    return hyp, decoder_out
+
+
+def create_streaming_feature_extractor(sample_rate) -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = Tokenizer.load(args.lang, args.lang_type)
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor(args.sample_rate)
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=args.sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    logging.info("Decoding started")
+    chunk_length = args.decode_chunk_len
+    assert encoder.decode_chunk_size == chunk_length // 2, (
+        encoder.decode_chunk_size,
+        chunk_length,
+    )
+
+    # we subsample features with ((x_len - 7) // 2 + 1) // 2
+    pad_length = 7
+    T = chunk_length + pad_length
+
+    logging.info(f"chunk_length: {chunk_length}")
+
+    states = encoder.get_init_state(device)
+
+    tail_padding = torch.zeros(int(0.3 * args.sample_rate), dtype=torch.float32)
+
+    wave_samples = torch.cat([wave_samples, tail_padding])
+
+    chunk = int(0.25 * args.sample_rate)  # 0.2 second
+    num_processed_frames = 0
+
+    hyp = None
+    decoder_out = None
+
+    start = 0
+    while start < wave_samples.numel():
+        logging.info(f"{start}/{wave_samples.numel()}")
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+        online_fbank.accept_waveform(
+            sampling_rate=args.sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= T:
+            frames = []
+            for i in range(T):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            frames = torch.cat(frames, dim=0).unsqueeze(0)
+            x_lens = torch.tensor([T], dtype=torch.int32)
+            encoder_out, out_lens, states = encoder(
+                x=frames,
+                x_lens=x_lens,
+                states=states,
+            )
+            num_processed_frames += chunk_length
+
+            hyp, decoder_out = greedy_search(
+                decoder, joiner, encoder_out.squeeze(0), decoder_out, hyp
+            )
+
+    context_size = 2
+    logging.info(args.sound_file)
+    logging.info(sp.decode(hyp[context_size:]))
+
+    logging.info("Decoding Done")
+
+
+torch.set_num_threads(4)
+torch.set_num_interop_threads(1)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/joiner.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/joiner.py
new file mode 120000
index 000000000..482ebcfef
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/model.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/model.py
new file mode 120000
index 000000000..16c2bf28d
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/model.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/optim.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/optim.py
new file mode 120000
index 000000000..522bbaff9
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/optim.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/pretrained.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/pretrained.py
new file mode 100644
index 000000000..932026868
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/pretrained.py
@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7_streaming/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7_streaming/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_streaming/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from tokenizer import Tokenizer
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/scaling.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/scaling.py
new file mode 120000
index 000000000..a7ef73bcb
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
new file mode 120000
index 000000000..566c317ff
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
new file mode 120000
index 000000000..92c3904af
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
new file mode 120000
index 000000000..2adf271c1
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
new file mode 100755
index 000000000..9700dd89e
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
@@ -0,0 +1,597 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding_method greedy_search \
+  --lang data/lang_char \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+import torch.nn as nn
+from asr_datamodule import CSJAsrDataModule
+from decode import save_results
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from tokenizer import Tokenizer
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--gpu",
+        type=int,
+        default=0,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-graph",
+        type=str,
+        default="",
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    parser.add_argument(
+        "--res-dir",
+        type=Path,
+        default=None,
+        help="The path to save results.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: Tokenizer,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].custom[params.transcript_mode]
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        sp.text2word(decode_streams[i].ground_truth),
+                        sp.text2word(sp.decode(decode_streams[i].decoding_result())),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    sp.text2word(decode_streams[i].ground_truth),
+                    sp.text2word(sp.decode(decode_streams[i].decoding_result())),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    CSJAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if not params.res_dir:
+        params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", params.gpu)
+
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> and <unk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_graph:
+        decoding_graph = k2.Fsa.from_dict(
+            torch.load(params.decoding_graph, map_location=device)
+        )
+    elif params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    args.return_cuts = True
+    csj_corpus = CSJAsrDataModule(args)
+
+    for subdir in ["eval1", "eval2", "eval3", "excluded", "valid"]:
+        results_dict = decode_dataset(
+            cuts=getattr(csj_corpus, f"{subdir}_cuts")(),
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+        tot_err = save_results(
+            params=params, test_set_name=subdir, results_dict=results_dict
+        )
+
+        with (
+            params.res_dir
+            / (
+                f"{subdir}-{params.decode_chunk_len}"
+                f"_{params.avg}_{params.epoch}.cer"
+            )
+        ).open("w") as fout:
+            if len(tot_err) == 1:
+                fout.write(f"{tot_err[0][1]}")
+            else:
+                fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/test_model.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/test_model.py
new file mode 100755
index 000000000..0a82ccfa4
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/test_model.py
@@ -0,0 +1,150 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/csj/ASR
+    python ./pruned_transducer_stateless7_streaming/test_model.py
+"""
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    # Test jit script
+    convert_scaled_to_non_scaled(model, inplace=True)
+    # We won't use the forward() method of the model in C++, so just ignore
+    # it here.
+    # Otherwise, one of its arguments is a ragged tensor and is not
+    # torch scriptabe.
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    print("Using torch.jit.script")
+    model = torch.jit.script(model)
+
+
+def test_model_jit_trace():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    # Test encoder
+    def _test_encoder():
+        encoder = model.encoder
+        assert encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+            encoder.decode_chunk_size,
+            params.decode_chunk_len,
+        )
+        T = params.decode_chunk_len + 7
+
+        x = torch.zeros(1, T, 80, dtype=torch.float32)
+        x_lens = torch.full((1,), T, dtype=torch.int32)
+        states = encoder.get_init_state(device=x.device)
+        encoder.__class__.forward = encoder.__class__.streaming_forward
+        traced_encoder = torch.jit.trace(encoder, (x, x_lens, states))
+
+        states1 = encoder.get_init_state(device=x.device)
+        states2 = traced_encoder.get_init_state(device=x.device)
+        for i in range(5):
+            x = torch.randn(1, T, 80, dtype=torch.float32)
+            x_lens = torch.full((1,), T, dtype=torch.int32)
+            y1, _, states1 = encoder.streaming_forward(x, x_lens, states1)
+            y2, _, states2 = traced_encoder(x, x_lens, states2)
+            assert torch.allclose(y1, y2, atol=1e-6), (i, (y1 - y2).abs().mean())
+
+    # Test decoder
+    def _test_decoder():
+        decoder = model.decoder
+        y = torch.zeros(10, decoder.context_size, dtype=torch.int64)
+        need_pad = torch.tensor([False])
+
+        traced_decoder = torch.jit.trace(decoder, (y, need_pad))
+        d1 = decoder(y, need_pad)
+        d2 = traced_decoder(y, need_pad)
+        assert torch.equal(d1, d2), (d1 - d2).abs().mean()
+
+    # Test joiner
+    def _test_joiner():
+        joiner = model.joiner
+        encoder_out_dim = joiner.encoder_proj.weight.shape[1]
+        decoder_out_dim = joiner.decoder_proj.weight.shape[1]
+        encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+        decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+        traced_joiner = torch.jit.trace(joiner, (encoder_out, decoder_out))
+        j1 = joiner(encoder_out, decoder_out)
+        j2 = traced_joiner(encoder_out, decoder_out)
+        assert torch.equal(j1, j2), (j1 - j2).abs().mean()
+
+    _test_encoder()
+    _test_decoder()
+    _test_joiner()
+
+
+def main():
+    test_model()
+    test_model_jit_trace()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/tokenizer.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/tokenizer.py
new file mode 120000
index 000000000..958c99e85
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/tokenizer.py
@@ -0,0 +1 @@
+../local/utils/tokenizer.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py
new file mode 100755
index 000000000..73fcd67aa
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -0,0 +1,1293 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import math
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import CSJAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from tokenizer import Tokenizer
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+LOG_EPS = math.log(1e-10)
+
+try:
+    from TelegramStreamIO import TelegramStreamIO
+
+    HAS_TELEGRAM = True
+except ImportError:
+    HAS_TELEGRAM = False
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument("--debug", action="store_true", help="Use hardcoded arguments")
+
+    parser.add_argument(
+        "--telegram-cred",
+        type=Path,
+        default=None,
+        help="Telegram credentials to report progress in telegram",
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--pad-feature",
+        type=int,
+        default=0,
+        help="""
+        Number of frames to pad at the end.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.pad_feature:
+        feature_lens += params.pad_feature
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.pad_feature),
+            value=LOG_EPS,
+        )
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: Tokenizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except Exception as e:  # noqa
+            logging.error(e, exc_info=True)
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise e
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            if HAS_TELEGRAM and batch_idx in [0, 500] and not rank:
+                logging.warning(
+                    f"Epoch {params.cur_epoch}, "
+                    f"batch {batch_idx}, loss[{loss_info}], "
+                    f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                    f"lr: {cur_lr:.2e}, "
+                    + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+                )
+            else:
+                logging.info(
+                    f"Epoch {params.cur_epoch}, "
+                    f"batch {batch_idx}, loss[{loss_info}], "
+                    f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                    f"lr: {cur_lr:.2e}, "
+                    + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+                )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            if (
+                HAS_TELEGRAM
+                and batch_idx % (params.valid_interval * 3) == 0
+                and not rank
+            ):
+                log_mode = logging.warning
+            else:
+                log_mode = logging.info
+            log_mode(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            log_mode(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, master_port=params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    if HAS_TELEGRAM and params.telegram_cred:
+        TelegramStreamIO.setup_logger(params)
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(args.lang, args.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 0.3 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.info(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    csj_corpus = CSJAsrDataModule(args)
+    train_cuts = csj_corpus.train_cuts()
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = csj_corpus.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = csj_corpus.valid_cuts()
+    valid_dl = csj_corpus.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: Tokenizer,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: Tokenizer,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    CSJAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/zipformer.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/zipformer.py
new file mode 120000
index 000000000..ec183baa7
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
new file mode 120000
index 000000000..d301e1f9b
--- /dev/null
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
\ No newline at end of file
diff --git a/egs/csj/ASR/shared b/egs/csj/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/csj/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/.gitignore b/egs/gigaspeech/ASR/.gitignore
new file mode 100644
index 000000000..8dec2d86d
--- /dev/null
+++ b/egs/gigaspeech/ASR/.gitignore
@@ -0,0 +1,2 @@
+log-*
+.DS_Store
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/README.md b/egs/gigaspeech/ASR/README.md
new file mode 100644
index 000000000..f0d60898c
--- /dev/null
+++ b/egs/gigaspeech/ASR/README.md
@@ -0,0 +1,22 @@
+# GigaSpeech
+GigaSpeech, an evolving, multi-domain English
+speech recognition corpus with 10,000 hours of high quality labeled
+audio, collected from audiobooks, podcasts
+and YouTube, covering both read and spontaneous speaking styles,
+and a variety of topics, such as arts, science, sports, etc. More details can be found: https://github.com/SpeechColab/GigaSpeech
+
+## Download
+
+Apply for the download credentials and download the dataset by following https://github.com/SpeechColab/GigaSpeech#download. Then create a symlink
+```bash
+ln -sfv /path/to/GigaSpeech download/GigaSpeech
+```
+
+## Performance Record
+|                                |  Dev  | Test  |
+|--------------------------------|-------|-------|
+|           `zipformer`          | 10.25 | 10.38 |
+|         `conformer_ctc`        | 10.47 | 10.58 |
+| `pruned_transducer_stateless2` | 10.40 | 10.51 |
+
+See [RESULTS](/egs/gigaspeech/ASR/RESULTS.md) for details.
diff --git a/egs/gigaspeech/ASR/RESULTS.md b/egs/gigaspeech/ASR/RESULTS.md
new file mode 100644
index 000000000..841ebdcfa
--- /dev/null
+++ b/egs/gigaspeech/ASR/RESULTS.md
@@ -0,0 +1,226 @@
+## Results
+### zipformer (zipformer + pruned stateless transducer)
+
+See <https://github.com/k2-fsa/icefall/pull/1254> for more details.
+
+[zipformer](./zipformer)
+
+- Non-streaming
+- normal-scaled model, number of model parameters: 65549011, i.e., 65.55 M
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17>
+
+The tensorboard log for training is available at
+<https://wandb.ai/yifanyeung/icefall-asr-gigaspeech-zipformer-2023-10-20>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 10.31      | 10.50      | --epoch 30 --avg 9 |
+| modified_beam_search | 10.25      | 10.38      | --epoch 30 --avg 9 |
+| fast_beam_search     | 10.26      | 10.48      | --epoch 30 --avg 9 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 0 \
+  --subset XL \
+  --max-duration 700 \
+  --use-transducer 1 \
+  --use-ctc 0 \
+  --lr-epochs 1 \
+  --master-port 12345
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES=0
+
+# greedy search
+./zipformer/decode.py \
+  --epoch 30 \
+  --avg 9 \
+  --exp-dir ./zipformer/exp \
+  --max-duration 1000 \
+  --decoding-method greedy_search
+
+# modified beam search
+./zipformer/decode.py \
+  --epoch 30 \
+  --avg 9 \
+  --exp-dir ./zipformer/exp \
+  --max-duration 1000 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+# fast beam search (one best)
+./zipformer/decode.py \
+  --epoch 30 \
+  --avg 9 \
+  --exp-dir ./zipformer/exp \
+  --max-duration 1000 \
+  --decoding-method fast_beam_search \
+  --beam 20.0 \
+  --max-contexts 8 \
+  --max-states 64
+```
+
+### GigaSpeech BPE training results (Pruned Transducer 2)
+
+#### 2022-05-12
+
+#### Conformer encoder + embedding decoder
+
+Conformer encoder + non-recurrent decoder. The encoder is a
+reworked version of the conformer encoder, with many changes. The
+decoder contains only an embedding layer, a Conv1d (with kernel
+size 2) and a linear layer (to transform tensor dim). k2 pruned
+RNN-T loss is used.
+
+The best WER, as of 2022-05-12, for the gigaspeech is below
+
+Results are:
+
+|                      |  Dev  | Test  |
+|----------------------|-------|-------|
+|    greedy search     | 10.51 | 10.73 |
+|   fast beam search   | 10.50 | 10.69 |
+| modified beam search | 10.40 | 10.51 |
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+cd egs/gigaspeech/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+./pruned_transducer_stateless2/train.py \
+  --max-duration 120 \
+  --num-workers 1 \
+  --world-size 8 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --use-fp16 True
+```
+
+and the following commands for decoding:
+
+```bash
+# greedy search
+./pruned_transducer_stateless2/decode.py \
+  --iter 3488000 \
+  --avg 20 \
+  --decoding-method greedy_search \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --max-duration 600
+
+# fast beam search
+./pruned_transducer_stateless2/decode.py \
+  --iter 3488000 \
+  --avg 20 \
+  --decoding-method fast_beam_search \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --max-duration 600
+
+# modified beam search
+./pruned_transducer_stateless2/decode.py \
+  --iter 3488000 \
+  --avg 15 \
+  --decoding-method modified_beam_search \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --max-duration 600
+```
+
+Pretrained model is available at
+<https://huggingface.co/wgb14/icefall-asr-gigaspeech-pruned-transducer-stateless2>
+
+The tensorboard log for training is available at
+<https://tensorboard.dev/experiment/zmmM0MLASnG1N2RmJ4MZBw/>
+
+### GigaSpeech BPE training results (Conformer-CTC)
+
+#### 2022-04-06
+
+The best WER, as of 2022-04-06, for the gigaspeech is below
+
+Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|     |  Dev  | Test  |
+|-----|-------|-------|
+| WER | 10.47 | 10.58 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.5       |       1.3       |
+
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+cd egs/gigaspeech/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+./conformer_ctc/train.py \
+  --max-duration 120 \
+  --num-workers 1 \
+  --world-size 8 \
+  --exp-dir conformer_ctc/exp_500 \
+  --lang-dir data/lang_bpe_500
+```
+
+and the following command for decoding:
+
+```bash
+./conformer_ctc/decode.py \
+  --epoch 18 \
+  --avg 6 \
+  --method attention-decoder \
+  --num-paths 1000 \
+  --exp-dir conformer_ctc/exp_500 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 20 \
+  --num-workers 1
+```
+
+Results using HLG decoding + whole lattice rescoring:
+
+|     |  Dev  | Test  |
+|-----|-------|-------|
+| WER | 10.51 | 10.62 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| lm_scale |
+|----------|
+|   0.2    |
+
+To reproduce the above result, use the training commands above, and the following command for decoding:
+
+```bash
+./conformer_ctc/decode.py \
+  --epoch 18 \
+  --avg 6 \
+  --method whole-lattice-rescoring \
+  --num-paths 1000 \
+  --exp-dir conformer_ctc/exp_500 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 20 \
+  --num-workers 1
+```
+Note: the `whole-lattice-rescoring` method is about twice as fast as the `attention-decoder` method, with slightly worse WER.
+
+Pretrained model is available at
+<https://huggingface.co/wgb14/icefall-asr-gigaspeech-conformer-ctc>
+
+The tensorboard log for training is available at
+<https://tensorboard.dev/experiment/rz63cmJXSK2fV9GceJtZXQ/>
diff --git a/egs/gigaspeech/ASR/conformer_ctc/__init__.py b/egs/gigaspeech/ASR/conformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py b/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py
new file mode 100644
index 000000000..a93e224d5
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py
@@ -0,0 +1,363 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class GigaSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+
+        # GigaSpeech specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="XL",
+            help="Select the GigaSpeech subset (XS|S|M|L|XL)",
+        )
+        group.add_argument(
+            "--small-dev",
+            type=str2bool,
+            default=False,
+            help="Should we use only 1000 utterances for dev (speeds up training)",
+        )
+
+    def train_dataloaders(self, cuts_train: CutSet) -> DataLoader:
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=2,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get train_{self.args.subset} cuts")
+        path = self.args.manifest_dir / f"cuts_{self.args.subset}.jsonl.gz"
+        cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(self.args.manifest_dir / "cuts_DEV.jsonl.gz")
+        if self.args.small_dev:
+            return cuts_valid.subset(first=1000)
+        else:
+            return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST.jsonl.gz")
diff --git a/egs/gigaspeech/ASR/conformer_ctc/conformer.py b/egs/gigaspeech/ASR/conformer_ctc/conformer.py
new file mode 100644
index 000000000..a1cfe6e75
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/conformer.py
@@ -0,0 +1,910 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: Union[float, bool] = 0.1,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        use_conv_batchnorm = True
+        if isinstance(use_feat_batchnorm, float):
+            use_conv_batchnorm = False
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+            use_conv_batchnorm,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self, x: Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        use_conv_batchnorm: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(
+            d_model, cnn_module_kernel, use_batchnorm=use_conv_batchnorm
+        )
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        use_batchnorm: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+        self.use_batchnorm = use_batchnorm
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        if self.use_batchnorm:
+            self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        if self.use_batchnorm:
+            x = self.norm(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/gigaspeech/ASR/conformer_ctc/decode.py b/egs/gigaspeech/ASR/conformer_ctc/decode.py
new file mode 100755
index 000000000..d7035a1f8
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/decode.py
@@ -0,0 +1,711 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+# Copyright 2022 Johns Hopkins University (Author: Guanbo Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from conformer import Conformer
+from gigaspeech_scoring import asr_text_post_processing
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=0,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=1000,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+        # TODO: pass `lattice` instead of `rescored_lattice` to
+        # `rescore_with_attention_decoder`
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for cut_id, ref_text in zip(cut_ids, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = post_processing(results)
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in ["whole-lattice-rescoring", "attention-decoder"]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/conformer_ctc/gigaspeech_scoring.py b/egs/gigaspeech/ASR/conformer_ctc/gigaspeech_scoring.py
new file mode 100755
index 000000000..ef53b77f8
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/gigaspeech_scoring.py
@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+# Copyright 2021 Jiayu Du
+# Copyright 2022 Johns Hopkins University (Author: Guanbo Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import os
+
+conversational_filler = [
+    "UH",
+    "UHH",
+    "UM",
+    "EH",
+    "MM",
+    "HM",
+    "AH",
+    "HUH",
+    "HA",
+    "ER",
+    "OOF",
+    "HEE",
+    "ACH",
+    "EEE",
+    "EW",
+]
+unk_tags = ["<UNK>", "<unk>"]
+gigaspeech_punctuations = [
+    "<COMMA>",
+    "<PERIOD>",
+    "<QUESTIONMARK>",
+    "<EXCLAMATIONPOINT>",
+]
+gigaspeech_garbage_utterance_tags = ["<SIL>", "<NOISE>", "<MUSIC>", "<OTHER>"]
+non_scoring_words = (
+    conversational_filler
+    + unk_tags
+    + gigaspeech_punctuations
+    + gigaspeech_garbage_utterance_tags
+)
+
+
+def asr_text_post_processing(text: str) -> str:
+    # 1. convert to uppercase
+    text = text.upper()
+
+    # 2. remove hyphen
+    #   "E-COMMERCE" -> "E COMMERCE", "STATE-OF-THE-ART" -> "STATE OF THE ART"
+    text = text.replace("-", " ")
+
+    # 3. remove non-scoring words from evaluation
+    remaining_words = []
+    for word in text.split():
+        if word in non_scoring_words:
+            continue
+        remaining_words.append(word)
+
+    return " ".join(remaining_words)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="This script evaluates GigaSpeech ASR result via"
+        "SCTK's tool sclite"
+    )
+    parser.add_argument(
+        "ref",
+        type=str,
+        help="sclite's standard transcription(trn) reference file",
+    )
+    parser.add_argument(
+        "hyp",
+        type=str,
+        help="sclite's standard transcription(trn) hypothesis file",
+    )
+    parser.add_argument(
+        "work_dir",
+        type=str,
+        help="working dir",
+    )
+    args = parser.parse_args()
+
+    if not os.path.isdir(args.work_dir):
+        os.mkdir(args.work_dir)
+
+    REF = os.path.join(args.work_dir, "REF")
+    HYP = os.path.join(args.work_dir, "HYP")
+    RESULT = os.path.join(args.work_dir, "RESULT")
+
+    for io in [(args.ref, REF), (args.hyp, HYP)]:
+        with open(io[0], "r", encoding="utf8") as fi:
+            with open(io[1], "w+", encoding="utf8") as fo:
+                for line in fi:
+                    line = line.strip()
+                    if line:
+                        cols = line.split()
+                        text = asr_text_post_processing(" ".join(cols[0:-1]))
+                        uttid_field = cols[-1]
+                        print(f"{text} {uttid_field}", file=fo)
+
+    # GigaSpeech's uttid comforms to swb
+    os.system(f"sclite -r {REF} trn -h {HYP} trn -i swb | tee {RESULT}")
diff --git a/egs/gigaspeech/ASR/conformer_ctc/label_smoothing.py b/egs/gigaspeech/ASR/conformer_ctc/label_smoothing.py
new file mode 100644
index 000000000..3b94f0c4b
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/label_smoothing.py
@@ -0,0 +1,95 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+class LabelSmoothingLoss(torch.nn.Module):
+    """
+    Implement the LabelSmoothingLoss proposed in the following paper
+    https://arxiv.org/pdf/1512.00567.pdf
+    (Rethinking the Inception Architecture for Computer Vision)
+
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -1,
+        label_smoothing: float = 0.1,
+        reduction: str = "sum",
+    ) -> None:
+        """
+        Args:
+          ignore_index:
+            ignored class id
+          label_smoothing:
+            smoothing rate (0.0 means the conventional cross entropy loss)
+          reduction:
+            It has the same meaning as the reduction in
+            `torch.nn.CrossEntropyLoss`. It can be one of the following three
+            values: (1) "none": No reduction will be applied. (2) "mean": the
+            mean of the output is taken. (3) "sum": the output will be summed.
+        """
+        super().__init__()
+        assert 0.0 <= label_smoothing < 1.0
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.reduction = reduction
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute loss between x and target.
+
+        Args:
+          x:
+            prediction of dimension
+            (batch_size, input_length, number_of_classes).
+          target:
+            target masked with self.ignore_index of
+            dimension (batch_size, input_length).
+
+        Returns:
+          A scalar tensor containing the loss without normalization.
+        """
+        assert x.ndim == 3
+        assert target.ndim == 2
+        assert x.shape[:2] == target.shape
+        num_classes = x.size(-1)
+        x = x.reshape(-1, num_classes)
+        # Now x is of shape (N*T, C)
+
+        # We don't want to change target in-place below,
+        # so we make a copy of it here
+        target = target.clone().reshape(-1)
+
+        ignored = target == self.ignore_index
+        target[ignored] = 0
+
+        true_dist = torch.nn.functional.one_hot(target, num_classes=num_classes).to(x)
+
+        true_dist = (
+            true_dist * (1 - self.label_smoothing) + self.label_smoothing / num_classes
+        )
+        # Set the value of ignored indexes to 0
+        true_dist[ignored] = 0
+
+        loss = -1 * (torch.log_softmax(x, dim=1) * true_dist)
+        if self.reduction == "sum":
+            return loss.sum()
+        elif self.reduction == "mean":
+            return loss.sum() / (~ignored).sum()
+        else:
+            return loss.sum(dim=-1)
diff --git a/egs/gigaspeech/ASR/conformer_ctc/subsampling.py b/egs/gigaspeech/ASR/conformer_ctc/subsampling.py
new file mode 100644
index 000000000..8e0f73d05
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/subsampling.py
@@ -0,0 +1,153 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x
diff --git a/egs/gigaspeech/ASR/conformer_ctc/train.py b/egs/gigaspeech/ASR/conformer_ctc/train.py
new file mode 100755
index 000000000..4883d04d8
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/train.py
@@ -0,0 +1,731 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.7,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Normalization for the input features, can be a
+                              boolean indicating whether to do batch
+                              normalization, or a float which means just scaling
+                              the input features with this float value.
+                              If given a float value, we will remove batchnorm
+                              layer in `ConvolutionModule` as well.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 500,
+            "reset_interval": 2000,
+            "valid_interval": 30000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            "num_decoder_layers": 6,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "warm_step": 100000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    token_ids = graph_compiler.texts_to_ids(texts)
+
+    decoding_graph = graph_compiler.compile(token_ids)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(42)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    GigaSpeech = GigaSpeechAsrDataModule(args)
+
+    train_cuts = GigaSpeech.train_cuts()
+    train_dl = GigaSpeech.train_dataloaders(train_cuts)
+
+    valid_cuts = GigaSpeech.dev_cuts()
+    valid_dl = GigaSpeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/conformer_ctc/transformer.py b/egs/gigaspeech/ASR/conformer_ctc/transformer.py
new file mode 100644
index 000000000..0566cfc81
--- /dev/null
+++ b/egs/gigaspeech/ASR/conformer_ctc/transformer.py
@@ -0,0 +1,928 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from label_smoothing import LabelSmoothingLoss
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: Union[float, bool] = 0.1,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+            Float value to scale the input layer.
+            False to do nothing.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        assert isinstance(use_feat_batchnorm, (float, bool))
+        if isinstance(use_feat_batchnorm, bool) and use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if isinstance(self.use_feat_batchnorm, bool) and self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        if isinstance(self.use_feat_batchnorm, float):
+            x *= self.use_feat_batchnorm
+        encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision)
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/gigaspeech/ASR/local/__init__.py b/egs/gigaspeech/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/gigaspeech/ASR/local/compile_hlg.py b/egs/gigaspeech/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_dev_test.py b/egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_dev_test.py
new file mode 100755
index 000000000..07beeb1f0
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_dev_test.py
@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_gigaspeech_dev_test():
+    in_out_dir = Path("data/fbank")
+    # number of workers in dataloader
+    num_workers = 20
+
+    # number of seconds in a batch
+    batch_duration = 600
+
+    subsets = ("DEV", "TEST")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    for partition in subsets:
+        cuts_path = in_out_dir / f"cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = in_out_dir / f"cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{in_out_dir}/feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            overwrite=True,
+        )
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_gigaspeech_dev_test()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_splits.py b/egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_splits.py
new file mode 100755
index 000000000..1c71be0f9
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_splits.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the XL subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+    return parser
+
+
+def compute_fbank_gigaspeech_splits(args):
+    num_splits = args.num_splits
+    output_dir = "data/fbank/XL_split"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = 8  # num_digits is fixed by lhotse split-lazy
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"cuts_XL.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"cuts_XL_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/feats_XL_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            overwrite=True,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+
+    log_filename = "log-compute_fbank_gigaspeech_splits"
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    log_filename = f"{log_filename}-{date_time}"
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=logging.INFO,
+        filemode="w",
+    )
+
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter(formatter))
+    logging.getLogger("").addHandler(console)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    compute_fbank_gigaspeech_splits(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/local/compute_fbank_musan.py b/egs/gigaspeech/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/local/convert_transcript_words_to_tokens.py b/egs/gigaspeech/ASR/local/convert_transcript_words_to_tokens.py
new file mode 120000
index 000000000..2ce13fd69
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/convert_transcript_words_to_tokens.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/local/generate_unique_lexicon.py b/egs/gigaspeech/ASR/local/generate_unique_lexicon.py
new file mode 120000
index 000000000..c0aea1403
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/generate_unique_lexicon.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/local/prepare_lang.py b/egs/gigaspeech/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/local/prepare_lang_bpe.py b/egs/gigaspeech/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py b/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py
new file mode 100755
index 000000000..31abe7fff
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py
@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import re
+from pathlib import Path
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Similar text filtering and normalization procedure as in:
+# https://github.com/SpeechColab/GigaSpeech/blob/main/toolkits/kaldi/gigaspeech_data_prep.sh
+
+
+def normalize_text(
+    utt: str,
+    punct_pattern=re.compile(r"<(COMMA|PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    whitespace_pattern=re.compile(r"\s\s+"),
+) -> str:
+    return whitespace_pattern.sub(" ", punct_pattern.sub("", utt))
+
+
+def has_no_oov(
+    sup: SupervisionSegment,
+    oov_pattern=re.compile(r"<(SIL|MUSIC|NOISE|OTHER)>"),
+) -> bool:
+    return oov_pattern.search(sup.text) is None
+
+
+def preprocess_giga_speech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    dataset_parts = (
+        "DEV",
+        "TEST",
+        "XL",
+    )
+
+    logging.info("Loading manifest (may take 4 minutes)")
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix="gigaspeech",
+        suffix="jsonl.gz",
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"cuts_{partition}_raw.jsonl.gz"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        # Note this step makes the recipe different than LibriSpeech:
+        # We must filter out some utterances and remove punctuation
+        # to be consistent with Kaldi.
+        logging.info("Filtering OOV utterances from supervisions")
+        m["supervisions"] = m["supervisions"].filter(has_no_oov)
+        logging.info(f"Normalizing text in {partition}")
+        for sup in m["supervisions"]:
+            sup.text = normalize_text(sup.text)
+
+        # Create long-recording cut manifests.
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        )
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        if partition not in ["DEV", "TEST"]:
+            logging.info(
+                f"Speed perturb for {partition} with factors 0.9 and 1.1 "
+                "(Perturbing may take 8 minutes and saving may take 20 minutes)"
+            )
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    preprocess_giga_speech()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/local/train_bpe_model.py b/egs/gigaspeech/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/gigaspeech/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/prepare.sh b/egs/gigaspeech/ASR/prepare.sh
new file mode 100755
index 000000000..a23b708d7
--- /dev/null
+++ b/egs/gigaspeech/ASR/prepare.sh
@@ -0,0 +1,328 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=0
+stop_stage=100
+
+# Split XL subset to a number of pieces (about 2000)
+# This is to avoid OOM during feature extraction.
+num_per_split=50
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/GigaSpeech
+#      You can find audio, dict, GigaSpeech.json inside it.
+#      You can apply for the download credentials by following
+#      https://github.com/SpeechColab/GigaSpeech#download
+#
+#  - $dl_dir/lm
+#      This directory contains the language model downloaded from
+#        https://huggingface.co/wgb14/gigaspeech_lm
+#
+#        - 3gram_pruned_1e7.arpa.gz
+#        - 4gram.arpa.gz
+#        - lexicon.txt
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "stage -1: Download LM"
+  # We assume that you have installed the git-lfs, if not, you could install it
+  # using: `sudo apt-get install git-lfs && git-lfs install`
+  [ ! -e $dl_dir/lm ] && mkdir -p $dl_dir/lm
+  git clone https://huggingface.co/wgb14/gigaspeech_lm $dl_dir/lm
+  gunzip -c $dl_dir/lm/3gram_pruned_1e7.arpa.gz > $dl_dir/lm/3gram_pruned_1e7.arpa
+  gunzip -c $dl_dir/lm/4gram.arpa.gz > $dl_dir/lm/4gram.arpa
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  [ ! -e $dl_dir/GigaSpeech ] && mkdir -p $dl_dir/GigaSpeech
+
+  # If you have pre-downloaded it to /path/to/GigaSpeech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/GigaSpeech $dl_dir/GigaSpeech
+  #
+  if [ ! -d $dl_dir/GigaSpeech/audio ] && [ ! -f $dl_dir/GigaSpeech.json ]; then
+    # Check credentials.
+    if [ ! -f $dl_dir/password ]; then
+      echo -n "$0: Please apply for the download credentials by following"
+      echo -n "https://github.com/SpeechColab/GigaSpeech#download"
+      echo " and save it to $dl_dir/password."
+      exit 1;
+    fi
+    PASSWORD=`cat $dl_dir/password 2>/dev/null`
+    if [ -z "$PASSWORD" ]; then
+      echo "$0: Error, $dl_dir/password is empty."
+      exit 1;
+    fi
+    PASSWORD_MD5=`echo $PASSWORD | md5sum | cut -d ' ' -f 1`
+    if [[ $PASSWORD_MD5 != "dfbf0cde1a3ce23749d8d81e492741b8" ]]; then
+      echo "$0: Error, invalid $dl_dir/password."
+      exit 1;
+    fi
+    # Download XL, DEV and TEST sets by default.
+    lhotse download gigaspeech --subset auto --host tsinghua \
+      $dl_dir/password $dl_dir/GigaSpeech
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare GigaSpeech manifest (may take 15 minutes)"
+  # We assume that you have downloaded the GigaSpeech corpus
+  # to $dl_dir/GigaSpeech
+  mkdir -p data/manifests
+  lhotse prepare gigaspeech --subset auto -j $nj \
+    $dl_dir/GigaSpeech data/manifests
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "State 3: Preprocess GigaSpeech manifest"
+  if [ ! -f data/fbank/.preprocess_complete ]; then
+   python3 ./local/preprocess_gigaspeech.py
+   touch data/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute features for DEV and TEST subsets of GigaSpeech (may take 2 minutes)"
+  python3 ./local/compute_fbank_gigaspeech_dev_test.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Split XL subset into pieces (may take 30 minutes)"
+  split_dir=data/fbank/XL_split
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split-lazy ./data/fbank/cuts_XL_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compute features for XL"
+  num_splits=$(find data/fbank/XL_split -name "cuts_XL_raw.*.jsonl.gz" | wc -l)
+  python3 ./local/compute_fbank_gigaspeech_splits.py \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Combine features for XL (may take 3 hours)"
+  if [ ! -f data/fbank/cuts_XL.jsonl.gz ]; then
+    pieces=$(find data/fbank/XL_split -name "cuts_XL.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_XL.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compute fbank for musan"
+  mkdir -p data/fbank
+  ./local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Prepare phone based lang"
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/lm/lexicon.txt |
+    sort | uniq > $lang_dir/lexicon.txt
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+
+  if [ ! -f $lang_dir/transcript_words.txt ]; then
+    gunzip -c "data/manifests/gigaspeech_supervisions_XL.jsonl.gz" \
+      | jq '.text' \
+      | sed 's/"//g' \
+      > $lang_dir/transcript_words.txt
+
+    # Delete utterances with garbage meta tags
+    garbage_utterance_tags="<SIL> <MUSIC> <NOISE> <OTHER>"
+    for tag in $garbage_utterance_tags; do
+      sed -i "/${tag}/d" $lang_dir/transcript_words.txt
+    done
+
+    # Delete punctuations in utterances
+    punctuation_tags="<COMMA> <EXCLAMATIONPOINT> <PERIOD> <QUESTIONMARK>"
+    for tag in $punctuation_tags; do
+      sed -i "s/${tag}//g" $lang_dir/transcript_words.txt
+    done
+
+    # Ensure space only appears once
+    sed -i 's/\t/ /g' $lang_dir/transcript_words.txt
+    sed -i 's/[ ][ ]*/ /g' $lang_dir/transcript_words.txt
+  fi
+
+  cat $lang_dir/transcript_words.txt | sed 's/ /\n/g' \
+    | sort -u | sed '/^$/d' > $lang_dir/words.txt
+  (echo '!SIL'; echo '<SPOKEN_NOISE>'; echo '<UNK>'; ) |
+    cat - $lang_dir/words.txt | sort | uniq | awk '
+    BEGIN {
+      print "<eps> 0";
+    }
+    {
+      if ($1 == "<s>") {
+        print "<s> is in the vocabulary!" | "cat 1>&2"
+        exit 1;
+      }
+      if ($1 == "</s>") {
+        print "</s> is in the vocabulary!" | "cat 1>&2"
+        exit 1;
+      }
+      printf("%s %d\n", $1, NR);
+    }
+    END {
+      printf("#0 %d\n", NR+1);
+      printf("<s> %d\n", NR+2);
+      printf("</s> %d\n", NR+3);
+    }' > $lang_dir/words || exit 1;
+  mv $lang_dir/words $lang_dir/words.txt
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang_phone/{words.txt,transcript_words.txt} $lang_dir
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+    fi
+  done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Prepare bigram P"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+      ./local/convert_transcript_words_to_tokens.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --transcript $lang_dir/transcript_words.txt \
+        --oov "<UNK>" \
+        > $lang_dir/transcript_tokens.txt
+    fi
+
+    if [ ! -f $lang_dir/P.arpa ]; then
+      ./shared/make_kn_lm.py \
+        -ngram-order 2 \
+        -text $lang_dir/transcript_tokens.txt \
+        -lm $lang_dir/P.arpa
+    fi
+
+    if [ ! -f $lang_dir/P.fst.txt ]; then
+      python3 -m kaldilm \
+        --read-symbol-table="$lang_dir/tokens.txt" \
+        --disambig-symbol='#0' \
+        --max-order=2 \
+        $lang_dir/P.arpa > $lang_dir/P.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $dl_dir/lm/3gram_pruned_1e7.arpa > data/lm/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      $dl_dir/lm/4gram.arpa > data/lm/G_4_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+  done
+fi
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/__init__.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 100644
index 000000000..b5b27ce95
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1,410 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class GigaSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+
+        # GigaSpeech specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="XL",
+            help="Select the GigaSpeech subset (XS|S|M|L|XL)",
+        )
+        group.add_argument(
+            "--small-dev",
+            type=str2bool,
+            default=False,
+            help="Should we use only 1000 utterances for dev (speeds up training)",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get train_{self.args.subset} cuts")
+        path = self.args.manifest_dir / f"gigaspeech_cuts_{self.args.subset}.jsonl.gz"
+        cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "gigaspeech_cuts_DEV.jsonl.gz"
+        )
+        if self.args.small_dev:
+            return cuts_valid.subset(first=1000)
+        else:
+            return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "gigaspeech_cuts_TEST.jsonl.gz"
+        )
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/compute_ppl.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/compute_ppl.py
new file mode 100755
index 000000000..76306fc4c
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/compute_ppl.py
@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corp. (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+./pruned_transducer_stateless7/compute_ppl.py \
+    --ngram-lm-path ./download/lm/3gram_pruned_1e7.arpa
+
+"""
+
+
+import argparse
+import logging
+import math
+from typing import Dict, List, Optional, Tuple
+
+import kenlm
+import torch
+from asr_datamodule import GigaSpeechAsrDataModule
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--ngram-lm-path",
+        type=str,
+        default="download/lm/3gram_pruned_1e7.arpa",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    return parser
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: kenlm.Model,
+) -> Dict[str, float]:
+    """
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        A ngram lm of kenlm.Model object.
+    Returns:
+      Return the perplexity of the giving dataset.
+    """
+    sum_score_log = 0
+    sum_n = 0
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        for text in texts:
+            sum_n += len(text.split()) + 1
+            sum_score_log += -1 * model.score(text)
+
+    ppl = math.pow(10.0, sum_score_log / sum_n)
+
+    return ppl
+
+
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    logging.info("About to load ngram LM")
+    model = kenlm.Model(args.ngram_lm_path)
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        ppl = decode_dataset(
+            dl=test_dl,
+            model=model,
+        )
+        logging.info(f"{test_set} PPL: {ppl}")
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/conformer.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/conformer.py
new file mode 120000
index 000000000..a65957180
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..72f74c968
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,628 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from gigaspeech_scoring import asr_text_post_processing
+from train import get_params, get_transducer_model
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=8,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = post_processing(results)
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/decoder.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py
new file mode 100755
index 000000000..b6190e8a6
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,209 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/gigaspeech_scoring.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/gigaspeech_scoring.py
new file mode 120000
index 000000000..a6a4d12b1
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/gigaspeech_scoring.py
@@ -0,0 +1 @@
+../conformer_ctc/gigaspeech_scoring.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/joiner.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/model.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/optim.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/scaling.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/train.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/train.py
new file mode 100755
index 000000000..a7772b62f
--- /dev/null
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,989 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --max-duration 120
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --use_fp16 1 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --max-duration 200
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 500,
+            "reset_interval": 2000,
+            "valid_interval": 20000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 20000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                model_avg=model_avg,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    train_cuts = gigaspeech.train_cuts()
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = gigaspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = gigaspeech.dev_cuts()
+    valid_dl = gigaspeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/shared b/egs/gigaspeech/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/gigaspeech/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/asr_datamodule.py b/egs/gigaspeech/ASR/zipformer/asr_datamodule.py
new file mode 100644
index 000000000..6adfdbfbb
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1,444 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2023  Xiaomi Corporation     (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import glob
+import inspect
+import logging
+import re
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import lhotse
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import AudioSamples, OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class GigaSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+        # GigaSpeech specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="XL",
+            help="Select the GigaSpeech subset (XS|S|M|L|XL)",
+        )
+        group.add_argument(
+            "--small-dev",
+            type=str2bool,
+            default=False,
+            help="Should we use only 1000 utterances for dev (speeds up training)",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get train {self.args.subset} cuts")
+        if self.args.subset == "XL":
+            filenames = glob.glob(
+                f"{self.args.manifest_dir}/XL_split/gigaspeech_cuts_XL.*.jsonl.gz"
+            )
+            pattern = re.compile(r"gigaspeech_cuts_XL.([0-9]+).jsonl.gz")
+            idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames)
+            idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
+            sorted_filenames = [f[1] for f in idx_filenames]
+            logging.info(
+                f"Loading GigaSpeech {len(sorted_filenames)} splits in lazy mode"
+            )
+            cuts_train = lhotse.combine(
+                lhotse.load_manifest_lazy(p) for p in sorted_filenames
+            )
+        else:
+            path = (
+                self.args.manifest_dir / f"gigaspeech_cuts_{self.args.subset}.jsonl.gz"
+            )
+            cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "gigaspeech_cuts_DEV.jsonl.gz"
+        )
+        if self.args.small_dev:
+            return cuts_valid.subset(first=1000)
+        else:
+            return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "gigaspeech_cuts_TEST.jsonl.gz"
+        )
diff --git a/egs/gigaspeech/ASR/zipformer/beam_search.py b/egs/gigaspeech/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/ctc_decode.py b/egs/gigaspeech/ASR/zipformer/ctc_decode.py
new file mode 100755
index 000000000..aa51036d5
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/ctc_decode.py
@@ -0,0 +1,847 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) ctc-decoding
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) 1best
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method 1best
+
+(3) nbest
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method nbest
+
+(4) nbest-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+
+(5) whole-lattice-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.6,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    test_clean_cuts = gigaspeech.test_clean_cuts()
+    test_other_cuts = gigaspeech.test_other_cuts()
+
+    test_clean_dl = gigaspeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = gigaspeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/decode.py b/egs/gigaspeech/ASR/zipformer/decode.py
new file mode 100755
index 000000000..3a0c71484
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/decode.py
@@ -0,0 +1,1065 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_rescore,
+    modified_beam_search_lm_rescore_LODR,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+)
+from gigaspeech_scoring import asr_text_post_processing
+from train import add_model_arguments, get_model, get_params
+
+from icefall import ContextGraph, LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - modified_beam_search_LODR
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding-method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding-method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=2,
+        help="""The order of the ngram lm.
+        """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="ID of the backoff symbol in the ngram LM",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding-method is modified_beam_search and
+        modified_beam_search_LODR.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding-method is modified_beam_search and
+        modified_beam_search_LODR.
+        """,
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding-method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network language model.
+      ngram_lm:
+        A ngram language model
+      ngram_lm_scale:
+        The scale for the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+            context_graph=context_graph,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_rescore":
+        lm_scale_list = [0.01 * i for i in range(10, 50)]
+        ans_dict = modified_beam_search_lm_rescore(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        lm_scale_list = [0.02 * i for i in range(2, 30)]
+        ans_dict = modified_beam_search_lm_rescore_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            LODR_lm=ngram_lm,
+            sp=sp,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    elif "modified_beam_search" in params.decoding_method:
+        prefix = f"beam_size_{params.beam_size}"
+        if params.decoding_method in (
+            "modified_beam_search_lm_rescore",
+            "modified_beam_search_lm_rescore_LODR",
+        ):
+            ans = dict()
+            assert ans_dict is not None
+            for key, hyps in ans_dict.items():
+                hyps = [sp.decode(hyp).split() for hyp in hyps]
+                ans[f"{prefix}_{key}"] = hyps
+            return ans
+        else:
+            if params.has_contexts:
+                prefix += f"-context-score-{params.context_score}"
+            return {prefix: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding-method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            word_table=word_table,
+            batch=batch,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = post_processing(results)
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.decoding_method in (
+            "modified_beam_search",
+            "modified_beam_search_LODR",
+        ):
+            if params.has_contexts:
+                params.suffix += f"-context-score-{params.context_score}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_shallow_fusion:
+        params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load the neural network LM if required
+    if params.use_shallow_fusion or params.decoding_method in (
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    ):
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+    else:
+        LM = None
+
+    # only load N-gram LM when needed
+    if params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        try:
+            import kenlm
+        except ImportError:
+            print("Please install kenlm first. You can use")
+            print(" pip install https://github.com/kpu/kenlm/archive/master.zip")
+            print("to install it")
+            import sys
+
+            sys.exit(-1)
+        ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa")
+        logging.info(f"lm filename: {ngram_file_name}")
+        ngram_lm = kenlm.Model(ngram_file_name)
+        ngram_lm_scale = None  # use a list to search
+
+    elif params.decoding_method == "modified_beam_search_LODR":
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"Loading token level lm: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    if "modified_beam_search" in params.decoding_method:
+        if os.path.exists(params.context_file):
+            contexts = []
+            for line in open(params.context_file).readlines():
+                contexts.append(line.strip())
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(sp.encode(contexts))
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/decode_stream.py b/egs/gigaspeech/ASR/zipformer/decode_stream.py
new file mode 120000
index 000000000..b8d8ddfc4
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decode_stream.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/decoder.py b/egs/gigaspeech/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/encoder_interface.py b/egs/gigaspeech/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/export-onnx-ctc.py b/egs/gigaspeech/ASR/zipformer/export-onnx-ctc.py
new file mode 120000
index 000000000..f9d756352
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export-onnx-ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/export-onnx-streaming.py b/egs/gigaspeech/ASR/zipformer/export-onnx-streaming.py
new file mode 120000
index 000000000..2962eb784
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/export-onnx.py b/egs/gigaspeech/ASR/zipformer/export-onnx.py
new file mode 100755
index 000000000..0f78cfe5b
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export-onnx.py
@@ -0,0 +1,620 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang)
+# Copyright 2023 Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/gigaspeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal False \
+  --chunk-size "16,32,64,-1" \
+  --left-context-frames "64,128,256,-1"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(
+        self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        x, x_lens = self.encoder_embed(x, x_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    encoder_model = torch.jit.trace(encoder_model, (x, x_lens))
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer2",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "non-streaming zipformer2",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/export.py b/egs/gigaspeech/ASR/zipformer/export.py
new file mode 100755
index 000000000..e45c96b57
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export.py
@@ -0,0 +1,522 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+Note: This is a example for gigaspeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+(1) Export to torchscript model using torch.jit.script()
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `jit_script.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("jit_script.pt")`.
+
+Check ./jit_pretrained.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`.
+You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`.
+
+Check ./jit_pretrained_streaming.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+- For non-streaming model:
+
+To use the generated file with `zipformer/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/gigaspeech/ASR
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+- For streaming model:
+
+To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/gigaspeech/ASR
+
+    # simulated streaming decoding
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+    # chunk-wise streaming decoding
+    ./zipformer/streaming_decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+- non-streaming model:
+https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+    # You will find the pre-trained models in exp dir
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor, nn
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class EncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+        """
+        x, x_lens = self.encoder_embed(features, feature_lengths)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+class StreamingEncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        assert len(encoder.chunk_size) == 1, encoder.chunk_size
+        assert len(encoder.left_context_frames) == 1, encoder.left_context_frames
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+
+        # The encoder_embed subsample features (T - 7) // 2
+        # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+        self.pad_length = 7 + 2 * 3
+
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor, states: List[Tensor]
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """Streaming forward for encoder_embed and encoder.
+
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+            states: a list of Tensors
+
+        Returns encoder outputs, output lengths, and updated states.
+        """
+        chunk_size = self.chunk_size
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        encoder_states = states[:-2]
+
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+        return encoder_out, encoder_out_lens, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+
+        # Wrap encoder and encoder_embed as a module
+        if params.causal:
+            model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed)
+            chunk_size = model.encoder.chunk_size
+            left_context_len = model.encoder.left_context_len
+            filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt"
+        else:
+            model.encoder = EncoderModel(model.encoder, model.encoder_embed)
+            filename = "jit_script.pt"
+
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        model.save(str(params.exp_dir / filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/gigaspeech_scoring.py b/egs/gigaspeech/ASR/zipformer/gigaspeech_scoring.py
new file mode 120000
index 000000000..a6a4d12b1
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/gigaspeech_scoring.py
@@ -0,0 +1 @@
+../conformer_ctc/gigaspeech_scoring.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/jit_pretrained.py b/egs/gigaspeech/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/jit_pretrained_ctc.py b/egs/gigaspeech/ASR/zipformer/jit_pretrained_ctc.py
new file mode 120000
index 000000000..9a8da5844
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/jit_pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/jit_pretrained_streaming.py b/egs/gigaspeech/ASR/zipformer/jit_pretrained_streaming.py
new file mode 120000
index 000000000..1962351e9
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/joiner.py b/egs/gigaspeech/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/model.py b/egs/gigaspeech/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_check.py b/egs/gigaspeech/ASR/zipformer/onnx_check.py
new file mode 120000
index 000000000..f3dd42004
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_check.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_decode.py b/egs/gigaspeech/ASR/zipformer/onnx_decode.py
new file mode 120000
index 000000000..0573b88c5
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_decode.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained-streaming.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 120000
index 000000000..cfea104c2
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc.py
new file mode 120000
index 000000000..a3183ebf6
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_H.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_H.py
new file mode 120000
index 000000000..a4fd76ac2
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_H.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HL.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
new file mode 120000
index 000000000..f805e3761
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
new file mode 120000
index 000000000..8343d5079
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/optim.py b/egs/gigaspeech/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/pretrained.py b/egs/gigaspeech/ASR/zipformer/pretrained.py
new file mode 120000
index 000000000..0bd71dde4
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/pretrained_ctc.py b/egs/gigaspeech/ASR/zipformer/pretrained_ctc.py
new file mode 120000
index 000000000..c2f6f6fc3
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/profile.py b/egs/gigaspeech/ASR/zipformer/profile.py
new file mode 120000
index 000000000..c93adbd14
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/profile.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/profile.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/scaling.py b/egs/gigaspeech/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/scaling_converter.py b/egs/gigaspeech/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/streaming_beam_search.py b/egs/gigaspeech/ASR/zipformer/streaming_beam_search.py
new file mode 120000
index 000000000..b1ed54557
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/streaming_decode.py b/egs/gigaspeech/ASR/zipformer/streaming_decode.py
new file mode 100755
index 000000000..a76788859
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/streaming_decode.py
@@ -0,0 +1,853 @@
+#!/usr/bin/env python3
+# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
+#                                                  Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 32 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+from asr_datamodule import GigaSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch import Tensor, nn
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_init_states(
+    model: nn.Module,
+    batch_size: int = 1,
+    device: torch.device = torch.device("cpu"),
+) -> List[torch.Tensor]:
+    """
+    Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+    is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+    states[-2] is the cached left padding for ConvNeXt module,
+    of shape (batch_size, num_channels, left_pad, num_freqs)
+    states[-1] is processed_lens of shape (batch,), which records the number
+    of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+    """
+    states = model.encoder.get_init_states(batch_size, device)
+
+    embed_states = model.encoder_embed.get_init_states(batch_size, device)
+    states.append(embed_states)
+
+    processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+    states.append(processed_lens)
+
+    return states
+
+
+def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance. For element-n,
+        state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
+        state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
+        cached_val2, cached_conv1, cached_conv2).
+        state_list[n][-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+        state_list[n][-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+    """
+    batch_size = len(state_list)
+    assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
+    tot_num_layers = (len(state_list[0]) - 2) // 6
+
+    batch_states = []
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key = torch.cat(
+            [state_list[i][layer_offset] for i in range(batch_size)], dim=1
+        )
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn = torch.cat(
+            [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1 = torch.cat(
+            [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2 = torch.cat(
+            [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1 = torch.cat(
+            [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2 = torch.cat(
+            [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
+        )
+        batch_states += [
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ]
+
+    cached_embed_left_pad = torch.cat(
+        [state_list[i][-2] for i in range(batch_size)], dim=0
+    )
+    batch_states.append(cached_embed_left_pad)
+
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
+    batch_states.append(processed_lens)
+
+    return batch_states
+
+
+def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+        batch_states: A list of cached tensors of all encoder layers. For layer-i,
+          states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+          cached_conv1, cached_conv2).
+          state_list[-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+          states[-1] is processed_lens of shape (batch,), which records the number
+          of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Returns:
+        state_list: A list of list. Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+    """
+    assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
+    tot_num_layers = (len(batch_states) - 2) // 6
+
+    processed_lens = batch_states[-1]
+    batch_size = processed_lens.shape[0]
+
+    state_list = [[] for _ in range(batch_size)]
+
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1_list = batch_states[layer_offset + 2].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2_list = batch_states[layer_offset + 3].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1_list = batch_states[layer_offset + 4].chunk(
+            chunks=batch_size, dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2_list = batch_states[layer_offset + 5].chunk(
+            chunks=batch_size, dim=0
+        )
+        for i in range(batch_size):
+            state_list[i] += [
+                cached_key_list[i],
+                cached_nonlin_attn_list[i],
+                cached_val1_list[i],
+                cached_val2_list[i],
+                cached_conv1_list[i],
+                cached_conv2_list[i],
+            ]
+
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(cached_embed_left_pad_list[i])
+
+    processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(processed_lens_list[i])
+
+    return state_list
+
+
+def streaming_forward(
+    features: Tensor,
+    feature_lens: Tensor,
+    model: nn.Module,
+    states: List[Tensor],
+    chunk_size: int,
+    left_context_len: int,
+) -> Tuple[Tensor, Tensor, List[Tensor]]:
+    """
+    Returns encoder outputs, output lengths, and updated states.
+    """
+    cached_embed_left_pad = states[-2]
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        cached_left_pad=cached_embed_left_pad,
+    )
+    assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    # processed_mask is used to mask out initial states
+    processed_mask = torch.arange(left_context_len, device=x.device).expand(
+        x.size(0), left_context_len
+    )
+    processed_lens = states[-1]  # (batch,)
+    # (batch, left_context_size)
+    processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+    # Update processed lengths
+    new_processed_lens = processed_lens + x_lens
+
+    # (batch, left_context_size + chunk_size)
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    encoder_states = states[:-2]
+    (
+        encoder_out,
+        encoder_out_lens,
+        new_encoder_states,
+    ) = model.encoder.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=encoder_states,
+        src_key_padding_mask=src_key_padding_mask,
+    )
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    new_states = new_encoder_states + [
+        new_cached_embed_left_pad,
+        new_processed_lens,
+    ]
+    return encoder_out, encoder_out_lens, new_states
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+    chunk_size = int(params.chunk_size)
+    left_context_len = int(params.left_context_frames)
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []  # Used in fast-beam-search
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(chunk_size * 2)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # Make sure the length after encoder_embed is at least 1.
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    tail_length = chunk_size * 2 + 7 + 2 * 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+
+    encoder_out, encoder_out_lens, new_states = streaming_forward(
+        features=features,
+        feature_lens=feature_lens,
+        model=model,
+        states=states,
+        chunk_size=chunk_size,
+        left_context_len=left_context_len,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = torch.tensor(processed_lens, device=device)
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = get_init_states(model=model, batch_size=1, device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=30)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    assert params.causal, params.causal
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
+    assert (
+        "," not in params.left_context_frames
+    ), "left_context_frames should be one value in decoding."
+    params.suffix += f"-chunk-{params.chunk_size}"
+    params.suffix += f"-left-context-{params.left_context_frames}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    test_sets = ["dev", "test"]
+    test_cuts = [dev_cuts, test_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/subsampling.py b/egs/gigaspeech/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/test_scaling.py b/egs/gigaspeech/ASR/zipformer/test_scaling.py
new file mode 120000
index 000000000..715798436
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/test_scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/test_scaling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/test_subsampling.py b/egs/gigaspeech/ASR/zipformer/test_subsampling.py
new file mode 120000
index 000000000..bf0ee3d11
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/test_subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/test_subsampling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/train.py b/egs/gigaspeech/ASR/zipformer/train.py
new file mode 100755
index 000000000..d93cc221c
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/train.py
@@ -0,0 +1,1345 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Yifan Yang,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=1,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 500,
+            "reset_interval": 2000,
+            "valid_interval": 20000,
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}"
+    )
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        loss = 0.0
+
+        if params.use_transducer:
+            s = params.simple_loss_scale
+            # take down the scale on the simple loss from 1.0 at the start
+            # to params.simple_loss scale by warm_step.
+            simple_loss_scale = (
+                s
+                if batch_idx_train >= warm_step
+                else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+            )
+            pruned_loss_scale = (
+                1.0
+                if batch_idx_train >= warm_step
+                else 0.1 + 0.9 * (batch_idx_train / warm_step)
+            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    if params.use_transducer:
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    train_cuts = gigaspeech.train_cuts()
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = gigaspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = gigaspeech.dev_cuts()
+    valid_dl = gigaspeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/zipformer.py b/egs/gigaspeech/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/README.md b/egs/libricss/SURT/README.md
new file mode 100644
index 000000000..10a1aaad1
--- /dev/null
+++ b/egs/libricss/SURT/README.md
@@ -0,0 +1,249 @@
+# Introduction
+
+This is a multi-talker ASR recipe for the LibriCSS dataset. We train a Streaming
+Unmixing and Recognition Transducer (SURT) model for the task. In this README,
+we will describe the task, the model, and the training process. We will also
+provide links to pre-trained models and training logs.
+
+## Task
+
+LibriCSS is a multi-talker meeting corpus formed from mixing together LibriSpeech utterances
+and replaying in a real meeting room. It consists of 10 1-hour sessions of audio, each
+recorded on a 7-channel microphone. The sessions are recorded at a sampling rate of 16 kHz.
+For more information, refer to the paper:
+Z. Chen et al., "Continuous speech separation: dataset and analysis,"
+ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
+Barcelona, Spain, 2020
+
+In this recipe, we perform the "continuous, streaming, multi-talker ASR" task on LibriCSS.
+
+* By "continuous", we mean that the model should be able to transcribe unsegmented audio
+without the need of an external VAD.
+* By "streaming", we mean that the model has limited right context. We use a right-context
+of at most 32 frames (320 ms).
+* By "multi-talker", we mean that the model should be able to transcribe overlapping speech
+from multiple speakers.
+
+For now, we do not care about speaker attribution, i.e., the transcription is speaker
+agnostic. The evaluation depends on the particular model type. In this case, we use
+the optimal reference combination WER (ORC-WER) metric as implemented in the
+[meeteval](https://github.com/fgnt/meeteval) toolkit.
+
+## Model
+
+We use the Streaming Unmixing and Recognition Transducer (SURT) model for this task.
+The model is based on the papers:
+
+- Lu, Liang et al. “Streaming End-to-End Multi-Talker Speech Recognition.” IEEE Signal Processing Letters 28 (2020): 803-807.
+- Raj, Desh et al. “Continuous Streaming Multi-Talker ASR with Dual-Path Transducers.” ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2021): 7317-7321.
+
+The model is a combination of a speech separation model and a speech recognition model,
+but trained end-to-end with a single loss function. The overall architecture is shown
+in the figure below. Note that this architecture is slightly different from the one
+in the above papers. A detailed description of the model can be found in the following
+paper: [SURT 2.0: Advanced in transducer-based multi-talker ASR](https://arxiv.org/abs/2306.10559).
+
+<p align="center">
+
+  <img src="surt.png">
+  Streaming Unmixing and Recognition Transducer
+
+</p>
+
+In the [dprnn_zipformer](./dprnn_zipformer) recipe, for example, we use a DPRNN-based masking network
+and a Zipfomer-based recognition network. But other combinations are possible as well.
+
+## Training objective
+
+We train the model using the pruned transducer loss, similar to other ASR recipes in
+icefall. However, an important consideration is how to assign references to the output
+channels (2 in this case). For this, we use the heuristic error assignment training (HEAT)
+strategy, which assigns references to the first available channel based on their start
+times. An illustrative example is shown in the figure below:
+
+<p align="center">
+
+  <img src="heat.png">
+  Illustration of HEAT-based reference assignment.
+
+</p>
+
+## Description of the recipe
+
+### Pre-requisites
+
+The recipes in this directory need the following packages to be installed:
+
+- [meeteval](https://github.com/fgnt/meeteval)
+- [einops](https://github.com/arogozhnikov/einops)
+
+Additionally, we initialize the "recognition" transducer with a pre-trained model,
+trained on LibriSpeech. For this, please run the following from within `egs/librispeech/ASR`:
+
+```bash
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+python pruned_transducer_stateless7_streaming/train.py \
+    --use-fp16 True \
+    --exp-dir pruned_transducer_stateless7_streaming/exp \
+    --world-size 4 \
+    --max-duration 800 \
+    --num-epochs 10 \
+    --keep-last-k 1 \
+    --manifest-dir data/manifests \
+    --enable-musan true \
+    --master-port 54321 \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --num-encoder-layers 2,2,2,2,2 \
+    --feedforward-dims 768,768,768,768,768 \
+    --nhead 8,8,8,8,8 \
+    --encoder-dims 256,256,256,256,256 \
+    --attention-dims 192,192,192,192,192 \
+    --encoder-unmasked-dims 192,192,192,192,192 \
+    --zipformer-downsampling-factors 1,2,4,8,2 \
+    --cnn-module-kernels 31,31,31,31,31 \
+    --decoder-dim 512 \
+    --joiner-dim 512
+```
+
+The above is for SURT-base (~26M). For SURT-large (~38M), use `--num-encoder-layers 2,4,3,2,4`.
+
+Once the above model is trained for 10 epochs, copy it to `egs/libricss/SURT/exp`:
+
+```bash
+cp -r pruned_transducer_stateless7_streaming/exp/epoch-10.pt exp/zipformer_base.pt
+```
+
+**NOTE:** We also provide this pre-trained checkpoint (see the section below), so you can skip
+the above step if you want.
+
+### Training
+
+To train the model, run the following from within `egs/libricss/SURT`:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+python dprnn_zipformer/train.py \
+    --use-fp16 True \
+    --exp-dir dprnn_zipformer/exp/surt_base \
+    --world-size 4 \
+    --max-duration 500 \
+    --max-duration-valid 250 \
+    --max-cuts 200 \
+    --num-buckets 50 \
+    --num-epochs 30 \
+    --enable-spec-aug True \
+    --enable-musan False \
+    --ctc-loss-scale 0.2 \
+    --heat-loss-scale 0.2 \
+    --base-lr 0.004 \
+    --model-init-ckpt exp/zipformer_base.pt \
+    --chunk-width-randomization True \
+    --num-mask-encoder-layers 4 \
+    --num-encoder-layers 2,2,2,2,2
+```
+
+The above is for SURT-base (~26M). For SURT-large (~38M), use:
+
+```bash
+    --num-mask-encoder-layers 6 \
+    --num-encoder-layers 2,4,3,2,4 \
+    --model-init-ckpt exp/zipformer_large.pt \
+```
+
+**NOTE:** You may need to decrease the `--max-duration` for SURT-large to avoid OOM.
+
+### Adaptation
+
+The training step above only trains on simulated mixtures. For best results, we also
+adapt the final model on the LibriCSS dev set. For this, run the following from within
+`egs/libricss/SURT`:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+
+python dprnn_zipformer/train_adapt.py \
+    --use-fp16 True \
+    --exp-dir dprnn_zipformer/exp/surt_base_adapt \
+    --world-size 1 \
+    --max-duration 500 \
+    --max-duration-valid 250 \
+    --max-cuts 200 \
+    --num-buckets 50 \
+    --num-epochs 8 \
+    --lr-epochs 2 \
+    --enable-spec-aug True \
+    --enable-musan False \
+    --ctc-loss-scale 0.2 \
+    --base-lr 0.0004 \
+    --model-init-ckpt dprnn_zipformer/exp/surt_base/epoch-30.pt \
+    --chunk-width-randomization True \
+    --num-mask-encoder-layers 4 \
+    --num-encoder-layers 2,2,2,2,2
+```
+
+For SURT-large, use the following config:
+
+```bash
+    --num-mask-encoder-layers 6 \
+    --num-encoder-layers 2,4,3,2,4 \
+    --model-init-ckpt dprnn_zipformer/exp/surt_large/epoch-30.pt \
+    --num-epochs 15 \
+    --lr-epochs 4 \
+```
+
+
+### Decoding
+
+To decode the model, run the following from within `egs/libricss/SURT`:
+
+#### Greedy search
+
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+
+python dprnn_zipformer/decode.py \
+    --epoch 8 --avg 1 --use-averaged-model False \
+    --exp-dir dprnn_zipformer/exp/surt_base_adapt \
+    --max-duration 250 \
+    --decoding-method greedy_search
+```
+
+#### Beam search
+
+```bash
+python dprnn_zipformer/decode.py \
+    --epoch 8 --avg 1 --use-averaged-model False \
+    --exp-dir dprnn_zipformer/exp/surt_base_adapt \
+    --max-duration 250 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+```
+
+## Results (using beam search)
+
+#### IHM-Mix
+
+| Model | # params | 0L | 0S | OV10 | OV20 | OV30 | OV40 | Avg. |
+|------------|:-------:|:----:|:---:|----:|:----:|:----:|:----:|:----:|
+| dprnn_zipformer (base)  | 26.7 | 5.1 | 4.2 | 13.7 | 18.7 | 20.5 | 20.6 | 13.8 |
+| dprnn_zipformer (large) | 37.9 | 4.6 | 3.8 | 12.7 | 14.3 | 16.7 | 21.2 | 12.2 |
+
+#### SDM
+
+| Model | # params | 0L | 0S | OV10 | OV20 | OV30 | OV40 | Avg. |
+|------------|:-------:|:----:|:---:|----:|:----:|:----:|:----:|:----:|
+| dprnn_zipformer (base)  | 26.7 | 6.8 | 7.2 | 21.4 | 24.5 | 28.6 | 31.2 | 20.0 |
+| dprnn_zipformer (large) | 37.9 | 6.4 | 6.9 | 17.9 | 19.7 | 25.2 | 25.5 | 16.9 |
+
+## Pre-trained models and logs
+
+* Pre-trained models: <https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer>
+
+* Training logs:
+    - surt_base: <https://tensorboard.dev/experiment/YLGJTkBETb2aqDQ61jbxvQ/>
+    - surt_base_adapt: <https://tensorboard.dev/experiment/pjXMFVL9RMej85rMHyd0EQ/>
+    - surt_large: <https://tensorboard.dev/experiment/82HvYqfrSOKZ4w8Jod2QMw/>
+    - surt_large_adapt: <https://tensorboard.dev/experiment/5oIdEgRqS9Wb6yVuxaExEw/>
diff --git a/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py
new file mode 100644
index 000000000..c1abdbdb5
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py
@@ -0,0 +1,372 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+# Copyright      2023  Johns Hopkins Univrtsity (Author: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutMix,
+    DynamicBucketingSampler,
+    K2SurtDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriCssAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--max-duration-valid",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--max-cuts",
+            type=int,
+            default=100,
+            help="Maximum number of cuts in a single batch. You can "
+            "reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help=(
+                "When enabled, use on-the-fly cut mixing and feature "
+                "extraction. Will drop existing precomputed feature manifests "
+                "if available."
+            ),
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        return_sources: bool = True,
+        strict: bool = True,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SurtDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+            return_sources=return_sources,
+            strict=strict,
+        )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                quadratic_duration=30.0,
+                max_cuts=self.args.max_cuts,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                max_cuts=self.args.max_cuts,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+
+        logging.info("About to create dev dataset")
+        validate = K2SurtDataset(
+            input_strategy=OnTheFlyFeatures(
+                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            )
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+            return_sources=False,
+            strict=False,
+        )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration_valid,
+            max_cuts=self.args.max_cuts,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SurtDataset(
+            input_strategy=OnTheFlyFeatures(
+                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            )
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+            return_sources=False,
+            strict=False,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration_valid,
+            max_cuts=self.args.max_cuts,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def lsmix_cuts(
+        self,
+        rvb_affix: str = "clean",
+        type_affix: str = "full",
+        sources: bool = True,
+    ) -> CutSet:
+        logging.info("About to get train cuts")
+        source_affix = "_sources" if sources else ""
+        cs = load_manifest_lazy(
+            self.args.manifest_dir
+            / f"cuts_train_{rvb_affix}_{type_affix}{source_affix}.jsonl.gz"
+        )
+        cs = cs.filter(lambda c: c.duration >= 1.0 and c.duration <= 30.0)
+        return cs
+
+    @lru_cache()
+    def libricss_cuts(self, split="dev", type="sdm") -> CutSet:
+        logging.info(f"About to get LibriCSS {split} {type} cuts")
+        cs = load_manifest_lazy(
+            self.args.manifest_dir / f"cuts_{split}_libricss-{type}.jsonl.gz"
+        )
+        return cs
diff --git a/egs/libricss/SURT/dprnn_zipformer/beam_search.py b/egs/libricss/SURT/dprnn_zipformer/beam_search.py
new file mode 100644
index 000000000..c8e4643d0
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/beam_search.py
@@ -0,0 +1,730 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Tuple, Union
+
+import k2
+import torch
+from model import SURT
+
+from icefall import NgramLmStateCost
+from icefall.utils import DecodingResults
+
+
+def greedy_search(
+    model: SURT,
+    encoder_out: torch.Tensor,
+    max_sym_per_frame: int,
+    return_timestamps: bool = False,
+) -> Union[List[int], DecodingResults]:
+    """Greedy search for a single utterance.
+    Args:
+      model:
+        An instance of `SURT`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 4
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    unk_id = getattr(model, "unk_id", blank_id)
+
+    device = next(model.parameters()).device
+
+    decoder_input = torch.tensor(
+        [-1] * (context_size - 1) + [blank_id], device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # timestamp[i] is the frame index after subsampling
+    # on which hyp[i] is decoded
+    timestamp = []
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y not in (blank_id, unk_id):
+            hyp.append(y)
+            timestamp.append(t)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    if not return_timestamps:
+        return hyp
+    else:
+        return DecodingResults(
+            hyps=[hyp],
+            timestamps=[timestamp],
+        )
+
+
+def greedy_search_batch(
+    model: SURT,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The SURT model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = next(model.parameters()).device
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[-1] * (context_size - 1) + [blank_id] for _ in range(N)]
+
+    # timestamp[n][i] is the frame index after subsampling
+    # on which hyp[n][i] is decoded
+    timestamps = [[] for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+    # decoder_out: (N, 1, decoder_out_dim)
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    for (t, batch_size) in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits'shape (batch_size, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (batch_size, vocab_size)
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v not in (blank_id, unk_id):
+                hyps[i].append(v)
+                timestamps[i].append(t)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    ans_timestamps = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(timestamps[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+        )
+
+
+def modified_beam_search(
+    model: SURT,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: int = 4,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The SURT model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for (t, batch_size) in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    new_timestamp.append(t)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(
+                    ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    sorted_timestamps = [h.timestamp for h in best_hyps]
+    ans = []
+    ans_timestamps = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(sorted_timestamps[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+        )
+
+
+def beam_search(
+    model: SURT,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Union[List[int], DecodingResults]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_SURT.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `SURT`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    device = next(model.parameters()).device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size,
+        device=device,
+        dtype=torch.int64,
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0, timestamp=[]))
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        # TODO(fangjun): Implement prefix search to update the `log_prob`
+        # of hypotheses in A
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            cached_key = y_star.key
+
+            if cached_key not in decoder_cache:
+                decoder_input = torch.tensor(
+                    [y_star.ys[-context_size:]],
+                    device=device,
+                    dtype=torch.int64,
+                ).reshape(1, context_size)
+
+                decoder_out = model.decoder(decoder_input, need_pad=False)
+                decoder_out = model.joiner.decoder_proj(decoder_out)
+                decoder_cache[cached_key] = decoder_out
+            else:
+                decoder_out = decoder_cache[cached_key]
+
+            cached_key += f"-t-{t}"
+            if cached_key not in joint_cache:
+                logits = model.joiner(
+                    current_encoder_out,
+                    decoder_out.unsqueeze(1),
+                    project_input=False,
+                )
+
+                # TODO(fangjun): Scale the blank posterior
+                log_prob = (logits / temperature).log_softmax(dim=-1)
+                # log_prob is (1, 1, 1, vocab_size)
+                log_prob = log_prob.squeeze()
+                # Now log_prob is (vocab_size,)
+                joint_cache[cached_key] = log_prob
+            else:
+                log_prob = joint_cache[cached_key]
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(
+                Hypothesis(
+                    ys=y_star.ys[:],
+                    log_prob=new_y_star_log_prob,
+                    timestamp=y_star.timestamp[:],
+                )
+            )
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for i, v in zip(indices.tolist(), values.tolist()):
+                if i in (blank_id, unk_id):
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                new_timestamp = y_star.timestamp + [t]
+                A.add(
+                    Hypothesis(
+                        ys=new_ys,
+                        log_prob=new_log_prob,
+                        timestamp=new_timestamp,
+                    )
+                )
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    if not return_timestamps:
+        return ys
+    else:
+        return DecodingResults(hyps=[ys], timestamps=[best_hyp.timestamp])
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    # timestamp[i] is the frame index after subsampling
+    # on which ys[i] is decoded
+    timestamp: List[int] = field(default_factory=list)
+
+    # the lm score for next token given the current ys
+    lm_score: Optional[torch.Tensor] = None
+
+    # the RNNLM states (h and c in LSTM)
+    state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
+
+    # N-gram LM state
+    state_cost: Optional[NgramLmStateCost] = None
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
diff --git a/egs/libricss/SURT/dprnn_zipformer/decode.py b/egs/libricss/SURT/dprnn_zipformer/decode.py
new file mode 100755
index 000000000..6abbffe00
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/decode.py
@@ -0,0 +1,654 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./dprnn_zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --use-averaged-model true \
+    --exp-dir ./dprnn_zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./dprnn_zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --use-averaged-model true \
+    --exp-dir ./dprnn_zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriCssAsrDataModule
+from beam_search import (
+    beam_search,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.utils import EPSILON
+from train import add_model_arguments, get_params, get_surt_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_surt_error_stats,
+)
+
+OVERLAP_RATIOS = ["0L", "0S", "OV10", "OV20", "OV30", "OV40"]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="dprnn_zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--save-masks",
+        type=str2bool,
+        default=False,
+        help="""If true, save masks generated by unmixing module.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    feature_lens = batch["input_lens"].to(device)
+
+    # Apply the mask encoder
+    B, T, F = feature.shape
+    processed = model.mask_encoder(feature)  # B,T,F*num_channels
+    masks = processed.view(B, T, F, params.num_channels).unbind(dim=-1)
+    x_masked = [feature * m for m in masks]
+
+    masks_dict = {}
+    if params.save_masks:
+        # To save the masks, we split them by batch and trim each mask to the length of
+        # the corresponding feature. We save them in a dict, where the key is the
+        # cut ID and the value is the mask.
+        for i in range(B):
+            mask = torch.cat(
+                [x_masked[j][i, : feature_lens[i]] for j in range(params.num_channels)],
+                dim=-1,
+            )
+            mask = mask.cpu().numpy()
+            masks_dict[batch["cuts"][i].id] = mask
+
+    # Recognition
+    # Concatenate the inputs along the batch axis
+    h = torch.cat(x_masked, dim=0)
+    h_lens = feature_lens.repeat(params.num_channels)
+    encoder_out, encoder_out_lens = model.encoder(x=h, x_lens=h_lens)
+
+    if model.joint_encoder_layer is not None:
+        encoder_out = model.joint_encoder_layer(encoder_out)
+
+    def _group_channels(hyps: List[str]) -> List[List[str]]:
+        """
+        Currently we have a batch of size M*B, where M is the number of
+        channels and B is the batch size. We need to group the hypotheses
+        into B groups, each of which contains M hypotheses.
+
+        Example:
+            hyps = ['a1', 'b1', 'c1', 'a2', 'b2', 'c2']
+            _group_channels(hyps) = [['a1', 'a2'], ['b1', 'b2'], ['c1', 'c2']]
+        """
+        assert len(hyps) == B * params.num_channels
+        out_hyps = []
+        for i in range(B):
+            out_hyps.append(hyps[i::B])
+        return out_hyps
+
+    hyps = []
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp)
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp)
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp))
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": _group_channels(hyps)}, masks_dict
+    else:
+        return {f"beam_size_{params.beam_size}": _group_channels(hyps)}, masks_dict
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    masks = {}
+    for batch_idx, batch in enumerate(dl):
+        cut_ids = [cut.id for cut in batch["cuts"]]
+        cuts_batch = batch["cuts"]
+
+        hyps_dict, masks_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+        )
+        masks.update(masks_dict)
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            for cut_id, hyp_words in zip(cut_ids, hyps):
+                # Reference is a list of supervision texts sorted by start time.
+                ref_words = [
+                    s.text.strip()
+                    for s in sorted(
+                        cuts_batch[cut_id].supervisions, key=lambda s: s.start
+                    )
+                ]
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(cut_ids)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results, masks_dict
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_surt_error_stats(
+                f,
+                f"{test_set_name}-{key}",
+                results,
+                enable_log=True,
+                num_channels=params.num_channels,
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+def save_masks(
+    params: AttributeDict,
+    test_set_name: str,
+    masks: List[torch.Tensor],
+):
+    masks_path = params.res_dir / f"masks-{test_set_name}.txt"
+    torch.save(masks, masks_path)
+    logging.info(f"The masks are stored in {masks_path}")
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LmScorer.add_arguments(parser)
+    LibriCssAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "modified_beam_search",
+    ), f"Decoding method {params.decoding_method} is not supported."
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libricss = LibriCssAsrDataModule(args)
+
+    dev_cuts = libricss.libricss_cuts(split="dev", type="ihm-mix").to_eager()
+    dev_cuts_grouped = [dev_cuts.filter(lambda x: ol in x.id) for ol in OVERLAP_RATIOS]
+    test_cuts = libricss.libricss_cuts(split="test", type="ihm-mix").to_eager()
+    test_cuts_grouped = [
+        test_cuts.filter(lambda x: ol in x.id) for ol in OVERLAP_RATIOS
+    ]
+
+    for dev_set, ol in zip(dev_cuts_grouped, OVERLAP_RATIOS):
+        dev_dl = libricss.test_dataloaders(dev_set)
+        results_dict, masks = decode_dataset(
+            dl=dev_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=f"dev_{ol}",
+            results_dict=results_dict,
+        )
+
+        if params.save_masks:
+            save_masks(
+                params=params,
+                test_set_name=f"dev_{ol}",
+                masks=masks,
+            )
+
+    for test_set, ol in zip(test_cuts_grouped, OVERLAP_RATIOS):
+        test_dl = libricss.test_dataloaders(test_set)
+        results_dict, masks = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=f"test_{ol}",
+            results_dict=results_dict,
+        )
+
+        if params.save_masks:
+            save_masks(
+                params=params,
+                test_set_name=f"test_{ol}",
+                masks=masks,
+            )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libricss/SURT/dprnn_zipformer/decoder.py b/egs/libricss/SURT/dprnn_zipformer/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/dprnn_zipformer/dprnn.py b/egs/libricss/SURT/dprnn_zipformer/dprnn.py
new file mode 100644
index 000000000..440dea885
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/dprnn.py
@@ -0,0 +1,305 @@
+import random
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from scaling import ActivationBalancer, BasicNorm, DoubleSwish, ScaledLinear, ScaledLSTM
+from torch.autograd import Variable
+
+EPS = torch.finfo(torch.get_default_dtype()).eps
+
+
+def _pad_segment(input, segment_size):
+    # Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py#L342
+    # input is the features: (B, N, T)
+    batch_size, dim, seq_len = input.shape
+    segment_stride = segment_size // 2
+
+    rest = segment_size - (segment_stride + seq_len % segment_size) % segment_size
+    if rest > 0:
+        pad = Variable(torch.zeros(batch_size, dim, rest)).type(input.type())
+        input = torch.cat([input, pad], 2)
+
+    pad_aux = Variable(torch.zeros(batch_size, dim, segment_stride)).type(input.type())
+    input = torch.cat([pad_aux, input, pad_aux], 2)
+
+    return input, rest
+
+
+def split_feature(input, segment_size):
+    # Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py#L358
+    # split the feature into chunks of segment size
+    # input is the features: (B, N, T)
+
+    input, rest = _pad_segment(input, segment_size)
+    batch_size, dim, seq_len = input.shape
+    segment_stride = segment_size // 2
+
+    segments1 = (
+        input[:, :, :-segment_stride]
+        .contiguous()
+        .view(batch_size, dim, -1, segment_size)
+    )
+    segments2 = (
+        input[:, :, segment_stride:]
+        .contiguous()
+        .view(batch_size, dim, -1, segment_size)
+    )
+    segments = (
+        torch.cat([segments1, segments2], 3)
+        .view(batch_size, dim, -1, segment_size)
+        .transpose(2, 3)
+    )
+
+    return segments.contiguous(), rest
+
+
+def merge_feature(input, rest):
+    # Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py#L385
+    # merge the splitted features into full utterance
+    # input is the features: (B, N, L, K)
+
+    batch_size, dim, segment_size, _ = input.shape
+    segment_stride = segment_size // 2
+    input = (
+        input.transpose(2, 3).contiguous().view(batch_size, dim, -1, segment_size * 2)
+    )  # B, N, K, L
+
+    input1 = (
+        input[:, :, :, :segment_size]
+        .contiguous()
+        .view(batch_size, dim, -1)[:, :, segment_stride:]
+    )
+    input2 = (
+        input[:, :, :, segment_size:]
+        .contiguous()
+        .view(batch_size, dim, -1)[:, :, :-segment_stride]
+    )
+
+    output = input1 + input2
+    if rest > 0:
+        output = output[:, :, :-rest]
+
+    return output.contiguous()  # B, N, T
+
+
+class RNNEncoderLayer(nn.Module):
+    """
+    RNNEncoderLayer is made up of lstm and feedforward networks.
+    Args:
+      input_size:
+        The number of expected features in the input (required).
+      hidden_size:
+        The hidden dimension of rnn layer.
+      dropout:
+        The dropout value (default=0.1).
+      layer_dropout:
+        The dropout value for model-level warmup (default=0.075).
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        dropout: float = 0.1,
+        bidirectional: bool = False,
+    ) -> None:
+        super(RNNEncoderLayer, self).__init__()
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+
+        assert hidden_size >= input_size, (hidden_size, input_size)
+        self.lstm = ScaledLSTM(
+            input_size=input_size,
+            hidden_size=hidden_size // 2 if bidirectional else hidden_size,
+            proj_size=0,
+            num_layers=1,
+            dropout=0.0,
+            batch_first=True,
+            bidirectional=bidirectional,
+        )
+        self.norm_final = BasicNorm(input_size)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).  # noqa
+        self.balancer = ActivationBalancer(
+            num_channels=input_size,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            max_abs=6.0,
+        )
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Pass the input through the encoder layer.
+        Args:
+          src:
+            The sequence to the encoder layer (required).
+            Its shape is (S, N, E), where S is the sequence length,
+            N is the batch size, and E is the feature number.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (1, N, input_size);
+            states[1] is the cell states of all layers,
+              with shape of (1, N, hidden_size).
+        """
+        src_orig = src
+
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        alpha = warmup if self.training else 1.0
+
+        # lstm module
+        src_lstm, new_states = self.lstm(src, states)
+        src = self.dropout(src_lstm) + src
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+# dual-path RNN
+class DPRNN(nn.Module):
+    """Deep dual-path RNN.
+    Source: https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/dprnn.py
+
+    args:
+        input_size: int, dimension of the input feature. The input should have shape
+                    (batch, seq_len, input_size).
+        hidden_size: int, dimension of the hidden state.
+        output_size: int, dimension of the output size.
+        dropout: float, dropout ratio. Default is 0.
+        num_blocks: int, number of stacked RNN layers. Default is 1.
+    """
+
+    def __init__(
+        self,
+        feature_dim,
+        input_size,
+        hidden_size,
+        output_size,
+        dropout=0.1,
+        num_blocks=1,
+        segment_size=50,
+        chunk_width_randomization=False,
+    ):
+        super().__init__()
+
+        self.input_size = input_size
+        self.output_size = output_size
+        self.hidden_size = hidden_size
+
+        self.segment_size = segment_size
+        self.chunk_width_randomization = chunk_width_randomization
+
+        self.input_embed = nn.Sequential(
+            ScaledLinear(feature_dim, input_size),
+            BasicNorm(input_size),
+            ActivationBalancer(
+                num_channels=input_size,
+                channel_dim=-1,
+                min_positive=0.45,
+                max_positive=0.55,
+            ),
+        )
+
+        # dual-path RNN
+        self.row_rnn = nn.ModuleList([])
+        self.col_rnn = nn.ModuleList([])
+        for _ in range(num_blocks):
+            # intra-RNN is non-causal
+            self.row_rnn.append(
+                RNNEncoderLayer(
+                    input_size, hidden_size, dropout=dropout, bidirectional=True
+                )
+            )
+            self.col_rnn.append(
+                RNNEncoderLayer(
+                    input_size, hidden_size, dropout=dropout, bidirectional=False
+                )
+            )
+
+        # output layer
+        self.out_embed = nn.Sequential(
+            ScaledLinear(input_size, output_size),
+            BasicNorm(output_size),
+            ActivationBalancer(
+                num_channels=output_size,
+                channel_dim=-1,
+                min_positive=0.45,
+                max_positive=0.55,
+            ),
+        )
+
+    def forward(self, input):
+        # input shape: B, T, F
+        input = self.input_embed(input)
+        B, T, D = input.shape
+
+        if self.chunk_width_randomization and self.training:
+            segment_size = random.randint(self.segment_size // 2, self.segment_size)
+        else:
+            segment_size = self.segment_size
+        input, rest = split_feature(input.transpose(1, 2), segment_size)
+        # input shape: batch, N, dim1, dim2
+        # apply RNN on dim1 first and then dim2
+        # output shape: B, output_size, dim1, dim2
+        # input = input.to(device)
+        batch_size, _, dim1, dim2 = input.shape
+        output = input
+        for i in range(len(self.row_rnn)):
+            row_input = (
+                output.permute(0, 3, 2, 1)
+                .contiguous()
+                .view(batch_size * dim2, dim1, -1)
+            )  # B*dim2, dim1, N
+            output = self.row_rnn[i](row_input)  # B*dim2, dim1, H
+            output = (
+                output.view(batch_size, dim2, dim1, -1).permute(0, 3, 2, 1).contiguous()
+            )  # B, N, dim1, dim2
+
+            col_input = (
+                output.permute(0, 2, 3, 1)
+                .contiguous()
+                .view(batch_size * dim1, dim2, -1)
+            )  # B*dim1, dim2, N
+            output = self.col_rnn[i](col_input)  # B*dim1, dim2, H
+            output = (
+                output.view(batch_size, dim1, dim2, -1).permute(0, 3, 1, 2).contiguous()
+            )  # B, N, dim1, dim2
+
+        output = merge_feature(output, rest)
+        output = output.transpose(1, 2)
+        output = self.out_embed(output)
+
+        # Apply ReLU to the output
+        output = torch.relu(output)
+
+        return output
+
+
+if __name__ == "__main__":
+
+    model = DPRNN(
+        80,
+        256,
+        256,
+        160,
+        dropout=0.1,
+        num_blocks=4,
+        segment_size=32,
+        chunk_width_randomization=True,
+    )
+    input = torch.randn(2, 1002, 80)
+    print(sum(p.numel() for p in model.parameters()))
+    print(model(input).shape)
diff --git a/egs/libricss/SURT/dprnn_zipformer/encoder_interface.py b/egs/libricss/SURT/dprnn_zipformer/encoder_interface.py
new file mode 120000
index 000000000..0c2673d46
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/dprnn_zipformer/export.py b/egs/libricss/SURT/dprnn_zipformer/export.py
new file mode 100755
index 000000000..f51f2a7ab
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/export.py
@@ -0,0 +1,306 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./dprnn_zipformer/export.py \
+  --exp-dir ./dprnn_zipformer/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./dprnn_zipformer/export.py \
+  --exp-dir ./dprnn_zipformer/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `dprnn_zipformer/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./dprnn_zipformer/decode.py \
+        --exp-dir ./dprnn_zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_surt_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="dprnn_zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/libricss/SURT/dprnn_zipformer/joiner.py b/egs/libricss/SURT/dprnn_zipformer/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/dprnn_zipformer/model.py b/egs/libricss/SURT/dprnn_zipformer/model.py
new file mode 100644
index 000000000..688e1e78d
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/model.py
@@ -0,0 +1,316 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+# Copyright    2023  Johns Hopkins University (author: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class SURT(nn.Module):
+    """It implements Streaming Unmixing and Recognition Transducer (SURT).
+    https://arxiv.org/abs/2011.13148
+    """
+
+    def __init__(
+        self,
+        mask_encoder: nn.Module,
+        encoder: EncoderInterface,
+        joint_encoder_layer: Optional[nn.Module],
+        decoder: nn.Module,
+        joiner: nn.Module,
+        num_channels: int,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          mask_encoder:
+            It is the masking network. It generates a mask for each channel of the
+            encoder. These masks are applied to the input features, and then passed
+            to the transcription network.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          num_channels:
+            It is the number of channels that the input features will be split into.
+            In general, it should be equal to the maximum number of simultaneously
+            active speakers. For most real scenarios, using 2 channels is sufficient.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.mask_encoder = mask_encoder
+        self.encoder = encoder
+        self.joint_encoder_layer = joint_encoder_layer
+        self.decoder = decoder
+        self.joiner = joiner
+        self.num_channels = num_channels
+
+        self.simple_am_proj = nn.Linear(
+            encoder_dim,
+            vocab_size,
+        )
+        self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
+
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+
+    def forward_helper(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        reduction: str = "sum",
+        beam_size: int = 10,
+        use_double_scores: bool = False,
+        subsampling_factor: int = 1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Compute transducer loss for one branch of the SURT model.
+        """
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        if self.joint_encoder_layer is not None:
+            encoder_out = self.joint_encoder_layer(encoder_out)
+
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(encoder_out)
+
+        # For the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction=reduction,
+            )
+
+        # Compute ctc loss
+        supervision_segments = torch.stack(
+            (
+                torch.arange(len(x_lens), device="cpu"),
+                torch.zeros_like(x_lens, device="cpu"),
+                torch.clone(x_lens).detach().cpu(),
+            ),
+            dim=1,
+        ).to(torch.int32)
+        # We need to sort supervision_segments in decreasing order of num_frames
+        indices = torch.argsort(supervision_segments[:, 2], descending=True)
+        supervision_segments = supervision_segments[indices]
+
+        # Works with a BPE model
+        decoding_graph = k2.ctc_graph(y, modified=False, device=x.device)
+        dense_fsa_vec = k2.DenseFsaVec(
+            ctc_output,
+            supervision_segments,
+            allow_truncate=subsampling_factor - 1,
+        )
+        ctc_loss = k2.ctc_loss(
+            decoding_graph=decoding_graph,
+            dense_fsa_vec=dense_fsa_vec,
+            output_beam=beam_size,
+            reduction="none",
+            use_double_scores=use_double_scores,
+        )
+
+        return (simple_loss, pruned_loss, ctc_loss)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        reduction: str = "sum",
+        beam_size: int = 10,
+        use_double_scores: bool = False,
+        subsampling_factor: int = 1,
+        return_masks: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor of shape (N*num_channels, S). It contains the labels
+            of the N utterances. The labels are in the range [0, vocab_size). All
+            the channels are concatenated together one after another.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          beam_size:
+            The beam size used in CTC decoding.
+          use_double_scores:
+            If True, use double precision for CTC decoding.
+          subsampling_factor:
+            The subsampling factor of the model. It is used to compute the
+            supervision segments for CTC loss.
+          return_masks:
+            If True, return the masks as well as masked features.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0), (x.size(), x_lens.size())
+
+        # Apply the mask encoder
+        B, T, F = x.shape
+        processed = self.mask_encoder(x)  # B,T,F*num_channels
+        masks = processed.view(B, T, F, self.num_channels).unbind(dim=-1)
+        x_masked = [x * m for m in masks]
+
+        # Recognition
+        # Stack the inputs along the batch axis
+        h = torch.cat(x_masked, dim=0)
+        h_lens = torch.cat([x_lens for _ in range(self.num_channels)], dim=0)
+
+        simple_loss, pruned_loss, ctc_loss = self.forward_helper(
+            h,
+            h_lens,
+            y,
+            prune_range,
+            am_scale,
+            lm_scale,
+            reduction=reduction,
+            beam_size=beam_size,
+            use_double_scores=use_double_scores,
+            subsampling_factor=subsampling_factor,
+        )
+
+        # Chunks the outputs into 2 parts along batch axis and then stack them along a new axis.
+        simple_loss = torch.stack(
+            torch.chunk(simple_loss, self.num_channels, dim=0), dim=0
+        )
+        pruned_loss = torch.stack(
+            torch.chunk(pruned_loss, self.num_channels, dim=0), dim=0
+        )
+        ctc_loss = torch.stack(torch.chunk(ctc_loss, self.num_channels, dim=0), dim=0)
+
+        if return_masks:
+            return (simple_loss, pruned_loss, ctc_loss, x_masked, masks)
+        else:
+            return (simple_loss, pruned_loss, ctc_loss, x_masked)
diff --git a/egs/libricss/SURT/dprnn_zipformer/optim.py b/egs/libricss/SURT/dprnn_zipformer/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/dprnn_zipformer/scaling.py b/egs/libricss/SURT/dprnn_zipformer/scaling.py
new file mode 100644
index 000000000..4040a7b89
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/scaling.py
@@ -0,0 +1,1576 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import logging
+import random
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.backends.cudnn.rnn as rnn
+import torch.nn as nn
+from torch import _VF, Tensor
+
+from icefall.utils import is_jit_tracing
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        scale_factor: Tensor,
+        sign_factor: Optional[Tensor],
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        if sign_factor is None:
+            ctx.save_for_backward(xgt0, scale_factor)
+        else:
+            ctx.save_for_backward(xgt0, scale_factor, sign_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        if len(ctx.saved_tensors) == 3:
+            xgt0, scale_factor, sign_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+                sign_factor = sign_factor.unsqueeze(-1)
+            factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        else:
+            xgt0, scale_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+            factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+def _compute_scale_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_abs: float,
+    max_abs: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32)
+
+    if min_abs == 0.0:
+        below_threshold = 0.0
+    else:
+        # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if
+        # x_abs)_mean , min_abs.
+        below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp(
+            min=0, max=max_factor
+        )
+
+    above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp(
+        min=0, max=max_factor
+    )
+
+    return below_threshold - above_threshold
+
+
+def _compute_sign_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_positive: float,
+    max_positive: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims)
+    if min_positive == 0.0:
+        factor1 = 0.0
+    else:
+        # 0 if proportion_positive >= min_positive, else can be
+        # as large as max_factor.
+        factor1 = (
+            (min_positive - proportion_positive) * (gain_factor / min_positive)
+        ).clamp_(min=0, max=max_factor)
+
+    if max_positive == 1.0:
+        factor2 = 0.0
+    else:
+        # 0 if self.proportion_positive <= max_positive, else can be
+        # as large as -max_factor.
+        factor2 = (
+            (proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive))
+        ).clamp_(min=0, max=max_factor)
+    sign_factor = factor1 - factor2
+    # require min_positive != 0 or max_positive != 1:
+    assert not isinstance(sign_factor, float)
+    return sign_factor
+
+
+class ActivationScaleBalancerFunction(torch.autograd.Function):
+    """
+    This object is used in class ActivationBalancer when the user specified
+    min_positive=0, max_positive=1, so there are no constraints on the signs
+    of the activations and only the absolute value has a constraint.
+    """
+
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        sign_factor: Tensor,
+        scale_factor: Tensor,
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        ctx.save_for_backward(xgt0, sign_factor, scale_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        xgt0, sign_factor, scale_factor = ctx.saved_tensors
+        for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+            sign_factor = sign_factor.unsqueeze(-1)
+            scale_factor = scale_factor.unsqueeze(-1)
+
+        factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+class RandomClampFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        min: Optional[float],
+        max: Optional[float],
+        prob: float,
+        reflect: float,
+    ) -> Tensor:
+        x_clamped = torch.clamp(x, min=min, max=max)
+        mask = torch.rand_like(x) < prob
+        ans = torch.where(mask, x_clamped, x)
+        if x.requires_grad:
+            ctx.save_for_backward(ans == x)
+            ctx.reflect = reflect
+        if reflect != 0.0:
+            ans = ans * (1.0 + reflect) - (x * reflect)
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, None, None, None]:
+        (is_same,) = ctx.saved_tensors
+        x_grad = ans_grad * is_same.to(ans_grad.dtype)
+        reflect = ctx.reflect
+        if reflect != 0.0:
+            x_grad = x_grad * (1.0 + reflect) - (ans_grad * reflect)
+        return x_grad, None, None, None, None
+
+
+def random_clamp(
+    x: Tensor,
+    min: Optional[float] = None,
+    max: Optional[float] = None,
+    prob: float = 0.5,
+    reflect: float = 0.0,
+):
+    return RandomClampFunction.apply(x, min, max, prob, reflect)
+
+
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
+    """
+    A randomized way of casting a floating point value to half precision.
+    """
+    if x.dtype == torch.float16:
+        return x
+    x_abs = x.abs()
+    is_too_small = x_abs < min_abs
+    # for elements where is_too_small is true, random_val will contain +-min_abs with
+    # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
+    # for those elements].
+    random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs)
+    return torch.where(is_too_small, random_val, x).to(torch.float16)
+
+
+class RandomGradFunction(torch.autograd.Function):
+    """
+    Does nothing in forward pass; in backward pass, gets rid of very small grads using
+    randomized approach that preserves expectations (intended to reduce roundoff).
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, min_abs: float) -> Tensor:
+        ctx.min_abs = min_abs
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None]:
+        if ans_grad.dtype == torch.float16:
+            return (
+                random_cast_to_half(ans_grad.to(torch.float32), min_abs=ctx.min_abs),
+                None,
+            )
+        else:
+            return ans_grad, None
+
+
+class RandomGrad(torch.nn.Module):
+    """
+    Gets rid of very small gradients using an expectation-preserving method, intended to increase
+    accuracy of training when using amp (automatic mixed precision)
+    """
+
+    def __init__(self, min_abs: float = 5.0e-06):
+        super(RandomGrad, self).__init__()
+        self.min_abs = min_abs
+
+    def forward(self, x: Tensor):
+        if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing():
+            return x
+        else:
+            return RandomGradFunction.apply(x, self.min_abs)
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    """
+    Tries to handle half-precision derivatives in a randomized way that should
+    be more accurate for training than the default behavior.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, dim: int):
+        ans = x.softmax(dim=dim)
+        # if x dtype is float16, x.softmax() returns a float32 because
+        # (presumably) that op does not support float16, and autocast
+        # is enabled.
+        if torch.is_autocast_enabled():
+            ans = ans.to(torch.float16)
+        ctx.save_for_backward(ans)
+        ctx.x_dtype = x.dtype
+        ctx.dim = dim
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        (ans,) = ctx.saved_tensors
+        with torch.cuda.amp.autocast(enabled=False):
+            ans_grad = ans_grad.to(torch.float32)
+            ans = ans.to(torch.float32)
+            x_grad = ans_grad * ans
+            x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True)
+            return x_grad, None
+
+
+def softmax(x: Tensor, dim: int):
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x.softmax(dim)
+
+    return SoftmaxFunction.apply(x, dim)
+
+
+class MaxEigLimiterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.channel_dim = channel_dim
+        ctx.grad_scale = grad_scale
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        with torch.enable_grad():
+            (x_orig, coeffs, new_direction) = ctx.saved_tensors
+            x_orig.requires_grad = True
+            num_channels = x_orig.shape[ctx.channel_dim]
+            x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
+            new_direction.requires_grad = False
+            x = x - x.mean(dim=0)
+            x_var = (x**2).mean()
+            x_residual = x - coeffs * new_direction
+            x_residual_var = (x_residual**2).mean()
+            # `variance_proportion` is the proportion of the variance accounted for
+            # by the top eigen-direction.  This is to be minimized.
+            variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+            variance_proportion.backward()
+        x_orig_grad = x_orig.grad
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
+        return x_grad + x_extra_grad.detach(), None, None, None, None
+
+
+class GradientFilterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        batch_dim: int,  # e.g., 1
+        threshold: float,  # e.g., 10.0
+        *params: Tensor,  # module parameters
+    ) -> Tuple[Tensor, ...]:
+        if x.requires_grad:
+            if batch_dim < 0:
+                batch_dim += x.ndim
+            ctx.batch_dim = batch_dim
+            ctx.threshold = threshold
+        return (x,) + params
+
+    @staticmethod
+    def backward(
+        ctx,
+        x_grad: Tensor,
+        *param_grads: Tensor,
+    ) -> Tuple[Tensor, ...]:
+        eps = 1.0e-20
+        dim = ctx.batch_dim
+        norm_dims = [d for d in range(x_grad.ndim) if d != dim]
+        norm_of_batch = (x_grad**2).mean(dim=norm_dims, keepdim=True).sqrt()
+        median_norm = norm_of_batch.median()
+
+        cutoff = median_norm * ctx.threshold
+        inv_mask = (cutoff + norm_of_batch) / (cutoff + eps)
+        mask = 1.0 / (inv_mask + eps)
+        x_grad = x_grad * mask
+
+        avg_mask = 1.0 / (inv_mask.mean() + eps)
+        param_grads = [avg_mask * g for g in param_grads]
+
+        return (x_grad, None, None) + tuple(param_grads)
+
+
+class GradientFilter(torch.nn.Module):
+    """This is used to filter out elements that have extremely large gradients
+    in batch and the module parameters with soft masks.
+
+    Args:
+      batch_dim (int):
+        The batch dimension.
+      threshold (float):
+        For each element in batch, its gradient will be
+        filtered out if the gradient norm is larger than
+        `grad_norm_threshold * median`, where `median` is the median
+        value of gradient norms of all elememts in batch.
+    """
+
+    def __init__(self, batch_dim: int = 1, threshold: float = 10.0):
+        super(GradientFilter, self).__init__()
+        self.batch_dim = batch_dim
+        self.threshold = threshold
+
+    def forward(self, x: Tensor, *params: Tensor) -> Tuple[Tensor, ...]:
+        if torch.jit.is_scripting() or is_jit_tracing():
+            return (x,) + params
+        else:
+            return GradientFilterFunction.apply(
+                x,
+                self.batch_dim,
+                self.threshold,
+                *params,
+            )
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    eps_min: float
+    eps_max: float
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+        eps_min: float = -3.0,
+        eps_max: float = 3.0,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+        self.eps_min = eps_min
+        self.eps_max = eps_max
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        eps = self.eps
+        if self.training and random.random() < 0.25:
+            # with probability 0.25, in training mode, clamp eps between the min
+            # and max; this will encourage it to learn parameters within the
+            # allowed range by making parameters that are outside the allowed
+            # range noisy.
+
+            # gradients to allow the parameter to get back into the allowed
+            # region if it happens to exit it.
+            eps = eps.clamp(min=self.eps_min, max=self.eps_max)
+        scales = (
+            torch.mean(x**2, dim=self.channel_dim, keepdim=True) + eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+class ScaledEmbedding(nn.Module):
+    r"""This is a modified version of nn.Embedding that introduces a learnable scale
+    on the parameters.  Note: due to how we initialize it, it's best used with
+    schedulers like Noam that have a warmup period.
+
+    It is a simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+        initial_speed (float, optional):  This affects how fast the parameter will
+           learn near the start of training; you can set it to a value less than
+           one if you suspect that a module is contributing to instability near
+           the start of training.  Note: regardless of the use of this option,
+           it's best to use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.  Must be greater than 0.
+
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+
+    """
+    __constants__ = [
+        "num_embeddings",
+        "embedding_dim",
+        "padding_idx",
+        "scale_grad_by_freq",
+        "sparse",
+    ]
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        scale_grad_by_freq: bool = False,
+        sparse: bool = False,
+        initial_speed: float = 1.0,
+    ) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert (
+                    padding_idx < self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+            elif padding_idx < 0:
+                assert (
+                    padding_idx >= -self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale = nn.Parameter(torch.zeros(()))  # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters(initial_speed)
+
+    def reset_parameters(self, initial_speed: float = 1.0) -> None:
+        std = 0.1 / initial_speed
+        nn.init.normal_(self.weight, std=std)
+        nn.init.constant_(self.scale, torch.tensor(1.0 / std).log())
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        scale = self.scale.exp()
+        if input.numel() < self.num_embeddings:
+            return (
+                F.embedding(
+                    input,
+                    self.weight,
+                    self.padding_idx,
+                    None,
+                    2.0,  # None, 2.0 relate to normalization
+                    self.scale_grad_by_freq,
+                    self.sparse,
+                )
+                * scale
+            )
+        else:
+            return F.embedding(
+                input,
+                self.weight * scale,
+                self.padding_idx,
+                None,
+                2.0,  # None, 2.0 relates to normalization
+                self.scale_grad_by_freq,
+                self.sparse,
+            )
+
+    def extra_repr(self) -> str:
+        # s = "{num_embeddings}, {embedding_dim}, scale={scale}"
+        s = "{num_embeddings}, {embedding_dim}"
+        if self.padding_idx is not None:
+            s += ", padding_idx={padding_idx}"
+        if self.scale_grad_by_freq is not False:
+            s += ", scale_grad_by_freq={scale_grad_by_freq}"
+        if self.sparse is not False:
+            s += ", sparse=True"
+        return s.format(**self.__dict__)
+
+
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
+    """
+    Behaves like a constructor of a modified version of nn.Linear
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Linear(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv1d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv1d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+class ScaledLSTM(nn.LSTM):
+    # See docs for ScaledLinear.
+    # This class implements LSTM with scaling mechanism, using `torch._VF.lstm`
+    # Please refer to https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        grad_norm_threshold: float = 10.0,
+        **kwargs,
+    ):
+        super(ScaledLSTM, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self._scales_names = []
+        self._scales = []
+        self.batch_dim = 0 if self.batch_first else 1
+        self.num_directions = 1 + int(self.bidirectional)
+        for name in self._flat_weights_names:
+            scale_name = name + "_scale"
+            self._scales_names.append(scale_name)
+            param = nn.Parameter(initial_scale.clone().detach())
+            setattr(self, scale_name, param)
+            self._scales.append(param)
+
+        self.grad_filter = GradientFilter(
+            batch_dim=self.batch_dim, threshold=grad_norm_threshold
+        )
+
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        scale = self.hidden_size**-0.5
+        v = scale / std
+        for idx, name in enumerate(self._flat_weights_names):
+            if "weight" in name:
+                nn.init.uniform_(self._flat_weights[idx], -a, a)
+                with torch.no_grad():
+                    self._scales[idx] += torch.tensor(v).log()
+            elif "bias" in name:
+                nn.init.constant_(self._flat_weights[idx], 0.0)
+
+    def _flatten_parameters(self, flat_weights) -> None:
+        """Resets parameter data pointer so that they can use faster code paths.
+
+        Right now, this works only if the module is on the GPU and cuDNN is enabled.
+        Otherwise, it's a no-op.
+
+        This function is modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py  # noqa
+        """
+        # Short-circuits if _flat_weights is only partially instantiated
+        if len(flat_weights) != len(self._flat_weights_names):
+            return
+
+        for w in flat_weights:
+            if not isinstance(w, Tensor):
+                return
+        # Short-circuits if any tensor in flat_weights is not acceptable to cuDNN
+        # or the tensors in flat_weights are of different dtypes
+
+        first_fw = flat_weights[0]
+        dtype = first_fw.dtype
+        for fw in flat_weights:
+            if (
+                not isinstance(fw.data, Tensor)
+                or not (fw.data.dtype == dtype)
+                or not fw.data.is_cuda
+                or not torch.backends.cudnn.is_acceptable(fw.data)
+            ):
+                return
+
+        # If any parameters alias, we fall back to the slower, copying code path. This is
+        # a sufficient check, because overlapping parameter buffers that don't completely
+        # alias would break the assumptions of the uniqueness check in
+        # Module.named_parameters().
+        unique_data_ptrs = set(p.data_ptr() for p in flat_weights)
+        if len(unique_data_ptrs) != len(flat_weights):
+            return
+
+        with torch.cuda.device_of(first_fw):
+
+            # Note: no_grad() is necessary since _cudnn_rnn_flatten_weight is
+            # an inplace operation on self._flat_weights
+            with torch.no_grad():
+                if torch._use_cudnn_rnn_flatten_weight():
+                    num_weights = 4 if self.bias else 2
+                    if self.proj_size > 0:
+                        num_weights += 1
+                    torch._cudnn_rnn_flatten_weight(
+                        flat_weights,
+                        num_weights,
+                        self.input_size,
+                        rnn.get_cudnn_mode(self.mode),
+                        self.hidden_size,
+                        self.proj_size,
+                        self.num_layers,
+                        self.batch_first,
+                        bool(self.bidirectional),
+                    )
+
+    def _get_flat_weights(self):
+        """Get scaled weights, and resets their data pointer."""
+        flat_weights = []
+        for idx in range(len(self._flat_weights_names)):
+            flat_weights.append(self._flat_weights[idx] * self._scales[idx].exp())
+        self._flatten_parameters(flat_weights)
+        return flat_weights
+
+    def forward(self, input: Tensor, hx: Optional[Tuple[Tensor, Tensor]] = None):
+        # This function is modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py  # noqa
+        # The change for calling `_VF.lstm()` is:
+        # self._flat_weights -> self._get_flat_weights()
+        if hx is None:
+            h_zeros = torch.zeros(
+                self.num_layers * self.num_directions,
+                input.size(self.batch_dim),
+                self.proj_size if self.proj_size > 0 else self.hidden_size,
+                dtype=input.dtype,
+                device=input.device,
+            )
+            c_zeros = torch.zeros(
+                self.num_layers * self.num_directions,
+                input.size(self.batch_dim),
+                self.hidden_size,
+                dtype=input.dtype,
+                device=input.device,
+            )
+            hx = (h_zeros, c_zeros)
+
+        self.check_forward_args(input, hx, None)
+
+        flat_weights = self._get_flat_weights()
+        input, *flat_weights = self.grad_filter(input, *flat_weights)
+
+        result = _VF.lstm(
+            input,
+            hx,
+            flat_weights,
+            self.bias,
+            self.num_layers,
+            self.dropout,
+            self.training,
+            self.bidirectional,
+            self.batch_first,
+        )
+
+        output = result[0]
+        hidden = result[1:]
+        return output, hidden
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           sign_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_positive and max_positive
+              are violated.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+          min_prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.  Early in training we may use
+             higher probabilities than this; it will decay to this value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.04,
+        sign_gain_factor: float = 0.01,
+        scale_gain_factor: float = 0.02,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+        min_prob: float = 0.1,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.min_prob = min_prob
+        self.sign_gain_factor = sign_gain_factor
+        self.scale_gain_factor = scale_gain_factor
+
+        # count measures how many times the forward() function has been called.
+        # We occasionally sync this to a tensor called `count`, that exists to
+        # make sure it is synced to disk when we load and save the model.
+        self.cpu_count = 0
+        self.register_buffer("count", torch.tensor(0, dtype=torch.int64))
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or not x.requires_grad or torch.jit.is_tracing():
+            return _no_op(x)
+
+        count = self.cpu_count
+        self.cpu_count += 1
+
+        if random.random() < 0.01:
+            # Occasionally sync self.cpu_count with self.count.
+            # count affects the decay of 'prob'.  don't do this on every iter,
+            # because syncing with the GPU is slow.
+            self.cpu_count = max(self.cpu_count, self.count.item())
+            self.count.fill_(self.cpu_count)
+
+        # the prob of doing some work exponentially decreases from 0.5 till it hits
+        # a floor at min_prob (==0.1, by default)
+        prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0)))
+
+        if random.random() < prob:
+            sign_gain_factor = 0.5
+            if self.min_positive != 0.0 or self.max_positive != 1.0:
+                sign_factor = _compute_sign_factor(
+                    x,
+                    self.channel_dim,
+                    self.min_positive,
+                    self.max_positive,
+                    gain_factor=self.sign_gain_factor / prob,
+                    max_factor=self.max_factor,
+                )
+            else:
+                sign_factor = None
+
+            scale_factor = _compute_scale_factor(
+                x.detach(),
+                self.channel_dim,
+                min_abs=self.min_abs,
+                max_abs=self.max_abs,
+                gain_factor=self.scale_gain_factor / prob,
+                max_factor=self.max_factor,
+            )
+            return ActivationBalancerFunction.apply(
+                x,
+                scale_factor,
+                sign_factor,
+                self.channel_dim,
+            )
+        else:
+            return _no_op(x)
+
+
+def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float) -> Tensor:
+    """
+    Returns x unmodified, but in backprop will put a penalty for the excess of
+    the absolute values of elements of x over the limit "limit".  E.g. if
+    limit == 10.0, then if x has any values over 10 it will get a penalty.
+
+    Caution: the value of this penalty will be affected by grad scaling used
+    in automatic mixed precision training.  For this reasons we use this,
+    it shouldn't really matter, or may even be helpful; we just use this
+    to disallow really implausible values of scores to be given to softmax.
+    """
+    x_sign = x.sign()
+    over_limit = (x.abs() - limit) > 0
+    # The following is a memory efficient way to penalize the absolute values of
+    # x that's over the limit.  (The memory efficiency comes when you think
+    # about which items torch needs to cache for the autograd, and which ones it
+    # can throw away).  The numerical value of aux_loss as computed here will
+    # actually be larger than it should be, by limit * over_limit.sum(), but it
+    # has the same derivative as the real aux_loss which is penalty * (x.abs() -
+    # limit).relu().
+    aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x)
+    # note: we don't do sum() here on aux)_loss, but it's as if we had done
+    # sum() due to how with_loss() works.
+    x = with_loss(x, aux_loss)
+    # you must use x for something, or this will be ineffective.
+    return x
+
+
+def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
+    if x.ndim == 2:
+        return x.diag()
+    else:
+        (batch, dim, dim) = x.shape
+        x = x.reshape(batch, dim * dim)
+        x = x[:, :: dim + 1]
+        assert x.shape == (batch, dim)
+        return x
+
+
+def _whitening_metric(x: Tensor, num_groups: int):
+    """
+    Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
+    of the centered feature covariance are the same within each group's covariance matrix
+    and also between groups.
+    Args:
+        x: a Tensor of shape (*, num_channels)
+     num_groups:  the number of groups of channels, a number >=1 that divides num_channels
+    Returns:
+        Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and
+    greater than 1.0 otherwise.
+    """
+    assert x.dtype != torch.float16
+    x = x.reshape(-1, x.shape[-1])
+    (num_frames, num_channels) = x.shape
+    assert num_channels % num_groups == 0
+    channels_per_group = num_channels // num_groups
+    x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1)
+    # x now has shape (num_groups, num_frames, channels_per_group)
+    # subtract the mean so we use the centered, not uncentered, covariance.
+    # My experience has been that when we "mess with the gradients" like this,
+    # it's better not do anything that tries to move the mean around, because
+    # that can easily cause instability.
+    x = x - x.mean(dim=1, keepdim=True)
+    # x_covar: (num_groups, channels_per_group, channels_per_group)
+    x_covar = torch.matmul(x.transpose(1, 2), x)
+    x_covar_mean_diag = _diag(x_covar).mean()
+    # the following expression is what we'd get if we took the matrix product
+    # of each covariance and measured the mean of its trace, i.e.
+    # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
+    x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group)
+    # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20)
+    return metric
+
+
+class WhiteningPenaltyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx, x: Tensor, num_groups: int, whitening_limit: float, grad_scale: float
+    ) -> Tensor:
+        ctx.save_for_backward(x)
+        ctx.num_groups = num_groups
+        ctx.whitening_limit = whitening_limit
+        ctx.grad_scale = grad_scale
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
+        with torch.enable_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                x_detached = x_orig.to(torch.float32).detach()
+                x_detached.requires_grad = True
+
+                metric = _whitening_metric(x_detached, ctx.num_groups)
+
+                if random.random() < 0.005 or __name__ == "__main__":
+                    logging.info(
+                        f"Whitening: num_groups={ctx.num_groups}, num_channels={x_orig.shape[-1]}, "
+                        f"metric={metric.item():.2f} vs. limit={ctx.whitening_limit}"
+                    )
+
+                (metric - ctx.whitening_limit).relu().backward()
+                penalty_grad = x_detached.grad
+                scale = ctx.grad_scale * (
+                    x_grad.to(torch.float32).norm() / (penalty_grad.norm() + 1.0e-20)
+                )
+                penalty_grad = penalty_grad * scale
+        return x_grad + penalty_grad.to(x_grad.dtype), None, None, None
+
+
+class Whiten(nn.Module):
+    def __init__(
+        self,
+        num_groups: int,
+        whitening_limit: float,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: float,
+    ):
+        """
+        Args:
+          num_groups: the number of groups to divide the channel dim into before
+            whitening.  We will attempt to make the feature covariance
+            within each group, after mean subtraction, as "white" as possible,
+            while having the same trace across all groups.
+         whitening_limit: a value greater than 1.0, that dictates how much
+           freedom we have to violate the constraints.  1.0 would mean perfectly
+           white, with exactly the same trace across groups; larger values
+           give more freedom.  E.g. 2.0.
+         prob: the probability with which we apply the gradient modification
+           (also affects the grad scale).  May be supplied as a float,
+           or as a pair (min_prob, max_prob)
+
+          grad_scale: determines the scale on the gradient term from this object,
+            relative to the rest of the gradient on the attention weights.
+            E.g. 0.02 (you may want to use smaller values than this if prob is large)
+        """
+        super(Whiten, self).__init__()
+        assert num_groups >= 1
+        assert whitening_limit >= 1
+        assert grad_scale >= 0
+        self.num_groups = num_groups
+        self.whitening_limit = whitening_limit
+        if isinstance(prob, float):
+            assert 0 < prob <= 1
+            self.prob = prob
+        else:
+            (self.min_prob, self.max_prob) = prob
+            assert 0 < self.min_prob < self.max_prob <= 1
+            self.prob = self.max_prob
+
+        self.grad_scale = grad_scale
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        In the forward pass, this function just returns the input unmodified.
+        In the backward pass, it will modify the gradients to ensure that the
+        distribution in each group has close to (lambda times I) as the covariance
+        after mean subtraction, with the same lambda across groups.
+        For whitening_limit > 1, there will be more freedom to violate this
+        constraint.
+
+        Args:
+           x: the input of shape (*, num_channels)
+
+        Returns:
+            x, unmodified.   You should make sure
+        you use the returned value, or the graph will be freed
+        and nothing will happen in backprop.
+        """
+        if not x.requires_grad or random.random() > self.prob or self.grad_scale == 0:
+            return _no_op(x)
+        else:
+            if hasattr(self, "min_prob") and random.random() < 0.25:
+                # occasionally switch between min_prob and max_prob, based on whether
+                # we are above or below the threshold.
+                if (
+                    _whitening_metric(x.to(torch.float32), self.num_groups)
+                    > self.whitening_limit
+                ):
+                    # there would be a change to the grad.
+                    self.prob = self.max_prob
+                else:
+                    self.prob = self.min_prob
+
+            return WhiteningPenaltyFunction.apply(
+                x, self.num_groups, self.whitening_limit, self.grad_scale
+            )
+
+
+class WithLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, y: Tensor):
+        ctx.y_shape = y.shape
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        return (
+            ans_grad,
+            torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
+        )
+
+
+def with_loss(x, y):
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    # returns x but adds y.sum() to the loss function.
+    return WithLoss.apply(x, y)
+
+
+def _no_op(x: Tensor) -> Tensor:
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    else:
+        # a no-op function that will have a node in the autograd graph,
+        # to avoid certain bugs relating to backward hooks
+        return x.chunk(1, dim=-1)[0]
+
+
+class Identity(torch.nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, x):
+        return _no_op(x)
+
+
+class MaxEig(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to discourage
+    that any given direction in activation space accounts for more than
+    a specified proportion of the covariance (e.g. 0.2).
+
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           max_var_per_eig:  the maximum proportion of the variance of the
+               features/channels, after mean subtraction, that can come from
+               any given eigenvalue.
+           min_prob: the minimum probability with which we apply this during any invocation
+               of forward(), assuming last time we applied the constraint it was
+               not active; supplied for speed.
+           scale: determines the scale with which we modify the gradients, relative
+               to the existing / unmodified gradients
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        max_var_per_eig: float = 0.2,
+        min_prob: float = 0.01,
+        scale: float = 0.01,
+    ):
+        super(MaxEig, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.scale = scale
+        assert max_var_per_eig == 0.0 or max_var_per_eig > 1.0 / num_channels
+        self.max_var_per_eig = max_var_per_eig
+
+        # we figure out the dominant direction using the power method: starting with
+        # a random vector, keep multiplying by the covariance and renormalizing.
+        with torch.no_grad():
+            # arbitrary.. would use randn() but want to leave the rest of the model's
+            # random parameters unchanged for comparison
+            direction = torch.arange(num_channels).to(torch.float)
+            direction = direction / direction.norm()
+            self.register_buffer("max_eig_direction", direction)
+
+        self.min_prob = min_prob
+        # cur_prob is the current probability we'll use to apply the ActivationBalancer.
+        # We'll regress this towards prob, each tiem we try to apply it and it is not
+        # active.
+        self.cur_prob = 1.0
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or self.max_var_per_eig <= 0
+            or random.random() > self.cur_prob
+            or torch.jit.is_tracing()
+        ):
+            return _no_op(x)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            eps = 1.0e-20
+            orig_x = x
+            x = x.to(torch.float32)
+            with torch.no_grad():
+                x = x.transpose(self.channel_dim, -1).reshape(-1, self.num_channels)
+                x = x - x.mean(dim=0)
+                new_direction, coeffs = self._find_direction_coeffs(
+                    x, self.max_eig_direction
+                )
+                x_var = (x**2).mean()
+                x_residual = x - coeffs * new_direction
+                x_residual_var = (x_residual**2).mean()
+
+                # `variance_proportion` is the proportion of the variance accounted for
+                # by the top eigen-direction.
+                variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+
+                # ensure new direction is nonzero even if x == 0, by including `direction`.
+                self._set_direction(0.1 * self.max_eig_direction + new_direction)
+
+            if random.random() < 0.01 or __name__ == "__main__":
+                logging.info(
+                    f"variance_proportion = {variance_proportion.item()}, shape={tuple(orig_x.shape)}, cur_prob={self.cur_prob}"
+                )
+
+            if variance_proportion >= self.max_var_per_eig:
+                # The constraint is active.  Note, we should quite rarely
+                # reach here, only near the beginning of training if we are
+                # starting to diverge, should this constraint be active.
+                cur_prob = self.cur_prob
+                self.cur_prob = 1.0  # next time, do the update with probability 1.0.
+                return MaxEigLimiterFunction.apply(
+                    orig_x, coeffs, new_direction, self.channel_dim, self.scale
+                )
+            else:
+                # let self.cur_prob exponentially approach self.min_prob, as
+                # long as the constraint is inactive.
+                self.cur_prob = 0.75 * self.cur_prob + 0.25 * self.min_prob
+                return orig_x
+
+    def _set_direction(self, direction: Tensor):
+        """
+        Sets self.max_eig_direction to a normalized version of `direction`
+        """
+        direction = direction.detach()
+        direction = direction / direction.norm()
+        direction_sum = direction.sum().item()
+        if direction_sum - direction_sum == 0:  # no inf/nan
+            self.max_eig_direction[:] = direction
+        else:
+            logging.info(
+                f"Warning: sum of direction in MaxEig is {direction_sum}, "
+                "num_channels={self.num_channels}, channel_dim={self.channel_dim}"
+            )
+
+    def _find_direction_coeffs(
+        self, x: Tensor, prev_direction: Tensor
+    ) -> Tuple[Tensor, Tensor, Tensor]:
+        """
+            Figure out (an approximation to) the proportion of the variance of a set of
+            feature vectors that can be attributed to the top eigen-direction.
+            Args:
+             x: a Tensor of shape (num_frames, num_channels), with num_frames > 1.
+          prev_direction:  a Tensor of shape (num_channels,), that is our previous estimate
+                   of the top eigen-direction, or a random direction if this is the first
+                   iteration.  Does not have to be normalized, but should be nonzero.
+
+        Returns: (cur_direction, coeffs), where:
+             cur_direction: a Tensor of shape (num_channels,) that is the current
+                estimate of the top eigen-direction.
+             coeffs: a Tensor of shape (num_frames, 1) that minimizes, or
+                approximately minimizes, (x - coeffs * cur_direction).norm()
+        """
+        (num_frames, num_channels) = x.shape
+        assert num_channels > 1 and num_frames > 1
+        assert prev_direction.shape == (num_channels,)
+        # `coeffs` are the coefficients of `prev_direction` in x.
+        # actually represent the coeffs up to a constant positive factor.
+        coeffs = (x * prev_direction).sum(dim=1, keepdim=True) + 1.0e-10
+        cur_direction = (x * coeffs).sum(dim=0) / ((coeffs**2).sum() + 1.0e-20)
+        return cur_direction, coeffs
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.043637 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.043637
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+def _test_max_eig():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = MaxEig(
+            num_channels, 1, 0.5, scale=0.1  # channel_dim  # max_var_per_eig
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad, atol=1.0e-02)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_whiten():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"_test_whiten(): proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_activation_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        probs.numel(),
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        magnitudes.numel(),
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+        min_prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x**2).mean().sqrt()
+    y_rms = (y**2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = DoubleSwish()
+
+    tol = (1.2 - (-0.043637)) / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_softmax():
+    a = torch.randn(2, 10, dtype=torch.float64)
+    b = a.clone()
+    a.requires_grad = True
+    b.requires_grad = True
+    a.softmax(dim=1)[:, 0].sum().backward()
+    print("a grad = ", a.grad)
+    softmax(b, dim=1)[:, 0].sum().backward()
+    print("b grad = ", b.grad)
+    assert torch.allclose(a.grad, b.grad)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_softmax()
+    _test_whiten()
+    _test_max_eig()
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()
diff --git a/egs/libricss/SURT/dprnn_zipformer/scaling_converter.py b/egs/libricss/SURT/dprnn_zipformer/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/dprnn_zipformer/train.py b/egs/libricss/SURT/dprnn_zipformer/train.py
new file mode 100755
index 000000000..6598f8b5d
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/train.py
@@ -0,0 +1,1452 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#              2023  Johns Hopkins University (author: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+cd egs/libricss/SURT
+./prepare.sh
+
+./dprnn_zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir dprnn_zipformer/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./dprnn_zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir dprnn_zipformer/exp \
+  --max-duration 550
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriCssAsrDataModule
+from decoder import Decoder
+from dprnn import DPRNN
+from einops.layers.torch import Rearrange
+from graph_pit.loss.optimized import optimized_graph_pit_mse_loss as gpit_mse
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import LOG_EPSILON, fix_random_seed
+from model import SURT
+from optim import Eden, ScaledAdam
+from scaling import ScaledLSTM
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-mask-encoder-layers",
+        type=int,
+        default=4,
+        help="Number of layers in the DPRNN based mask encoder.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-dim",
+        type=int,
+        default=256,
+        help="Hidden dimension of the LSTM blocks in DPRNN.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-segment-size",
+        type=int,
+        default=32,
+        help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the "
+        "decode-chunk-length of the zipformer encoder.",
+    )
+
+    parser.add_argument(
+        "--chunk-width-randomization",
+        type=bool,
+        default=False,
+        help="Whether to randomize the chunk width in DPRNN.",
+    )
+
+    # Zipformer config is based on:
+    # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,2,2,2",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="768,768,768,768,768",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--use-joint-encoder-layer",
+        type=str,
+        default="lstm",
+        choices=["linear", "lstm", "none"],
+        help="Whether to use a joint layer to combine all branches.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conv_lstm_transducer_stateless_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--model-init-ckpt",
+        type=str,
+        default=None,
+        help="""The model checkpoint to initialize the model (either full or part).
+        If not specified, the model is randomly initialized.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.004, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--heat-loss-scale",
+        type=float,
+        default=0.0,
+        help="Scale for HEAT loss on separated sources.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=1,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 2000,
+            # parameters for SURT
+            "num_channels": 2,
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed
+            # parameters for Noam
+            "model_warm_step": 5000,  # arg given to model, not for lrate
+            # parameters for ctc loss
+            "beam_size": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_mask_encoder_model(params: AttributeDict) -> nn.Module:
+    mask_encoder = DPRNN(
+        feature_dim=params.feature_dim,
+        input_size=params.mask_encoder_dim,
+        hidden_size=params.mask_encoder_dim,
+        output_size=params.feature_dim * params.num_channels,
+        segment_size=params.mask_encoder_segment_size,
+        num_blocks=params.num_mask_encoder_layers,
+        chunk_width_randomization=params.chunk_width_randomization,
+    )
+    return mask_encoder
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_joint_encoder_layer(params: AttributeDict) -> nn.Module:
+    class TakeFirst(nn.Module):
+        def forward(self, x):
+            return x[0]
+
+    if params.use_joint_encoder_layer == "linear":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            nn.Linear(
+                params.num_channels * encoder_dim, params.num_channels * encoder_dim
+            ),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "lstm":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            ScaledLSTM(
+                input_size=params.num_channels * encoder_dim,
+                hidden_size=params.num_channels * encoder_dim,
+                num_layers=1,
+                bias=True,
+                batch_first=True,
+                dropout=0.0,
+                bidirectional=False,
+            ),
+            TakeFirst(),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "none":
+        joint_layer = None
+    else:
+        raise ValueError(
+            f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}"
+        )
+    return joint_layer
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_surt_model(
+    params: AttributeDict,
+) -> nn.Module:
+    mask_encoder = get_mask_encoder_model(params)
+    encoder = get_encoder_model(params)
+    joint_layer = get_joint_encoder_layer(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = SURT(
+        mask_encoder=mask_encoder,
+        encoder=encoder,
+        joint_encoder_layer=joint_layer,
+        decoder=decoder,
+        joiner=joiner,
+        num_channels=params.num_channels,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor:
+    """
+    Compute HEAT loss for separated sources using the output of mask encoder.
+    Args:
+      x_masked:
+        The output of mask encoder. It is a tensor of shape (B, T, C).
+      batch:
+        A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()`
+        for the content in it.
+      num_channels:
+        The number of output branches in the SURT model.
+    """
+    B, T, D = x_masked[0].shape
+    device = x_masked[0].device
+
+    # Create training targets for each channel.
+    targets = []
+    for i in range(num_channels):
+        target = torch.ones_like(x_masked[i]) * LOG_EPSILON
+        targets.append(target)
+
+    source_feats = batch["source_feats"]
+    source_boundaries = batch["source_boundaries"]
+    input_lens = batch["input_lens"].to(device)
+    # Assign sources to channels based on the HEAT criteria
+    for b in range(B):
+        cut_source_feats = source_feats[b]
+        cut_source_boundaries = source_boundaries[b]
+        last_seg_end = [0 for _ in range(num_channels)]
+        for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries):
+            assigned = False
+            for i in range(num_channels):
+                if start >= last_seg_end[i]:
+                    targets[i][b, start:end, :] += source_feat.to(device)
+                    last_seg_end[i] = max(end, last_seg_end[i])
+                    assigned = True
+                    break
+            if not assigned:
+                min_end_channel = last_seg_end.index(min(last_seg_end))
+                targets[min_end_channel][b, start:end, :] += source_feat
+                last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel])
+
+    # Get padding mask based on input lengths
+    pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1)
+    pad_mask = pad_mask.unsqueeze(-1)
+
+    # Compute masked loss for each channel
+    losses = torch.zeros((num_channels, B, T, D), device=device)
+    for i in range(num_channels):
+        loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none")
+        # Apply padding mask to loss
+        loss.masked_fill_(pad_mask, 0)
+        losses[i] = loss
+
+    # loss: C x B x T x D. pad_mask: B x T x 1
+    # We want to compute loss for each item in the batch. Each item has loss given
+    # by the sum over C, and average over T and D. For T, we need to use the padding.
+    loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device)
+    return loss
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"].to(device)
+    feature_lens = batch["input_lens"].to(device)
+
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+
+    # The dataloader returns text as a list of cuts, each of which is a list of channel
+    # text. We flatten this to a list where all channels are together, i.e., it looks like
+    # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2].
+    text = [val for tup in zip(*batch["text"]) for val in tup]
+    assert len(text) == len(feature) * params.num_channels
+
+    # Convert all channel texts to token IDs and create a ragged tensor.
+    y = sp.encode(text, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.model_warm_step
+
+    with torch.set_grad_enabled(is_training):
+        (simple_loss, pruned_loss, ctc_loss, x_masked) = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            reduction="none",
+            subsampling_factor=params.subsampling_factor,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        ctc_loss_is_finite = torch.isfinite(ctc_loss)
+
+        # Compute HEAT loss
+        if is_training and params.heat_loss_scale > 0.0:
+            heat_loss = compute_heat_loss(
+                x_masked, batch, num_channels=params.num_channels
+            )
+        else:
+            heat_loss = torch.tensor(0.0, device=device)
+
+        heat_loss_is_finite = torch.isfinite(heat_loss)
+        is_finite = (
+            simple_loss_is_finite
+            & pruned_loss_is_finite
+            & ctc_loss_is_finite
+            & heat_loss_is_finite
+        )
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_losses: {simple_loss}\n"
+                f"pruned_losses: {pruned_loss}\n"
+                f"ctc_losses: {ctc_loss}\n"
+                f"heat_losses: {heat_loss}\n"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+            ctc_loss = ctc_loss[ctc_loss_is_finite]
+            heat_loss = heat_loss[heat_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if (
+                torch.all(~simple_loss_is_finite)
+                or torch.all(~pruned_loss_is_finite)
+                or torch.all(~ctc_loss_is_finite)
+                or torch.all(~heat_loss_is_finite)
+            ):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss_sum = simple_loss.sum()
+        pruned_loss_sum = pruned_loss.sum()
+        ctc_loss_sum = ctc_loss.sum()
+        heat_loss_sum = heat_loss.sum()
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+        loss = (
+            simple_loss_scale * simple_loss_sum
+            + pruned_loss_scale * pruned_loss_sum
+            + params.ctc_loss_scale * ctc_loss_sum
+            + params.heat_loss_scale * heat_loss_sum
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss_sum.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss_sum.detach().cpu().item()
+    if params.ctc_loss_scale > 0.0:
+        info["ctc_loss"] = ctc_loss_sum.detach().cpu().item()
+    if params.heat_loss_scale > 0.0:
+        info["heat_loss"] = heat_loss_sum.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    train_dl_warmup: Optional[torch.utils.data.DataLoader],
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      train_dl_warmup:
+        Dataloader for the training dataset with 2 speakers. This is used during the
+        warmup stage.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    torch.cuda.empty_cache()
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    iter_train = iter(train_dl)
+    iter_train_warmup = iter(train_dl_warmup) if train_dl_warmup is not None else None
+
+    batch_idx = 0
+
+    while True:
+        # We first sample a batch from the main dataset. This is because we want to
+        # make sure all epochs have the same number of batches.
+        try:
+            batch = next(iter_train)
+        except StopIteration:
+            break
+
+        # If we are in warmup stage, get the batch from the warmup dataset.
+        if (
+            params.batch_idx_train <= params.model_warm_step
+            and iter_train_warmup is not None
+        ):
+            try:
+                batch = next(iter_train_warmup)
+            except StopIteration:
+                iter_train_warmup = iter(train_dl_warmup)
+                batch = next(iter_train_warmup)
+
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = batch["inputs"].shape[0]
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+
+    if checkpoints is None and params.model_init_ckpt is not None:
+        logging.info(
+            f"Initializing model with checkpoint from {params.model_init_ckpt}"
+        )
+        init_ckpt = torch.load(params.model_init_ckpt, map_location=device)
+        model.load_state_dict(init_ckpt["model"], strict=False)
+
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    libricss = LibriCssAsrDataModule(args)
+
+    train_cuts = libricss.lsmix_cuts(rvb_affix="comb", type_affix="full", sources=True)
+    train_cuts_ov40 = libricss.lsmix_cuts(
+        rvb_affix="comb", type_affix="ov40", sources=True
+    )
+    dev_cuts = libricss.libricss_cuts(split="dev", type="sdm")
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = libricss.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+    )
+    train_dl_ov40 = libricss.train_dataloaders(train_cuts_ov40)
+    valid_dl = libricss.valid_dataloaders(dev_cuts)
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            train_dl_warmup=train_dl_ov40,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = [sp.encode(text_ch) for text_ch in batch["text"]]
+    num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y]
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def main():
+    parser = get_parser()
+    LibriCssAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libricss/SURT/dprnn_zipformer/train_adapt.py b/egs/libricss/SURT/dprnn_zipformer/train_adapt.py
new file mode 100755
index 000000000..1c1b0c28c
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/train_adapt.py
@@ -0,0 +1,1343 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES=0
+
+./dprnn_zipformer/train.py \
+  --world-size 1 \
+  --num-epochs 15 \
+  --start-epoch 1 \
+  --exp-dir dprnn_zipformer/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./dprnn_zipformer/train.py \
+  --world-size 1 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir dprnn_zipformer/exp \
+  --max-duration 550
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from itertools import chain
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriCssAsrDataModule
+from decoder import Decoder
+from dprnn import DPRNN
+from einops.layers.torch import Rearrange
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import LOG_EPSILON, fix_random_seed
+from model import SURT
+from optim import Eden, ScaledAdam
+from scaling import ScaledLinear, ScaledLSTM
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-mask-encoder-layers",
+        type=int,
+        default=4,
+        help="Number of layers in the DPRNN based mask encoder.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-dim",
+        type=int,
+        default=256,
+        help="Hidden dimension of the LSTM blocks in DPRNN.",
+    )
+
+    parser.add_argument(
+        "--mask-encoder-segment-size",
+        type=int,
+        default=32,
+        help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the "
+        "decode-chunk-length of the zipformer encoder.",
+    )
+
+    parser.add_argument(
+        "--chunk-width-randomization",
+        type=bool,
+        default=False,
+        help="Whether to randomize the chunk width in DPRNN.",
+    )
+
+    # Zipformer config is based on:
+    # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,2,2,2",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="768,768,768,768,768",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--use-joint-encoder-layer",
+        type=str,
+        default="lstm",
+        choices=["linear", "lstm", "none"],
+        help="Whether to use a joint layer to combine all branches.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=15,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conv_lstm_transducer_stateless_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--model-init-ckpt",
+        type=str,
+        default=None,
+        help="""The model checkpoint to initialize the model (either full or part).
+        If not specified, the model is randomly initialized.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.0004, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=1000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=2,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=1000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=5,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 100,
+            # parameters for SURT
+            "num_channels": 2,
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed
+            # parameters for Noam
+            "model_warm_step": 5000,  # arg given to model, not for lrate
+            # parameters for ctc loss
+            "beam_size": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_mask_encoder_model(params: AttributeDict) -> nn.Module:
+    mask_encoder = DPRNN(
+        feature_dim=params.feature_dim,
+        input_size=params.mask_encoder_dim,
+        hidden_size=params.mask_encoder_dim,
+        output_size=params.feature_dim * params.num_channels,
+        segment_size=params.mask_encoder_segment_size,
+        num_blocks=params.num_mask_encoder_layers,
+        chunk_width_randomization=params.chunk_width_randomization,
+    )
+    return mask_encoder
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_joint_encoder_layer(params: AttributeDict) -> nn.Module:
+    class TakeFirst(nn.Module):
+        def forward(self, x):
+            return x[0]
+
+    if params.use_joint_encoder_layer == "linear":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            nn.Linear(
+                params.num_channels * encoder_dim, params.num_channels * encoder_dim
+            ),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "lstm":
+        encoder_dim = int(params.encoder_dims.split(",")[-1])
+        joint_layer = nn.Sequential(
+            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
+            ScaledLSTM(
+                input_size=params.num_channels * encoder_dim,
+                hidden_size=params.num_channels * encoder_dim,
+                num_layers=1,
+                bias=True,
+                batch_first=True,
+                dropout=0.0,
+                bidirectional=False,
+            ),
+            TakeFirst(),
+            nn.ReLU(),
+            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
+        )
+    elif params.use_joint_encoder_layer == "none":
+        joint_layer = None
+    else:
+        raise ValueError(
+            f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}"
+        )
+    return joint_layer
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_surt_model(
+    params: AttributeDict,
+) -> nn.Module:
+    mask_encoder = get_mask_encoder_model(params)
+    encoder = get_encoder_model(params)
+    joint_layer = get_joint_encoder_layer(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = SURT(
+        mask_encoder=mask_encoder,
+        encoder=encoder,
+        joint_encoder_layer=joint_layer,
+        decoder=decoder,
+        joiner=joiner,
+        num_channels=params.num_channels,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"].to(device)
+    feature_lens = batch["input_lens"].to(device)
+
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+
+    # The dataloader returns text as a list of cuts, each of which is a list of channel
+    # text. We flatten this to a list where all channels are together, i.e., it looks like
+    # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2].
+    text = [val for tup in zip(*batch["text"]) for val in tup]
+    assert len(text) == len(feature) * params.num_channels
+
+    # Convert all channel texts to token IDs and create a ragged tensor.
+    y = sp.encode(text, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.model_warm_step
+
+    with torch.set_grad_enabled(is_training):
+        (simple_loss, pruned_loss, ctc_loss, x_masked) = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            reduction="none",
+            subsampling_factor=params.subsampling_factor,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        ctc_loss_is_finite = torch.isfinite(ctc_loss)
+
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite & ctc_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_losses: {simple_loss}\n"
+                f"pruned_losses: {pruned_loss}\n"
+                f"ctc_losses: {ctc_loss}\n"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+            ctc_loss = ctc_loss[ctc_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if (
+                torch.all(~simple_loss_is_finite)
+                or torch.all(~pruned_loss_is_finite)
+                or torch.all(~ctc_loss_is_finite)
+            ):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss_sum = simple_loss.sum()
+        pruned_loss_sum = pruned_loss.sum()
+        ctc_loss_sum = ctc_loss.sum()
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+        loss = (
+            simple_loss_scale * simple_loss_sum
+            + pruned_loss_scale * pruned_loss_sum
+            + params.ctc_loss_scale * ctc_loss_sum
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss_sum.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss_sum.detach().cpu().item()
+    if params.ctc_loss_scale > 0.0:
+        info["ctc_loss"] = ctc_loss_sum.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      train_dl_warmup:
+        Dataloader for the training dataset with 2 speakers. This is used during the
+        warmup stage.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    torch.cuda.empty_cache()
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = batch["inputs"].shape[0]
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_surt_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+
+    if checkpoints is None and params.model_init_ckpt is not None:
+        logging.info(
+            f"Initializing model with checkpoint from {params.model_init_ckpt}"
+        )
+        init_ckpt = torch.load(params.model_init_ckpt, map_location=device)
+        model.load_state_dict(init_ckpt["model"], strict=True)
+
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    libricss = LibriCssAsrDataModule(args)
+
+    train_cuts_ihm = libricss.libricss_cuts(split="dev", type="ihm-mix")
+    train_cuts_sdm = libricss.libricss_cuts(split="dev", type="sdm")
+    train_cuts = train_cuts_ihm + train_cuts_sdm
+
+    # This will create 2 copies of the sessions with different segmentation
+    train_cuts = train_cuts.trim_to_supervision_groups(
+        max_pause=0.1
+    ) + train_cuts.trim_to_supervision_groups(max_pause=0.5)
+    dev_cuts = libricss.libricss_cuts(split="dev", type="sdm")
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = libricss.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        return_sources=False,
+        strict=False,
+    )
+    valid_dl = libricss.valid_dataloaders(dev_cuts)
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = [sp.encode(text_ch) for text_ch in batch["text"]]
+    num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y]
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def main():
+    parser = get_parser()
+    LibriCssAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libricss/SURT/dprnn_zipformer/zipformer.py b/egs/libricss/SURT/dprnn_zipformer/zipformer.py
new file mode 120000
index 000000000..ec183baa7
--- /dev/null
+++ b/egs/libricss/SURT/dprnn_zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/heat.png b/egs/libricss/SURT/heat.png
new file mode 100644
index 000000000..ac7ecfff4
Binary files /dev/null and b/egs/libricss/SURT/heat.png differ
diff --git a/egs/libricss/SURT/local/add_source_feats.py b/egs/libricss/SURT/local/add_source_feats.py
new file mode 100755
index 000000000..c9775561f
--- /dev/null
+++ b/egs/libricss/SURT/local/add_source_feats.py
@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file adds source features as temporal arrays to the mixture manifests.
+It looks for manifests in the directory data/manifests.
+"""
+import logging
+from pathlib import Path
+
+import numpy as np
+from lhotse import CutSet, LilcomChunkyWriter, load_manifest, load_manifest_lazy
+from tqdm import tqdm
+
+
+def add_source_feats(num_jobs=1):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    for type_affix in ["full", "ov40"]:
+        logging.info(f"Adding source features for {type_affix}")
+        mixed_name_clean = f"train_clean_{type_affix}"
+        mixed_name_rvb = f"train_rvb_{type_affix}"
+
+        logging.info("Reading mixed cuts")
+        mixed_cuts_clean = load_manifest_lazy(
+            src_dir / f"cuts_{mixed_name_clean}.jsonl.gz"
+        )
+        mixed_cuts_rvb = load_manifest_lazy(src_dir / f"cuts_{mixed_name_rvb}.jsonl.gz")
+
+        logging.info("Reading source cuts")
+        source_cuts = load_manifest(src_dir / "librispeech_cuts_train_trimmed.jsonl.gz")
+
+        logging.info("Adding source features to the mixed cuts")
+        with tqdm() as pbar, CutSet.open_writer(
+            src_dir / f"cuts_{mixed_name_clean}_sources.jsonl.gz"
+        ) as cut_writer_clean, CutSet.open_writer(
+            src_dir / f"cuts_{mixed_name_rvb}_sources.jsonl.gz"
+        ) as cut_writer_rvb, LilcomChunkyWriter(
+            output_dir / f"feats_train_{type_affix}_sources"
+        ) as source_feat_writer:
+            for cut_clean, cut_rvb in zip(mixed_cuts_clean, mixed_cuts_rvb):
+                assert cut_rvb.id == cut_clean.id + "_rvb"
+                # Create source_feats and source_feat_offsets
+                # (See `lhotse.datasets.K2SurtDataset` for details)
+                source_feats = []
+                source_feat_offsets = []
+                cur_offset = 0
+                for sup in sorted(
+                    cut_clean.supervisions, key=lambda s: (s.start, s.speaker)
+                ):
+                    source_cut = source_cuts[sup.id]
+                    source_feats.append(source_cut.load_features())
+                    source_feat_offsets.append(cur_offset)
+                    cur_offset += source_cut.num_frames
+                cut_clean.source_feats = source_feat_writer.store_array(
+                    cut_clean.id, np.concatenate(source_feats, axis=0)
+                )
+                cut_clean.source_feat_offsets = source_feat_offsets
+                cut_writer_clean.write(cut_clean)
+                cut_rvb.source_feats = cut_clean.source_feats
+                cut_rvb.source_feat_offsets = cut_clean.source_feat_offsets
+                cut_writer_rvb.write(cut_rvb)
+                pbar.update(1)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    add_source_feats()
diff --git a/egs/libricss/SURT/local/compute_fbank_libricss.py b/egs/libricss/SURT/local/compute_fbank_libricss.py
new file mode 100755
index 000000000..afd66899c
--- /dev/null
+++ b/egs/libricss/SURT/local/compute_fbank_libricss.py
@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriCSS dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+from pathlib import Path
+
+import pyloudnorm as pyln
+import torch
+import torch.multiprocessing
+from lhotse import LilcomChunkyWriter, load_manifest_lazy
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_libricss():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    cuts_ihm_mix = load_manifest_lazy(
+        src_dir / "libricss-ihm-mix_segments_all.jsonl.gz"
+    )
+    cuts_sdm = load_manifest_lazy(src_dir / "libricss-sdm_segments_all.jsonl.gz")
+
+    for name, cuts in [("ihm-mix", cuts_ihm_mix), ("sdm", cuts_sdm)]:
+        dev_cuts = cuts.filter(lambda c: "session0" in c.id)
+        test_cuts = cuts.filter(lambda c: "session0" not in c.id)
+
+        # If SDM cuts, apply loudness normalization
+        if name == "sdm":
+            dev_cuts = dev_cuts.normalize_loudness(target=-23.0)
+            test_cuts = test_cuts.normalize_loudness(target=-23.0)
+
+        logging.info(f"Extracting fbank features for {name} dev cuts")
+        _ = dev_cuts.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / f"libricss-{name}_feats_dev",
+            manifest_path=src_dir / f"cuts_dev_libricss-{name}.jsonl.gz",
+            batch_duration=500,
+            num_workers=2,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Extracting fbank features for {name} test cuts")
+        _ = test_cuts.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / f"libricss-{name}_feats_test",
+            manifest_path=src_dir / f"cuts_test_libricss-{name}.jsonl.gz",
+            batch_duration=2000,
+            num_workers=4,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_libricss()
diff --git a/egs/libricss/SURT/local/compute_fbank_librispeech.py b/egs/libricss/SURT/local/compute_fbank_librispeech.py
new file mode 100755
index 000000000..5c8aece9c
--- /dev/null
+++ b/egs/libricss/SURT/local/compute_fbank_librispeech.py
@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, LilcomChunkyWriter
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_librispeech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "train-clean-100",
+        "train-clean-360",
+        "train-other-500",
+    )
+    prefix = "librispeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=16000),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    for partition, m in manifests.items():
+        cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{partition} already exists - skipping.")
+            continue
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        )
+
+        cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+            manifest_path=f"{src_dir}/{cuts_filename}",
+            batch_duration=4000,
+            num_workers=2,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_librispeech()
diff --git a/egs/libricss/SURT/local/compute_fbank_lsmix.py b/egs/libricss/SURT/local/compute_fbank_lsmix.py
new file mode 100755
index 000000000..da42f8ba1
--- /dev/null
+++ b/egs/libricss/SURT/local/compute_fbank_lsmix.py
@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the synthetically mixed LibriSpeech
+train and dev sets.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+import logging
+import random
+import warnings
+from pathlib import Path
+
+import torch
+import torch.multiprocessing
+from lhotse import LilcomChunkyWriter, load_manifest
+from lhotse.cut import MixedCut, MixTrack, MultiCut
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+from lhotse.utils import fix_random_seed, uuid4
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_lsmix():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    logging.info("Reading manifests")
+    manifests = read_manifests_if_cached(
+        dataset_parts=["train_clean_full", "train_clean_ov40"],
+        types=["cuts"],
+        output_dir=src_dir,
+        prefix="lsmix",
+        suffix="jsonl.gz",
+        lazy=True,
+    )
+
+    cs = {}
+    cs["clean_full"] = manifests["train_clean_full"]["cuts"]
+    cs["clean_ov40"] = manifests["train_clean_ov40"]["cuts"]
+
+    # only uses RIRs and noises from REVERB challenge
+    real_rirs = load_manifest(src_dir / "real-rir_recordings_all.jsonl.gz").filter(
+        lambda r: "RVB2014" in r.id
+    )
+    noises = load_manifest(src_dir / "iso-noise_recordings_all.jsonl.gz").filter(
+        lambda r: "RVB2014" in r.id
+    )
+
+    # Apply perturbation to the training cuts
+    logging.info("Applying perturbation to the training cuts")
+    cs["rvb_full"] = cs["clean_full"].map(
+        lambda c: augment(
+            c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True
+        )
+    )
+    cs["rvb_ov40"] = cs["clean_ov40"].map(
+        lambda c: augment(
+            c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True
+        )
+    )
+
+    for type_affix in ["full", "ov40"]:
+        for rvb_affix in ["clean", "rvb"]:
+            logging.info(
+                f"Extracting fbank features for {type_affix} {rvb_affix} training cuts"
+            )
+            cuts = cs[f"{rvb_affix}_{type_affix}"]
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                _ = cuts.compute_and_store_features_batch(
+                    extractor=extractor,
+                    storage_path=output_dir
+                    / f"lsmix_feats_train_{rvb_affix}_{type_affix}",
+                    manifest_path=src_dir
+                    / f"cuts_train_{rvb_affix}_{type_affix}.jsonl.gz",
+                    batch_duration=5000,
+                    num_workers=4,
+                    storage_type=LilcomChunkyWriter,
+                    overwrite=True,
+                )
+
+
+def augment(cut, perturb_snr=False, rirs=None, noises=None, perturb_loudness=False):
+    """
+    Given a mixed cut, this function optionally applies the following augmentations:
+    - Perturbing the SNRs of the tracks (in range [-5, 5] dB)
+    - Reverberation using a randomly selected RIR
+    - Adding noise
+    - Perturbing the loudness (in range [-20, -25] dB)
+    """
+    out_cut = cut.drop_features()
+
+    # Perturb the SNRs (optional)
+    if perturb_snr:
+        snrs = [random.uniform(-5, 5) for _ in range(len(cut.tracks))]
+        for i, (track, snr) in enumerate(zip(out_cut.tracks, snrs)):
+            if i == 0:
+                # Skip the first track since it is the reference
+                continue
+            track.snr = snr
+
+    # Reverberate the cut (optional)
+    if rirs is not None:
+        # Select an RIR at random
+        rir = random.choice(rirs)
+        # Select a channel at random
+        rir_channel = random.choice(list(range(rir.num_channels)))
+        # Reverberate the cut
+        out_cut = out_cut.reverb_rir(rir_recording=rir, rir_channels=[rir_channel])
+
+    # Add noise (optional)
+    if noises is not None:
+        # Select a noise recording at random
+        noise = random.choice(noises).to_cut()
+        if isinstance(noise, MultiCut):
+            noise = noise.to_mono()[0]
+        # Select an SNR at random
+        snr = random.uniform(10, 30)
+        # Repeat the noise to match the duration of the cut
+        noise = repeat_cut(noise, out_cut.duration)
+        out_cut = MixedCut(
+            id=out_cut.id,
+            tracks=[
+                MixTrack(cut=out_cut, type="MixedCut"),
+                MixTrack(cut=noise, type="DataCut", snr=snr),
+            ],
+        )
+
+    # Perturb the loudness (optional)
+    if perturb_loudness:
+        target_loudness = random.uniform(-20, -25)
+        out_cut = out_cut.normalize_loudness(target_loudness, mix_first=True)
+    return out_cut
+
+
+def repeat_cut(cut, duration):
+    while cut.duration < duration:
+        cut = cut.mix(cut, offset_other_by=cut.duration)
+    return cut.truncate(duration=duration)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    fix_random_seed(42)
+    compute_fbank_lsmix()
diff --git a/egs/libricss/SURT/local/compute_fbank_musan.py b/egs/libricss/SURT/local/compute_fbank_musan.py
new file mode 100755
index 000000000..1fcf951f9
--- /dev/null
+++ b/egs/libricss/SURT/local/compute_fbank_musan.py
@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, LilcomChunkyWriter, combine
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    musan_cuts_path = src_dir / "musan_cuts.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    # create chunks of Musan with duration 5 - 10 seconds
+    _ = (
+        CutSet.from_manifests(
+            recordings=combine(part["recordings"] for part in manifests.values())
+        )
+        .cut_into_windows(10.0)
+        .filter(lambda c: c.duration > 5)
+        .compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / "musan_feats",
+            manifest_path=musan_cuts_path,
+            batch_duration=500,
+            num_workers=4,
+            storage_type=LilcomChunkyWriter,
+        )
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()
diff --git a/egs/libricss/SURT/prepare.sh b/egs/libricss/SURT/prepare.sh
new file mode 100755
index 000000000..b2d37f949
--- /dev/null
+++ b/egs/libricss/SURT/prepare.sh
@@ -0,0 +1,216 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/librispeech
+#      You can find audio and transcripts for LibriSpeech in this path.
+#
+#  - $dl_dir/libricss
+#      You can find audio and transcripts for LibriCSS in this path.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+#
+#  - $dl_dir/rirs_noises
+#      This directory contains the RIRS_NOISES corpus downloaded from https://openslr.org/28/.
+#
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+vocab_size=500
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/librispeech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/librispeech $dl_dir/librispeech
+  #
+  if [ ! -d $dl_dir/librispeech ]; then
+    lhotse download librispeech $dl_dir/librispeech
+  fi
+
+  # If you have pre-downloaded it to /path/to/libricss,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/libricss $dl_dir/libricss
+  #
+  if [ ! -d $dl_dir/libricss ]; then
+    lhotse download libricss $dl_dir/libricss
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/rirs_noises,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/rirs_noises $dl_dir/
+  #
+  if [ ! -d $dl_dir/rirs_noises ]; then
+    lhotse download rir-noise $dl_dir/rirs_noises
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LibriSpeech manifests"
+  # We assume that you have downloaded the LibriSpeech corpus
+  # to $dl_dir/librispeech. We perform text normalization for the transcripts.
+  # NOTE: Alignments are required for this recipe.
+  mkdir -p data/manifests
+
+  log "This recipe uses mfa alignment for trimming"
+  if [ ! -d $dl_dir/libri_alignments/LibriSpeech ]; then
+    log "No alignment provided. please refer to ../../librispeech/ASR/add_alignments.sh \n \
+        for mfa alignments. Once you have downloaded and unzipped the .zip file containing \n \
+        all alignments, the folder should be renamed to libri_alignments and moved to your $dl_dir ."
+    exit 0
+  fi
+
+  lhotse prepare librispeech -p train-clean-100 -p train-clean-360 -p train-other-500 -p dev-clean \
+    -j 4 --alignments-dir $dl_dir/libri_alignments/LibriSpeech $dl_dir/librispeech data/manifests/
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare LibriCSS manifests"
+  # We assume that you have downloaded the LibriCSS corpus
+  # to $dl_dir/libricss. We perform text normalization for the transcripts.
+  mkdir -p data/manifests
+  for mic in sdm ihm-mix; do
+    lhotse prepare libricss --type $mic --segmented $dl_dir/libricss data/manifests/
+  done
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare musan manifest and RIRs"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+
+  # We assume that you have downloaded the RIRS_NOISES corpus
+  # to $dl_dir/rirs_noises
+  lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises/RIRS_NOISES data/manifests
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Extract features for LibriSpeech, trim to alignments, and shuffle the cuts"
+  # python local/compute_fbank_librispeech.py
+  lhotse combine data/manifests/librispeech_cuts_train* data/manifests/librispeech_cuts_train_all.jsonl.gz
+  lhotse cut trim-to-alignments --type word --max-pause 0.2 \
+    data/manifests/librispeech_cuts_train_all.jsonl.gz \
+    data/manifests/librispeech_cuts_train_all_trimmed.jsonl.gz
+  cat <(gunzip -c data/manifests/librispeech_cuts_train_all_trimmed.jsonl.gz) | \
+    shuf | gzip -c > data/manifests/librispeech_cuts_train_trimmed.jsonl.gz
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Create simulated mixtures from LibriSpeech (train and dev). This may take a while."
+  # We create a high overlap set which will be used during the model warmup phase, and a
+  # full training set that will be used for the subsequent training.
+
+  gunzip -c data/manifests/libricss-sdm_supervisions_all.jsonl.gz |\
+    grep -v "0L" | grep -v "OV10" |\
+    gzip -c > data/manifests/libricss-sdm_supervisions_all_v1.jsonl.gz
+
+  gunzip -c data/manifests/libricss-sdm_supervisions_all.jsonl.gz |\
+    grep "OV40" |\
+    gzip -c > data/manifests/libricss-sdm_supervisions_ov40.jsonl.gz
+
+  # Warmup mixtures (100k) based on high overlap (OV40)
+  log "Generating 100k anechoic train mixtures for warmup"
+  lhotse workflows simulate-meetings \
+    --method conversational \
+    --fit-to-supervisions data/manifests/libricss-sdm_supervisions_ov40.jsonl.gz \
+    --num-meetings 100000 \
+    --num-speakers-per-meeting 2,3 \
+    --max-duration-per-speaker 15.0 \
+    --max-utterances-per-speaker 3 \
+    --seed 1234 \
+    --num-jobs 4 \
+    data/manifests/librispeech_cuts_train_trimmed.jsonl.gz \
+    data/manifests/lsmix_cuts_train_clean_ov40.jsonl.gz
+
+  # Full training set (2,3 speakers) anechoic
+  log "Generating anechoic set (full)"
+  lhotse workflows simulate-meetings \
+    --method conversational \
+    --fit-to-supervisions data/manifests/libricss-sdm_supervisions_all_v1.jsonl.gz \
+    --num-repeats 1 \
+    --num-speakers-per-meeting 2,3 \
+    --max-duration-per-speaker 15.0 \
+    --max-utterances-per-speaker 3 \
+    --seed 1234 \
+    --num-jobs 4 \
+    data/manifests/librispeech_cuts_train_trimmed.jsonl.gz \
+    data/manifests/lsmix_cuts_train_clean_full.jsonl.gz
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compute fbank features for musan"
+  mkdir -p data/fbank
+  python local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compute fbank features for simulated Libri-mix"
+  mkdir -p data/fbank
+  python local/compute_fbank_lsmix.py
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Add source feats to mixtures (useful for auxiliary tasks)"
+  python local/add_source_feats.py
+
+  log "Combining lsmix-clean and lsmix-rvb"
+  for type in full ov40; do
+    cat <(gunzip -c data/manifests/cuts_train_clean_${type}_sources.jsonl.gz) \
+      <(gunzip -c data/manifests/cuts_train_rvb_${type}_sources.jsonl.gz) |\
+      shuf | gzip -c > data/manifests/cuts_train_comb_${type}_sources.jsonl.gz
+  done
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compute fbank features for LibriCSS"
+  mkdir -p data/fbank
+  python local/compute_fbank_libricss.py
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Download LibriSpeech BPE model from HuggingFace."
+  mkdir -p data/lang_bpe_500
+  pushd data/lang_bpe_500
+  wget https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/resolve/main/data/lang_bpe_500/bpe.model
+  popd
+fi
diff --git a/egs/libricss/SURT/shared b/egs/libricss/SURT/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/libricss/SURT/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/libricss/SURT/surt.png b/egs/libricss/SURT/surt.png
new file mode 100644
index 000000000..fcc8119d4
Binary files /dev/null and b/egs/libricss/SURT/surt.png differ
diff --git a/egs/libriheavy/ASR/README.md b/egs/libriheavy/ASR/README.md
new file mode 100644
index 000000000..2498d017f
--- /dev/null
+++ b/egs/libriheavy/ASR/README.md
@@ -0,0 +1,6 @@
+# Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context
+
+Libriheavy is a labeled version of [Librilight](https://arxiv.org/pdf/1912.07875.pdf). Please refer to our repository [k2-fsa/libriheavy](https://github.com/k2-fsa/libriheavy) for more details. We also have a paper: *Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context*, [Preprint available on arxiv](https://arxiv.org/abs/2309.08105).
+
+
+See [RESULTS](./RESULTS.md) for the results for icefall recipes.
diff --git a/egs/libriheavy/ASR/RESULTS.md b/egs/libriheavy/ASR/RESULTS.md
new file mode 100644
index 000000000..513bbf72e
--- /dev/null
+++ b/egs/libriheavy/ASR/RESULTS.md
@@ -0,0 +1,315 @@
+# Results
+
+## zipformer (zipformer + pruned stateless transducer)
+
+See <https://github.com/k2-fsa/icefall/pull/1261> for more details.
+
+[zipformer](./zipformer)
+
+### Non-streaming
+
+#### Training on normalized text, i.e. Upper case without punctuation
+
+##### normal-scaled model, number of model parameters: 65805511, i.e., 65.81 M
+
+You can find a pretrained model, training logs at:
+<https://www.modelscope.cn/models/pkufool/icefall-asr-zipformer-libriheavy-20230926/summary>
+
+Note: The repository above contains three models trained on different subset of libriheavy exp(large set), exp_medium_subset(medium set),
+exp_small_subset(small set).
+
+Results of models:
+
+| training set  |  decoding method    | librispeech clean | librispeech other | libriheavy clean | libriheavy other | comment            |
+|---------------|---------------------|-------------------|-------------------|------------------|------------------|--------------------|
+| small         |  greedy search      | 4.19              | 9.99              | 4.75             | 10.25            |--epoch 90 --avg 20 |
+| small         | modified beam search| 4.05              | 9.89              | 4.68             | 10.01            |--epoch 90 --avg 20 |
+| medium        |  greedy search      | 2.39              | 4.85              | 2.90             | 6.6              |--epoch 60 --avg 20 |
+| medium        | modified beam search| 2.35              | 4.82              | 2.90             | 6.57             |--epoch 60 --avg 20 |
+| large         |  greedy search      | 1.67              | 3.32              | 2.24             | 5.61             |--epoch 16 --avg 3  |
+| large         | modified beam search| 1.62              | 3.36              | 2.20             | 5.57             |--epoch 16 --avg 3  |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+python ./zipformer/train.py \
+    --world-size 4 \
+    --master-port 12365 \
+    --exp-dir zipformer/exp \
+    --num-epochs 60 \   # 16 for large; 90 for small
+    --lr-hours 15000 \  # 20000 for large; 5000 for small
+    --use-fp16 1 \
+    --start-epoch 1 \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --max-duration 1000 \
+    --subset medium
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search; do
+  ./zipformer/decode.py \
+      --epoch 16 \
+      --avg 3 \
+      --exp-dir zipformer/exp \
+      --max-duration 1000 \
+      --causal 0 \
+      --decoding-method $m
+done
+```
+
+#### Training on full formatted text, i.e. with casing and punctuation
+
+##### normal-scaled model, number of model parameters: 66074067 , i.e., 66M
+
+You can find a pretrained model, training logs at:
+<https://www.modelscope.cn/models/pkufool/icefall-asr-zipformer-libriheavy-punc-20230830/summary>
+
+Note: The repository above contains three models trained on different subset of libriheavy exp(large set), exp_medium_subset(medium set),
+exp_small_subset(small set).
+
+Results of models:
+
+| training set  |  decoding method    | libriheavy clean (WER) | libriheavy other (WER) | libriheavy clean (CER) | libriheavy other (CER) | comment            |
+|---------------|---------------------|-------------------|-------------------|------------------|------------------|--------------------|
+| small         | modified beam search| 13.04             | 19.54             | 4.51             | 7.90             |--epoch 88 --avg 41 |
+| medium        | modified beam search| 9.84              | 13.39             | 3.02             | 5.10             |--epoch 50 --avg 15 |
+| large         | modified beam search| 7.76              | 11.32             | 2.41             | 4.22             |--epoch 16 --avg 2  |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+python ./zipformer/train.py \
+    --world-size 4 \
+    --master-port 12365 \
+    --exp-dir zipformer/exp \
+    --num-epochs 60 \   # 16 for large; 90 for small
+    --lr-hours 15000 \  # 20000 for large; 10000 for small
+    --use-fp16 1 \
+    --train-with-punctuation 1 \
+    --start-epoch 1 \
+    --bpe-model data/lang_punc_bpe_756/bpe.model \
+    --max-duration 1000 \
+    --subset medium
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search; do
+  ./zipformer/decode.py \
+      --epoch 16 \
+      --avg 3 \
+      --exp-dir zipformer/exp \
+      --max-duration 1000 \
+      --causal 0 \
+      --decoding-method $m
+done
+```
+
+## Zipformer PromptASR (zipformer + PromptASR + BERT text encoder)
+
+#### [zipformer_prompt_asr](./zipformer_prompt_asr)
+
+See <https://github.com/k2-fsa/icefall/pull/1250> for commit history and
+our paper <https://arxiv.org/abs/2309.07414> for more details.
+
+
+
+##### Training on the medium subset, with content & style prompt, **no** context list
+
+You can find a pre-trained model, training logs, decoding logs, and decoding results at: <https://huggingface.co/marcoyang/icefall-promptasr-libriheavy-zipformer-BERT-2023-10-10>
+
+The training command is:
+
+```bash
+causal=0
+subset=medium
+memory_dropout_rate=0.05
+text_encoder_type=BERT
+
+python ./zipformer_prompt_asr/train_bert_encoder.py \
+    --world-size 4 \
+    --start-epoch 1 \
+    --num-epochs 60 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --use-fp16 True \
+    --memory-dropout-rate $memory_dropout_rate \
+    --causal $causal \
+    --subset $subset \
+    --manifest-dir data/fbank \
+    --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+    --max-duration 1000 \
+    --text-encoder-type $text_encoder_type \
+    --text-encoder-dim 768 \
+    --use-context-list 0 \
+    --top-k $top_k \
+    --use-style-prompt 1
+```
+
+The decoding results using utterance-level context (epoch-60-avg-10):
+
+| decoding method      | lh-test-clean | lh-test-other | comment             |
+|----------------------|---------------|---------------|---------------------|
+| modified_beam_search        |   3.13        |    6.78       |    --use-pre-text False --use-style-prompt False      |
+| modified_beam_search        |   2.86        |    5.93       |    --pre-text-transform upper-no-punc --style-text-transform upper-no-punc       |
+| modified_beam_search        |   2.6        |    5.5       |    --pre-text-transform mixed-punc --style-text-transform mixed-punc       |
+
+
+The decoding command is:
+
+```bash
+for style in mixed-punc upper-no-punc; do
+    python ./zipformer_prompt_asr/decode_bert.py \
+        --epoch 60 \
+        --avg 10 \
+        --use-averaged-model True \
+        --post-normalization True \
+        --causal False \
+        --exp-dir ./zipformer_prompt_asr/exp \
+        --manifest-dir data/fbank \
+        --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+        --max-duration 1000 \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        --text-encoder-type BERT \
+        --text-encoder-dim 768 \
+        --memory-layer 0 \
+        --use-ls-test-set False \
+        --use-ls-context-list False \
+        --max-prompt-lens 1000 \
+        --use-pre-text True \
+        --use-style-prompt True \
+        --style-text-transform $style \
+        --pre-text-transform $style \
+        --compute-CER 0
+done
+```
+
+##### Training on the medium subset, with content & style prompt, **with** context list
+
+You can find a pre-trained model, training logs, decoding logs, and decoding results at: <https://huggingface.co/marcoyang/icefall-promptasr-with-context-libriheavy-zipformer-BERT-2023-10-10>
+
+This model is trained with an extra type of content prompt (context words), thus it does better
+on **word-level** context biasing. Note that to train this model, please first run `prepare_prompt_asr.sh`
+to prepare a manifest containing context words.
+
+The training command is:
+
+```bash
+
+causal=0
+subset=medium
+memory_dropout_rate=0.05
+text_encoder_type=BERT
+use_context_list=True
+
+# prepare the required data for context biasing
+./prepare_prompt_asr.sh --stage 0 --stop_stage 1
+
+python ./zipformer_prompt_asr/train_bert_encoder.py \
+    --world-size 4 \
+    --start-epoch 1 \
+    --num-epochs 50 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --use-fp16 True \
+    --memory-dropout-rate $memory_dropout_rate \
+    --causal $causal \
+    --subset $subset \
+    --manifest-dir data/fbank \
+    --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+    --max-duration 1000 \
+    --text-encoder-type $text_encoder_type \
+    --text-encoder-dim 768 \
+    --use-context-list $use_context_list \
+    --top-k 10000 \
+    --use-style-prompt 1
+```
+
+*Utterance-level biasing:*
+
+| decoding method      | lh-test-clean | lh-test-other | comment             |
+|----------------------|---------------|---------------|---------------------|
+| modified_beam_search        |   3.17        |    6.72       |    --use-pre-text 0 --use-style-prompt 0    |
+| modified_beam_search        |   2.91        |    6.24       |    --pre-text-transform upper-no-punc --style-text-transform upper-no-punc       |
+| modified_beam_search        |   2.72        |    5.72       |    --pre-text-transform mixed-punc --style-text-transform mixed-punc       |
+
+
+The decoding command for the table above is:
+
+```bash
+for style in mixed-punc upper-no-punc; do
+    python ./zipformer_prompt_asr/decode_bert.py \
+        --epoch 50 \
+        --avg 10 \
+        --use-averaged-model True \
+        --post-normalization True \
+        --causal False \
+        --exp-dir ./zipformer_prompt_asr/exp \
+        --manifest-dir data/fbank \
+        --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+        --max-duration 1000 \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        --text-encoder-type BERT \
+        --text-encoder-dim 768 \
+        --memory-layer 0 \
+        --use-ls-test-set False \
+        --use-ls-context-list False \
+        --max-prompt-lens 1000 \
+        --use-pre-text True \
+        --use-style-prompt True \
+        --style-text-transform $style \
+        --pre-text-transform $style \
+        --compute-CER 0
+done
+```
+
+*Word-level biasing:*
+
+The results are reported on LibriSpeech test-sets using the biasing list provided from <https://arxiv.org/abs/2104.02194>.
+You need to set `--use-ls-test-set True` so that the LibriSpeech test sets are used.
+
+| decoding method      | ls-test-clean | ls-test-other | comment             |
+|----------------------|---------------|---------------|---------------------|
+| modified_beam_search        |   2.4        |    5.08       |    --use-pre-text 0 --use-style-prompt 0    |
+| modified_beam_search        |   2.14        |    4.62       |    --use-ls-context-list 1 --pre-text-transform mixed-punc --style-text-transform mixed-punc --ls-distractors 0       |
+| modified_beam_search        |   2.14        |    4.64       |    --use-ls-context-list 1 --pre-text-transform mixed-punc --style-text-transform mixed-punc --ls-distractors 100       |
+
+The decoding command is for the table above is:
+
+```bash
+use_ls_test_set=1
+use_ls_context_list=1
+
+for ls_distractors in 0 100; do
+    python ./zipformer_prompt_asr/decode_bert.py \
+        --epoch 50 \
+        --avg 10 \
+        --use-averaged-model True \
+        --post-normalization True \
+        --causal False \
+        --exp-dir ./zipformer_prompt_asr/exp \
+        --manifest-dir data/fbank \
+        --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+        --max-duration 1000 \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        --text-encoder-type BERT \
+        --text-encoder-dim 768 \
+        --memory-layer 0 \
+        --use-ls-test-set $use_ls_test_setse \
+        --use-ls-context-list $use_ls_context_list \
+        --ls-distractors $ls_distractors \
+        --max-prompt-lens 1000 \
+        --use-pre-text True \
+        --use-style-prompt True \
+        --style-text-transform mixed-punc \
+        --pre-text-transform mixed-punc \
+        --compute-CER 0
+done
+
+```
diff --git a/egs/libriheavy/ASR/local/compute_fbank_libriheavy.py b/egs/libriheavy/ASR/local/compute_fbank_libriheavy.py
new file mode 100755
index 000000000..010531db2
--- /dev/null
+++ b/egs/libriheavy/ASR/local/compute_fbank_libriheavy.py
@@ -0,0 +1,242 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the Libriheavy dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+)
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest-dir",
+        type=str,
+        help="""The source directory that contains raw manifests.
+        """,
+        default="data/manifests",
+    )
+
+    parser.add_argument(
+        "--fbank-dir",
+        type=str,
+        help="""Fbank output dir
+        """,
+        default="data/fbank",
+    )
+
+    parser.add_argument(
+        "--subset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Whether to use speed perturbation.",
+    )
+
+    parser.add_argument(
+        "--use-splits",
+        type=str2bool,
+        default=False,
+        help="Whether to compute fbank on splits.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        help="""The number of splits of the medium and large subset.
+        Only needed when --use-splits is true.""",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="""Process pieces starting from this number (inclusive).
+        Only needed when --use-splits is true.""",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="""Stop processing pieces until this number (exclusive).
+        Only needed when --use-splits is true.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_libriheavy(args):
+    src_dir = Path(args.manifest_dir)
+    output_dir = Path(args.fbank_dir)
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+    subset = args.subset
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        output_cuts_path = output_dir / f"libriheavy_cuts_{subset}.jsonl.gz"
+        if output_cuts_path.exists():
+            logging.info(f"{output_cuts_path} exists - skipping")
+            return
+
+        input_cuts_path = src_dir / f"libriheavy_cuts_{subset}.jsonl.gz"
+        assert input_cuts_path.exists(), f"{input_cuts_path} does not exist!"
+        logging.info(f"Loading {input_cuts_path}")
+        cut_set = CutSet.from_file(input_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/libriheavy_feats_{subset}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+
+        logging.info(f"Saving to {output_cuts_path}")
+        cut_set.to_file(output_cuts_path)
+
+
+def compute_fbank_libriheavy_splits(args):
+    num_splits = args.num_splits
+    subset = args.subset
+    src_dir = f"{args.manifest_dir}/libriheavy_{subset}_split"
+    src_dir = Path(src_dir)
+    output_dir = f"{args.fbank_dir}/libriheavy_{subset}_split"
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    num_digits = 8  # num_digits is fixed by lhotse split-lazy
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"libriheavy_cuts_{subset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = src_dir / f"libriheavy_cuts_{subset}.{idx}.jsonl.gz"
+        if not raw_cuts_path.is_file():
+            logging.info(f"{raw_cuts_path} does not exist - skipping it")
+            continue
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+        if (output_dir / f"libriheavy_feats_{subset}_{idx}.lca").exists():
+            logging.info(f"Removing {output_dir}/libriheavy_feats_{subset}_{idx}.lca")
+            os.remove(output_dir / f"libriheavy_feats_{subset}_{idx}.lca")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/libriheavy_feats_{subset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            overwrite=True,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+
+    if args.use_splits:
+        assert args.num_splits is not None, "Please provide num_splits"
+        compute_fbank_libriheavy_splits(args)
+    else:
+        compute_fbank_libriheavy(args)
diff --git a/egs/libriheavy/ASR/local/compute_fbank_musan.py b/egs/libriheavy/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/libriheavy/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/local/norm_text.py b/egs/libriheavy/ASR/local/norm_text.py
new file mode 100755
index 000000000..c2fc0d92d
--- /dev/null
+++ b/egs/libriheavy/ASR/local/norm_text.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import codecs
+import sys
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="""Path to the input text.
+        """,
+    )
+    return parser.parse_args()
+
+
+def remove_punc_to_upper(text: str) -> str:
+    text = text.replace("‘", "'")
+    text = text.replace("’", "'")
+    tokens = set("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789'")
+    s_list = [x.upper() if x in tokens else " " for x in text]
+    s = " ".join("".join(s_list).split()).strip()
+    return s
+
+
+def main():
+    args = get_args()
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(sys.stdin.buffer)
+
+    sys.stdout = codecs.getwriter("utf-8")(sys.stdout.buffer)
+    line = f.readline()
+    while line:
+        print(remove_punc_to_upper(line))
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/local/prepare_manifest.py b/egs/libriheavy/ASR/local/prepare_manifest.py
new file mode 100755
index 000000000..42f392cae
--- /dev/null
+++ b/egs/libriheavy/ASR/local/prepare_manifest.py
@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gzip
+import json
+import sys
+from pathlib import Path
+
+
+def simple_cleanup(text: str) -> str:
+    table = str.maketrans("’‘，。；？！（）：-《》、“”【】", "'',.;?!(): <>/\"\"[]")
+    text = text.translate(table)
+    return text.strip()
+
+
+# Assign text of the supervisions and remove unnecessary entries.
+def main():
+    assert len(sys.argv) == 3, "Usage: ./local/prepare_manifest.py INPUT OUTPUT_DIR"
+    fname = Path(sys.argv[1]).name
+    oname = Path(sys.argv[2]) / fname
+    with gzip.open(sys.argv[1], "r") as fin, gzip.open(oname, "w") as fout:
+        for line in fin:
+            cut = json.loads(line)
+            cut["supervisions"][0]["text"] = simple_cleanup(
+                cut["supervisions"][0]["custom"]["texts"][0]
+            )
+            del cut["supervisions"][0]["custom"]
+            del cut["custom"]
+            fout.write((json.dumps(cut) + "\n").encode())
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/local/train_bpe_model.py b/egs/libriheavy/ASR/local/train_bpe_model.py
new file mode 100755
index 000000000..19caf43ab
--- /dev/null
+++ b/egs/libriheavy/ASR/local/train_bpe_model.py
@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--byte-fallback",
+        action="store_true",
+        help="""Whether to enable byte_fallback when training bpe.""",
+    )
+
+    parser.add_argument(
+        "--character-coverage",
+        type=float,
+        default=1.0,
+        help="Character coverage in vocabulary.",
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=args.character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            byte_fallback=args.byte_fallback,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/prepare.sh b/egs/libriheavy/ASR/prepare.sh
new file mode 100755
index 000000000..af7e3c5b0
--- /dev/null
+++ b/egs/libriheavy/ASR/prepare.sh
@@ -0,0 +1,314 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+export CUDA_VISIBLE_DEVICES=""
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/librilight
+#      You can find small, medium, large, etc. inside it.
+#
+#  - $dl_dir/libriheavy
+#      You can find libriheavy_cuts_small.jsonl.gz, libriheavy_cuts_medium.jsonl.gz, etc. inside it.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+fbank_dir=data/fbank
+manifests_dir=data/manifests
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download audio data."
+  # If you have pre-downloaded it to /path/to/librilight,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/librilight $dl_dir/librilight
+  #
+  mkdir -p $dl_dir/librilight
+  for subset in small medium large; do
+    log "Downloading ${subset} subset."
+    if [ ! -d $dl_dir/librilight/${subset} ]; then
+      wget -P $dl_dir/librilight -c https://dl.fbaipublicfiles.com/librilight/data/${subset}.tar 
+      tar xf $dl_dir/librilight/${subset}.tar -C $dl_dir/librilight
+    else
+      log "Skipping download, ${subset} subset exists."
+    fi
+  done
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download manifests from huggingface."
+
+  # If you have pre-downloaded it to /path/to/libriheavy,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/libriheavy $dl_dir/libriheavy
+  #
+  mkdir -p $dl_dir/libriheavy
+  for subset in small medium large dev test_clean test_other; do
+    if [ ! -e $dl_dir/libriheavy/libriheavy_cuts_${subset}.jsonl.gz ]; then
+      log "Downloading ${subset} subset."
+      wget -P $dl_dir/libriheavy -c https://huggingface.co/datasets/pkufool/libriheavy/resolve/main/libriheavy_cuts_${subset}.jsonl.gz
+    else
+      log "Skipping download, ${subset} subset exists."
+    fi
+  done
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Download manifests from modelscope"
+  mkdir -p $dl_dir/libriheavy
+  if [ ! -e $dl_dir/libriheavy/libriheavy_cuts_small.jsonl.gz ]; then
+      cd $dl_dir/libriheavy
+      GIT_LFS_SKIP_SMUDGE=1 git clone https://www.modelscope.cn/datasets/pkufool/Libriheavy.git
+      cd Libriheavy
+      git lfs pull --exclude "raw/*"
+      mv *.jsonl.gz ../
+      cd ..
+      rm -rf Libriheavy
+      cd ../../
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p $manifests_dir
+  if [ ! -e $manifests_dir/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan $manifests_dir
+    touch $manifests_dir/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare Libriheavy manifests"
+  mkdir -p $manifests_dir
+  for subset in small medium large dev test_clean test_other; do
+    if [ ! -e $manifests_dir/libriheavy_cuts_${subset}.jsonl.gz ]; then
+      log "Prepare manifest for subset : ${subset}"
+      ./local/prepare_manifest.py $dl_dir/libriheavy/libriheavy_cuts_${subset}.jsonl.gz $manifests_dir
+    fi
+  done
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p $fbank_dir
+  if [ ! -e $fbank_dir/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch $fbank_dir/.musan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for small subset and validation subsets"
+  for subset in test_clean test_other dev small; do
+    log "Computing $subset subset."
+    if [ ! -e $fbank_dir/.libriheavy.${subset}.done ]; then
+      ./local/compute_fbank_libriheavy.py \
+        --manifest-dir ${manifests_dir} \
+        --subset ${subset} \
+        --fbank-dir $fbank_dir \
+        --num-workers $nj
+    fi
+  done
+fi
+
+num_per_split=8000
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Split medium and large subsets."
+  for subset in medium large; do
+    log "Spliting subset : $subset"
+    split_dir=$manifests_dir/libriheavy_${subset}_split
+    mkdir -p $split_dir
+    if [ ! -e $split_dir/.split_completed ]; then
+      lhotse split-lazy $manifests_dir/libriheavy_cuts_${subset}.jsonl.gz $split_dir $num_per_split
+      touch $split_dir/.split_completed
+    fi
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compute fbank for medium and large subsets"
+  mkdir -p $fbank_dir
+  chunk_size=20
+  for subset in medium large; do
+    if [ $subset == "large" ]; then
+      chunk_size=200
+    fi
+    num_splits=$(find $manifests_dir/libriheavy_${subset}_split -name "libriheavy_cuts_${subset}.*.jsonl.gz" | wc -l)
+    if [ ! -e $fbank_dir/.libriheavy.${subset}.done ]; then
+      for i in $(seq 0 1 6); do
+        start=$(( i * $chunk_size ))
+        end=$(( (i+1) * $chunk_size ))
+        ./local/compute_fbank_libriheavy.py \
+          --manifest-dir ${manifests_dir} \
+          --use-splits 1 \
+          --subset ${subset} \
+          --fbank-dir $fbank_dir \
+          --num-splits $num_splits \
+          --num-workers $nj \
+          --start $start \
+          --stop $end &
+      done
+      wait
+      touch $fbank_dir/.libriheavy.${subset}.done
+    fi
+  done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Combine features for medium and large subsets."
+  for subset in medium large; do
+    log "Combining $subset subset."
+    if [ ! -f $fbank_dir/libriheavy_cuts_${subset}.jsonl.gz ]; then
+      pieces=$(find $fbank_dir/libriheavy_${subset}_split -name "libriheavy_cuts_${subset}.*.jsonl.gz")
+      lhotse combine $pieces $fbank_dir/libriheavy_cuts_${subset}.jsonl.gz
+    fi
+  done
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Train BPE model for normalized text"
+
+  if [ ! -f data/texts ]; then
+    gunzip -c $manifests_dir/libriheavy_cuts_medium.jsonl.gz \
+      | jq '.supervisions[].text' | sed 's/"//;s/\\//g;s/"$//' \
+      | ./local/norm_text.py > data/texts
+  fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    cp data/texts $lang_dir/text
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/text
+    fi
+  done
+fi
+
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Train BPE model for unnormalized text"
+  if [ ! -f data/punc_texts ]; then
+    gunzip -c $manifests_dir/libriheavy_cuts_medium.jsonl.gz \
+      | jq '.supervisions[].text' | sed 's/"//;s/\\//g;s/"$//' > data/punc_texts
+  fi
+  for vocab_size in ${vocab_sizes[@]}; do
+    new_vacab_size = $(($vocab_size + 256))
+    lang_dir=data/lang_punc_bpe_${new_vocab_size}
+    mkdir -p $lang_dir
+
+    cp data/punc_texts $lang_dir/text
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --byte-fallback \
+        --vocab-size ${new_vocab_size} \
+        --byte-fallback \
+        --character-coverage 0.99 \
+        --transcript $lang_dir/text
+    fi
+  done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Prepare language model for normalized text"
+
+  for subset in small medium large; do
+    if [ ! -f $manifests_dir/texts_${subset} ]; then
+      gunzip -c $manifests_dir/libriheavy_cuts_${subset}.jsonl.gz \
+        | jq '.supervisions[].text' | sed 's/"//;s/\\//g;s/"$//' \
+        | ./local/norm_text.py > $manifests_dir/texts_${subset}
+    fi
+  done
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/text ]; then
+    cat $manifests_dir/texts_small $manifests_dir/texts_medium $manifests_dir/texts_large > data/lm/text
+  fi
+
+  (echo '<eps> 0'; echo '!SIL 1'; echo '<SPOKEN_NOISE> 2'; echo '<UNK> 3';) \
+    > data/lm/words.txt
+
+  cat data/lm/text | sed 's/ /\n/g' | sort -u | sed '/^$/d' \
+     | awk '{print $1" "NR+3}' >> data/lm/words.txt
+
+  num_lines=$(< data/lm/words.txt wc -l)
+  (echo "#0 $num_lines"; echo "<s> $(($num_lines + 1))"; echo "</s> $(($num_lines + 2))";) \
+    >> data/lm/words.txt
+
+  # Train LM on transcripts
+  if [ ! -f data/lm/3-gram.unpruned.arpa ]; then
+    python3 ./shared/make_kn_lm.py \
+      -ngram-order 3 \
+      -text data/lm/text \
+      -lm data/lm/3-gram.unpruned.arpa
+  fi
+
+  # We assume you have install kaldilm, if not, please install
+  # it using: pip install kaldilm
+  if [ ! -f data/lm/G_3_gram_char.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table=data/lm/words.txt \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram.fst.txt
+  fi
+fi
+
diff --git a/egs/libriheavy/ASR/prepare_prompt_asr.sh b/egs/libriheavy/ASR/prepare_prompt_asr.sh
new file mode 100755
index 000000000..b931cea26
--- /dev/null
+++ b/egs/libriheavy/ASR/prepare_prompt_asr.sh
@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+# This is the preparation recipe for PromptASR: https://arxiv.org/pdf/2309.07414
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+stage=-1
+stop_stage=100
+manifest_dir=data/fbank
+subset=medium
+topk=10000
+
+. shared/parse_options.sh || exit 1
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+    log "Stage 0: Download the meta biasing list for LibriSpeech"
+    mkdir -p data/context_biasing
+    cd data/context_biasing
+    git clone https://github.com/facebookresearch/fbai-speech.git
+    cd ../..
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+    log "Stage 1: Add rare-words for context biasing to the manifest"
+    python zipformer_prompt_asr/utils.py \
+        --manifest-dir $manifest_dir \
+        --subset $subset \
+        --top-k $topk
+
+fi
diff --git a/egs/libriheavy/ASR/shared b/egs/libriheavy/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/libriheavy/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/asr_datamodule.py b/egs/libriheavy/ASR/zipformer/asr_datamodule.py
new file mode 100644
index 000000000..df761c1b8
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1,443 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022-2023  Xiaomi Corporation     (Authors: Mingshuang Luo,
+#                                                            Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriHeavyAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="S",
+            help="""The subset to be used. Should be S, M or L. Note: S subset
+            includes libriheavy_cuts_small.jsonl.gz, M subset includes
+            libriheavy_cuts_small.jsonl.gz and libriheavy_cuts_medium.jsonl.gz,
+            L subset includes libriheavy_cuts_small.jsonl.gz,
+            libriheavy_cuts_medium.jsonl.gz and libriheavy_cuts_large.jsonl.gz.
+            """,
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_small_cuts(self) -> CutSet:
+        logging.info("About to get small subset cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_small.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_medium_cuts(self) -> CutSet:
+        logging.info("About to get medium subset cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_medium.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_large_cuts(self) -> CutSet:
+        logging.info("About to get large subset cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_large.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_dev.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get the test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test_clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get the test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test_other.jsonl.gz"
+        )
diff --git a/egs/libriheavy/ASR/zipformer/beam_search.py b/egs/libriheavy/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/decode.py b/egs/libriheavy/ASR/zipformer/decode.py
new file mode 100644
index 000000000..1928e2635
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/decode.py
@@ -0,0 +1,794 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from text_normalization import remove_punc_to_upper
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--train-with-punctuation",
+        type=str2bool,
+        default=False,
+        help="""Set to True, if the model was trained on texts with casing
+        and punctuation.""",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=False,
+        help="""Upper case and remove all chars except ' and -
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        for batch_idx, batch in enumerate(dl):
+            texts = batch["supervisions"]["text"]
+
+            cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+            hyps_dict = decode_one_batch(
+                params=params,
+                model=model,
+                sp=sp,
+                decoding_graph=decoding_graph,
+                batch=batch,
+            )
+
+            for name, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[name].extend(this_batch)
+
+                this_batch = []
+                if params.post_normalization and params.train_with_punctuation:
+                    for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                        ref_words = remove_punc_to_upper(ref_text).split()
+                        hyp_words = remove_punc_to_upper(" ".join(hyp_words)).split()
+                        this_batch.append((cut_id, ref_words, hyp_words))
+
+                    results[f"{name}_norm"].extend(this_batch)
+
+            num_cuts += len(texts)
+
+            if batch_idx % log_interval == 0:
+                batch_str = f"{batch_idx}/{num_batches}"
+
+                logging.info(
+                    f"batch {batch_str}, cuts processed until now is {num_cuts}"
+                )
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    def normalize_text(c: Cut):
+        text = remove_punc_to_upper(c.supervisions[0].text)
+        c.supervisions[0].text = text
+        return c
+
+    test_clean_cuts = libriheavy.test_clean_cuts()
+    test_other_cuts = libriheavy.test_other_cuts()
+
+    if not params.train_with_punctuation:
+        test_clean_cuts = test_clean_cuts.map(normalize_text)
+        test_other_cuts = test_other_cuts.map(normalize_text)
+
+    test_clean_dl = libriheavy.test_dataloaders(test_clean_cuts)
+    test_other_dl = libriheavy.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer/decoder.py b/egs/libriheavy/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/encoder_interface.py b/egs/libriheavy/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..c2eaca671
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/encoder_interface.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/export-onnx.py b/egs/libriheavy/ASR/zipformer/export-onnx.py
new file mode 120000
index 000000000..70a15683c
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/export.py b/egs/libriheavy/ASR/zipformer/export.py
new file mode 120000
index 000000000..dfc1bec08
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/jit_pretrained.py b/egs/libriheavy/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/joiner.py b/egs/libriheavy/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/model.py b/egs/libriheavy/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/onnx_decode.py b/egs/libriheavy/ASR/zipformer/onnx_decode.py
new file mode 120000
index 000000000..0573b88c5
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/onnx_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_decode.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/onnx_pretrained.py b/egs/libriheavy/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/optim.py b/egs/libriheavy/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/pretrained.py b/egs/libriheavy/ASR/zipformer/pretrained.py
new file mode 120000
index 000000000..0bd71dde4
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/scaling.py b/egs/libriheavy/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/scaling_coverter.py b/egs/libriheavy/ASR/zipformer/scaling_coverter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/scaling_coverter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/subsampling.py b/egs/libriheavy/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer/text_normalization.py b/egs/libriheavy/ASR/zipformer/text_normalization.py
new file mode 100644
index 000000000..92590769c
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/text_normalization.py
@@ -0,0 +1,50 @@
+from num2words import num2words
+
+
+def remove_punc_to_upper(text: str) -> str:
+    text = text.replace("‘", "'")
+    text = text.replace("’", "'")
+    tokens = set("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789'")
+    s_list = [x.upper() if x in tokens else " " for x in text]
+    s = " ".join("".join(s_list).split()).strip()
+    return s
+
+
+def word_normalization(word: str) -> str:
+    # 1. Use full word for some abbreviation
+    # 2. Convert digits to english words
+    # 3. Convert ordinal number to english words
+    if word == "MRS":
+        return "MISSUS"
+    if word == "MR":
+        return "MISTER"
+    if word == "ST":
+        return "SAINT"
+    if word == "ECT":
+        return "ET CETERA"
+
+    if word[-2:] in ("ST", "ND", "RD", "TH") and word[:-2].isnumeric():  #  e.g 9TH, 6TH
+        word = num2words(word[:-2], to="ordinal")
+        word = word.replace("-", " ")
+
+    if word.isnumeric():
+        num = int(word)
+        if num > 1500 and num < 2030:
+            word = num2words(word, to="year")
+        else:
+            word = num2words(word)
+        word = word.replace("-", " ")
+    return word.upper()
+
+
+def text_normalization(text: str) -> str:
+    text = text.upper()
+    return " ".join([word_normalization(x) for x in text.split()])
+
+
+if __name__ == "__main__":
+    assert remove_punc_to_upper("I like this 《book>") == "I LIKE THIS BOOK"
+    assert (
+        text_normalization("Hello Mrs st 21st world 3rd she 99th MR")
+        == "HELLO MISSUS SAINT TWENTY FIRST WORLD THIRD SHE NINETY NINTH MISTER"
+    )
diff --git a/egs/libriheavy/ASR/zipformer/train.py b/egs/libriheavy/ASR/zipformer/train.py
new file mode 100644
index 000000000..c97da4a11
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/train.py
@@ -0,0 +1,1415 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey,
+#                                                       Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --full-libri 1 \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --full-libri 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from text_normalization import remove_punc_to_upper
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-hours",
+        type=float,
+        default=30000,
+        help="""Number of hours that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--train-with-punctuation",
+        type=str2bool,
+        default=False,
+        help="If True, the training text will include casing and punctuation.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}"
+    )
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        loss = 0.0
+
+        if params.use_transducer:
+            s = params.simple_loss_scale
+            # take down the scale on the simple loss from 1.0 at the start
+            # to params.simple_loss scale by warm_step.
+            simple_loss_scale = (
+                s
+                if batch_idx_train >= warm_step
+                else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+            )
+            pruned_loss_scale = (
+                1.0
+                if batch_idx_train >= warm_step
+                else 0.1 + 0.9 * (batch_idx_train / warm_step)
+            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    if params.use_transducer:
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            # Use the number of hours of speech to adjust the learning rate
+            scheduler.step_epoch(
+                params.batch_idx_train * params.max_duration * params.world_size / 3600
+            )
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_hours)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def normalize_text(c: Cut):
+        text = remove_punc_to_upper(c.supervisions[0].text)
+        c.supervisions[0].text = text
+        return c
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 2.0 or c.duration > 30.0:
+            # logging.warning(
+            #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    train_cuts = libriheavy.train_small_cuts()
+    if params.subset == "M" or params.subset == "L":
+        train_cuts += libriheavy.train_medium_cuts()
+    if params.subset == "L":
+        train_cuts += libriheavy.train_large_cuts()
+
+    if not params.train_with_punctuation:
+        train_cuts = train_cuts.map(normalize_text)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = libriheavy.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = libriheavy.dev_cuts()
+
+    if not params.train_with_punctuation:
+        valid_cuts = valid_cuts.map(normalize_text)
+
+    valid_dl = libriheavy.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         sp=sp,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer/zipformer.py b/egs/libriheavy/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/__init__.py b/egs/libriheavy/ASR/zipformer_prompt_asr/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py
new file mode 100644
index 000000000..690003377
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py
@@ -0,0 +1,520 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional
+
+import torch
+from dataset import PromptASRDataset
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # SingleCutSampler,
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    ExtraPadding,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriHeavyAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+        if args.use_context_list:
+            assert args.rare_word_file is not None
+            with open(args.rare_word_file, "r") as f:
+                self.rare_word_list = (
+                    f.read().lower().split()
+                )  # Use lower-cased for easier style transform
+        else:
+            self.rare_word_list = None
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+
+        # Libriheavy specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="small",
+            help="Select the Libriheavy subset (small|medium|large)",
+        )
+
+        group.add_argument(
+            "--use-context-list",
+            type=str2bool,
+            default=False,
+            help="Use the context list of libri heavy",
+        )
+
+        group.add_argument(
+            "--top-k",
+            type=int,
+            default=10000,
+            help="""The top-k words are identified as common words,
+            the rest as rare words""",
+        )
+
+        group.add_argument(
+            "--with-decoding",
+            type=str2bool,
+            default=False,
+            help="If the texts field contain decoding",
+        )
+
+        group.add_argument(
+            "--random-left-padding",
+            type=str2bool,
+        )
+
+        group.add_argument(
+            "--rare-word-file",
+            type=str,
+        )
+
+        group.add_argument(
+            "--long-audio-cuts",
+            type=str,
+            default="data/manifest_npr/npr1_cuts_all_guids_0.jsonl.gz",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        text_sampling_func: Callable[[List[str]], str] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = PromptASRDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+            text_sampling_func=text_sampling_func,
+            rare_word_list=self.rare_word_list,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = PromptASRDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+                rare_word_list=self.rare_word_list,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            raise NotImplementedError(
+                "SingleCutSampler is no longer supported by lhotse"
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(
+        self,
+        cuts_valid: CutSet,
+        text_sampling_func: Callable[[List[str]], str] = None,
+    ) -> DataLoader:
+        transforms = []
+        if self.args.random_left_padding:
+            logging.info("Enable random left padding")
+            transforms.append(
+                ExtraPadding(extra_frames=16, randomized=True, direction="left")
+            )
+
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = PromptASRDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+                rare_word_list=self.rare_word_list,
+            )
+        else:
+            validate = PromptASRDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+                rare_word_list=self.rare_word_list,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get {self.args.subset} cuts")
+
+        if self.args.use_context_list:
+            path = (
+                self.args.manifest_dir
+                / f"libriheavy_cuts_{self.args.subset}_with_context_list_topk_{self.args.top_k}.jsonl.gz"
+            )
+        elif self.args.with_decoding:
+            path = (
+                self.args.manifest_dir
+                / f"libriheavy_cuts_{self.args.subset}_with_decoding.jsonl.gz"
+            )
+        else:
+            path = (
+                self.args.manifest_dir / f"libriheavy_cuts_{self.args.subset}.jsonl.gz"
+            )
+
+        logging.info(f"Loading manifest from {path}.")
+        cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_dev.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test-clean_official.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test-other_official.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def librispeech_test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def librispeech_test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def long_audio_cuts(self) -> CutSet:
+        logging.info("About to get long audio cuts")
+        cuts = load_manifest_lazy(
+            self.args.long_audio_cuts,
+        )
+        return cuts
+
+    @lru_cache()
+    def test_dev_cuts(self) -> CutSet:
+        logging.info("About to get test dev cuts")
+        cuts = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test_dev.jsonl.gz"
+        )
+        return cuts
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py b/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py b/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py
new file mode 100644
index 000000000..e0bf8f73d
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py
@@ -0,0 +1,586 @@
+#      Copyright      2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from lhotse import validate
+from lhotse.cut import CutSet
+from lhotse.dataset import K2SpeechRecognitionDataset
+from lhotse.dataset.input_strategies import BatchIO, PrecomputedFeatures
+from lhotse.utils import compute_num_frames, ifnone
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from torch.utils.data.dataloader import DataLoader, default_collate
+
+
+class PromptASRDataset(torch.utils.data.Dataset):
+    """This is a dataset for Prompt ASR. It supports the following features:
+    1. Select a tuple of (text, pre_text, style_text) randomly from a
+    list of texts as supervisions.
+
+    """
+
+    def __init__(
+        self,
+        return_cuts: bool = False,
+        cut_transforms: List[Callable[[CutSet], CutSet]] = None,
+        input_transforms: List[Callable[[torch.Tensor], torch.Tensor]] = None,
+        input_strategy: BatchIO = PrecomputedFeatures(),
+        text_sampling_func: Optional[Callable[[List[str]], str]] = None,
+        rare_word_list: Optional[List[str]] = None,
+    ):
+        """
+        Icefall ASR IterableDataset constructor. See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py
+        for more details.
+
+        :param return_cuts: When ``True``, will additionally return a "cut" field in each batch with the Cut
+            objects used to create that batch.
+        :param cut_transforms: A list of transforms to be applied on each sampled batch,
+            before converting cuts to an input representation (audio/features).
+            Examples: cut concatenation, noise cuts mixing, etc.
+        :param input_transforms: A list of transforms to be applied on each sampled batch,
+            after the cuts are converted to audio/features.
+            Examples: normalization, SpecAugment, etc.
+        :param input_strategy: Converts cuts into a collated batch of audio/features.
+            By default, reads pre-computed features from disk.
+        :param text_sampling_func: Sampling a text as transcription from a list of texts.
+        """
+        super().__init__()
+        # Initialize the fields
+        self.return_cuts = return_cuts
+        self.cut_transforms = ifnone(cut_transforms, [])
+        self.input_transforms = ifnone(input_transforms, [])
+        self.input_strategy = input_strategy
+
+        # a text sampling function
+        self.text_sampling_func = text_sampling_func
+        self.rare_word_list = rare_word_list
+
+    def __getitem__(self, cuts: CutSet) -> Dict[str, Union[torch.Tensor, List[str]]]:
+        """
+        Return a new batch, with the batch size automatically determined using the constraints
+        of max_frames and max_cuts.
+        """
+        validate_for_asr(cuts)
+
+        # Sort the cuts by duration so that the first one determines the batch time dimensions.
+        cuts = cuts.sort_by_duration(ascending=False)
+
+        # Optional CutSet transforms - e.g. padding, or speed perturbation that adjusts
+        # the supervision boundaries.
+        for tnfm in self.cut_transforms:
+            cuts = tnfm(cuts)
+
+        # Sort the cuts again after transforms
+        cuts = cuts.sort_by_duration(ascending=False)
+
+        # Get a tensor with batched feature matrices, shape (B, T, F)
+        # Collation performs auto-padding, if necessary.
+        input_tpl = self.input_strategy(cuts)
+        if len(input_tpl) == 3:
+            # An input strategy with fault tolerant audio reading mode.
+            # "cuts" may be a subset of the original "cuts" variable,
+            # that only has cuts for which we succesfully read the audio.
+            inputs, _, cuts = input_tpl
+        else:
+            inputs, _ = input_tpl
+
+        # Get a dict of tensors that encode the positional information about supervisions
+        # in the batch of feature matrices. The tensors are named "sequence_idx",
+        # "start_frame/sample" and "num_frames/samples".
+        supervision_intervals = self.input_strategy.supervision_intervals(cuts)
+
+        # Apply all available transforms on the inputs, i.e. either audio or features.
+        # This could be feature extraction, global MVN, SpecAugment, etc.
+        segments = torch.stack(list(supervision_intervals.values()), dim=1)
+        for tnfm in self.input_transforms:
+            inputs = tnfm(inputs, supervision_segments=segments)
+
+        batch = {
+            "inputs": inputs,
+            "supervisions": default_collate(
+                [
+                    self.text_sampling_func(
+                        texts=supervision.texts,
+                        pre_texts=supervision.pre_texts,
+                        context_list=supervision.context_list
+                        if "context_list" in supervision.custom
+                        else None,
+                        rare_word_list=self.rare_word_list,
+                    )
+                    if self.text_sampling_func is not None
+                    else {
+                        "text": train_text_normalization(supervision.texts[0]),
+                        "pre_text": train_text_normalization(supervision.pre_texts[0]),
+                        "style_text": train_text_normalization(
+                            supervision.pre_texts[0]
+                        ),
+                        "transform_ids": 0,
+                    }
+                    for sequence_idx, cut in enumerate(cuts)
+                    for supervision in cut.supervisions
+                ]
+            ),
+        }
+        # Update the 'supervisions' field with sequence_idx and start/num frames/samples
+        batch["supervisions"].update(supervision_intervals)
+        if self.return_cuts:
+            batch["supervisions"]["cut"] = [
+                cut for cut in cuts for sup in cut.supervisions
+            ]
+
+        has_word_alignments = all(
+            s.alignment is not None and "word" in s.alignment
+            for c in cuts
+            for s in c.supervisions
+        )
+
+        return batch
+
+
+def validate_for_asr(cuts: CutSet) -> None:
+    validate(cuts)
+    tol = 2e-3  # 1ms
+    for cut in cuts:
+        for supervision in cut.supervisions:
+            assert supervision.start >= -tol, (
+                f"Supervisions starting before the cut are not supported for ASR"
+                f" (sup id: {supervision.id}, cut id: {cut.id})"
+            )
+
+            # Supervision start time is relative to Cut ...
+            # https://lhotse.readthedocs.io/en/v0.10_e/cuts.html
+            #
+            # 'supervision.end' is end of supervision inside the Cut
+            assert supervision.end <= cut.duration + tol, (
+                f"Supervisions ending after the cut "
+                f"are not supported for ASR"
+                f" (sup id: {supervision.id}, cut id: {cut.id})"
+            )
+
+
+def get_substring(s: str, min_len: int = 40, max_len: int = 250) -> str:
+    """A helper function that generates a random substring from a given string
+
+    Args:
+        s (str): Input string
+
+    Returns:
+        str: Returned substring
+    """
+    min_len = min(len(s), min_len)
+
+    start = random.randint(0, len(s) - min_len)
+    end = min(start + max_len, random.randint(start + min_len, len(s)))
+
+    return s[start:end]
+
+
+def triplet_text_sampling(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: Optional[str] = None,
+    rare_word_list: Optional[List[str]] = None,
+    transforms: Optional[List[Callable[[str], str]]] = None,
+    min_len_style: Optional[int] = 80,
+) -> Dict[str, str]:
+    """This function generates a triplet of
+    (pre_text, style_text, ref_text). The style of style_text and ref_text
+    should **always** match, whereas the style of pre_text is arbitrary.
+    Suppose we have 2 different transforms A,B, and the preceding text is
+    referred to as pre_text. The following three tuples are all valid:
+
+    (A(pre_text), A(style_text), A(ref_text))
+    (A(pre_text), B(style_text), B(ref_text))
+    (A(pre_text), A(style_text), A(ref_text))
+    (B(pre_text), B(style_text), B(ref_text))
+
+    If transforms is not given, the following pre-defined transforms
+    are available:
+    0: original (mixed-cased, with punc)
+    1: upper_only_alpha (upper-cased, no punc)
+
+    When the transform of text and pre_text match, we can use the whole
+    pre_text as the prompt text.
+
+    Args:
+        texts (List[str]):
+            A list of ref_texts whose first item is the ground truth
+            text from books.
+        pre_texts (List[str]):
+            A list of pre_texts, whose first item is the groundtruth
+            pre_text from books.
+        context_list: Optional[str] = None,
+            A list of biasing words separated by space
+        rare_word_list: Optional[str] = None,
+            A list of rare-words separated by space (used as distractors)
+        transforms (List[Callable[[str], str]]): A list of possible transforms to be applied
+
+    Returns:
+        A dictionary of ref_text, pre_text, style_text
+    """
+    assert len(texts) == len(pre_texts)
+    assert len(texts) == 2
+
+    # we assume the first item to be ground truth
+    gt_text = texts[0]
+    gt_pre_text = pre_texts[0]
+
+    if transforms is None:
+        transforms = [
+            lambda x: x,  # return it self
+            upper_only_alpha,
+            lower_only_alpha,
+            lower_all_char,
+        ]
+
+    sampling_weight = [
+        0.7,
+        0.3,
+        0.0,
+        0.0,
+    ]  # Mixed-punc should have the largest sampling prob
+
+    total_transforms = len(transforms)  # do not use the recognized trans
+
+    # Randomly sample transforms
+    i_text, i_pre_text = np.random.choice(total_transforms, 2, p=sampling_weight)
+
+    # get the normalized text and pre_text
+    text = transforms[i_text](gt_text)
+    pre_text = transforms[i_pre_text](gt_pre_text)
+
+    if i_text == i_pre_text:
+        style_text = get_substring(pre_text, min_len=min_len_style, max_len=150)
+    else:
+        # get the pre_text of same style as text
+        # For now, **don't** do transform to the style text, because we do it after the dataloader
+        style_text = gt_pre_text
+        # style_text = pre_texts[i_text] if i_text <= 1 else transforms[i_text-2](gt_pre_text)
+        style_text = get_substring(style_text, min_len=min_len_style, max_len=150)
+
+    return {
+        "text": train_text_normalization(text),
+        "pre_text": train_text_normalization(pre_text),
+        "style_text": train_text_normalization(style_text),
+        "transform_ids": i_text,
+    }
+
+
+def triplet_text_sampling_with_context_list(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: str,
+    rare_word_list: List[str],
+    transforms: Optional[List[Callable[[str], str]]] = None,
+    min_len_style: Optional[int] = 80,
+) -> Dict[str, str]:
+    """This function generates a triplet of
+    (pre_text, style_text, ref_text). The pre_text is either the preceding text
+    or a list of words (context words + distractors).
+    The style of style_text and ref_text should **always** match, whereas
+    the style of pre_text is arbitrary.
+    Suppose we have 2 different transforms A,B, and the preceding text is
+    referred to as pre_text. The following three tuples are all valid:
+
+    (A(pre_text), A(style_text), A(ref_text))
+    (A(pre_text), B(style_text), B(ref_text))
+    (A(pre_text), A(style_text), A(ref_text))
+    (B(pre_text), B(style_text), B(ref_text))
+
+    If transforms is not given, the following pre-defined transforms
+    are available:
+    0: original (mixed-cased, with punc)
+    1: upper_only_alpha (upper-cased, no punc)
+
+    When the transform of text and pre_text match, we can use the whole
+    pre_text as the prompt text.
+
+    Args:
+        texts (List[str]):
+            A list of ref_texts whose first item is the ground truth
+            text from books.
+        pre_texts (List[str]):
+            A list of pre_texts, whose first item is the groundtruth
+            pre_text from books.
+        context_list: Optional[str] = None,
+            A list of biasing words separated by space
+        rare_word_list: Optional[str] = None,
+            A list of rare-words separated by space (used as distractors)
+        transforms (List[Callable[[str], str]]): A list of possible transforms to be applied
+
+    Returns:
+        A dictionary of ref_text, pre_text, style_text
+    Returns:
+        str: A dictionary
+    """
+    # import pdb; pdb.set_trace()
+    assert len(texts) == len(pre_texts)
+    assert len(texts) == 2
+
+    if context_list is not None:
+        context_list = context_list.lower()
+
+    # we assume the first item to be ground truth
+    gt_text = texts[0]
+    gt_pre_text = pre_texts[0]
+
+    if transforms is None:
+        transforms = [
+            lambda x: x,  # return it self
+            upper_only_alpha,
+            lower_only_alpha,
+            lower_all_char,
+        ]
+
+    sampling_weight = [
+        0.7,
+        0.3,
+        0.0,
+        0.0,
+    ]  # Mixed-punc should have the largest sampling prob
+
+    total_transforms = len(transforms)  # do not use the recognized trans
+
+    # Select a transformation randomly
+    i_text, i_pre_text = np.random.choice(total_transforms, 2, p=sampling_weight)
+
+    # get the normalized text and pre_text
+    text = transforms[i_text](gt_text)
+    pre_text = get_pre_text_with_context_list2(
+        text=gt_text,
+        pre_text=gt_pre_text,
+        context_list=context_list,
+        rare_words_list=rare_word_list,
+    )
+    pre_text = transforms[i_pre_text](pre_text)
+
+    if i_text == i_pre_text:
+        style_text = get_substring(pre_text, min_len=min_len_style, max_len=150)
+    else:
+        # get the pre_text of same style as text
+        # For now, **don't** do transform to the style text
+        style_text = gt_pre_text
+        # style_text = pre_texts[i_text] if i_text <= 1 else transforms[i_text-2](gt_pre_text)
+        style_text = get_substring(style_text, min_len=min_len_style, max_len=150)
+
+    return {
+        "text": train_text_normalization(text),
+        "pre_text": train_text_normalization(pre_text),
+        "style_text": train_text_normalization(style_text),
+        "transform_ids": i_text,
+    }
+
+
+def get_pre_text_with_context_list(
+    text: str,
+    pre_text: str,
+    context_list: str,
+    rare_words_list: List[str] = None,
+) -> str:
+    # Always get the first one, which is the gt (mixed-cased trans), but with upper_only_alpha
+    # By a small proportion of time, use the substring of ref_text as pre_text
+
+    if context_list != "" and context_list is not None:
+        v = random.random()
+        if v < 0.5:
+            # correct + distractors
+            # sample distractors
+            num_distractors = random.randint(0, 50)
+            distractors = random.sample(rare_words_list, num_distractors)
+            # sample correct
+            correct = context_list.split()
+            i = random.randint(1, len(correct))
+            correct = random.sample(correct, i)
+            # combine correct and distractors
+            pre_text = distractors + correct
+            random.shuffle(pre_text)
+            pre_text = " ".join(pre_text)
+        elif v < 0.7:
+            splitted = text.split()
+            sampling_weights = [len(w) ** 1.2 for w in splitted]
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            num_distractors = random.randint(0, 70)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+            pre_text = " ".join(splitted)
+        else:
+            pre_text = pre_text
+    else:
+        v = random.random()
+        if v < 0.1:
+            splitted = text.split()
+            sampling_weights = [len(w) ** 1.2 for w in splitted]
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            pre_text = " ".join(splitted)
+            num_distractors = random.randint(0, 70)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+        elif v < 0.2:
+            # full distractors
+            num_distractors = random.randint(5, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            pre_text = " ".join(distractors)
+
+        elif v < 0.3:
+            pre_text = get_substring(text, min_len=15, max_len=150)
+        else:
+            pre_text = pre_text
+
+    return pre_text
+
+
+def get_pre_text_with_context_list2(
+    text: str,
+    pre_text: str,
+    context_list: str,
+    rare_words_list: List[str] = None,
+) -> str:
+    # Get the pre_text, either the ground truth preceding text or
+    # a list of words consisting of biasing words and distrators
+    # By a small proportion of time, use the substring of ref_text as pre_text
+
+    if context_list != "" and context_list is not None:
+        v = random.random()
+        if v < 0.4:
+            # sample distractors
+            num_distractors = random.randint(50, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            # sample correct
+            correct = context_list.split()
+            i = random.randint(1, len(correct))
+            correct = random.sample(correct, i)
+            # combine correct and distractors
+            pre_text = distractors + correct
+            random.shuffle(pre_text)
+            pre_text = " ".join(pre_text)
+        elif v < 0.55:
+            splitted = text.split()
+            sampling_weights = [
+                len(w) ** 1.2 for w in splitted
+            ]  # longer words with higher weights
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            num_distractors = random.randint(50, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+            pre_text = " ".join(splitted)
+        else:
+            pre_text = pre_text
+    else:
+        v = random.random()
+        if v < 0.3:
+            splitted = text.split()
+            sampling_weights = [len(w) ** 1.2 for w in splitted]
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            pre_text = " ".join(splitted)
+            num_distractors = random.randint(50, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+        elif v < 0.4:
+            # full distractors
+            num_distractors = random.randint(5, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            pre_text = " ".join(distractors)
+        elif v < 0.6:
+            pre_text = get_substring(text, min_len=15, max_len=150)
+        else:
+            pre_text = pre_text
+
+    return pre_text
+
+
+def naive_triplet_text_sampling(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: str = None,
+    rare_word_list: List[str] = None,
+    min_len_style: Optional[int] = 120,
+):
+    # The most simplest text sampling function, used only for
+    # evaluation, use a fixed sentence as the style text
+
+    return {
+        "text": train_text_normalization(texts[0]),
+        "pre_text": train_text_normalization(pre_texts[0]),
+        "style_text": "Mixed-case English transcription, with punctuation. Actually, it is fully not related. What do you think?",
+        "transform_ids": 0,
+    }
+
+
+def random_shuffle_subset(
+    data: List[str],
+    p: float = 0.2,
+    p_mask: float = 0.05,
+) -> List[str]:
+    """
+    Randomly shuffle the subset by probability `p`, which means that p% of the samples
+    in the original batch are shuffled, the others are kept in the original order.
+
+    With a probability of `p_mask`, replace the original string with an empty string.
+
+    """
+
+    num_to_shuffle = int(len(data) * p)
+    id_to_shuffle = np.random.choice(len(data), num_to_shuffle, replace=False)
+    item_to_shuffle = [data[id] for id in id_to_shuffle]
+    random.shuffle(item_to_shuffle)
+
+    for id, item in zip(id_to_shuffle, item_to_shuffle):
+        data[id] = item
+
+    # Randomly mask a proportion of the data to empty string
+    if p_mask > 0:
+        for i in range(len(data)):
+            if random.random() < p_mask:
+                data[i] = ""
+
+    return data
+
+
+if __name__ == "__main__":
+    texts = [
+        "AA, BB, cC, dD!",
+        "AA BB CC DD",
+    ]
+
+    pre_texts = [
+        "EE, Ff, Gg? EE, Ff, Gg? EE, Ff, Gg? EE, Ff, Gg?",
+        "EE FF GG EE FF GG EE FF GG EE FF GG EE FF GG",
+    ]
+    for i in range(10):
+        print(f"Run: {i}")
+        print(triplet_text_sampling(texts, pre_texts))
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py
new file mode 100644
index 000000000..6a3bab3c8
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py
@@ -0,0 +1,791 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import greedy_search, greedy_search_batch, modified_beam_search
+from ls_text_normalization import word_normalization
+from text_normalization import (
+    ref_text_normalization,
+    remove_non_alphabetic,
+    upper_only_alpha,
+)
+from train_baseline import add_model_arguments, get_params, get_transducer_model
+from utils import write_error_stats
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, setup_logger, store_transcripts, str2bool
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+        help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ",
+    )
+
+    parser.add_argument(
+        "--long-audio-recog",
+        type=str2bool,
+        default=False,
+    )
+
+    parser.add_argument(
+        "--use-ls-test-set",
+        type=str2bool,
+        default=False,
+        help="Use librispeech test set for evaluation.",
+    )
+
+    parser.add_argument(
+        "--compute-CER",
+        type=str2bool,
+        default=True,
+        help="Reports CER. By default, only reports WER",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    texts = batch["supervisions"]["text"]
+    batch_size = feature.size(0)
+
+    # Get the transducer encoder output
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=feature,
+            feature_lens=feature_lens,
+        )
+
+    hyps = []
+
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        if not params.use_ls_test_set:
+            book_names = [
+                cut.text_path.split("/")[-2] for cut in batch["supervisions"]["cut"]
+            ]
+        else:
+            book_names = ["" for _ in cut_ids]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, book_name, hyp_words, ref_text in zip(
+                cut_ids, book_names, hyps, texts
+            ):
+                ref_text = ref_text_normalization(ref_text)
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+            # if not params.use_ls_test_set:
+            #     results[name + " " + book_name].extend(this_batch)
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+    biasing_words: List[str] = None,
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f,
+                f"{test_set_name}-{key}",
+                results,
+                enable_log=True,
+                biasing_words=biasing_words,
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        if params.compute_CER:
+            # Write CER statistics
+            recog_path = (
+                params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt"
+            )
+            store_transcripts(filename=recog_path, texts=results, char_level=True)
+            errs_filename = (
+                params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt"
+            )
+            with open(errs_filename, "w") as f:
+                cer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{key}",
+                    results,
+                    enable_log=True,
+                    compute_CER=params.compute_CER,
+                )
+                test_set_cers[key] = cer
+
+            logging.info("Wrote detailed CER stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    if params.compute_CER:
+        test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+        errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tcER", file=f)
+            for key, val in test_set_cers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_cers:
+            s += "{} CER\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "modified_beam_search",
+    )
+
+    if params.long_audio_recog:
+        params.res_dir = params.exp_dir / (params.decoding_method + "long_audio")
+    else:
+        params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    LM = None
+
+    decoding_graph = None
+    word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    test_clean_cuts = libriheavy.test_clean_cuts()
+    test_other_cuts = libriheavy.test_other_cuts()
+    ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts()
+    ls_test_other_cuts = libriheavy.librispeech_test_other_cuts()
+    long_audio_cuts = libriheavy.long_audio_cuts()
+
+    test_clean_dl = libriheavy.valid_dataloaders(
+        test_clean_cuts,
+    )
+    test_other_dl = libriheavy.valid_dataloaders(
+        test_other_cuts,
+    )
+    ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts)
+    ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts)
+    long_audio_dl = libriheavy.valid_dataloaders(
+        long_audio_cuts,
+    )
+
+    if params.use_ls_test_set:
+        test_sets = ["ls-test-clean", "ls-test-other"]
+        test_dl = [ls_test_clean_dl, ls_test_other_dl]
+    else:
+        test_sets = ["test-clean", "test-other"]
+        test_dl = [test_clean_dl, test_other_dl]
+
+    if params.long_audio_recog:
+        test_sets = ["long-audio"]
+        test_dl = [long_audio_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        if params.use_ls_test_set:
+            f = open(
+                "data/context_biasing/LibriSpeechBiasingLists/all_rare_words.txt", "r"
+            )
+            biasing_words = f.read().strip().split()
+            f.close()
+        else:
+            biasing_words = None
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+        if params.post_normalization:
+            if "-post-normalization" not in params.suffix:
+                params.suffix += "-post-normalization"
+
+            new_res = {}
+            for k in results_dict:
+                new_ans = []
+                for item in results_dict[k]:
+                    id, ref, hyp = item
+                    if params.use_ls_test_set:
+                        hyp = (
+                            " ".join(hyp).replace("-", " ").split()
+                        )  # handle the hypens
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        hyp = [word_normalization(w.upper()) for w in hyp]
+                        hyp = " ".join(hyp).split()
+                        hyp = [w for w in hyp if w != ""]
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    else:
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    new_ans.append((id, ref, hyp))
+                new_res[k] = new_ans
+
+            save_results(
+                params=params,
+                test_set_name=test_set,
+                results_dict=new_res,
+                biasing_words=biasing_words,
+            )
+
+            if params.suffix.endswith("-post-normalization"):
+                params.suffix = params.suffix.replace("-post-normalization", "")
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py
new file mode 100755
index 000000000..e71999b0a
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py
@@ -0,0 +1,1025 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method greedy_search \
+    --text-encoder-type BERT \
+    --memory-layer 0 \
+    --use-pre-text True \
+    --use-style-prompt True \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc \
+    --compute-CER 0
+
+
+(2) modified beam search
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    --text-encoder-type BERT \
+    --memory-layer 0 \
+    --use-pre-text True \
+    --use-style-prompt True \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc \
+    --compute-CER 0
+
+(3) Decode LibriSpeech
+
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method modified_beam_search \
+    --use-ls-test-set True \
+    --beam-size 4 \
+    --text-encoder-type BERT \
+    --memory-layer 0 \
+    --use-pre-text True \
+    --use-style-prompt True \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc \
+    --compute-CER 0
+
+(4) Decode LibriSpeech + biasing list
+
+biasing_list=100 # could also be 0
+
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    --use-ls-test-set True \
+    --use-ls-context-list True \
+    --biasing-level utterance \
+    --ls-distractors $biasing_list \
+    --post-normalization True \
+    --text-encoder-type BERT \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc
+
+
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import greedy_search, greedy_search_batch, modified_beam_search
+from dataset import naive_triplet_text_sampling, random_shuffle_subset
+from ls_text_normalization import word_normalization
+from text_normalization import (
+    _apply_style_transform,
+    lower_all_char,
+    lower_only_alpha,
+    ref_text_normalization,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from train_bert_encoder import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+from transformers import BertModel, BertTokenizer
+from utils import brian_biasing_list, get_facebook_biasing_list, write_error_stats
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, setup_logger, store_transcripts, str2bool
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=True,
+        help="Use pre-text is available during decoding",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt when evaluation",
+    )
+
+    parser.add_argument(
+        "--max-prompt-lens",
+        type=int,
+        default=1000,
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+        help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ",
+    )
+
+    parser.add_argument(
+        "--compute-CER",
+        type=str2bool,
+        default=False,
+        help="Reports CER. By default, only reports WER",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--use-ls-test-set",
+        type=str2bool,
+        default=False,
+        help="Use librispeech test set for evaluation.",
+    )
+
+    parser.add_argument(
+        "--use-ls-context-list",
+        type=str2bool,
+        default=False,
+        help="If use a fixed context list for LibriSpeech decoding",
+    )
+
+    parser.add_argument(
+        "--biasing-level",
+        type=str,
+        default="utterance",
+        choices=["utterance", "Book", "Chapter"],
+    )
+
+    parser.add_argument(
+        "--ls-distractors",
+        type=int,
+        default=0,
+        help="The number of distractors into context list for LibriSpeech decoding",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer: spm.SentencePieceProcessor,
+    batch: dict,
+    biasing_dict: dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      tokenizer:
+        Tokenizer for the text encoder
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      biasing_dict:
+        A dictionary in the form `{cut_id: :w1 w2"}` that contains a list
+        of biasing words (separated with space)
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    cuts = batch["supervisions"]["cut"]
+    cut_ids = [c.supervisions[0].id for c in cuts]
+    batch_size = feature.size(0)
+
+    if "pre_text" in batch["supervisions"] and params.use_pre_text:
+        pre_texts = batch["supervisions"]["pre_text"]
+        pre_texts = [train_text_normalization(t) for t in pre_texts]
+    else:
+        pre_texts = ["" for _ in range(batch_size)]
+
+    # get the librispeech biasing data
+    if params.use_pre_text and (params.use_ls_context_list and params.use_ls_test_set):
+        if params.biasing_level == "utterance":
+            pre_texts = [biasing_dict[id] for id in cut_ids]
+        elif params.biasing_level == "Chapter":
+            chapter_ids = [c.split("-")[1] for c in cut_ids]
+            pre_texts = [biasing_dict[id] for id in chapter_ids]
+        elif params.biasing_level == "Book":
+            chapter_ids = [c.split("-")[1] for c in cut_ids]
+            pre_texts = [biasing_dict[id] for id in chapter_ids]
+        else:
+            raise ValueError(f"Unseen biasing level: {params.biasing_level}")
+        if params.pre_text_transform == "mixed-punc":
+            pre_texts = [t.lower() for t in pre_texts]
+
+    # get style_text
+    if params.use_style_prompt:
+        fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it's fully not related."
+        style_texts = batch["supervisions"].get(
+            "style_text", [fixed_sentence for _ in range(batch_size)]
+        )
+        style_texts = [train_text_normalization(t) for t in style_texts]
+    else:
+        style_texts = ["" for _ in range(batch_size)]  # use empty string
+
+    # Get the text embedding
+    if params.use_pre_text or params.use_style_prompt:
+        # apply style transform to the pre_text and style_text
+        pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform)
+        if not params.use_ls_context_list:
+            pre_texts = [t[-params.max_prompt_lens :] for t in pre_texts]
+
+        if params.use_style_prompt:
+            style_texts = _apply_style_transform(
+                style_texts, params.style_text_transform
+            )
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+
+            # Use tokenizer to prepare input for text encoder
+            encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+                pre_texts=pre_texts,
+                style_texts=style_texts,
+                tokenizer=tokenizer,
+                device=device,
+                no_limit=True,
+            )
+            logging.info(
+                f"Shape of the encoded prompts: {encoded_inputs['input_ids'].shape}"
+            )
+
+            memory, memory_key_padding_mask = model.encode_text(
+                encoded_inputs=encoded_inputs,
+                style_lens=style_lens,
+            )  # (T,B,C)
+    else:
+        memory = None
+        memory_key_padding_mask = None
+
+    # Get the transducer encoder output
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=feature,
+            feature_lens=feature_lens,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+
+    hyps = []
+
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer: spm.SentencePieceProcessor,
+    biasing_dict: Dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      tokenizer:
+        Tokenizer for the text encoder
+      biasing_dict:
+        A dictionary in the form `{cut_id: :w1 w2"}` that contains a list
+        of biasing words (separated with space)
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"][
+            "text"
+        ]  # By default, this should be in mixed-punc format
+
+        # the style of ref_text should match style_text
+        texts = _apply_style_transform(texts, params.style_text_transform)
+        if params.use_style_prompt:
+            texts = _apply_style_transform(texts, params.style_text_transform)
+
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        if not params.use_ls_test_set:
+            try:
+                book_names = [
+                    cut.text_path.split("/")[-2] for cut in batch["supervisions"]["cut"]
+                ]
+            except AttributeError:
+                book_names = [
+                    cut.id.split("/")[0] for cut in batch["supervisions"]["cut"]
+                ]
+        else:
+            book_names = ["" for _ in cut_ids]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, book_name, hyp_words, ref_text in zip(
+                cut_ids, book_names, hyps, texts
+            ):
+                ref_text = ref_text_normalization(
+                    ref_text
+                )  # remove full-width symbols & some book marks
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+    biasing_words: List[str] = None,
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        if params.compute_CER:
+            # Write CER statistics
+            recog_path = (
+                params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt"
+            )
+            store_transcripts(filename=recog_path, texts=results, char_level=True)
+            errs_filename = (
+                params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt"
+            )
+            with open(errs_filename, "w") as f:
+                cer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{key}",
+                    results,
+                    enable_log=True,
+                    compute_CER=params.compute_CER,
+                )
+                test_set_cers[key] = cer
+
+            logging.info("Wrote detailed CER stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    if params.compute_CER:
+        test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+        errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tCER", file=f)
+            for key, val in test_set_cers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_cers:
+            s += "{} CER\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "modified_beam_search",
+    )
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_pre_text:
+        params.suffix += (
+            f"-pre-text-{params.pre_text_transform}-len-{params.max_prompt_lens}"
+        )
+
+    if params.use_style_prompt:
+        params.suffix += f"-style-prompt-{params.style_text_transform}"
+
+    if params.use_ls_context_list:
+        assert (
+            params.use_pre_text
+        ), "Must set --use-pre-text to True if using context list"
+        params.suffix += f"-use-{params.biasing_level}-level-ls-context-list"
+        if params.biasing_level == "utterance" and params.ls_distractors:
+            params.suffix += f"-ls-context-distractors-{params.ls_distractors}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    LM = None
+
+    decoding_graph = None
+    word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    test_clean_cuts = libriheavy.test_clean_cuts()
+    test_other_cuts = libriheavy.test_other_cuts()
+    ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts()
+    ls_test_other_cuts = libriheavy.librispeech_test_other_cuts()
+
+    test_clean_dl = libriheavy.valid_dataloaders(
+        test_clean_cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+    test_other_dl = libriheavy.valid_dataloaders(
+        test_other_cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+    ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts)
+    ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts)
+
+    if params.use_ls_test_set:
+        test_sets = ["ls-test-clean", "ls-test-other"]
+        test_dl = [ls_test_clean_dl, ls_test_other_dl]
+    else:
+        test_sets = ["test-clean", "test-other"]
+        test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        biasing_dict = None
+        if params.use_ls_context_list:
+            if test_set == "ls-test-clean":
+                biasing_dict = get_facebook_biasing_list(
+                    test_set="test-clean",
+                    num_distractors=params.ls_distractors,
+                )
+            elif test_set == "ls-test-other":
+                biasing_dict = get_facebook_biasing_list(
+                    test_set="test-other",
+                    num_distractors=params.ls_distractors,
+                )
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+        if params.post_normalization:
+            if "-post-normalization" not in params.suffix:
+                params.suffix += "-post-normalization"
+
+            new_res = {}
+            for k in results_dict:
+                new_ans = []
+                for item in results_dict[k]:
+                    id, ref, hyp = item
+                    if params.use_ls_test_set:
+                        hyp = (
+                            " ".join(hyp).replace("-", " ").split()
+                        )  # handle the hypens
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        hyp = [word_normalization(w.upper()) for w in hyp]
+                        hyp = " ".join(hyp).split()
+                        hyp = [w for w in hyp if w != ""]
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    else:
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    new_ans.append((id, ref, hyp))
+                new_res[k] = new_ans
+
+            save_results(
+                params=params,
+                test_set_name=test_set,
+                results_dict=new_res,
+            )
+
+            if params.suffix.endswith("-post-normalization"):
+                params.suffix = params.suffix.replace("-post-normalization", "")
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py
new file mode 100755
index 000000000..4559ebb6d
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py
@@ -0,0 +1,963 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import (
+    greedy_search,
+    greedy_search_batch,
+    greedy_search_batch_with_context,
+    greedy_search_with_context,
+    modified_beam_search,
+)
+from dataset import naive_triplet_text_sampling, random_shuffle_subset
+from lhotse import load_manifest_lazy
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    ref_text_normalization,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from train_bert_encoder_with_style import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+from transformers import BertModel, BertTokenizer
+from utils import get_facebook_biasing_list
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--log-dir",
+        type=str,
+        required=True,
+        help="Where to store the logs",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--input-manifest",
+        type=str,
+        required=True,
+        help="The input manifest to be decoded",
+    )
+
+    parser.add_argument(
+        "--output-manifest",
+        type=str,
+        required=True,
+        help="Where to store the output manifest (directory)",
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=True,
+        help="Use pre-text is available during decoding",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt when evaluation",
+    )
+
+    parser.add_argument(
+        "--use-context-embedding",
+        type=str2bool,
+        default=False,
+        help="Use context fuser when evaluation",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+        help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ",
+    )
+
+    parser.add_argument(
+        "--compute-CER",
+        type=str2bool,
+        default=True,
+        help="Reports CER. By default, only reports WER",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--use-ls-test-set",
+        type=str2bool,
+        default=False,
+        help="Use librispeech test set for evaluation.",
+    )
+
+    parser.add_argument(
+        "--use-ls-context-list",
+        type=str2bool,
+        default=False,
+        help="If use a fixed context list for LibriSpeech decoding",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _apply_style_transform(text: List[str], transform: str) -> List[str]:
+    """Apply transform to a list of text. By default, the text are in
+    ground truth format, i.e mixed-punc.
+
+    Args:
+        text (List[str]): Input text string
+        transform (str): Transform to be applied
+
+    Returns:
+        List[str]: _description_
+    """
+    if transform == "mixed-punc":
+        return text
+    elif transform == "upper-no-punc":
+        return [upper_only_alpha(s) for s in text]
+    elif transform == "lower-no-punc":
+        return [lower_only_alpha(s) for s in text]
+    elif transform == "lower-punc":
+        return [lower_all_char(s) for s in text]
+    else:
+        raise NotImplementedError(f"Unseen transform: {transform}")
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    batch: dict,
+    biasing_dict: dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    cuts = batch["supervisions"]["cut"]
+    cut_ids = [c.supervisions[0].id for c in cuts]
+    batch_size = feature.size(0)
+
+    # get pre_text
+    if "pre_text" in batch["supervisions"] and params.use_pre_text:
+        pre_texts = batch["supervisions"][
+            "text"
+        ]  # use the ground truth ref text as pre_text
+        pre_texts = [train_text_normalization(t) for t in pre_texts]
+    else:
+        pre_texts = ["" for _ in range(batch_size)]
+
+    if params.use_ls_context_list:
+        pre_texts = [biasing_dict[id] for id in cut_ids]
+
+    # get style_text
+    if params.use_style_prompt:
+        fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it's fully not related."
+        style_texts = batch["supervisions"].get(
+            "style_text", [fixed_sentence for _ in range(batch_size)]
+        )
+        style_texts = [train_text_normalization(t) for t in style_texts]
+    else:
+        style_texts = ["" for _ in range(batch_size)]  # use empty string
+
+    # Get the text embedding input
+    if params.use_pre_text or params.use_style_prompt:
+
+        # apply style transform to the pre_text and style_text
+        pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform)
+        # pre_texts = random_shuffle_subset(pre_texts, p=1.0, p_mask=0.0)
+        if params.use_style_prompt:
+            style_texts = _apply_style_transform(
+                style_texts, params.style_text_transform
+            )
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+
+            # Use tokenizer to prepare input for text encoder
+            encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+                pre_texts=pre_texts,
+                style_texts=style_texts,
+                tokenizer=tokenizer,
+                device=device,
+            )
+
+            memory, memory_key_padding_mask = model.encode_text(
+                encoded_inputs=encoded_inputs,
+                style_lens=style_lens,
+            )  # (T,B,C)
+    else:
+        memory = None
+        memory_key_padding_mask = None
+
+    # Get the transducer encoder output
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=feature,
+            feature_lens=feature_lens,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+
+    hyps = []
+
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        if memory is None or not params.use_context_embedding:
+            hyp_tokens = greedy_search_batch(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+            )
+        else:
+            memory = memory.permute(1, 0, 2)  # (T,N,C) -> (N,T,C)
+            context = model.context_fuser(
+                memory, padding_mask=memory_key_padding_mask
+            )  # (N,C)
+            context = model.joiner.context_proj(context)  # (N,C)
+            hyp_tokens = greedy_search_batch_with_context(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                context=context,
+            )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                if memory is None or not params.use_context_embedding:
+                    hyp = greedy_search(
+                        model=model,
+                        encoder_out=encoder_out_i,
+                        max_sym_per_frame=params.max_sym_per_frame,
+                    )
+                else:
+                    cur_context = context[i : i + 1, :]
+                    hyp = greedy_search_with_context(
+                        model=model,
+                        encoder_out=encoder_out_i,
+                        context=cur_context,
+                        max_sym_per_frame=params.max_sym_per_frame,
+                    )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    biasing_dict: Dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 40
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"][
+            "text"
+        ]  # By default, this should be in mixed-punc format
+
+        # the style of ref_text should match style_text
+        texts = _apply_style_transform(texts, params.style_text_transform)
+        if params.use_style_prompt:
+            texts = _apply_style_transform(texts, params.style_text_transform)
+
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_text = ref_text_normalization(
+                    ref_text
+                )  # remove full-width symbols & some book marks
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        if params.compute_CER:
+            # Write CER statistics
+            recog_path = (
+                params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt"
+            )
+            store_transcripts(filename=recog_path, texts=results, char_level=True)
+            errs_filename = (
+                params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt"
+            )
+            with open(errs_filename, "w") as f:
+                cer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{key}",
+                    results,
+                    enable_log=True,
+                    compute_CER=params.compute_CER,
+                )
+                test_set_cers[key] = cer
+
+            logging.info("Wrote detailed CER stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    if params.compute_CER:
+        test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+        errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tCER", file=f)
+            for key, val in test_set_cers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_cers:
+            s += "{} CER\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+def add_decoding_result_to_manifest(
+    in_manifest,
+    out_manifest: str,
+    results_dict: Dict,
+):
+    # write the decoding results with prompt to the manifest as an
+    # extra ref text
+    new_ans = {}
+    for key, value in results_dict.items():
+        for items in value:
+            id, ref, hyp = items
+            new_ans[id] = " ".join(hyp)
+
+    def _add_decoding(c):
+        key = c.supervisions[0].id
+        c.supervisions[0].texts.append(new_ans[key])
+        return c
+
+    in_manifest = in_manifest.map(_add_decoding)
+    logging.info(f"Saving manifest to {out_manifest}")
+    in_manifest.to_file(out_manifest)
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    cuts = load_manifest_lazy(args.input_manifest)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        splitted_cuts = cuts.split(num_splits=world_size)
+        mp.spawn(
+            run, args=(world_size, args, splitted_cuts), nprocs=world_size, join=True
+        )
+    else:
+        run(rank=0, world_size=1, args=args, cuts=cuts)
+
+
+@torch.no_grad()
+def run(rank, world_size, args, cuts):
+    params = get_params()
+    params.update(vars(args))
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_pre_text:
+        params.suffix += f"-pre-text-{params.pre_text_transform}"
+
+    if params.use_style_prompt:
+        params.suffix += f"-style-prompt-{params.style_text_transform}"
+
+    params.suffix += f"-{rank}"
+
+    world_size = params.world_size
+
+    params.output_manifest = Path(params.output_manifest)
+    if world_size > 1:
+        cuts = cuts[rank]
+        out_name = params.output_manifest / f"with_decoding_job_{rank}.jsonl.gz"
+    else:
+        out_name = params.output_manifest / "with_decoding.jsonl.gz"
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    setup_logger(f"{params.log_dir}/log-get-manifest-with-decoding-{rank}")
+    logging.info("Decoding started")
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    LM = None
+
+    decoding_graph = None
+    word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    dl = libriheavy.valid_dataloaders(
+        cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+
+    test_sets = ["test"]
+    test_dl = [dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        biasing_dict = None
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        # save_results(
+        #     params=params,
+        #     test_set_name=test_set,
+        #     results_dict=results_dict,
+        # )
+
+        add_decoding_result_to_manifest(
+            in_manifest=cuts,
+            out_manifest=out_name,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+# torch.set_num_threads(1)
+# torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py
new file mode 100644
index 000000000..91f167204
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py
@@ -0,0 +1,135 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import Balancer
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+            padding_idx=blank_id,
+        )
+        # the balancers are to avoid any drift in the magnitude of the
+        # embeddings, which would interact badly with parameter averaging.
+        self.balancer = Balancer(
+            decoder_dim,
+            channel_dim=-1,
+            min_positive=0.0,
+            max_positive=1.0,
+            min_abs=0.5,
+            max_abs=1.0,
+            prob=0.05,
+        )
+
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim // 4,  # group size == 4
+                bias=False,
+            )
+            self.balancer2 = Balancer(
+                decoder_dim,
+                channel_dim=-1,
+                min_positive=0.0,
+                max_positive=1.0,
+                min_abs=0.5,
+                max_abs=1.0,
+                prob=0.05,
+            )
+        else:
+            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
+            # when inference with torch.jit.script and context_size == 1
+            self.conv = nn.Identity()
+            self.balancer2 = nn.Identity()
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+
+        embedding_out = self.balancer(embedding_out)
+
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+            embedding_out = F.relu(embedding_out)
+            embedding_out = self.balancer2(embedding_out)
+
+        return embedding_out
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py b/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py
new file mode 100644
index 000000000..257facce4
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py
@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py b/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py
new file mode 100644
index 000000000..e0bc556a8
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py
@@ -0,0 +1,255 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+"""
+Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer_prompt_asr/export_PromptASR.py \
+  --exp-dir ./zipformer_prompt_asr/exp \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --epoch 50 \
+  --avg 10
+
+- For streaming model:
+
+./zipformer_prompt_asr/export_PromptASR.py \
+  --exp-dir ./zipformer_prompt_asr/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --epoch 50 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from torch import Tensor, nn
+from train_bert_encoder import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    assert params.jit is False, "Jit is not supported yet"
+
+    logging.info("Not using torchscript. Export model.state_dict()")
+    # Save it using a format so that it can be loaded
+    # by :func:`load_checkpoint`
+    filename = params.exp_dir / "pretrained.pt"
+    torch.save({"model": model.state_dict()}, str(filename))
+    logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py b/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py
new file mode 100644
index 000000000..59f822748
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py
@@ -0,0 +1,86 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        context_dim: int = 512,
+        context_injection: bool = False,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim, initial_scale=0.25)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim, initial_scale=0.25)
+        self.output_linear = nn.Linear(joiner_dim, vocab_size)
+        if context_injection:
+            self.context_proj = ScaledLinear(
+                context_dim, joiner_dim, initial_scale=0.25
+            )
+        else:
+            self.context_proj = None
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        context: torch.Tensor = None,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+          context:
+            An embedding vector representing the previous context information
+          project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.shape[:-1] == decoder_out.shape[:-1]
+
+        if project_input:
+            if context:
+                logit = (
+                    self.encoder_proj(encoder_out)
+                    + self.decoder_proj(decoder_out)
+                    + self.context_proj(context)
+                )
+            else:
+                logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            if context is not None:
+                logit = encoder_out + decoder_out + context.unsqueeze(1).unsqueeze(1)
+            else:
+                logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py b/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py
new file mode 100644
index 000000000..9a693ca4f
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py
@@ -0,0 +1,153 @@
+import re
+
+words = {
+    0: "zero",
+    1: "one",
+    2: "two",
+    3: "three",
+    4: "four",
+    5: "five",
+    6: "six",
+    7: "seven",
+    8: "eight",
+    9: "nine",
+    10: "ten",
+    11: "eleven",
+    12: "twelve",
+    13: "thirteen",
+    14: "fourteen",
+    15: "fifteen",
+    16: "sixteen",
+    17: "seventeen",
+    18: "eighteen",
+    19: "nineteen",
+    20: "twenty",
+    30: "thirty",
+    40: "forty",
+    50: "fifty",
+    60: "sixty",
+    70: "seventy",
+    80: "eighty",
+    90: "ninety",
+}
+ordinal_nums = [
+    "zeroth",
+    "first",
+    "second",
+    "third",
+    "fourth",
+    "fifth",
+    "sixth",
+    "seventh",
+    "eighth",
+    "ninth",
+    "tenth",
+    "eleventh",
+    "twelfth",
+    "thirteenth",
+    "fourteenth",
+    "fifteenth",
+    "sixteenth",
+    "seventeenth",
+    "eighteenth",
+    "nineteenth",
+    "twentieth",
+]
+
+num_ordinal_dict = {num: ordinal_nums[num] for num in range(21)}
+
+
+def year_to_words(num: int):
+    assert isinstance(num, int), num
+    # check if a num is representing a year
+    if num > 1500 and num < 2000:
+        return words[num // 100] + " " + num_to_words(num % 100)
+    elif num == 2000:
+        return "TWO THOUSAND"
+    elif num > 2000:
+        return "TWO THOUSAND AND " + num_to_words(num % 100)
+    else:
+        return num_to_words(num)
+
+
+def num_to_words(num: int):
+    # Return the English words of a integer number
+
+    # If this is a year number
+    if num > 1500 and num < 2030:
+        return year_to_words(num)
+
+    if num < 20:
+        return words[num]
+    if num < 100:
+        if num % 10 == 0:
+            return words[num // 10 * 10]
+        else:
+            return words[num // 10 * 10] + " " + words[num % 10]
+    if num < 1000:
+        return words[num // 100] + " hundred and " + num_to_words(num % 100)
+    if num < 1000000:
+        return num_to_words(num // 1000) + " thousand " + num_to_words(num % 1000)
+    return num
+
+
+def num_to_ordinal_word(num: int):
+
+    return num_ordinal_dict.get(num, num_to_words(num)).upper()
+
+
+def replace_full_width_symbol(s: str) -> str:
+    # replace full-width symbol with theri half width counterpart
+    s = s.replace("“", '"')
+    s = s.replace("”", '"')
+    s = s.replace("‘", "'")
+    s = s.replace("’", "'")
+
+    return s
+
+
+def decoding_normalization(text: str) -> str:
+    text = replace_full_width_symbol(text)
+
+    # Only keep all alpha-numeric characters, hypen and apostrophe
+    text = text.replace("-", " ")
+    text = re.sub(r"[^a-zA-Z0-9\s']+", "", text)
+    return text
+
+
+def word_normalization(word: str) -> str:
+    # 1 .Use full word for some abbreviation
+    # 2. Convert digits to english words
+    # 3. Convert ordinal number to english words
+    if word == "MRS":
+        return "MISSUS"
+    if word == "MR":
+        return "MISTER"
+    if word == "ST":
+        return "SAINT"
+    if word == "ECT":
+        return "ET CETERA"
+    if word.isnumeric():
+        word = num_to_words(int(word))
+        return str(word).upper()
+    #  e.g 9TH, 6TH
+    if word[-2:] == "TH" and word[0].isnumeric():
+        return num_to_ordinal_word(int(word[:-2])).upper()
+    if word[0] == "'":
+        return word[1:]
+
+    return word
+
+
+def simple_normalization(text: str) -> str:
+    text = replace_full_width_symbol(text)
+    text = text.replace("--", " ")
+
+    return text
+
+
+if __name__ == "__main__":
+
+    s = str(1830)
+    out = word_normalization(s)
+    print(s, out)
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py
new file mode 100644
index 000000000..77b4057c4
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py
@@ -0,0 +1,262 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+import warnings
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear, penalize_abs_values_gt
+from torch import Tensor
+
+from icefall.utils import add_sos, make_pad_mask
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+        self.simple_lm_proj = ScaledLinear(
+            decoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          text:
+            A 2-D tensor of integer dtype containing prompt text, of shape (N, T).
+            It is exptected to contain the style prompt (first) and then the content
+            prompt.
+          text_lens:
+            A 1-D tensor of shape (N,). It contains the number of elements (bytes)
+            in `text` before padding, which will include the lengths of the
+            style plus the content prompt.
+          style_lens:
+            A 1-D tensor of shape (N,), containing the number of elements (bytes)
+            within each row of `text` that correspond to the style prompt (these
+            are expected to come first).
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        x, x_lens = self.encoder_embed(x, x_lens)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, x_lens = self.encoder(
+            x,
+            x_lens,
+            src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
+
+    def encode_audio(
+        self,
+        feature: Tensor,
+        feature_lens: Tensor,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Encode the input audio features
+
+        Args:
+            feature (Tensor): Input audio (N,T,C)
+            feature_lens (Tensor): Length of input audio (N,)
+        Returns:
+            Tuple[Tensor, Tensor]: Encoded acoustic features and length
+        """
+        x, x_lens = self.encoder_embed(feature, feature_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(
+            x=x,
+            x_lens=x_lens,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py b/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py
new file mode 100644
index 000000000..21c7b4fac
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py
@@ -0,0 +1,392 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+import warnings
+from typing import Dict, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear, penalize_abs_values_gt
+from torch import Tensor
+
+from icefall.utils import add_sos, make_pad_mask
+
+
+class PromptedTransducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        text_encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        use_BERT: bool = True,
+        text_encoder_type: str = "BERT",
+        text_encoder_adapter: bool = False,
+        freeze_text_encoder: bool = True,
+        context_fuser: nn.Module = None,
+    ):
+        """
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          text_encoder:
+            This is a encoder that processes text information (e.g content prompt
+            and style prompt). The input is `x` of (N,T) and `x_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          text_encoder_type:
+            The type of the text_encoder. Supported are (BERT, DistilBERT)
+          context_fuser
+            A optional module that fuses the embeddings of text encoder. The fused embedding
+            will be added to the joiner.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+        self.text_encoder = text_encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+        self.simple_lm_proj = ScaledLinear(
+            decoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+
+        self.use_BERT = use_BERT  # if the text encoder is a pre-trained BERT
+        self.context_fuser = context_fuser
+
+        assert text_encoder_type in (
+            "BERT",
+            "DistilBERT",
+            "BERT-UNCASED",
+        ), f"Unseen text_encoder type {text_encoder_type}"
+        self.text_encoder_dim = (
+            self.text_encoder.config.hidden_size
+            if text_encoder_type in ("BERT", "BERT-UNCASED")
+            else self.text_encoder.config.dim
+        )
+        self.freeze_text_encoder = freeze_text_encoder
+
+        if text_encoder_adapter:
+            self.text_encoder_adapter = nn.Sequential(
+                nn.Linear(self.text_encoder_dim, self.text_encoder_dim, bias=False),
+                nn.Tanh(),
+            )
+        else:
+            self.text_encoder_adapter = None
+
+        self.style_prompt_embedding = nn.Parameter(
+            torch.full((self.text_encoder_dim,), 0.5)
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        encoded_inputs: Dict,
+        style_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        use_pre_text: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          text:
+            A 2-D tensor of integer dtype containing prompt text, of shape (N, T).
+            It is exptected to contain the style prompt (first) and then the content
+            prompt.
+          text_lens:
+            A 1-D tensor of shape (N,). It contains the number of elements (bytes)
+            in `text` before padding, which will include the lengths of the
+            style plus the content prompt.
+          style_lens:
+            A 1-D tensor of shape (N,), containing the number of elements (bytes)
+            within each row of `text` that correspond to the style prompt (these
+            are expected to come first).
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        if self.freeze_text_encoder:
+            self.text_encoder.eval()
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        x, x_lens = self.encoder_embed(x, x_lens)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # freeze the BERT text encoder
+
+        if use_pre_text:
+            memory, memory_key_padding_mask = self.encode_text(
+                encoded_inputs, style_lens=style_lens
+            )
+        else:
+            memory = None
+            memory_key_padding_mask = None
+
+        encoder_out, x_lens = self.encoder(
+            x,
+            x_lens,
+            src_key_padding_mask,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        if self.context_fuser is not None and memory is not None:
+            memory = memory.permute(1, 0, 2)  # (T,N,C) -> (N,T,C)
+            context = self.context_fuser(memory, padding_mask=memory_key_padding_mask)
+            context = self.joiner.context_proj(context)
+        else:
+            context = None
+
+        logits = self.joiner(am_pruned, lm_pruned, context=context, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
+
+    def _add_style_indicator(self, memory: Tensor, style_lens: Tensor):
+        """
+        Adds to `memory` an indicator that is 1.0 for positions that correspond to
+        the `style prompt` and 0 elsewhere.  The scale can be fixed because the
+        scale of the embedding vector can adjust to compensate.
+
+        Args:
+            memory: (memory_len, batch_size, embed_dim)
+            style_lens: (batch_size,),  a vector of lengths of the style prompt.
+        """
+
+        (memory_len, batch_size, embed_dim) = memory.shape
+
+        indicator = (
+            torch.arange(memory_len, device=memory.device).unsqueeze(-1) < style_lens
+        )
+        indicator = indicator.to(memory.dtype)
+
+        extra_term = torch.zeros_like(memory)
+        extra_term += indicator.unsqueeze(-1) * self.style_prompt_embedding.expand(
+            memory_len, batch_size, self.text_encoder_dim
+        )
+
+        return memory + extra_term
+
+    def encode_text(
+        self,
+        encoded_inputs: Dict,
+        style_lens: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Get the embeddings of text
+
+        Args:
+            encoded_inputs: The encoded inputs generated by a tokenizer (Dict)
+
+        Returns:
+            Tuple[Tensor, Tensor]: Returns the text embeddings encoded by the
+            text_encoder and the attention mask
+        """
+        text_lens = encoded_inputs.pop("length")  # need to use pop to remove this item
+
+        # Freeze the pre-trained text encoder
+        with torch.no_grad():
+            memory = self.text_encoder(**encoded_inputs)["last_hidden_state"]  # (B,T,C)
+            memory = memory.permute(1, 0, 2)
+
+        # Text encoder adapter
+        if self.text_encoder_adapter is not None:
+            memory = self.text_encoder_adapter(memory)
+
+        memory = self._add_style_indicator(memory, style_lens)
+
+        memory_key_padding_mask = make_pad_mask(text_lens)
+
+        return memory, memory_key_padding_mask
+
+    def encode_audio(
+        self,
+        feature: Tensor,
+        feature_lens: Tensor,
+        memory: Optional[Tensor],
+        memory_key_padding_mask: Optional[Tensor],
+    ) -> Tuple[Tensor, Tensor]:
+        """Encode the input audio features
+
+        Args:
+            feature (Tensor): Input audio (N,T,C)
+            feature_lens (Tensor): Length of input audio (N,)
+            memory (Tensor): Embeddings from the text encoder
+            memory_key_padding_mask (Tensor): _description_
+
+        Returns:
+            Tuple[Tensor, Tensor]: _description_
+        """
+        x, x_lens = self.encoder_embed(feature, feature_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(
+            x=x,
+            x_lens=x_lens,
+            src_key_padding_mask=src_key_padding_mask,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+Transducer = PromptedTransducer  # for decoding
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py b/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py
new file mode 100644
index 000000000..159e363c7
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py
@@ -0,0 +1,1164 @@
+#      Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import contextlib
+import logging
+import random
+from collections import defaultdict
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.optim import Optimizer
+
+
+class BatchedOptimizer(Optimizer):
+    """
+    This class adds to class Optimizer the capability to optimize parameters in batches:
+    it will stack the parameters and their grads for you so the optimizer can work
+    on tensors with an extra leading dimension.  This is intended for speed with GPUs,
+    as it reduces the number of kernels launched in the optimizer.
+
+    Args:
+      params:
+    """
+
+    def __init__(self, params, defaults):
+        super(BatchedOptimizer, self).__init__(params, defaults)
+
+    @contextlib.contextmanager
+    def batched_params(self, param_group, group_params_names):
+        """
+        This function returns (technically, yields) a list of
+          of tuples (p, state), where
+        p is a `fake` parameter that is stacked (over axis 0) from real parameters
+        that share the same shape, and its gradient is also stacked;
+        `state` is the state corresponding to this batch of parameters
+        (it will be physically located in the "state" for one of the real
+        parameters, the last one that has any particular shape and dtype).
+
+        This function is decorated as a context manager so that it can
+        write parameters back to their "real" locations.
+
+        The idea is, instead of doing:
+        <code>
+          for p in group["params"]:
+             state = self.state[p]
+             ...
+        </code>
+        you can do:
+        <code>
+          with self.batched_params(group["params"]) as batches:
+             for p, state, p_names in batches:
+                 ...
+        </code>
+
+        Args:
+          group: a parameter group, which is a list of parameters; should be
+                one of self.param_groups.
+          group_params_names: name for each parameter in group,
+                which is List[str].
+        """
+        batches = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
+        batches_names = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of str
+
+        assert len(param_group) == len(group_params_names)
+        for p, named_p in zip(param_group, group_params_names):
+            key = (str(p.dtype), *p.shape)
+            batches[key].append(p)
+            batches_names[key].append(named_p)
+
+        batches_names_keys = list(batches_names.keys())
+        sorted_idx = sorted(
+            range(len(batches_names)), key=lambda i: batches_names_keys[i]
+        )
+        batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx]
+        batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
+
+        stacked_params_dict = dict()
+
+        # turn batches into a list, in deterministic order.
+        # tuples will contain tuples of (stacked_param, state, stacked_params_names),
+        # one for each batch in `batches`.
+        tuples = []
+
+        for batch, batch_names in zip(batches, batches_names):
+            p = batch[0]
+            # we arbitrarily store the state in the
+            # state corresponding to the 1st parameter in the
+            # group.  class Optimizer will take care of saving/loading state.
+            state = self.state[p]
+            p_stacked = torch.stack(batch)
+            grad = torch.stack(
+                [torch.zeros_like(p) if p.grad is None else p.grad for p in batch]
+            )
+            p_stacked.grad = grad
+            stacked_params_dict[key] = p_stacked
+            tuples.append((p_stacked, state, batch_names))
+
+        yield tuples  # <-- calling code will do the actual optimization here!
+
+        for (stacked_params, _state, _names), batch in zip(tuples, batches):
+            for i, p in enumerate(batch):  # batch is list of Parameter
+                p.copy_(stacked_params[i])
+
+
+class ScaledAdam(BatchedOptimizer):
+    """
+     Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
+     proportional to the norm of that parameter; and also learn the scale of the parameter,
+     in log space, subject to upper and lower limits (as if we had factored each parameter as
+     param = underlying_param * log_scale.exp())
+
+
+     Args:
+          params:  The parameters or param_groups to optimize (like other Optimizer subclasses)
+                   Unlike common optimizers, which accept model.parameters() or groups of parameters(),
+                   this optimizer could accept model.named_parameters() or groups of named_parameters().
+                   See comments of function _get_names_of_parameters for its 4 possible cases.
+              lr:  The learning rate.  We will typically use a learning rate schedule that starts
+                   at 0.03 and decreases over time, i.e. much higher than other common
+                   optimizers.
+     clipping_scale: (e.g. 2.0)
+                   A scale for gradient-clipping: if specified, the normalized gradients
+                   over the whole model will be clipped to have 2-norm equal to
+                   `clipping_scale` times the median 2-norm over the most recent period
+                   of `clipping_update_period` minibatches.  By "normalized gradients",
+                   we mean after multiplying by the rms parameter value for this tensor
+                   [for non-scalars]; this is appropriate because our update is scaled
+                   by this quantity.
+            betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
+                   Must satisfy 0 < beta <= beta2 < 1.
+     scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
+                   scale of each parameter tensor and scalar parameters of the mode..
+                   If each parameter were decomposed
+                   as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
+                   would be a the scaling factor on the learning rate of p_scale.
+              eps:  A general-purpose epsilon to prevent division by zero
+    param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be >= this value)
+    param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be <= this value)
+       scalar_max: Maximum absolute value for scalar parameters (applicable if your
+                   model has any parameters with numel() == 1).
+    size_update_period: The periodicity, in steps, with which we update the size (scale)
+                   of the parameter tensor.  This is provided to save a little time
+                   in the update.
+     clipping_update_period: if clipping_scale is specified, this is the period
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=3e-02,
+        clipping_scale=None,
+        betas=(0.9, 0.98),
+        scalar_lr_scale=0.1,
+        eps=1.0e-08,
+        param_min_rms=1.0e-05,
+        param_max_rms=3.0,
+        scalar_max=10.0,
+        size_update_period=4,
+        clipping_update_period=100,
+    ):
+        defaults = dict(
+            lr=lr,
+            clipping_scale=clipping_scale,
+            betas=betas,
+            scalar_lr_scale=scalar_lr_scale,
+            eps=eps,
+            param_min_rms=param_min_rms,
+            param_max_rms=param_max_rms,
+            scalar_max=scalar_max,
+            size_update_period=size_update_period,
+            clipping_update_period=clipping_update_period,
+        )
+
+        # If params only contains parameters or group of parameters,
+        # i.e when parameter names are not given,
+        # this flag will be set to False in funciton _get_names_of_parameters.
+        self.show_dominant_parameters = True
+        param_groups, parameters_names = self._get_names_of_parameters(params)
+        super(ScaledAdam, self).__init__(param_groups, defaults)
+        assert len(self.param_groups) == len(parameters_names)
+        self.parameters_names = parameters_names
+
+    def _get_names_of_parameters(
+        self, params_or_named_params
+    ) -> Tuple[List[Dict], List[List[str]]]:
+        """
+        Args:
+          params_or_named_params: according to the way ScaledAdam is initialized in train.py,
+            this argument could be one of following 4 cases,
+            case 1, a generator of parameter, e.g.:
+              optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=3.0)
+
+            case 2, a list of parameter groups with different config, e.g.:
+              model_param_groups = [
+                      {'params': model.encoder.parameters(), 'lr': 0.05},
+                      {'params': model.decoder.parameters(), 'lr': 0.01},
+                      {'params': model.joiner.parameters(), 'lr': 0.03},
+                      ]
+              optimizer = ScaledAdam(model_param_groups, lr=params.base_lr, clipping_scale=3.0)
+
+            case 3, a generator of named_parameter, e.g.:
+              optimizer = ScaledAdam(model.named_parameters(), lr=params.base_lr, clipping_scale=3.0)
+
+            case 4, a list of named_parameter groups with different config, e.g.:
+              model_named_param_groups = [
+                      {'named_params': model.encoder.named_parameters(), 'lr': 0.05},
+                      {'named_params': model.decoder.named_parameters(), 'lr': 0.01},
+                      {'named_params': model.joiner.named_parameters(), 'lr': 0.03},
+                      ]
+              optimizer = ScaledAdam(model_named_param_groups, lr=params.base_lr, clipping_scale=3.0)
+
+          For case 1 and case 2, input params is used to initialize the underlying torch.optimizer.
+          For case 3 and case 4, firstly, names and params are extracted from input named_params,
+            then, these extracted params are used to initialize the underlying torch.optimizer,
+            and these extracted names are mainly used by function
+            `_show_gradient_dominating_parameter`
+
+        Returns:
+          Returns a tuple containing 2 elements:
+            - `param_groups` with type List[Dict], each Dict element is a parameter group.
+              An example of `param_groups` could be:
+              [
+                  {'params': `one iterable of Parameter`, 'lr': 0.05},
+                  {'params': `another iterable of Parameter`, 'lr': 0.08},
+                  {'params': `a third iterable of Parameter`, 'lr': 0.1},
+              ]
+            - `param_gruops_names` with type List[List[str]],
+               each `List[str]` is for a group['params'] in param_groups,
+               and each `str` is the name of a parameter.
+               A dummy name "foo" is related to each parameter,
+               if input are params without names, i.e. case 1 or case 2.
+        """
+        # variable naming convention in this function:
+        #   p is short for param.
+        #   np is short for named_param.
+        #   p_or_np is short for param_or_named_param.
+        #   cur is short for current.
+        #   group is a dict, e.g. {'params': iterable of parameter, 'lr': 0.05, other fields}.
+        #   groups is a List[group]
+
+        iterable_or_groups = list(params_or_named_params)
+        if len(iterable_or_groups) == 0:
+            raise ValueError("optimizer got an empty parameter list")
+
+        # The first value of returned tuple.  A list of dicts containing at
+        # least 'params' as a key.
+        param_groups = []
+
+        # The second value of returned tuple,
+        # a List[List[str]], each sub-List is for a group.
+        param_groups_names = []
+
+        if not isinstance(iterable_or_groups[0], dict):
+            # case 1 or case 3,
+            # the input is an iterable of parameter or named parameter.
+            param_iterable_cur_group = []
+            param_names_cur_group = []
+            for p_or_np in iterable_or_groups:
+                if isinstance(p_or_np, tuple):
+                    # case 3
+                    name, param = p_or_np
+                else:
+                    # case 1
+                    assert isinstance(p_or_np, torch.Tensor)
+                    param = p_or_np
+                    # Assign a dummy name as a placeholder
+                    name = "foo"
+                    self.show_dominant_parameters = False
+                param_iterable_cur_group.append(param)
+                param_names_cur_group.append(name)
+            param_groups.append({"params": param_iterable_cur_group})
+            param_groups_names.append(param_names_cur_group)
+        else:
+            # case 2 or case 4
+            # the input is groups of parameter or named parameter.
+            for cur_group in iterable_or_groups:
+                assert "named_params" in cur_group
+                name_list = [x[0] for x in cur_group["named_params"]]
+                p_list = [x[1] for x in cur_group["named_params"]]
+                del cur_group["named_params"]
+                cur_group["params"] = p_list
+                param_groups.append(cur_group)
+                param_groups_names.append(name_list)
+
+        return param_groups, param_groups_names
+
+    def __setstate__(self, state):
+        super(ScaledAdam, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        batch = True
+
+        for group, group_params_names in zip(self.param_groups, self.parameters_names):
+            with self.batched_params(group["params"], group_params_names) as batches:
+                # batches is list of pairs (stacked_param, state).  stacked_param is like
+                # a regular parameter, and will have a .grad, but the 1st dim corresponds to
+                # a stacking dim, it is not a real dim.
+
+                if (
+                    len(batches[0][1]) == 0
+                ):  # if len(first state) == 0: not yet initialized
+                    clipping_scale = 1
+                else:
+                    clipping_scale = self._get_clipping_scale(group, batches)
+
+                for p, state, _ in batches:
+                    # Perform optimization step.
+                    # grad is not going to be None, we handled that when creating the batches.
+                    grad = p.grad
+                    if grad.is_sparse:
+                        raise RuntimeError(
+                            "ScaledAdam optimizer does not support sparse gradients"
+                        )
+                    # State initialization
+                    if len(state) == 0:
+                        self._init_state(group, p, state)
+
+                    self._step_one_batch(group, p, state, clipping_scale)
+
+        return loss
+
+    def _init_state(self, group: dict, p: Tensor, state: dict):
+        """
+        Initializes state dict for parameter 'p'.  Assumes that dim 0 of tensor p
+        is actually the batch dimension, corresponding to batched-together
+        parameters of a given shape.
+
+
+        Args:
+           group:   Dict to look up configuration values.
+               p: The parameter that we are initializing the state for
+           state: Dict from string to whatever state we are initializing
+        """
+        size_update_period = group["size_update_period"]
+
+        state["step"] = 0
+
+        kwargs = {"device": p.device, "dtype": p.dtype}
+
+        # 'delta' implements conventional momentum.  There are
+        # several different kinds of update going on, so rather than
+        # compute "exp_avg" like in Adam, we store and decay a
+        # parameter-change "delta", which combines all forms of
+        # update.  this is equivalent to how it's done in Adam,
+        # except for the first few steps.
+        state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        numel = p.numel()
+
+        if numel > 1:
+            # "param_rms" just periodically records the scalar root-mean-square value of
+            # the parameter tensor.
+            # it has a shape like (batch_size, 1, 1, 1, 1)
+            param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+            state["param_rms"] = param_rms
+
+            state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
+            state["scale_grads"] = torch.zeros(
+                size_update_period, *param_rms.shape, **kwargs
+            )
+
+        # exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
+        state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+    def _get_clipping_scale(
+        self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]]
+    ) -> float:
+        """
+        Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
+        by this amount before applying the rest of the update.
+
+        Args:
+           group: the parameter group, an item in self.param_groups
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+        """
+        assert len(tuples) >= 1
+        clipping_scale = group["clipping_scale"]
+        (first_p, first_state, _) = tuples[0]
+        step = first_state["step"]
+        if clipping_scale is None or step == 0:
+            # no clipping.  return early on step == 0 because the other
+            # parameters' state won't have been initialized yet.
+            return 1.0
+        clipping_update_period = group["clipping_update_period"]
+
+        tot_sumsq = torch.tensor(0.0, device=first_p.device)
+        for p, state, param_names in tuples:
+            grad = p.grad
+            if grad.is_sparse:
+                raise RuntimeError(
+                    "ScaledAdam optimizer does not support sparse gradients"
+                )
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                tot_sumsq += (grad**2).sum()  # sum() to change shape [1] to []
+            else:
+                tot_sumsq += ((grad * state["param_rms"]) ** 2).sum()
+
+        tot_norm = tot_sumsq.sqrt()
+        if "model_norms" not in first_state:
+            first_state["model_norms"] = torch.zeros(
+                clipping_update_period, device=p.device
+            )
+        first_state["model_norms"][step % clipping_update_period] = tot_norm
+
+        if step % clipping_update_period == 0:
+            # Print some stats.
+            # We don't reach here if step == 0 because we would have returned
+            # above.
+            sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
+            quartiles = []
+            for n in range(0, 5):
+                index = min(
+                    clipping_update_period - 1, (clipping_update_period // 4) * n
+                )
+                quartiles.append(sorted_norms[index].item())
+
+            median = quartiles[2]
+            threshold = clipping_scale * median
+            first_state["model_norm_threshold"] = threshold
+            percent_clipped = (
+                first_state["num_clipped"] * 100.0 / clipping_update_period
+                if "num_clipped" in first_state
+                else 0.0
+            )
+            first_state["num_clipped"] = 0
+            quartiles = " ".join(["%.3e" % x for x in quartiles])
+            logging.info(
+                f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
+                f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
+            )
+
+        if step < clipping_update_period:
+            return 1.0  # We have not yet estimated a norm to clip to.
+        else:
+            try:
+                model_norm_threshold = first_state["model_norm_threshold"]
+            except KeyError:
+                logging.info(
+                    "Warning: model_norm_threshold not in state: possibly "
+                    "you changed config when restarting, adding clipping_scale option?"
+                )
+                return 1.0
+            ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
+            if ans < 1.0:
+                first_state["num_clipped"] += 1
+            if ans < 0.1:
+                logging.warn(
+                    f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
+                )
+                if self.show_dominant_parameters:
+                    assert p.shape[0] == len(param_names)
+                    self._show_gradient_dominating_parameter(tuples, tot_sumsq)
+            return ans
+
+    def _show_gradient_dominating_parameter(
+        self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor
+    ):
+        """
+        Show information of parameter which dominates tot_sumsq.
+
+        Args:
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+            tot_sumsq: sumsq of all parameters. Though it's could be calculated
+                from tuples, we still pass it to save some time.
+        """
+        all_sumsq_orig = {}
+        for p, state, batch_param_names in tuples:
+            # p is a stacked batch parameters.
+            batch_grad = p.grad
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                batch_sumsq_orig = batch_grad**2
+                # Dummy values used by following `zip` statement.
+                batch_rms_orig = torch.ones(p.shape[0])
+            else:
+                batch_rms_orig = state["param_rms"]
+                batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum(
+                    dim=list(range(1, batch_grad.ndim))
+                )
+
+            for name, sumsq_orig, rms, grad in zip(
+                batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
+            ):
+                proportion_orig = sumsq_orig / tot_sumsq
+                all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
+
+        assert torch.isclose(
+            sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
+            torch.tensor(1.0),
+        )
+        sorted_by_proportion = {
+            k: v
+            for k, v in sorted(
+                all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True
+            )
+        }
+        dominant_param_name = next(iter(sorted_by_proportion))
+        (
+            dominant_proportion,
+            dominant_sumsq,
+            dominant_rms,
+            dominant_grad,
+        ) = sorted_by_proportion[dominant_param_name]
+        logging.info(
+            f"Parameter dominating tot_sumsq {dominant_param_name}"
+            f" with proportion {dominant_proportion:.2f},"
+            f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
+            f"={dominant_sumsq:.3e},"
+            f" grad_sumsq={(dominant_grad**2).sum():.3e},"
+            f" orig_rms_sq={(dominant_rms**2).item():.3e}"
+        )
+
+    def _step_one_batch(
+        self, group: dict, p: Tensor, state: dict, clipping_scale: float
+    ):
+        """
+        Do the step for one parameter, which is actually going to be a batch of
+        `real` parameters, with dim 0 as the batch dim.
+        Args:
+                  group:  dict to look up configuration values
+                    p: parameter to update (actually multiple parameters stacked together
+                       as a batch)
+                  state: state-dict for p, to look up the optimizer state
+        """
+        lr = group["lr"]
+        size_update_period = group["size_update_period"]
+        beta1 = group["betas"][0]
+
+        grad = p.grad
+        if clipping_scale != 1.0:
+            grad = grad * clipping_scale
+        step = state["step"]
+        delta = state["delta"]
+
+        delta.mul_(beta1)
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        if numel > 1:
+            # Update the size/scale of p, and set param_rms
+            scale_grads = state["scale_grads"]
+            scale_grads[step % size_update_period] = (p * grad).sum(
+                dim=list(range(1, p.ndim)), keepdim=True
+            )
+            if step % size_update_period == size_update_period - 1:
+                param_rms = state["param_rms"]  # shape: (batch_size, 1, 1, ..)
+                param_rms.copy_(
+                    (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+                )
+                if step > 0:
+                    # self._size_update() learns the overall scale on the
+                    # parameter, by shrinking or expanding it.
+                    self._size_update(group, scale_grads, p, state)
+
+        if numel == 1:
+            # For parameters with 1 element we just use regular Adam.
+            # Updates delta.
+            self._step_scalar(group, p, state)
+        else:
+            self._step(group, p, state)
+
+        state["step"] = step + 1
+
+    def _size_update(
+        self, group: dict, scale_grads: Tensor, p: Tensor, state: dict
+    ) -> None:
+        """
+               Called only where p.numel() > 1, this updates the scale of the parameter.
+               If we imagine: p =  underlying_param * scale.exp(), and we are doing
+               gradient descent on underlying param and on scale, this function does the update
+               on `scale`.
+
+               Args:
+              group: dict to look up configuration values
+        scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
+                      grads w.r.t. the scales.
+                  p:  The parameter to update
+               state: The state-dict of p
+        """
+
+        param_rms = state["param_rms"]
+        beta1, beta2 = group["betas"]
+        size_lr = group["lr"] * group["scalar_lr_scale"]
+        param_min_rms = group["param_min_rms"]
+        param_max_rms = group["param_max_rms"]
+        eps = group["eps"]
+        step = state["step"]
+        batch_size = p.shape[0]
+
+        size_update_period = scale_grads.shape[0]
+        # correct beta2 for the size update period: we will have
+        # faster decay at this level.
+        beta2_corr = beta2**size_update_period
+
+        scale_exp_avg_sq = state["scale_exp_avg_sq"]  # shape: (batch_size, 1, 1, ..)
+        scale_exp_avg_sq.mul_(beta2_corr).add_(
+            (scale_grads**2).mean(dim=0),  # mean over dim `size_update_period`
+            alpha=1 - beta2_corr,
+        )  # shape is (batch_size, 1, 1, ...)
+
+        # The 1st time we reach here is when size_step == 1.
+        size_step = (step + 1) // size_update_period
+        bias_correction2 = 1 - beta2_corr**size_step
+        # we don't bother with bias_correction1; this will help prevent divergence
+        # at the start of training.
+
+        denom = scale_exp_avg_sq.sqrt() + eps
+
+        scale_step = (
+            -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom
+        )
+
+        is_too_small = param_rms < param_min_rms
+
+        # when the param gets too small, just don't shrink it any further.
+        scale_step.masked_fill_(is_too_small, 0.0)
+
+        # and ensure the parameter rms after update never exceeds param_max_rms.
+        # We have to look at the trained model for parameters at or around the
+        # param_max_rms, because sometimes they can indicate a problem with the
+        # topology or settings.
+        scale_step = torch.minimum(scale_step, (param_max_rms - param_rms) / param_rms)
+
+        delta = state["delta"]
+        # the factor of (1-beta1) relates to momentum.
+        delta.add_(p * scale_step, alpha=(1 - beta1))
+
+    def _step(self, group: dict, p: Tensor, state: dict):
+        """
+        This function does the core update of self.step(), in the case where the members of
+        the batch have more than 1 element.
+
+        Args:
+            group: A dict which will be used to look up configuration values
+                p: The parameter to be updated
+             grad: The grad of p
+            state: The state-dict corresponding to parameter p
+
+        This function modifies p.
+        """
+        grad = p.grad
+        lr = group["lr"]
+        beta1, beta2 = group["betas"]
+        eps = group["eps"]
+        param_min_rms = group["param_min_rms"]
+        step = state["step"]
+
+        exp_avg_sq = state["exp_avg_sq"]
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
+
+        this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0)
+        bias_correction2 = 1 - beta2 ** (this_step + 1)
+        if bias_correction2 < 0.99:
+            # note: not in-place.
+            exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
+
+        denom = exp_avg_sq.sqrt()
+        denom += eps
+        grad = grad / denom
+
+        alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
+
+        delta = state["delta"]
+        delta.add_(grad * alpha)
+        p.add_(delta)
+
+    def _step_scalar(self, group: dict, p: Tensor, state: dict):
+        """
+        A simplified form of the core update for scalar tensors, where we cannot get a good
+        estimate of the parameter rms.
+        """
+        beta1, beta2 = group["betas"]
+        scalar_max = group["scalar_max"]
+        eps = group["eps"]
+        lr = group["lr"] * group["scalar_lr_scale"]
+        grad = p.grad
+
+        exp_avg_sq = state["exp_avg_sq"]  # shape: (batch_size,)
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+        # bias_correction2 is like in Adam.  Don't bother with bias_correction1;
+        # slower update at the start will help stability anyway.
+        bias_correction2 = 1 - beta2 ** (state["step"] + 1)
+        denom = (exp_avg_sq / bias_correction2).sqrt() + eps
+
+        delta = state["delta"]
+        delta.add_(grad / denom, alpha=-lr * (1 - beta1))
+        p.clamp_(min=-scalar_max, max=scalar_max)
+        p.add_(delta)
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("base_lr", group["lr"])
+
+        self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[Union[int, float]] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            logging.info(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+    The basic formula (before warmup) is:
+      lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                     (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup
+    where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
+    and then stays constant at 1.
+
+
+     E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        warmup_batches: Union[int, float] = 500.0,
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+        self.warmup_batches = warmup_batches
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        warmup_factor = (
+            1.0
+            if self.batch >= self.warmup_batches
+            else 0.5 + 0.5 * (self.batch / self.warmup_batches)
+        )
+
+        return [x * factor * warmup_factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = ScaledAdam(m.parameters(), lr=0.03)
+
+    scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+
+    logging.info(f"last lr = {scheduler.get_last_lr()}")
+    logging.info(f"state dict = {scheduler.state_dict()}")
+
+
+# This is included mostly as a baseline for ScaledAdam.
+class Eve(Optimizer):
+    """
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                if random.random() < 0.0005:
+                    step = (exp_avg / denom) * step_size
+                    logging.info(
+                        f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}"
+                    )
+
+        return loss
+
+
+def _test_scaled_adam(hidden_dim: int):
+    import timeit
+
+    from scaling import ScaledLinear
+
+    E = 100
+    B = 4
+    T = 2
+    logging.info("in test_eve_cain")
+    # device = torch.device('cuda')
+    device = torch.device("cpu")
+    dtype = torch.float32
+
+    fix_random_seed(42)
+    # these input_magnitudes and output_magnitudes are to test that
+    # Abel is working as we expect and is able to adjust scales of
+    # different dims differently.
+    input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+    output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+
+    for iter in [1, 0]:
+        fix_random_seed(42)
+        Linear = torch.nn.Linear if iter == 0 else ScaledLinear
+
+        m = torch.nn.Sequential(
+            Linear(E, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, E),
+        ).to(device)
+
+        train_pairs = [
+            (
+                100.0
+                * torch.randn(B, T, E, device=device, dtype=dtype)
+                * input_magnitudes,
+                torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes,
+            )
+            for _ in range(20)
+        ]
+
+        if iter == 0:
+            optim = Eve(m.parameters(), lr=0.003)
+        elif iter == 1:
+            optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0)
+        scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False)
+
+        start = timeit.default_timer()
+        avg_loss = 0.0
+        for epoch in range(180):
+            scheduler.step_epoch()
+            # if epoch == 100 and iter in [2,3]:
+            #    optim.reset_speedup()  # check it doesn't crash.
+
+            # if epoch == 130:
+            #    opts = diagnostics.TensorDiagnosticOptions(
+            #        512
+            #    )  # allow 4 megabytes per sub-module
+            #    diagnostic = diagnostics.attach_diagnostics(m, opts)
+
+            for n, (x, y) in enumerate(train_pairs):
+                y_out = m(x)
+                loss = ((y_out - y) ** 2).mean() * 100.0
+                if epoch == 0 and n == 0:
+                    avg_loss = loss.item()
+                else:
+                    avg_loss = 0.98 * avg_loss + 0.02 * loss.item()
+                if n == 0 and epoch % 5 == 0:
+                    # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item()
+                    # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item()
+                    # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item()
+                    # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item()
+                    # scale1 = '%.2e' % (m[0].weight_scale.exp().item())
+                    # scale1b = '%.2e' % (m[0].bias_scale.exp().item())
+                    # scale2 = '%.2e' % (m[2].weight_scale.exp().item())
+                    # scale2b = '%.2e' % (m[2].bias_scale.exp().item())
+                    lr = scheduler.get_last_lr()[0]
+                    logging.info(
+                        f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}"
+                    )  # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
+                loss.log().backward()
+                optim.step()
+                optim.zero_grad()
+                scheduler.step_batch()
+
+        # diagnostic.print_diagnostics()
+
+        stop = timeit.default_timer()
+        logging.info(f"Iter={iter}, Time taken: {stop - start}")
+
+        logging.info(f"last lr = {scheduler.get_last_lr()}")
+        # logging.info("state dict = ", scheduler.state_dict())
+        # logging.info("optim state_dict = ", optim.state_dict())
+        logging.info(f"input_magnitudes = {input_magnitudes}")
+        logging.info(f"output_magnitudes = {output_magnitudes}")
+
+
+if __name__ == "__main__":
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    logging.getLogger().setLevel(logging.INFO)
+    import subprocess
+
+    s = subprocess.check_output(
+        "git status -uno .; git log -1; git diff HEAD .", shell=True
+    )
+    logging.info(s)
+    import sys
+
+    if len(sys.argv) > 1:
+        hidden_dim = int(sys.argv[1])
+    else:
+        hidden_dim = 200
+
+    _test_scaled_adam(hidden_dim)
+    _test_eden()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py b/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py
new file mode 100644
index 000000000..48fd2612a
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py
@@ -0,0 +1,359 @@
+#!/usr/bin/env python3
+# Copyright      2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script loads a checkpoint (`pretrained.pt`) and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./zipformer/export_PromptASR.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --epoch 50 \
+  --avg 10
+
+Utterance level context biasing:
+
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --method modified_beam_search \
+  --use-pre-text True \
+  --content-prompt "bessy random words hello k2 ASR" \
+  --use-style-prompt True \
+  librispeech.flac
+
+
+Word level context biasing:
+
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --method modified_beam_search \
+  --use-pre-text True \
+  --content-prompt "The topic is about horses." \
+  --use-style-prompt True \
+  test.wav
+
+
+"""
+
+import argparse
+import logging
+import math
+import warnings
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import greedy_search_batch, modified_beam_search
+from text_normalization import _apply_style_transform, train_text_normalization
+from torch.nn.utils.rnn import pad_sequence
+from train_bert_encoder import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500_fallback_coverage_0.99/bpe.model",
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=True,
+        help="Use content prompt during decoding",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt during decoding",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--content-prompt", type=str, default="", help="The content prompt for decoding"
+    )
+
+    parser.add_argument(
+        "--style-prompt",
+        type=str,
+        default="Mixed-cased English text with punctuations, feel free to change it.",
+        help="The style prompt for decoding",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)  # for text encoder
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    assert (
+        len(params.sound_files) == 1
+    ), "Only support decoding one audio at this moment"
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    # encode prompts
+    if params.use_pre_text:
+        pre_text = [train_text_normalization(params.content_prompt)]
+        pre_text = _apply_style_transform(pre_text, params.pre_text_transform)
+    else:
+        pre_text = [""]
+
+    if params.use_style_prompt:
+        style_text = [params.style_prompt]
+        style_text = _apply_style_transform(style_text, params.style_text_transform)
+    else:
+        style_text = [""]
+
+    if params.use_pre_text or params.use_style_prompt:
+        encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+            pre_texts=pre_text,
+            style_texts=style_text,
+            tokenizer=tokenizer,
+            device=device,
+            no_limit=True,
+        )
+
+        memory, memory_key_padding_mask = model.encode_text(
+            encoded_inputs=encoded_inputs,
+            style_lens=style_lens,
+        )  # (T,B,C)
+    else:
+        memory = None
+        memory_key_padding_mask = None
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=features,
+            feature_lens=feature_lengths,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+
+    hyps = []
+    msg = f"Using {params.method}"
+    logging.info(msg)
+
+    if params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        hyps.append(sp.decode(hyp_tokens)[0])
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        hyps.append(sp.decode(hyp_tokens)[0])
+    else:
+        raise ValueError(f"Unsupported method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py b/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py
new file mode 100644
index 000000000..0e6764ba0
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py
@@ -0,0 +1,1872 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+import logging
+import math
+import random
+from functools import reduce
+from itertools import repeat
+from typing import Optional, Tuple, Union
+
+import k2
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from torch.cuda.amp import custom_bwd, custom_fwd
+from torch.nn import Embedding as ScaledEmbedding
+
+
+class PiecewiseLinear(object):
+    """
+    Piecewise linear function, from float to float, specified as nonempty list of (x,y) pairs with
+    the x values in order.  x values <[initial x] or >[final x] are map to [initial y], [final y]
+    respectively.
+    """
+
+    def __init__(self, *args):
+        assert len(args) >= 1
+        if len(args) == 1 and isinstance(args[0], PiecewiseLinear):
+            self.pairs = list(args[0].pairs)
+        else:
+            self.pairs = [(float(x), float(y)) for x, y in args]
+        for (x, y) in self.pairs:
+            assert isinstance(x, float) or isinstance(x, int)
+            assert isinstance(y, float) or isinstance(y, int)
+
+        for i in range(len(self.pairs) - 1):
+            assert self.pairs[i + 1][0] > self.pairs[i][0], self.pairs
+
+    def __str__(self):
+        # e.g. 'PiecewiseLinear((0., 10.), (100., 0.))'
+        return f"PiecewiseLinear({str(self.pairs)[1:-1]})"
+
+    def __call__(self, x):
+        if x <= self.pairs[0][0]:
+            return self.pairs[0][1]
+        elif x >= self.pairs[-1][0]:
+            return self.pairs[-1][1]
+        else:
+            cur_x, cur_y = self.pairs[0]
+            for i in range(1, len(self.pairs)):
+                next_x, next_y = self.pairs[i]
+                if x >= cur_x and x <= next_x:
+                    return cur_y + (next_y - cur_y) * (x - cur_x) / (next_x - cur_x)
+                cur_x, cur_y = next_x, next_y
+            assert False
+
+    def __mul__(self, alpha):
+        return PiecewiseLinear(*[(x, y * alpha) for x, y in self.pairs])
+
+    def __add__(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return PiecewiseLinear(*[(p[0], p[1] + x) for p in self.pairs])
+        s, x = self.get_common_basis(x)
+        return PiecewiseLinear(
+            *[(sp[0], sp[1] + xp[1]) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def max(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            x = PiecewiseLinear((0, x))
+        s, x = self.get_common_basis(x, include_crossings=True)
+        return PiecewiseLinear(
+            *[(sp[0], max(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def min(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            x = PiecewiseLinear((0, x))
+        s, x = self.get_common_basis(x, include_crossings=True)
+        return PiecewiseLinear(
+            *[(sp[0], min(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def __eq__(self, other):
+        return self.pairs == other.pairs
+
+    def get_common_basis(self, p: "PiecewiseLinear", include_crossings: bool = False):
+        """
+        Returns (self_mod, p_mod) which are equivalent piecewise lienar
+        functions to self and p, but with the same x values.
+
+          p: the other piecewise linear function
+          include_crossings: if true, include in the x values positions
+              where the functions indicate by this and p crosss.
+        """
+        assert isinstance(p, PiecewiseLinear)
+
+        # get sorted x-values without repetition.
+        x_vals = sorted(set([x for x, y in self.pairs] + [x for x, y in p.pairs]))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+
+        if include_crossings:
+            extra_x_vals = []
+            for i in range(len(x_vals) - 1):
+                if (y_vals1[i] > y_vals2[i]) != (y_vals1[i + 1] > y_vals2[i + 1]):
+                    # if the two lines in this subsegment potentially cross each other..
+                    diff_cur = abs(y_vals1[i] - y_vals2[i])
+                    diff_next = abs(y_vals1[i + 1] - y_vals2[i + 1])
+                    # `pos`, between 0 and 1, gives the relative x position,
+                    # with 0 being x_vals[i] and 1 being x_vals[i+1].
+                    pos = diff_cur / (diff_cur + diff_next)
+                    extra_x_val = x_vals[i] + pos * (x_vals[i + 1] - x_vals[i])
+                    extra_x_vals.append(extra_x_val)
+            if len(extra_x_vals) > 0:
+                x_vals = sorted(set(x_vals + extra_x_vals))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+        return (
+            PiecewiseLinear(*zip(x_vals, y_vals1)),
+            PiecewiseLinear(*zip(x_vals, y_vals2)),
+        )
+
+
+class ScheduledFloat(torch.nn.Module):
+    """
+    This object is a torch.nn.Module only because we want it to show up in [top_level module].modules();
+    it does not have a working forward() function.  You are supposed to cast it to float, as
+    in, float(parent_module.whatever), and use it as something like a dropout prob.
+
+    It is a floating point value whose value changes depending on the batch count of the
+    training loop.  It is a piecewise linear function where you specifiy the (x,y) pairs
+    in sorted order on x; x corresponds to the batch index.  For batch-index values before the
+    first x or after the last x, we just use the first or last y value.
+
+    Example:
+       self.dropout = ScheduledFloat((0.0, 0.2), (4000.0, 0.0), default=0.0)
+
+    `default` is used when self.batch_count is not set or in training or mode or in
+     torch.jit scripting mode.
+    """
+
+    def __init__(self, *args, default: float = 0.0):
+        super().__init__()
+        # self.batch_count and self.name will be written to in the training loop.
+        self.batch_count = None
+        self.name = None
+        self.default = default
+        self.schedule = PiecewiseLinear(*args)
+
+    def extra_repr(self) -> str:
+        return (
+            f"batch_count={self.batch_count}, schedule={str(self.schedule.pairs[1:-1])}"
+        )
+
+    def __float__(self):
+        batch_count = self.batch_count
+        if batch_count is None or not self.training or torch.jit.is_scripting():
+            return float(self.default)
+        else:
+            ans = self.schedule(self.batch_count)
+            if random.random() < 0.0002:
+                logging.info(
+                    f"ScheduledFloat: name={self.name}, batch_count={self.batch_count}, ans={ans}"
+                )
+            return ans
+
+    def __add__(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return ScheduledFloat(self.schedule + x, default=self.default)
+        else:
+            return ScheduledFloat(
+                self.schedule + x.schedule, default=self.default + x.default
+            )
+
+    def max(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return ScheduledFloat(self.schedule.max(x), default=self.default)
+        else:
+            return ScheduledFloat(
+                self.schedule.max(x.schedule), default=max(self.default, x.default)
+            )
+
+
+FloatLike = Union[float, ScheduledFloat]
+
+
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
+    """
+    A randomized way of casting a floating point value to half precision.
+    """
+    if x.dtype == torch.float16:
+        return x
+    x_abs = x.abs()
+    is_too_small = x_abs < min_abs
+    # for elements where is_too_small is true, random_val will contain +-min_abs with
+    # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
+    # for those elements].
+    random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs)
+    return torch.where(is_too_small, random_val, x).to(torch.float16)
+
+
+class CutoffEstimator:
+    """
+    Estimates cutoffs of an arbitrary numerical quantity such that a specified
+    proportion of items will be above the cutoff on average.
+
+      p is the proportion of items that should be above the cutoff.
+    """
+
+    def __init__(self, p: float):
+        self.p = p
+        # total count of items
+        self.count = 0
+        # total count of items that were above the cutoff
+        self.count_above = 0
+        # initial cutoff value
+        self.cutoff = 0
+
+    def __call__(self, x: float) -> bool:
+        """
+        Returns true if x is above the cutoff.
+        """
+        ans = x > self.cutoff
+        self.count += 1
+        if ans:
+            self.count_above += 1
+        cur_p = self.count_above / self.count
+        delta_p = cur_p - self.p
+        if (delta_p > 0) == ans:
+            q = abs(delta_p)
+            self.cutoff = x * q + self.cutoff * (1 - q)
+        return ans
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    """
+    Tries to handle half-precision derivatives in a randomized way that should
+    be more accurate for training than the default behavior.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, dim: int):
+        ans = x.softmax(dim=dim)
+        # if x dtype is float16, x.softmax() returns a float32 because
+        # (presumably) that op does not support float16, and autocast
+        # is enabled.
+        if torch.is_autocast_enabled():
+            ans = ans.to(torch.float16)
+        ctx.save_for_backward(ans)
+        ctx.x_dtype = x.dtype
+        ctx.dim = dim
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        (ans,) = ctx.saved_tensors
+        with torch.cuda.amp.autocast(enabled=False):
+            ans_grad = ans_grad.to(torch.float32)
+            ans = ans.to(torch.float32)
+            x_grad = ans_grad * ans
+            x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True)
+            return x_grad, None
+
+
+def softmax(x: Tensor, dim: int):
+    return SoftmaxFunction.apply(x, dim)
+
+
+class MaxEigLimiterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.channel_dim = channel_dim
+        ctx.grad_scale = grad_scale
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        with torch.enable_grad():
+            (x_orig, coeffs, new_direction) = ctx.saved_tensors
+            x_orig.requires_grad = True
+            num_channels = x_orig.shape[ctx.channel_dim]
+            x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
+            new_direction.requires_grad = False
+            x = x - x.mean(dim=0)
+            x_var = (x**2).mean()
+            x_residual = x - coeffs * new_direction
+            x_residual_var = (x_residual**2).mean()
+            # `variance_proportion` is the proportion of the variance accounted for
+            # by the top eigen-direction.  This is to be minimized.
+            variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+            variance_proportion.backward()
+        x_orig_grad = x_orig.grad
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
+        return x_grad + x_extra_grad.detach(), None, None, None, None
+
+
+class BiasNormFunction(torch.autograd.Function):
+    # This computes:
+    #   scales = (torch.mean((x - bias) ** 2, keepdim=True)) ** -0.5 * log_scale.exp()
+    #   return (x - bias) * scales
+    # (after unsqueezing the bias), but it does it in a memory-efficient way so that
+    # it can just store the returned value (chances are, this will also be needed for
+    # some other reason, related to the next operation, so we can save memory).
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        bias: Tensor,
+        log_scale: Tensor,
+        channel_dim: int,
+        store_output_for_backprop: bool,
+    ) -> Tensor:
+        assert bias.ndim == 1
+        if channel_dim < 0:
+            channel_dim = channel_dim + x.ndim
+        ctx.store_output_for_backprop = store_output_for_backprop
+        ctx.channel_dim = channel_dim
+        for _ in range(channel_dim + 1, x.ndim):
+            bias = bias.unsqueeze(-1)
+        scales = (
+            torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+        ) * log_scale.exp()
+        ans = x * scales
+        ctx.save_for_backward(
+            ans.detach() if store_output_for_backprop else x,
+            scales.detach(),
+            bias.detach(),
+            log_scale.detach(),
+        )
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tensor:
+        ans_or_x, scales, bias, log_scale = ctx.saved_tensors
+        if ctx.store_output_for_backprop:
+            x = ans_or_x / scales
+        else:
+            x = ans_or_x
+        x = x.detach()
+        x.requires_grad = True
+        bias.requires_grad = True
+        log_scale.requires_grad = True
+        with torch.enable_grad():
+            # recompute scales from x, bias and log_scale.
+            scales = (
+                torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) ** -0.5
+            ) * log_scale.exp()
+            ans = x * scales
+            ans.backward(gradient=ans_grad)
+        return x.grad, bias.grad.flatten(), log_scale.grad, None, None
+
+
+class BiasNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    Instead, we give the BiasNorm a trainable bias that it can use when
+    computing the scale for normalization.  We also give it a (scalar)
+    trainable scale on the output.
+
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+       channel_dim: the axis/dimension corresponding to the channel,
+         interprted as an offset from the input's ndim if negative.
+         shis is NOT the num_channels; it should typically be one of
+         {-2, -1, 0, 1, 2, 3}.
+      log_scale: the initial log-scale that we multiply the output by; this
+         is learnable.
+      log_scale_min: FloatLike, minimum allowed value of log_scale
+      log_scale_max: FloatLike, maximum allowed value of log_scale
+      store_output_for_backprop: only possibly affects memory use; recommend
+         to set to True if you think the output of this module is more likely
+         than the input of this module to be required to be stored for the
+         backprop.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        log_scale: float = 1.0,
+        log_scale_min: float = -1.5,
+        log_scale_max: float = 1.5,
+        store_output_for_backprop: bool = False,
+    ) -> None:
+        super(BiasNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.log_scale = nn.Parameter(torch.tensor(log_scale))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+
+        self.log_scale_min = log_scale_min
+        self.log_scale_max = log_scale_max
+
+        self.store_output_for_backprop = store_output_for_backprop
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+
+        if torch.jit.is_scripting():
+            channel_dim = self.channel_dim
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            bias = self.bias
+            for _ in range(channel_dim + 1, x.ndim):
+                bias = bias.unsqueeze(-1)
+            scales = (
+                torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+            ) * self.log_scale.exp()
+            return x * scales
+
+        log_scale = limit_param_value(
+            self.log_scale,
+            min=float(self.log_scale_min),
+            max=float(self.log_scale_max),
+            training=self.training,
+        )
+
+        return BiasNormFunction.apply(
+            x, self.bias, log_scale, self.channel_dim, self.store_output_for_backprop
+        )
+
+
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
+    """
+    Behaves like a constructor of a modified version of nn.Linear
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Linear(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv1d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv1d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv2d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv2d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv2d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False, but:
+    NO PADDING-RELATED ARGS.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv2d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+class ChunkCausalDepthwiseConv1d(torch.nn.Module):
+    """
+    Behaves like a depthwise 1d convolution, except that it is causal in
+    a chunkwise way, as if we had a block-triangular attention mask.
+    The chunk size is provided at test time (it should probably be
+    kept in sync with the attention mask).
+
+    This has a little more than twice the parameters of a conventional
+    depthwise conv1d module: we implement it by having one
+    depthwise convolution, of half the width, that is causal (via
+    right-padding); and one depthwise convolution that is applied only
+    within chunks, that we multiply by a scaling factor which depends
+    on the position within the chunk.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        initial_scale: float = 1.0,
+        bias: bool = True,
+    ):
+        super().__init__()
+        assert kernel_size % 2 == 1
+
+        half_kernel_size = (kernel_size + 1) // 2
+        # will pad manually, on one side.
+        self.causal_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=half_kernel_size,
+            padding=0,
+            bias=True,
+        )
+
+        self.chunkwise_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            bias=bias,
+        )
+
+        # first row is correction factors added to the scale near the left edge of the chunk,
+        # second row is correction factors added to the scale near the right edge of the chunk,
+        # both of these are added to a default scale of 1.0.
+        self.chunkwise_conv_scale = nn.Parameter(torch.zeros(2, channels, kernel_size))
+        self.kernel_size = kernel_size
+
+        with torch.no_grad():
+            self.causal_conv.weight[:] *= initial_scale
+            self.chunkwise_conv.weight[:] *= initial_scale
+            if bias:
+                torch.nn.init.uniform_(
+                    self.causal_conv.bias, -0.1 * initial_scale, 0.1 * initial_scale
+                )
+
+    def forward(self, x: Tensor, chunk_size: int = -1) -> Tensor:
+        """
+             Forward function.  Args:
+               x: a Tensor of shape (batch_size, channels, seq_len)
+        chunk_size: the chunk size, in frames; does not have to divide seq_len exactly.
+        """
+        (batch_size, num_channels, seq_len) = x.shape
+
+        half_kernel_size = self.kernel_size + 1 // 2
+        # left_pad is half_kernel_size - 1 where half_kernel_size is the size used
+        # in the causal conv.  It's the amount by which we must pad on the left,
+        # to make the convolution causal.
+        left_pad = self.kernel_size // 2
+
+        if chunk_size < 0 or chunk_size > seq_len:
+            chunk_size = seq_len
+        right_pad = -seq_len % chunk_size
+
+        x = torch.nn.functional.pad(x, (left_pad, right_pad))
+
+        x_causal = self.causal_conv(x[..., : left_pad + seq_len])
+        assert x_causal.shape == (batch_size, num_channels, seq_len)
+
+        x_chunk = x[..., left_pad:]
+        num_chunks = x_chunk.shape[2] // chunk_size
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks, chunk_size)
+        x_chunk = x_chunk.permute(0, 2, 1, 3).reshape(
+            batch_size * num_chunks, num_channels, chunk_size
+        )
+        x_chunk = self.chunkwise_conv(x_chunk)  # does not change shape
+
+        chunk_scale = self._get_chunk_scale(chunk_size)
+
+        x_chunk = x_chunk * chunk_scale
+        x_chunk = x_chunk.reshape(
+            batch_size, num_chunks, num_channels, chunk_size
+        ).permute(0, 2, 1, 3)
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks * chunk_size)[
+            ..., :seq_len
+        ]
+
+        return x_chunk + x_causal
+
+    def _get_chunk_scale(self, chunk_size: int):
+        """Returns tensor of shape (num_channels, chunk_size) that will be used to
+        scale the output of self.chunkwise_conv."""
+        left_edge = self.chunkwise_conv_scale[0]
+        right_edge = self.chunkwise_conv_scale[1]
+        if chunk_size < self.kernel_size:
+            left_edge = left_edge[:, :chunk_size]
+            right_edge = right_edge[:, -chunk_size:]
+        else:
+            t = chunk_size - self.kernel_size
+            channels = left_edge.shape[0]
+            pad = torch.zeros(
+                channels, t, device=left_edge.device, dtype=left_edge.dtype
+            )
+            left_edge = torch.cat((left_edge, pad), dim=-1)
+            right_edge = torch.cat((pad, right_edge), dim=-1)
+        return 1.0 + (left_edge + right_edge)
+
+
+class BalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        min_mean: float,
+        max_mean: float,
+        min_rms: float,
+        max_rms: float,
+        grad_scale: float,
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        ctx.save_for_backward(x)
+        ctx.config = (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None, None, None]:
+        (x,) = ctx.saved_tensors
+        (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim) = ctx.config
+
+        try:
+            with torch.enable_grad():
+                with torch.cuda.amp.autocast(enabled=False):
+                    x = x.to(torch.float32)
+                    x = x.detach()
+                    x.requires_grad = True
+                    mean_dims = [i for i in range(x.ndim) if i != channel_dim]
+                    uncentered_var = (x**2).mean(dim=mean_dims, keepdim=True)
+                    mean = x.mean(dim=mean_dims, keepdim=True)
+                    stddev = (uncentered_var - (mean * mean)).clamp(min=1.0e-20).sqrt()
+                    rms = uncentered_var.clamp(min=1.0e-20).sqrt()
+
+                    m = mean / stddev
+                    # part of loss that relates to mean / stddev
+                    m_loss = (m - m.clamp(min=min_mean, max=max_mean)).abs()
+
+                    # put a much larger scale on the RMS-max-limit loss, so that if both it and the
+                    # m_loss are violated we fix the RMS loss first.
+                    rms_clamped = rms.clamp(min=min_rms, max=max_rms)
+                    r_loss = (rms_clamped / rms).log().abs()
+
+                    loss = m_loss + r_loss
+
+                    loss.backward(gradient=torch.ones_like(loss))
+                    loss_grad = x.grad
+                    loss_grad_rms = (
+                        (loss_grad**2)
+                        .mean(dim=mean_dims, keepdim=True)
+                        .sqrt()
+                        .clamp(min=1.0e-20)
+                    )
+
+                    loss_grad = loss_grad * (grad_scale / loss_grad_rms)
+
+                    x_grad_float = x_grad.to(torch.float32)
+                    # scale each element of loss_grad by the absolute value of the corresponding
+                    # element of x_grad, which we view as a noisy estimate of its magnitude for that
+                    # (frame and dimension).  later we can consider factored versions.
+                    x_grad_mod = x_grad_float + (x_grad_float.abs() * loss_grad)
+                    x_grad = x_grad_mod.to(x_grad.dtype)
+        except Exception as e:
+            logging.info(
+                f"Caught exception in Balancer backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
+
+        return x_grad, None, None, None, None, None, None
+
+
+class Balancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+         prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: FloatLike = 0.05,
+        max_positive: FloatLike = 0.95,
+        min_abs: FloatLike = 0.2,
+        max_abs: FloatLike = 100.0,
+        grad_scale: FloatLike = 0.04,
+        prob: Optional[FloatLike] = None,
+    ):
+        super().__init__()
+
+        if prob is None:
+            prob = ScheduledFloat((0.0, 0.5), (8000.0, 0.125), default=0.4)
+        self.prob = prob
+        # 5% of the time we will return and do nothing because memory usage is
+        # too high.
+        self.mem_cutoff = CutoffEstimator(0.05)
+
+        # actually self.num_channels is no longer needed except for an assertion.
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.grad_scale = grad_scale
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))
+        ):
+            return _no_op(x)
+
+        prob = float(self.prob)
+        if random.random() < prob:
+            # The following inner-functions convert from the way we historically specified
+            # these limitations, as limits on the absolute value and the proportion of positive
+            # values, to limits on the RMS value and the (mean / stddev).
+            def _abs_to_rms(x):
+                # for normally distributed data, if the expected absolute value is x, the
+                # expected rms value will be sqrt(pi/2) * x.
+                return 1.25331413732 * x
+
+            def _proportion_positive_to_mean(x):
+                def _atanh(x):
+                    eps = 1.0e-10
+                    # eps is to prevent crashes if x is exactly 0 or 1.
+                    # we'll just end up returning a fairly large value.
+                    return (math.log(1 + x + eps) - math.log(1 - x + eps)) / 2.0
+
+                def _approx_inverse_erf(x):
+                    # 1 / (sqrt(pi) * ln(2)),
+                    # see https://math.stackexchange.com/questions/321569/approximating-the-error-function-erf-by-analytical-functions
+                    # this approximation is extremely crude and gets progressively worse for
+                    # x very close to -1 or +1, but we mostly care about the "middle" region
+                    # e.g. _approx_inverse_erf(0.05) = 0.0407316414078772,
+                    # and math.erf(0.0407316414078772) = 0.045935330944660666,
+                    # which is pretty close to 0.05.
+                    return 0.8139535143 * _atanh(x)
+
+                # first convert x from the range 0..1 to the range -1..1 which the error
+                # function returns
+                x = -1 + (2 * x)
+                return _approx_inverse_erf(x)
+
+            min_mean = _proportion_positive_to_mean(float(self.min_positive))
+            max_mean = _proportion_positive_to_mean(float(self.max_positive))
+            min_rms = _abs_to_rms(float(self.min_abs))
+            max_rms = _abs_to_rms(float(self.max_abs))
+            grad_scale = float(self.grad_scale)
+
+            assert x.shape[self.channel_dim] == self.num_channels
+
+            return BalancerFunction.apply(
+                x, min_mean, max_mean, min_rms, max_rms, grad_scale, self.channel_dim
+            )
+        else:
+            return _no_op(x)
+
+
+def penalize_abs_values_gt(
+    x: Tensor, limit: float, penalty: float, name: str = None
+) -> Tensor:
+    """
+    Returns x unmodified, but in backprop will put a penalty for the excess of
+    the absolute values of elements of x over the limit "limit".  E.g. if
+    limit == 10.0, then if x has any values over 10 it will get a penalty.
+
+    Caution: the value of this penalty will be affected by grad scaling used
+    in automatic mixed precision training.  For this reasons we use this,
+    it shouldn't really matter, or may even be helpful; we just use this
+    to disallow really implausible values of scores to be given to softmax.
+
+    The name is for randomly printed debug info.
+    """
+    x_sign = x.sign()
+    over_limit = (x.abs() - limit) > 0
+    # The following is a memory efficient way to penalize the absolute values of
+    # x that's over the limit.  (The memory efficiency comes when you think
+    # about which items torch needs to cache for the autograd, and which ones it
+    # can throw away).  The numerical value of aux_loss as computed here will
+    # actually be larger than it should be, by limit * over_limit.sum(), but it
+    # has the same derivative as the real aux_loss which is penalty * (x.abs() -
+    # limit).relu().
+    aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x)
+    # note: we don't do sum() here on aux)_loss, but it's as if we had done
+    # sum() due to how with_loss() works.
+    x = with_loss(x, aux_loss, name)
+    # you must use x for something, or this will be ineffective.
+    return x
+
+
+def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
+    if x.ndim == 2:
+        return x.diag()
+    else:
+        (batch, dim, dim) = x.shape
+        x = x.reshape(batch, dim * dim)
+        x = x[:, :: dim + 1]
+        assert x.shape == (batch, dim)
+        return x
+
+
+def _whitening_metric(x: Tensor, num_groups: int):
+    """
+    Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
+    of the centered feature covariance are the same within each group's covariance matrix
+    and also between groups.
+    Args:
+        x: a Tensor of shape (*, num_channels)
+     num_groups:  the number of groups of channels, a number >=1 that divides num_channels
+    Returns:
+        Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and
+    greater than 1.0 otherwise.
+    """
+    assert x.dtype != torch.float16
+    x = x.reshape(-1, x.shape[-1])
+    (num_frames, num_channels) = x.shape
+    assert num_channels % num_groups == 0
+    channels_per_group = num_channels // num_groups
+    x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1)
+    # x now has shape (num_groups, num_frames, channels_per_group)
+    # subtract the mean so we use the centered, not uncentered, covariance.
+    # My experience has been that when we "mess with the gradients" like this,
+    # it's better not do anything that tries to move the mean around, because
+    # that can easily cause instability.
+    x = x - x.mean(dim=1, keepdim=True)
+    # x_covar: (num_groups, channels_per_group, channels_per_group)
+    x_covar = torch.matmul(x.transpose(1, 2), x)
+    x_covar_mean_diag = _diag(x_covar).mean()
+    # the following expression is what we'd get if we took the matrix product
+    # of each covariance and measured the mean of its trace, i.e.
+    # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
+    x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group)
+    # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20)
+    return metric
+
+
+class WhiteningPenaltyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, module: nn.Module) -> Tensor:
+        ctx.save_for_backward(x)
+        ctx.module = module
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
+        w = ctx.module
+
+        try:
+            with torch.enable_grad():
+                with torch.cuda.amp.autocast(enabled=False):
+                    x_detached = x_orig.to(torch.float32).detach()
+                    x_detached.requires_grad = True
+
+                    metric = _whitening_metric(x_detached, w.num_groups)
+
+                    if random.random() < 0.005 or __name__ == "__main__":
+                        logging.info(
+                            f"Whitening: name={w.name}, num_groups={w.num_groups}, num_channels={x_orig.shape[-1]}, "
+                            f"metric={metric.item():.2f} vs. limit={float(w.whitening_limit)}"
+                        )
+
+                    if metric < float(w.whitening_limit):
+                        w.prob = w.min_prob
+                        return x_grad, None
+                    else:
+                        w.prob = w.max_prob
+                        metric.backward()
+                        penalty_grad = x_detached.grad
+                        scale = w.grad_scale * (
+                            x_grad.to(torch.float32).norm()
+                            / (penalty_grad.norm() + 1.0e-20)
+                        )
+                        penalty_grad = penalty_grad * scale
+                        return x_grad + penalty_grad.to(x_grad.dtype), None
+        except Exception as e:
+            logging.info(
+                f"Caught exception in Whiten backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
+        return x_grad, None
+
+
+class Whiten(nn.Module):
+    def __init__(
+        self,
+        num_groups: int,
+        whitening_limit: FloatLike,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: FloatLike,
+    ):
+        """
+        Args:
+          num_groups: the number of groups to divide the channel dim into before
+            whitening.  We will attempt to make the feature covariance
+            within each group, after mean subtraction, as "white" as possible,
+            while having the same trace across all groups.
+         whitening_limit: a value greater than 1.0, that dictates how much
+           freedom we have to violate the constraints.  1.0 would mean perfectly
+           white, with exactly the same trace across groups; larger values
+           give more freedom.  E.g. 2.0.
+         prob: the probability with which we apply the gradient modification
+           (also affects the grad scale).  May be supplied as a float,
+           or as a pair (min_prob, max_prob)
+
+          grad_scale: determines the scale on the gradient term from this object,
+            relative to the rest of the gradient on the attention weights.
+            E.g. 0.02 (you may want to use smaller values than this if prob is large)
+        """
+        super(Whiten, self).__init__()
+        assert num_groups >= 1
+        assert float(whitening_limit) >= 1
+        assert grad_scale >= 0
+        self.num_groups = num_groups
+        self.whitening_limit = whitening_limit
+        self.grad_scale = grad_scale
+
+        if isinstance(prob, float):
+            prob = (prob, prob)
+        (self.min_prob, self.max_prob) = prob
+        assert 0 < self.min_prob <= self.max_prob <= 1
+        self.prob = self.max_prob
+        self.name = None  # will be set in training loop
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        In the forward pass, this function just returns the input unmodified.
+        In the backward pass, it will modify the gradients to ensure that the
+        distribution in each group has close to (lambda times I) as the covariance
+        after mean subtraction, with the same lambda across groups.
+        For whitening_limit > 1, there will be more freedom to violate this
+        constraint.
+
+        Args:
+           x: the input of shape (*, num_channels)
+
+        Returns:
+            x, unmodified.   You should make sure
+        you use the returned value, or the graph will be freed
+        and nothing will happen in backprop.
+        """
+        grad_scale = float(self.grad_scale)
+        if not x.requires_grad or random.random() > self.prob or grad_scale == 0:
+            return _no_op(x)
+        else:
+            return WhiteningPenaltyFunction.apply(x, self)
+
+
+class WithLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, y: Tensor, name: str):
+        ctx.y_shape = y.shape
+        if random.random() < 0.002 and name is not None:
+            loss_sum = y.sum().item()
+            logging.info(f"WithLoss: name={name}, loss-sum={loss_sum:.3e}")
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        return (
+            ans_grad,
+            torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
+            None,
+        )
+
+
+def with_loss(x, y, name):
+    # returns x but adds y.sum() to the loss function.
+    return WithLoss.apply(x, y, name)
+
+
+class ScaleGradFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, alpha: float) -> Tensor:
+        ctx.alpha = alpha
+        return x
+
+    @staticmethod
+    def backward(ctx, grad: Tensor):
+        return grad * ctx.alpha, None
+
+
+def scale_grad(x: Tensor, alpha: float):
+    return ScaleGradFunction.apply(x, alpha)
+
+
+class ScaleGrad(nn.Module):
+    def __init__(self, alpha: float):
+        super().__init__()
+        self.alpha = alpha
+
+    def forward(self, x: Tensor) -> Tensor:
+        return scale_grad(x, self.alpha)
+
+
+class LimitParamValue(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, min: float, max: float):
+        ctx.save_for_backward(x)
+        assert max >= min
+        ctx.min = min
+        ctx.max = max
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x,) = ctx.saved_tensors
+        # where x < ctx.min, ensure all grads are negative (this will tend to make
+        # x more positive).
+        x_grad = x_grad * torch.where(
+            torch.logical_and(x_grad > 0, x < ctx.min), -1.0, 1.0
+        )
+        # where x > ctx.max, ensure all grads are positive (this will tend to make
+        # x more negative).
+        x_grad *= torch.where(torch.logical_and(x_grad < 0, x > ctx.max), -1.0, 1.0)
+        return x_grad, None, None
+
+
+def limit_param_value(
+    x: Tensor, min: float, max: float, prob: float = 0.6, training: bool = True
+):
+    # You apply this to (typically) an nn.Parameter during training to ensure that its
+    # (elements mostly) stays within a supplied range.  This is done by modifying the
+    # gradients in backprop.
+    # It's not necessary to do this on every batch: do it only some of the time,
+    # to save a little time.
+    if training and random.random() < prob:
+        return LimitParamValue.apply(x, min, max)
+    else:
+        return x
+
+
+def _no_op(x: Tensor) -> Tensor:
+    if torch.jit.is_scripting():
+        return x
+    else:
+        # a no-op function that will have a node in the autograd graph,
+        # to avoid certain bugs relating to backward hooks
+        return x.chunk(1, dim=-1)[0]
+
+
+class Identity(torch.nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, x):
+        return _no_op(x)
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.044 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.044
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+# Dropout2 is just like normal dropout, except it supports schedules on the dropout rates.
+class Dropout2(nn.Module):
+    def __init__(self, p: FloatLike):
+        super().__init__()
+        self.p = p
+
+    def forward(self, x: Tensor) -> Tensor:
+        return torch.nn.functional.dropout(x, p=float(self.p), training=self.training)
+
+
+class MulForDropout3(torch.autograd.Function):
+    # returns (x * y * alpha) where alpha is a float and y doesn't require
+    # grad and is zero-or-one.
+    @staticmethod
+    @custom_fwd
+    def forward(ctx, x, y, alpha):
+        assert not y.requires_grad
+        ans = x * y * alpha
+        ctx.save_for_backward(ans)
+        ctx.alpha = alpha
+        return ans
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad):
+        (ans,) = ctx.saved_tensors
+        x_grad = ctx.alpha * ans_grad * (ans != 0)
+        return x_grad, None, None
+
+
+# Dropout3 is just like normal dropout, except it supports schedules on the dropout rates,
+# and it lets you choose one dimension to share the dropout mask over
+class Dropout3(nn.Module):
+    def __init__(self, p: FloatLike, shared_dim: int):
+        super().__init__()
+        self.p = p
+        self.shared_dim = shared_dim
+
+    def forward(self, x: Tensor) -> Tensor:
+        p = float(self.p)
+        if not self.training or p == 0:
+            return _no_op(x)
+        scale = 1.0 / (1 - p)
+        rand_shape = list(x.shape)
+        rand_shape[self.shared_dim] = 1
+        mask = torch.rand(*rand_shape, device=x.device) > p
+        ans = MulForDropout3.apply(x, mask, scale)
+        return ans
+
+
+class SwooshLFunction(torch.autograd.Function):
+    """
+    swoosh(x) =  log(1 + exp(x-4)) - 0.08*x - 0.035
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+
+        coeff = -0.08
+
+        with torch.cuda.amp.autocast(enabled=False):
+            with torch.enable_grad():
+                x = x.detach()
+                x.requires_grad = True
+                y = torch.logaddexp(zero, x - 4.0) + coeff * x - 0.035
+
+                if not requires_grad:
+                    return y
+                y.backward(gradient=torch.ones_like(y))
+
+                grad = x.grad
+                floor = coeff
+                ceil = 1.0 + coeff + 0.005
+
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
+                if __name__ == "__main__":
+                    # for self-testing only.
+                    assert d_scaled.min() >= 0.0
+                    assert d_scaled.max() < 256.0
+
+                d_int = d_scaled.to(torch.uint8)
+                ctx.save_for_backward(d_int)
+                if x.dtype == torch.float16 or torch.is_autocast_enabled():
+                    y = y.to(torch.float16)
+                return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+
+        coeff = -0.08
+        floor = coeff
+        ceil = 1.0 + coeff + 0.005
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class SwooshL(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-L activation."""
+        if torch.jit.is_scripting():
+            zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+            return torch.logaddexp(zero, x - 4.0) - 0.08 * x - 0.035
+        if not x.requires_grad:
+            return k2.swoosh_l_forward(x)
+        else:
+            return k2.swoosh_l(x)
+        # return SwooshLFunction.apply(x)
+
+
+class SwooshRFunction(torch.autograd.Function):
+    """
+     swoosh(x) =  log(1 + exp(x-1)) - 0.08*x - 0.313261687
+
+    derivatives are between -0.08 and 0.92.
+
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            with torch.enable_grad():
+                x = x.detach()
+                x.requires_grad = True
+                y = torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
+
+                if not requires_grad:
+                    return y
+                y.backward(gradient=torch.ones_like(y))
+
+                grad = x.grad
+                floor = -0.08
+                ceil = 0.925
+
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
+                if __name__ == "__main__":
+                    # for self-testing only.
+                    assert d_scaled.min() >= 0.0
+                    assert d_scaled.max() < 256.0
+                d_int = d_scaled.to(torch.uint8)
+                ctx.save_for_backward(d_int)
+                if x.dtype == torch.float16 or torch.is_autocast_enabled():
+                    y = y.to(torch.float16)
+                return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.08
+        ceil = 0.925
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class SwooshR(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-R activation."""
+        if torch.jit.is_scripting():
+            zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+            return torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
+        if not x.requires_grad:
+            return k2.swoosh_r_forward(x)
+        else:
+            return k2.swoosh_r(x)
+        # return SwooshRFunction.apply(x)
+
+
+# simple version of SwooshL that does not redefine the backprop, used in
+# ActivationDropoutAndLinearFunction.
+def SwooshLForward(x: Tensor):
+    x_offset = x - 4.0
+    log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
+    return log_sum - 0.08 * x - 0.035
+
+
+# simple version of SwooshR that does not redefine the backprop, used in
+# ActivationDropoutAndLinearFunction.
+def SwooshRForward(x: Tensor):
+    x_offset = x - 1.0
+    log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
+    return log_sum - 0.08 * x - 0.313261687
+
+
+class ActivationDropoutAndLinearFunction(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(
+        ctx,
+        x: Tensor,
+        weight: Tensor,
+        bias: Optional[Tensor],
+        activation: str,
+        dropout_p: float,
+        dropout_shared_dim: Optional[int],
+    ):
+        if dropout_p != 0.0:
+            dropout_shape = list(x.shape)
+            if dropout_shared_dim is not None:
+                dropout_shape[dropout_shared_dim] = 1
+            # else it won't be very memory efficient.
+            dropout_mask = (1.0 / (1.0 - dropout_p)) * (
+                torch.rand(*dropout_shape, device=x.device, dtype=x.dtype) > dropout_p
+            )
+        else:
+            dropout_mask = None
+
+        ctx.save_for_backward(x, weight, bias, dropout_mask)
+
+        ctx.activation = activation
+
+        forward_activation_dict = {
+            "SwooshL": k2.swoosh_l_forward,
+            "SwooshR": k2.swoosh_r_forward,
+        }
+        # it will raise a KeyError if this fails.  This will be an error.  We let it
+        # propagate to the user.
+        activation_func = forward_activation_dict[activation]
+        x = activation_func(x)
+        if dropout_mask is not None:
+            x = x * dropout_mask
+        x = torch.nn.functional.linear(x, weight, bias)
+        return x
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad: Tensor):
+        saved = ctx.saved_tensors
+        (x, weight, bias, dropout_mask) = saved
+
+        forward_and_deriv_activation_dict = {
+            "SwooshL": k2.swoosh_l_forward_and_deriv,
+            "SwooshR": k2.swoosh_r_forward_and_deriv,
+        }
+        # the following lines a KeyError if the activation is unrecognized.
+        # This will be an error.  We let it propagate to the user.
+        func = forward_and_deriv_activation_dict[ctx.activation]
+
+        y, func_deriv = func(x)
+        if dropout_mask is not None:
+            y = y * dropout_mask
+        # now compute derivative of y w.r.t. weight and bias..
+        # y: (..., in_channels), ans_grad: (..., out_channels),
+        (out_channels, in_channels) = weight.shape
+
+        in_channels = y.shape[-1]
+        g = ans_grad.reshape(-1, out_channels)
+        weight_deriv = torch.matmul(g.t(), y.reshape(-1, in_channels))
+        y_deriv = torch.matmul(ans_grad, weight)
+        bias_deriv = None if bias is None else g.sum(dim=0)
+        x_deriv = y_deriv * func_deriv
+        if dropout_mask is not None:
+            # order versus func_deriv does not matter
+            x_deriv = x_deriv * dropout_mask
+
+        return x_deriv, weight_deriv, bias_deriv, None, None, None
+
+
+class ActivationDropoutAndLinear(torch.nn.Module):
+    """
+     This merges an activation function followed by dropout and then a nn.Linear module;
+     it does so in a memory efficient way so that it only stores the input to the whole
+     module.  If activation == SwooshL and dropout_shared_dim != None, this will be
+     equivalent to:
+       nn.Sequential(SwooshL(),
+                     Dropout3(dropout_p, shared_dim=dropout_shared_dim),
+                     ScaledLinear(in_channels, out_channels, bias=bias,
+                                  initial_scale=initial_scale))
+    If dropout_shared_dim is None, the dropout would be equivalent to
+    Dropout2(dropout_p).  Note: Dropout3 will be more memory efficient as the dropout
+    mask is smaller.
+
+     Args:
+        in_channels: number of input channels, e.g. 256
+        out_channels: number of output channels, e.g. 256
+        bias: if true, have a bias
+        activation: the activation function, for now just support SwooshL.
+        dropout_p: the dropout probability or schedule (happens after nonlinearity).
+        dropout_shared_dim: the dimension, if any, across which the dropout mask is
+             shared (e.g. the time dimension).  If None, this may be less memory
+             efficient if there are modules before this one that cache the input
+             for their backprop (e.g. Balancer or Whiten).
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        bias: bool = True,
+        activation: str = "SwooshL",
+        dropout_p: FloatLike = 0.0,
+        dropout_shared_dim: Optional[int] = -1,
+        initial_scale: float = 1.0,
+    ):
+        super().__init__()
+        # create a temporary module of nn.Linear that we'll steal the
+        # weights and bias from
+        layer = ScaledLinear(
+            in_channels, out_channels, bias=bias, initial_scale=initial_scale
+        )
+
+        self.weight = layer.weight
+        # register_parameter properly handles making it a parameter when l.bias
+        # is None. I think there is some reason for doing it this way rather
+        # than just setting it to None but I don't know what it is, maybe
+        # something to do with exporting the module..
+        self.register_parameter("bias", layer.bias)
+
+        self.activation = activation
+        self.dropout_p = dropout_p
+        self.dropout_shared_dim = dropout_shared_dim
+
+    def forward(self, x: Tensor):
+        if torch.jit.is_scripting():
+            if self.activation == "SwooshL":
+                x = SwooshLForward(x)
+            elif self.activation == "SwooshR":
+                x = SwooshRForward(x)
+            else:
+                assert False, self.activation
+            return torch.nn.functional.linear(x, self.weight, self.bias)
+
+        return ActivationDropoutAndLinearFunction.apply(
+            x,
+            self.weight,
+            self.bias,
+            self.activation,
+            float(self.dropout_p),
+            self.dropout_shared_dim,
+        )
+
+
+class ClipGradFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, limit: float):
+        ctx.limit = limit
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        return x_grad.clamp(-ctx.limit, ctx.limit), None
+
+
+def clip_grad(x: Tensor, limit: float):
+    return ClipGradFunction.apply(x, limit)
+
+
+class AbsValuePenalizer(nn.Module):
+    """
+    This module adds a penalty to the loss function when ever the absolute value of
+    any element of the input tensor exceeds a certain limit.
+    """
+
+    def __init__(self, limit: float, prob: float = 0.1, penalty: float = 1.0e-04):
+        super().__init__()
+        self.limit = limit
+        self.penalty = penalty
+
+        self.prob = prob
+        self.name = None  # will be set in training loop
+
+        # 20% of the time we will return and do nothing because memory usage is
+        # too high.
+        self.mem_cutoff = CutoffEstimator(0.2)
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or not self.training
+            or random.random() > self.prob
+        ):
+            # or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))
+            return _no_op(x)  # the _no_op op is to make our diagnostics code work.
+
+        x = penalize_abs_values_gt(
+            x, limit=self.limit, penalty=self.penalty, name=self.name
+        )
+        return x
+
+
+def convert_num_channels(x: Tensor, num_channels: int) -> Tensor:
+    if num_channels <= x.shape[-1]:
+        return x[..., :num_channels]
+    else:
+        shape = list(x.shape)
+        shape[-1] = num_channels - shape[-1]
+        zeros = torch.zeros(*shape, dtype=x.dtype, device=x.device)
+        return torch.cat((x, zeros), dim=-1)
+
+
+def _test_whiten():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"_test_whiten(): proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0)
+    x = x.detach()
+    x.requires_grad = True
+    m = Balancer(
+        probs.numel(),
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        min_abs=0.0,
+        prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_balancer_sign: x = ", x)
+    print("_test_balancer_sign: y grad = ", y_grad)
+    print("_test_balancer_sign: x grad = ", x.grad)
+
+
+def _test_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = Balancer(
+        magnitudes.numel(),
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        min_abs=0.2,
+        max_abs=0.7,
+        prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_balancer_magnitude: x = ", x)
+    print("_test_balancer_magnitude: y grad = ", y_grad)
+    print("_test_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = DoubleSwish()
+
+    tol = (1.2 - (-0.043637)) / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_swooshl_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = SwooshL()
+
+    tol = 1.0 / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_swooshr_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = SwooshR()
+
+    tol = 1.0 / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_softmax():
+    a = torch.randn(2, 10, dtype=torch.float64)
+    b = a.clone()
+    a.requires_grad = True
+    b.requires_grad = True
+    a.softmax(dim=1)[:, 0].sum().backward()
+    print("a grad = ", a.grad)
+    softmax(b, dim=1)[:, 0].sum().backward()
+    print("b grad = ", b.grad)
+    assert torch.allclose(a.grad, b.grad)
+
+
+def _test_piecewise_linear():
+    p = PiecewiseLinear((0, 10.0))
+    for x in [-100, 0, 100]:
+        assert p(x) == 10.0
+    p = PiecewiseLinear((0, 10.0), (1, 0.0))
+    for x, y in [(-100, 10.0), (0, 10.0), (0.5, 5.0), (1, 0.0), (2, 0.0)]:
+        print("x, y = ", x, y)
+        assert p(x) == y, (x, p(x), y)
+
+    q = PiecewiseLinear((0.5, 15.0), (0.6, 1.0))
+    x_vals = [-1.0, 0.0, 0.1, 0.2, 0.5, 0.6, 0.7, 0.9, 1.0, 2.0]
+    pq = p.max(q)
+    for x in x_vals:
+        y1 = max(p(x), q(x))
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+    pq = p.min(q)
+    for x in x_vals:
+        y1 = min(p(x), q(x))
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+    pq = p + q
+    for x in x_vals:
+        y1 = p(x) + q(x)
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+
+
+def _test_activation_dropout_and_linear():
+    in_channels = 20
+    out_channels = 30
+
+    for bias in [True, False]:
+        # actually we don't test for dropout_p != 0.0 because forward functions will give
+        # different answers.  This is because we are using the k2 implementation of
+        # swoosh_l an swoosh_r inside SwooshL() and SwooshR(), and they call randn()
+        # internally, messing up the random state.
+        for dropout_p in [0.0]:
+            for activation in ["SwooshL", "SwooshR"]:
+                m1 = nn.Sequential(
+                    SwooshL() if activation == "SwooshL" else SwooshR(),
+                    Dropout3(p=dropout_p, shared_dim=-1),
+                    ScaledLinear(
+                        in_channels, out_channels, bias=bias, initial_scale=0.5
+                    ),
+                )
+                m2 = ActivationDropoutAndLinear(
+                    in_channels,
+                    out_channels,
+                    bias=bias,
+                    initial_scale=0.5,
+                    activation=activation,
+                    dropout_p=dropout_p,
+                )
+                with torch.no_grad():
+                    m2.weight[:] = m1[2].weight
+                    if bias:
+                        m2.bias[:] = m1[2].bias
+                # make sure forward gives same result.
+                x1 = torch.randn(10, in_channels)
+                x1.requires_grad = True
+
+                # TEMP.
+                assert torch.allclose(
+                    SwooshRFunction.apply(x1), SwooshRForward(x1), atol=1.0e-03
+                )
+
+                x2 = x1.clone().detach()
+                x2.requires_grad = True
+                seed = 10
+                torch.manual_seed(seed)
+                y1 = m1(x1)
+                y_grad = torch.randn_like(y1)
+                y1.backward(gradient=y_grad)
+                torch.manual_seed(seed)
+                y2 = m2(x2)
+                y2.backward(gradient=y_grad)
+
+                print(
+                    f"bias = {bias}, dropout_p = {dropout_p}, activation = {activation}"
+                )
+                print("y1 = ", y1)
+                print("y2 = ", y2)
+                assert torch.allclose(y1, y2, atol=0.02)
+                assert torch.allclose(m1[2].weight.grad, m2.weight.grad, atol=1.0e-05)
+                if bias:
+                    assert torch.allclose(m1[2].bias.grad, m2.bias.grad, atol=1.0e-05)
+                print("x1.grad = ", x1.grad)
+                print("x2.grad = ", x2.grad)
+
+                def isclose(a, b):
+                    # return true if cosine similarity is > 0.9.
+                    return (a * b).sum() > 0.9 * (
+                        (a**2).sum() * (b**2).sum()
+                    ).sqrt()
+
+                # the SwooshL() implementation has a noisy gradient due to 1-byte
+                # storage of it.
+                assert isclose(x1.grad, x2.grad)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_piecewise_linear()
+    _test_softmax()
+    _test_whiten()
+    _test_balancer_sign()
+    _test_balancer_magnitude()
+    _test_swooshl_deriv()
+    _test_swooshr_deriv()
+    _test_activation_dropout_and_linear()
+    _test_double_swish_deriv()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py b/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py
new file mode 100644
index 000000000..7acbc1808
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py
@@ -0,0 +1,276 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import Tuple
+
+import torch
+from scaling import (
+    Balancer,
+    BiasNorm,
+    Dropout3,
+    FloatLike,
+    Optional,
+    ScaledConv2d,
+    ScaleGrad,
+    ScheduledFloat,
+    SwooshL,
+    SwooshR,
+    Whiten,
+)
+from torch import Tensor, nn
+
+
+class ConvNeXt(nn.Module):
+    """
+    Our interpretation of the ConvNeXt module as used in https://arxiv.org/pdf/2206.14747.pdf
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        hidden_ratio: int = 3,
+        kernel_size: Tuple[int, int] = (7, 7),
+        layerdrop_rate: FloatLike = None,
+    ):
+        super().__init__()
+        padding = ((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2)
+        hidden_channels = channels * hidden_ratio
+        if layerdrop_rate is None:
+            layerdrop_rate = ScheduledFloat((0.0, 0.2), (20000.0, 0.015))
+        self.layerdrop_rate = layerdrop_rate
+
+        self.depthwise_conv = nn.Conv2d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=padding,
+        )
+
+        self.pointwise_conv1 = nn.Conv2d(
+            in_channels=channels, out_channels=hidden_channels, kernel_size=1
+        )
+
+        self.hidden_balancer = Balancer(
+            hidden_channels,
+            channel_dim=1,
+            min_positive=0.3,
+            max_positive=1.0,
+            min_abs=0.75,
+            max_abs=5.0,
+        )
+
+        self.activation = SwooshL()
+        self.pointwise_conv2 = ScaledConv2d(
+            in_channels=hidden_channels,
+            out_channels=channels,
+            kernel_size=1,
+            initial_scale=0.01,
+        )
+
+        self.out_balancer = Balancer(
+            channels,
+            channel_dim=1,
+            min_positive=0.4,
+            max_positive=0.6,
+            min_abs=1.0,
+            max_abs=6.0,
+        )
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=5.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or not self.training:
+            return self.forward_internal(x)
+        layerdrop_rate = float(self.layerdrop_rate)
+
+        if layerdrop_rate != 0.0:
+            batch_size = x.shape[0]
+            mask = (
+                torch.rand((batch_size, 1, 1, 1), dtype=x.dtype, device=x.device)
+                > layerdrop_rate
+            )
+        else:
+            mask = None
+        # turns out this caching idea does not work with --world-size > 1
+        # return caching_eval(self.forward_internal, x, mask)
+        return self.forward_internal(x, mask)
+
+    def forward_internal(
+        self, x: Tensor, layer_skip_mask: Optional[Tensor] = None
+    ) -> Tensor:
+        """
+        x layout: (N, C, H, W), i.e. (batch_size, num_channels, num_frames, num_freqs)
+
+        The returned value has the same shape as x.
+        """
+        bypass = x
+        x = self.depthwise_conv(x)
+        x = self.pointwise_conv1(x)
+        x = self.hidden_balancer(x)
+        x = self.activation(x)
+        x = self.pointwise_conv2(x)
+
+        if layer_skip_mask is not None:
+            x = x * layer_skip_mask
+
+        x = bypass + x
+        x = self.out_balancer(x)
+        x = x.transpose(1, 3)  # (N, W, H, C); need channel dim to be last
+        x = self.out_whiten(x)
+        x = x.transpose(1, 3)  # (N, C, H, W)
+
+        return x
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/2 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = (T-3)//2 - 2 == (T-7)//2
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        dropout: FloatLike = 0.1,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, (T-3)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+          bottleneck:
+            bottleneck dimension for 1d squeeze-excite
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        # The ScaleGrad module is there to prevent the gradients
+        # w.r.t. the weight or bias of the first Conv2d module in self.conv from
+        # exceeding the range of fp16 when using automatic mixed precision (amp)
+        # training.  (The second one is necessary to stop its bias from getting
+        # a too-large gradient).
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=(0, 1),  # (time, freq)
+            ),
+            ScaleGrad(0.2),
+            Balancer(layer1_channels, channel_dim=1, max_abs=1.0),
+            SwooshR(),
+            nn.Conv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+                padding=0,
+            ),
+            Balancer(layer2_channels, channel_dim=1, max_abs=4.0),
+            SwooshR(),
+            nn.Conv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=(1, 2),  # (time, freq)
+            ),
+            Balancer(layer3_channels, channel_dim=1, max_abs=4.0),
+            SwooshR(),
+        )
+
+        # just one convnext layer
+        self.convnext = ConvNeXt(layer3_channels, kernel_size=(7, 7))
+
+        out_width = (((in_channels - 1) // 2) - 1) // 2
+
+        self.out = nn.Linear(out_width * layer3_channels, out_channels)
+        # use a larger than normal grad_scale on this whitening module; there is
+        # only one such module, so there is not a concern about adding together
+        # many copies of this extra gradient term.
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=ScheduledFloat((0.0, 4.0), (20000.0, 8.0), default=4.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.02,
+        )
+
+        # max_log_eps=0.0 is to prevent both eps and the output of self.out from
+        # getting large, there is an unnecessary degree of freedom.
+        self.out_norm = BiasNorm(out_channels)
+        self.dropout = Dropout3(dropout, shared_dim=1)
+
+    def forward(self, x: torch.Tensor, x_lens: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+
+        Returns:
+          - a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+          - output lengths, of shape (batch_size,)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        # scaling x by 0.1 allows us to use a larger grad-scale in fp16 "amp" (automatic mixed precision)
+        # training, since the weights in the first convolution are otherwise the limiting factor for getting infinite
+        # gradients.
+        x = self.conv(x)
+        x = self.convnext(x)
+
+        # Now x is of shape (N, odim, ((T-3)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+
+        x = x.transpose(1, 2).reshape(b, t, c * f)
+        # now x: (N, ((T-1)//2 - 1))//2, out_width * layer3_channels))
+
+        x = self.out(x)
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_whiten(x)
+        x = self.out_norm(x)
+        x = self.dropout(x)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            x_lens = (x_lens - 7) // 2
+        assert x.size(1) == x_lens.max().item()
+
+        return x, x_lens
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py b/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py
new file mode 100755
index 000000000..13483637d
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless4/test_model.py
+"""
+
+from scaling import ScheduledFloat
+from train_subformer import get_params, get_text_encoder, get_transducer_model
+from zipformer import Zipformer2
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,15,15"
+
+    params.text_encoder_dim = (192, 192, 256, 384)
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = Zipformer2(
+        output_downsampling_factor=8,
+        downsampling_factor=(1, 2, 4, 8),
+        num_encoder_layers=(2, 4, 4, 4),
+        encoder_dim=(192, 192, 256, 384),
+        encoder_unmasked_dim=(192, 192, 256, 256),
+        query_head_dim=(32, 32, 32, 32),
+        pos_head_dim=(4, 4, 4, 4),
+        value_head_dim=(12, 12, 12, 12),
+        pos_dim=48,
+        num_heads=(4, 4, 4, 8),
+        feedforward_dim=(
+            384,
+            512,
+            768,
+            1024,
+        ),  # could increase this if there is nough data
+        cnn_module_kernel=(31, 31, 15, 15),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=False,
+    )
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    model = Zipformer2(
+        output_downsampling_factor=8,
+        downsampling_factor=(1, 2, 4, 8),
+        num_encoder_layers=(2, 4, 6, 6),
+        encoder_dim=(256, 256, 384, 512),
+        encoder_unmasked_dim=(196, 196, 256, 256),
+        query_head_dim=(32, 32, 32, 32),
+        pos_head_dim=(4, 4, 4, 4),
+        value_head_dim=(12, 12, 12, 12),
+        pos_dim=48,
+        num_heads=(4, 4, 4, 8),
+        feedforward_dim=(
+            384,
+            512,
+            768,
+            1024,
+        ),  # could increase this if there is nough data
+        cnn_module_kernel=(31, 31, 15, 15),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=False,
+    )
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py b/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py
new file mode 100644
index 000000000..efb4acc3c
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py
@@ -0,0 +1,101 @@
+#      Copyright      2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from typing import List
+
+
+def train_text_normalization(s: str) -> str:
+    # replace full-width with half-width
+    s = s.replace("“", '"')
+    s = s.replace("”", '"')
+    s = s.replace("‘", "'")
+    s = s.replace("’", "'")
+    if s[:2] == '" ':  # remove the starting double quote
+        s = s[2:]
+
+    return s
+
+
+def ref_text_normalization(ref_text: str) -> str:
+    # Rule 1: Remove the [FN#[]]
+    p = r"[FN#[0-9]*]"
+    pattern = re.compile(p)
+
+    res = pattern.findall(ref_text)
+    ref_text = re.sub(p, "", ref_text)
+
+    ref_text = train_text_normalization(ref_text)
+
+    return ref_text
+
+
+def remove_non_alphabetic(text: str, strict: bool = True) -> str:
+    # Recommend to set strict to False
+    if not strict:
+        # Note, this also keeps space, single quote(') and hypen (-)
+        text = text.replace("-", " ")
+        text = text.replace("—", " ")
+        return re.sub(r"[^a-zA-Z0-9\s']+", "", text)
+    else:
+        # only keeps space
+        return re.sub(r"[^a-zA-Z\s]+", "", text)
+
+
+def upper_only_alpha(text: str) -> str:
+    return remove_non_alphabetic(text.upper(), strict=False)
+
+
+def lower_only_alpha(text: str) -> str:
+    return remove_non_alphabetic(text.lower(), strict=False)
+
+
+def lower_all_char(text: str) -> str:
+    return text.lower()
+
+
+def upper_all_char(text: str) -> str:
+    return text.upper()
+
+
+def _apply_style_transform(text: List[str], transform: str) -> List[str]:
+    """Apply transform to a list of text. By default, the text are in
+    ground truth format, i.e mixed-punc.
+
+    Args:
+        text (List[str]): Input text string
+        transform (str): Transform to be applied
+
+    Returns:
+        List[str]: _description_
+    """
+    if transform == "mixed-punc":
+        return text
+    elif transform == "upper-no-punc":
+        return [upper_only_alpha(s) for s in text]
+    elif transform == "lower-no-punc":
+        return [lower_only_alpha(s) for s in text]
+    elif transform == "lower-punc":
+        return [lower_all_char(s) for s in text]
+    else:
+        raise NotImplementedError(f"Unseen transform: {transform}")
+
+
+if __name__ == "__main__":
+    ref_text = "Mixed-case English transcription, with punctuation. Actually, it is fully not related."
+    print(ref_text)
+    res = upper_only_alpha(ref_text)
+    print(res)
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py
new file mode 100644
index 000000000..c8b20d021
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py
@@ -0,0 +1,1419 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo
+#                                                       Zengwei Yao,
+#                                                       Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+
+# For mix precision training, using MCP style transcript:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_prompt_asr/train_baseline.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_prompt_asr/exp \
+  --transcript-style MCP \
+  --max-duration 1000
+
+# For mix precision training, using UC style transcript:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_prompt_asr/train_baseline.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_prompt_asr/exp \
+  --transcript-style UC \
+  --max-duration 1000
+
+# To train a streaming model
+
+./zipformer_prompt_asr/train_baseline.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --causal 1
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model_baseline import Transducer
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from text_normalization import train_text_normalization, upper_only_alpha
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_mixed_cased_with_punc(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: Optional[str] = None,
+    rare_word_list: Optional[List[str]] = None,
+) -> str:
+    # Always get the first one, which is the mixed-cased text with punc
+    out = {"text": texts[0], "pre_text": pre_texts[0]}
+    return out
+
+
+def get_upper_only_alpha(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: Optional[str] = None,
+    rare_word_list: Optional[List[str]] = None,
+) -> str:
+    # Always get the first one, which is the mixed-cased text with punc,
+    # but with upper case it and remove punctuation
+    out = {
+        "text": upper_only_alpha(texts[0]),
+        "pre_text": upper_only_alpha(pre_texts[0]),
+    }
+    return out
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--text-encoder-dim",
+        type=str,
+        default="256,256,384,512",
+        help="Embedding dimension in text encoder stacks: a comma-separated list of 4 elements, "
+        "or you should change other configs in the code.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--transcript-style",
+        type=str,
+        default="UC",
+        choices=["UC", "MCP"],
+        help="""The transcript style used for training. UC stands for upper-cased text w/o punctuations,
+        MCP stands for mix-cased text with punctuation.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    texts = [train_text_normalization(t) for t in texts]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    if random.random() < 0.02:
+        logging.info(f"Ref texts: {texts[0]}")
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    train_cuts = libriheavy.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 30.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].texts[0], out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].texts[0]}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    if params.transcript_style == "UC":
+        text_sampling_func = get_upper_only_alpha
+    else:
+        text_sampling_func = get_mixed_cased_with_punc
+    logging.info(f"Using {params.transcript_style} style for training.")
+    logging.info(f"Text sampling func: {text_sampling_func}")
+    train_dl = libriheavy.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        text_sampling_func=text_sampling_func,
+    )
+
+    valid_cuts = libriheavy.dev_cuts()
+    valid_dl = libriheavy.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py
new file mode 100755
index 000000000..9822b99c1
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py
@@ -0,0 +1,1798 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Xiaoyu Yang,
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For mix precision training:
+
+(1) Non-streaming model, **without** context list
+
+./zipformer_prompt_asr/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --subset medium \
+  --causal False \
+  --exp-dir zipformer_prompt_asr/exp \
+  --max-duration 1000 \
+  --memory-layer 0 \
+  --text-encoder-type BERT \
+  --text-encoder-dim 768 \
+  --use-style-prompt True \
+  --use-context-list False
+
+(2) Non-streaming model, **with** context list
+
+./zipformer_prompt_asr/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --subset medium \
+  --causal False \
+  --exp-dir zipformer_prompt_asr/exp \
+  --max-duration 1000 \
+  --memory-layer 0 \
+  --text-encoder-type BERT \
+  --text-encoder-dim 768 \
+  --use-style-prompt True \
+  --use-context-list True \
+  --top-k 10000 \
+  --rare-word-file data/context_biasing/small_rare_words_topk_10000.txt
+
+
+"""
+
+
+import argparse
+import copy
+import logging
+import os
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import numpy
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from dataset import (
+    naive_triplet_text_sampling,
+    random_shuffle_subset,
+    triplet_text_sampling,
+    triplet_text_sampling_with_context_list,
+)
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model_with_BERT import PromptedTransducer
+from optim import Eden, ScaledAdam
+from scaling import Balancer, BiasNorm, Dropout3, ScaleGrad, ScheduledFloat, SwooshR
+from subsampling import Conv2dSubsampling
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+style_transforms = [
+    lambda x: x,  # return it self
+    upper_only_alpha,
+    lower_only_alpha,
+    lower_all_char,
+]
+
+
+def get_first(texts: List[str], pre_texts: List[str]) -> str:
+    out = {
+        "text": texts[0],
+        "pre_text": pre_texts[0],
+        "style_text": "",
+        "transform_ids": 0,
+    }
+    return out
+
+
+def get_upper_only_alpha(texts: List[str], pre_texts: List[str]) -> str:
+    # Always get the first one, which is the gt (mixed-cased trans), but with upper_only_alpha
+    out = {
+        "text": upper_only_alpha(texts[0]),
+        "pre_text": upper_only_alpha(pre_texts[0]),
+        "style_text": "",
+        "transform_ids": 0,
+    }
+    return out
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--memory-dropout-rate",
+        type=float,
+        default=0.05,
+        help="By which probability, dropout the memory when doing cross-attention.",
+    )
+
+    parser.add_argument(
+        "--memory-layer",
+        type=int,
+        default=0,
+        help="Start doing cross-attention from which layer. Zero-indexed",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--freeze-text-encoder",
+        type=str2bool,
+        default=True,
+    )
+
+    parser.add_argument(
+        "--text-encoder-type",
+        type=str,
+        default="BERT",
+        choices=["BERT", "DistilBERT"],
+        help="Type of the text encoder",
+    )
+
+    parser.add_argument(
+        "--text-encoder-dim",
+        type=int,
+        default=768,
+        help="Dimension of the text encoder",
+    )
+
+    parser.add_argument(
+        "--text-encoder-adapter",
+        type=str2bool,
+        default=False,
+        help="An adapter for pre-trained BERT",
+    )
+
+    parser.add_argument(
+        "--context-injection",
+        type=str2bool,
+        default=False,
+        help="Inject context embedding into the joiner",
+    )
+
+    parser.add_argument(
+        "--context-dropout-rate",
+        type=float,
+        default=0.05,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Whether to use style prompt.",
+    )
+
+    # arguments for using prompt
+    parser.add_argument(
+        "--pre-text-shuffle-prob",
+        type=float,
+        default=0.05,
+        help="The proportion of pre_text to be shuffled with in a batch",
+    )
+
+    parser.add_argument(
+        "--style-text-shuffle-prob",
+        type=float,
+        default=0.2,
+        help="The proportion of style_text to be shuffled with in a batch",
+    )
+
+    parser.add_argument(
+        "--prompt-mask-prob",
+        type=float,
+        default=0.05,
+        help="The probability of masking prompts",
+    )
+
+    parser.add_argument(
+        "--forced-upper-pre-text",
+        type=str2bool,
+        default=False,
+        help="Forced format of pre-text",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+class TextEmbedding(nn.Module):
+    def __init__(
+        self,
+        num_embeddings: int = 256,
+        embedding_dim: int = 256,
+        kernel_size: int = 3,
+        layer1_channels: int = 256,
+        layer2_channels: int = 256,
+        bias: bool = True,
+        dropout: float = 0.1,
+    ):
+        super().__init__()
+        self.embed = nn.Embedding(
+            num_embeddings=num_embeddings,  # we encode the text as UTF-8 bytes
+            embedding_dim=embedding_dim,  #
+        )
+
+        assert embedding_dim == layer1_channels  # for depth wise convolution
+        self.conv = nn.Sequential(
+            nn.Conv1d(
+                embedding_dim,
+                layer1_channels,  # depthwise convolution
+                kernel_size=kernel_size,
+                stride=1,
+                padding=(kernel_size - 1) // 2,
+                groups=layer1_channels,
+                bias=True,
+            ),
+            ScaleGrad(0.2),
+            Balancer(layer1_channels, channel_dim=1, min_positive=0.1, max_abs=1.0),
+            nn.ReLU(),
+            nn.Conv1d(
+                layer1_channels,
+                layer2_channels,
+                kernel_size=1,  # pointwise convolution
+                stride=1,
+                padding=0,
+                bias=True,
+            ),
+            Balancer(layer2_channels, channel_dim=1, min_positive=0.1, max_abs=1.0),
+            nn.ReLU(),
+        )
+
+        self.out_norm = BiasNorm(layer2_channels)
+        self.dropout = Dropout3(dropout, shared_dim=1)
+
+    def forward(self, text: torch.Tensor) -> torch.Tensor:
+        """Forward function of the text embedding
+
+        Args:
+            text (torch.Tensor): Text in UTF-8 bytes (T,N)
+        Returns:
+            The embeddings of text (T,N,C)
+        """
+        text = self.embed(text)  # (T,N,C)
+
+        # src = text
+        text = text.permute(1, 2, 0)  # (T,N,C) -> (N,C,T)
+        text = self.conv(text)
+        text = text.permute(2, 0, 1)  # (N,C,T) -> (T,N,C)
+        # src = src + text
+
+        text = self.out_norm(text)
+        text = self.dropout(text)
+
+        return text
+
+
+def get_text_encoder(params: AttributeDict) -> nn.Module:
+    # Return a text encoder
+    if params.text_encoder_type == "BERT":  # This is a BERT-base-cased
+        from transformers import BertModel
+
+        logging.info("Loading pre-trained BERT-base-cased as text encoder")
+        if os.path.exists("data/models/bert-base-cased"):
+            model = BertModel.from_pretrained("data/models/bert-base-cased")
+        else:
+            model = BertModel.from_pretrained("bert-base-cased")
+        assert params.text_encoder_dim == 768
+    elif params.text_encoder_type == "BERT-large":
+        from transformers import BertModel
+
+        logging.info("Loading pre-trained BERT-large-uncased as text encoder")
+        if os.path.exists("data/models/bert-large-uncased"):
+            model = BertModel.from_pretrained("data/models/bert-large-uncased")
+        else:
+            model = BertModel.from_pretrained("bert-large-uncased")
+        assert params.text_encoder_dim == 1024
+    elif params.text_encoder_type == "DistilBERT":
+        from transformers import DistilBertModel  # This is a DistilBERT-base-cased
+
+        logging.info("Loading pre-trained DistilBERT-base-cased as text encoder")
+        model = DistilBertModel.from_pretrained("distilbert-base-cased")
+        assert params.text_encoder_dim == 768
+    else:
+        raise ValueError()
+
+    return model
+
+
+def get_tokenizer(params: AttributeDict):
+
+    if params.text_encoder_type == "BERT":
+        from transformers import BertTokenizer
+
+        # This is a BERT-base-cased
+        if os.path.exists("data/models/bert-base-cased"):
+            tokenizer = BertTokenizer.from_pretrained("data/models/bert-base-cased")
+        else:
+            tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    elif params.text_encoder_type == "BERT-large":
+        from transformers import BertTokenizer
+
+        # This is a BERT-large-uncased
+        if os.path.exists("data/models/bert-large-uncased"):
+            tokenizer = BertTokenizer.from_pretrained("data/models/bert-large-uncased")
+        else:
+            tokenizer = BertTokenizer.from_pretrained("bert-large-uncased")
+    elif params.text_encoder_type == "DistilBERT":
+        from transformers import DistilBertTokenizer
+
+        tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-cased")
+    else:
+        raise ValueError()
+
+    return tokenizer
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+        memory_dim=params.text_encoder_dim,  # This is fixed as the BERT base model is 768-D
+        memory_layer=params.memory_layer,
+        memory_dropout_rate=params.memory_dropout_rate,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        context_dim=4 * 768
+        if params.context_injection
+        else -1,  # the output dim of text encoder
+        context_injection=params.context_injection,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    text_encoder = get_text_encoder(params)  # This should be a cased BERT base model
+    num_param = sum([p.numel() for p in text_encoder.parameters()])
+    logging.info(f"Num params in text encoder: {num_param}")
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = PromptedTransducer(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        text_encoder=text_encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        text_encoder_type=params.text_encoder_type,
+        text_encoder_adapter=params.text_encoder_adapter,
+        freeze_text_encoder=params.freeze_text_encoder,
+        context_fuser=None,
+    )
+
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def _encode_texts_as_bytes_with_tokenizer(
+    pre_texts: List[str],
+    style_texts: List[str],
+    tokenizer,
+    device: torch.device,
+    max_len: int = 500,
+    no_limit: bool = False,
+) -> Tuple[Dict, Tensor]:
+    """
+    Encode texts as bytes and then integer tensors.
+    Note that the style text will be added to the beginning of texts.
+    """
+    batch_size = len(pre_texts)
+    max_len = min(max_len, 500)
+
+    if no_limit:
+        allowed_lens = [5000 - len(s) for s in style_texts]
+    else:
+        allowed_lens = [1000 - len(s) for s in style_texts]
+    truncated_pre_texts = [pre_texts[i][-allowed_lens[i] :] for i in range(batch_size)]
+    combined_text = [
+        style_texts[i] + " [SEP] " + truncated_pre_texts[i] for i in range(batch_size)
+    ]
+
+    encoded_style_texts = tokenizer(
+        style_texts,
+        return_tensors="pt",
+        padding=True,
+        truncation=True,
+        return_length=True,
+        max_length=max_len,
+    )
+    style_lens = encoded_style_texts["length"].to(device)
+
+    # Use tokenizer to prepare input for text encoder
+    encoded_inputs = tokenizer(
+        combined_text,
+        return_tensors="pt",
+        padding=True,
+        truncation=True,
+        return_length=True,
+        max_length=max_len,
+    ).to(device)
+
+    return encoded_inputs, style_lens
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    batch_size = feature.size(0)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    pre_texts = batch["supervisions"]["pre_text"]
+    style_texts = batch["supervisions"][
+        "style_text"
+    ]  # the style texts are in gt format
+    transform_ids = batch["supervisions"]["transform_ids"]
+
+    # This is to replace full-width symbols with half-width symbols
+    texts = [train_text_normalization(t) for t in texts]
+    pre_texts = [train_text_normalization(t) for t in pre_texts]
+    style_texts = [train_text_normalization(t) for t in style_texts]
+
+    y = sp.encode(
+        texts, out_type=int
+    )  # sp.encode treats consecutive space as a single space
+    y = k2.RaggedTensor(y).to(device)
+
+    if params.forced_upper_pre_text:
+        pre_texts = [upper_only_alpha(p) for p in pre_texts]
+
+    # only shuffle the pre_text and style texts if during training, and use style prompt
+    if is_training:
+        # randomly shuffle&mask the pre_text
+        pre_texts = random_shuffle_subset(
+            pre_texts,
+            p=params.pre_text_shuffle_prob,
+            p_mask=params.prompt_mask_prob,
+        )
+
+        if params.use_style_prompt:
+            if random.random() < 0.5:
+                # randomly shuffle the style_text
+                # now the style_texts are all in gt format
+                style_texts = random_shuffle_subset(
+                    style_texts,
+                    p=params.style_text_shuffle_prob,
+                    p_mask=params.prompt_mask_prob,
+                )
+
+            assert len(transform_ids) == len(style_texts)
+
+            for i in range(len(style_texts)):
+                t = transform_ids[i]  # get the transform id
+                style_texts[i] = style_transforms[t](style_texts[i])
+
+    if not params.use_style_prompt:
+        style_texts = [
+            "" for _ in style_texts
+        ]  # use empty string for style texts if don't use style prompt
+
+    if random.random() < 0.05:
+        logging.info(f"Pre texts: {pre_texts[0]}")
+        logging.info(f"Ref texts: {texts[0]}")
+        logging.info(f"Style texts: {style_texts[0]}")
+
+    encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+        pre_texts=pre_texts,
+        style_texts=style_texts,
+        tokenizer=tokenizer,
+        device=device,
+    )
+
+    if random.random() < 0.02:
+        logging.info(f"Shape of encoded texts: {encoded_inputs['input_ids'].shape} ")
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            encoded_inputs=encoded_inputs,
+            style_lens=style_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    tokenizer=tokenizer,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                tokenizer=tokenizer,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if not params.use_style_prompt:
+        if params.pre_text_shuffle_prob == 0.0:
+            logging.info(
+                f"Pre_text shuffle prob is set to: {params.pre_text_shuffle_prob}"
+            )
+            logging.info(
+                "If style prompt is not used, you should be careful when shuffling the pre_text within the same batch"
+            )
+            logging.info("Hard set this probability to 0.0!")
+            params.pre_text_shuffle_prob = 0.0
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    if params.freeze_text_encoder:
+        freeze_modules = ["text_encoder"]
+        logging.info(
+            "Freeze the parameters of text encoder and don't include them in the optimizer"
+        )
+    else:
+        freeze_modules = []
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(
+            model, lr=params.base_lr, include_names=True, freeze_modules=freeze_modules
+        ),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        args.max_duration = 100
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    train_cuts = libriheavy.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 30.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].texts[0], out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].texts[0]}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    if params.use_context_list:
+        text_sampling_func = triplet_text_sampling_with_context_list
+    else:
+        text_sampling_func = triplet_text_sampling
+
+    logging.info(f"Text sampling: {text_sampling_func}")
+
+    train_dl = libriheavy.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        text_sampling_func=text_sampling_func,
+    )
+
+    # For fair comparison, use fixed sampling in valid dataloaders
+    valid_cuts = libriheavy.dev_cuts()
+    valid_dl = libriheavy.valid_dataloaders(
+        valid_cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            tokenizer=tokenizer,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            tokenizer=tokenizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    tokenizer: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    tokenizer=tokenizer,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py b/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py
new file mode 100644
index 000000000..ef0c48e8a
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py
@@ -0,0 +1,515 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+python ./zipformer_prompt_asr/transcribe_bert.py \
+    --epoch 50 \
+    --avg 10 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --manifest-dir data/long_audios/long_audio.jsonl.gz \
+    --pre-text-transform mixed-punc \
+    --style-text-transform mixed-punc \
+    --num-history 5 \
+    --use-pre-text True \
+    --use-gt-pre-text False
+
+
+"""
+
+import argparse
+import logging
+import math
+import warnings
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from decode_bert import _apply_style_transform
+from lhotse import Fbank, load_manifest
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    ref_text_normalization,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from tqdm import tqdm
+from train_bert_encoder import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+    )
+
+    parser.add_argument(
+        "--manifest-dir",
+        type=str,
+        default="data/long_audios/long_audio.jsonl.gz",
+        help="""This is the manfiest for long audio transcription.
+        The cust are intended to be sorted, i.e first sort by recording ID and
+        then sort by start timestamp""",
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=False,
+        help="Whether use pre-text when decoding the current chunk",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt when evaluation",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--num-history",
+        type=int,
+        default=2,
+        help="How many previous chunks to look if using pre-text for decoding",
+    )
+
+    parser.add_argument(
+        "--use-gt-pre-text",
+        type=str2bool,
+        default=False,
+        help="Whether use gt pre text when using content prompt",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.res_dir = params.exp_dir / "long_audio_transcribe"
+    params.res_dir.mkdir(exist_ok=True)
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "beam_search" in params.method:
+        params.suffix += f"-{params.method}-beam-size-{params.beam_size}"
+
+    if params.use_pre_text:
+        if params.use_gt_pre_text:
+            params.suffix += f"-use-gt-pre-text-{params.pre_text_transform}-history-{params.num_history}"
+        else:
+            params.suffix += (
+                f"-pre-text-{params.pre_text_transform}-history-{params.num_history}"
+            )
+
+    book_name = params.manifest_dir.split("/")[-1].replace(".jsonl.gz", "")
+    setup_logger(
+        f"{params.res_dir}/log-decode-{book_name}-{params.suffix}", log_level="info"
+    )
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg + 1
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for" f" --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg + 1:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        filename_start = filenames[-1]
+        filename_end = filenames[0]
+        logging.info(
+            "Calculating the averaged model over iteration checkpoints"
+            f" from {filename_start} (excluded) to {filename_end}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+    else:
+        assert params.avg > 0, params.avg
+        start = params.epoch - params.avg
+        assert start >= 1, start
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        logging.info(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    # load manifest
+    manifest = load_manifest(params.manifest_dir)
+
+    results = []
+    count = 0
+
+    last_recording = ""
+    last_end = -1
+    history = []
+    num_pre_texts = []
+
+    for cut in manifest:
+        if cut.has_features:
+            feat = cut.load_features()
+            feat_lens = cut.num_frames
+        else:
+            feat = cut.compute_features(extractor=Fbank())
+            feat_lens = feat.shape[0]
+
+        cur_recording = cut.recording.id
+
+        if cur_recording != last_recording:
+            last_recording = cur_recording
+            history = []  # clean up the history
+            last_end = -1
+            logging.info("Moving on to the next recording")
+        else:
+            if cut.start < last_end - 0.2:  # overlap with the previous cuts
+                logging.warning("An overlap exists between current cut and last cut")
+                logging.warning("Skipping this cut!")
+                continue
+            if cut.start > last_end + 10:
+                logging.warning(
+                    f"Large time gap between the current and previous utterance: {cut.start - last_end}."
+                )
+
+        # prepare input
+        x = torch.tensor(feat, device=device).unsqueeze(0)
+        x_lens = torch.tensor(
+            [
+                feat_lens,
+            ],
+            device=device,
+        )
+
+        if params.use_pre_text:
+            if params.num_history > 0:
+                pre_texts = history[-params.num_history :]
+            else:
+                pre_texts = []
+            num_pre_texts.append(len(pre_texts))
+            pre_texts = [train_text_normalization(" ".join(pre_texts))]
+            fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it is fully not related."
+            style_texts = [fixed_sentence]
+
+            pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform)
+            if params.use_style_prompt:
+                style_texts = _apply_style_transform(
+                    style_texts, params.style_text_transform
+                )
+
+            # encode prompts
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+
+                encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+                    pre_texts=pre_texts,
+                    style_texts=style_texts,
+                    tokenizer=tokenizer,
+                    device=device,
+                    no_limit=True,
+                )
+                if params.num_history > 5:
+                    logging.info(
+                        f"Shape of encoded texts: {encoded_inputs['input_ids'].shape} "
+                    )
+
+                memory, memory_key_padding_mask = model.encode_text(
+                    encoded_inputs=encoded_inputs,
+                    style_lens=style_lens,
+                )  # (T,B,C)
+        else:
+            memory = None
+            memory_key_padding_mask = None
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            encoder_out, encoder_out_lens = model.encode_audio(
+                feature=x,
+                feature_lens=x_lens,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+            )
+
+        if params.method == "greedy_search":
+            hyp_tokens = greedy_search_batch(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+            )
+        elif params.method == "modified_beam_search":
+            hyp_tokens = modified_beam_search(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                beam=params.beam_size,
+            )
+
+        hyp = sp.decode(hyp_tokens)[0]  # in string format
+        ref_text = ref_text_normalization(
+            cut.supervisions[0].texts[0]
+        )  # required to match the training
+
+        # extend the history
+        if params.use_gt_pre_text:
+            history.append(ref_text)
+        else:
+            history.append(hyp)
+        last_end = cut.end  # update the last end timestamp
+
+        # append the current decoding result
+        hyp = hyp.split()
+        ref = ref_text.split()
+        results.append((cut.id, ref, hyp))
+
+        count += 1
+        if count % 100 == 0:
+            logging.info(f"Cuts processed until now: {count}/{len(manifest)}")
+            logging.info(
+                f"Averaged context numbers of last 100 samples is: {sum(num_pre_texts[-100:])/100}"
+            )
+
+    logging.info(f"A total of {count} cuts")
+    logging.info(
+        f"Averaged context numbers of whole set is: {sum(num_pre_texts)/len(num_pre_texts)}"
+    )
+
+    results = sorted(results)
+    recog_path = (
+        params.res_dir / f"recogs-long-audio-{params.method}-{params.suffix}.txt"
+    )
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    errs_filename = (
+        params.res_dir / f"errs-long-audio-{params.method}-{params.suffix}.txt"
+    )
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(
+            f,
+            f"long-audio-{params.method}",
+            results,
+            enable_log=True,
+            compute_CER=False,
+        )
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    if params.post_normalization:
+        params.suffix += "-post-normalization"
+
+        new_res = []
+        for item in results:
+            id, ref, hyp = item
+            hyp = upper_only_alpha(" ".join(hyp)).split()
+            ref = upper_only_alpha(" ".join(ref)).split()
+            new_res.append((id, ref, hyp))
+
+    new_res = sorted(new_res)
+    recog_path = (
+        params.res_dir
+        / f"recogs-long-audio-{params.method}-{params.suffix}-post-normalization.txt"
+    )
+    store_transcripts(filename=recog_path, texts=new_res)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    errs_filename = (
+        params.res_dir
+        / f"errs-long-audio-{params.method}-{params.suffix}-post-normalization.txt"
+    )
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(
+            f,
+            f"long-audio-{params.method}",
+            new_res,
+            enable_log=True,
+            compute_CER=False,
+        )
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py b/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py
new file mode 100644
index 000000000..533982519
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py
@@ -0,0 +1,439 @@
+import argparse
+import ast
+import glob
+import logging
+import os
+from collections import defaultdict
+from typing import Dict, Iterable, List, TextIO, Tuple, Union
+
+import kaldialign
+from lhotse import load_manifest, load_manifest_lazy
+from lhotse.cut import Cut, CutSet
+from text_normalization import remove_non_alphabetic
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--manifest-dir",
+        type=str,
+        default="data/fbank",
+        help="Where are the manifest stored",
+    )
+
+    parser.add_argument(
+        "--subset", type=str, default="medium", help="Which subset to work with"
+    )
+
+    parser.add_argument(
+        "--top-k",
+        type=int,
+        default=10000,
+        help="How many words to keep",
+    )
+
+    return parser
+
+
+def get_facebook_biasing_list(
+    test_set: str,
+    num_distractors: int = 100,
+) -> Dict:
+    # Get the biasing list from the meta paper: https://arxiv.org/pdf/2104.02194.pdf
+    assert num_distractors in (0, 100, 500, 1000, 2000), num_distractors
+    if num_distractors == 0:
+        if test_set == "test-clean":
+            biasing_file = "data/context_biasing/fbai-speech/is21_deep_bias/ref/test-clean.biasing_100.tsv"
+        elif test_set == "test-other":
+            biasing_file = "data/context_biasing/fbai-speech/is21_deep_bias/ref/test-other.biasing_100.tsv"
+        else:
+            raise ValueError(f"Unseen test set {test_set}")
+    else:
+        if test_set == "test-clean":
+            biasing_file = f"data/context_biasing/fbai-speech/is21_deep_bias/ref/test-clean.biasing_{num_distractors}.tsv"
+        elif test_set == "test-other":
+            biasing_file = f"data/context_biasing/fbai-speech/is21_deep_bias/ref/test-other.biasing_{num_distractors}.tsv"
+        else:
+            raise ValueError(f"Unseen test set {test_set}")
+
+    f = open(biasing_file, "r")
+    data = f.readlines()
+    f.close()
+
+    output = dict()
+    for line in data:
+        id, _, l1, l2 = line.split("\t")
+        if num_distractors > 0:  # use distractors
+            biasing_list = ast.literal_eval(l2)
+        else:
+            biasing_list = ast.literal_eval(l1)
+        biasing_list = [w.strip().upper() for w in biasing_list]
+        output[id] = " ".join(biasing_list)
+
+    return output
+
+
+def brian_biasing_list(level: str):
+    # The biasing list from Brian's paper: https://arxiv.org/pdf/2109.00627.pdf
+    root_dir = f"data/context_biasing/LibriSpeechBiasingLists/{level}Level"
+    all_files = glob.glob(root_dir + "/*")
+    biasing_dict = {}
+    for f in all_files:
+        k = f.split("/")[-1]
+        fin = open(f, "r")
+        data = fin.read().strip().split()
+        biasing_dict[k] = " ".join(data)
+        fin.close()
+
+    return biasing_dict
+
+
+def get_rare_words(
+    subset: str = "medium",
+    top_k: int = 10000,
+    # min_count: int = 10000,
+):
+    """Get a list of rare words appearing less than `min_count` times
+
+    Args:
+        subset: The dataset
+        top_k (int): How many frequent words
+    """
+    txt_path = f"data/tmp/transcript_words_{subset}.txt"
+    rare_word_file = f"data/context_biasing/{subset}_rare_words_topk_{top_k}.txt"
+
+    if os.path.exists(rare_word_file):
+        print("File exists, do not proceed!")
+        return
+
+    print("---Identifying rare words in the manifest---")
+    count_file = f"data/tmp/transcript_words_{subset}_count.txt"
+    if not os.path.exists(count_file):
+        with open(txt_path, "r") as file:
+            words = file.read().upper().split()
+            word_count = {}
+            for word in words:
+                word = remove_non_alphabetic(word, strict=False)
+                word = word.split()
+                for w in word:
+                    if w not in word_count:
+                        word_count[w] = 1
+                    else:
+                        word_count[w] += 1
+
+        word_count = list(word_count.items())  # convert to a list of tuple
+        word_count = sorted(word_count, key=lambda w: int(w[1]), reverse=True)
+        with open(count_file, "w") as fout:
+            for w, count in word_count:
+                fout.write(f"{w}\t{count}\n")
+
+    else:
+        word_count = {}
+        with open(count_file, "r") as fin:
+            word_count = fin.read().strip().split("\n")
+            word_count = [pair.split("\t") for pair in word_count]
+            word_count = sorted(word_count, key=lambda w: int(w[1]), reverse=True)
+
+    print(f"A total of {len(word_count)} words appeared!")
+    rare_words = []
+    for word, count in word_count[top_k:]:
+        rare_words.append(word + "\n")
+    print(f"A total of {len(rare_words)} are identified as rare words.")
+
+    with open(rare_word_file, "w") as f:
+        f.writelines(rare_words)
+
+
+def add_context_list_to_manifest(
+    manifest_dir: str,
+    subset: str = "medium",
+    top_k: int = 10000,
+):
+    """Generate a context list of rare words for each utterance in the manifest
+
+    Args:
+        manifest_dir: Where to store the manifest with context list
+        subset (str): Subset
+        top_k (int): How many frequent words
+
+    """
+    orig_manifest_dir = f"{manifest_dir}/libriheavy_cuts_{subset}.jsonl.gz"
+    target_manifest_dir = orig_manifest_dir.replace(
+        ".jsonl.gz", f"_with_context_list_topk_{top_k}.jsonl.gz"
+    )
+    if os.path.exists(target_manifest_dir):
+        print(f"Target file exits at {target_manifest_dir}!")
+        return
+
+    rare_words_file = f"data/context_biasing/{subset}_rare_words_topk_{top_k}.txt"
+    print(f"---Reading rare words from {rare_words_file}---")
+    with open(rare_words_file, "r") as f:
+        rare_words = f.read()
+    rare_words = rare_words.split("\n")
+    rare_words = set(rare_words)
+    print(f"A total of {len(rare_words)} rare words!")
+
+    cuts = load_manifest_lazy(orig_manifest_dir)
+    print(f"Loaded manifest from {orig_manifest_dir}")
+
+    def _add_context(c: Cut):
+        splits = (
+            remove_non_alphabetic(c.supervisions[0].texts[0], strict=False)
+            .upper()
+            .split()
+        )
+        found = []
+        for w in splits:
+            if w in rare_words:
+                found.append(w)
+        c.supervisions[0].context_list = " ".join(found)
+        return c
+
+    cuts = cuts.map(_add_context)
+    print(f"---Saving manifest with context list to {target_manifest_dir}---")
+    cuts.to_file(target_manifest_dir)
+    print("Finished")
+
+
+def check(
+    manifest_dir: str,
+    subset: str = "medium",
+    top_k: int = 10000,
+):
+    # Show how many samples in the training set have a context list
+    # and the average length of context list
+    print("--- Calculating the stats over the manifest ---")
+
+    manifest_dir = f"{manifest_dir}/libriheavy_cuts_{subset}_with_context_list_topk_{top_k}.jsonl.gz"
+    cuts = load_manifest_lazy(manifest_dir)
+    total_cuts = len(cuts)
+    has_context_list = [c.supervisions[0].context_list != "" for c in cuts]
+    context_list_len = [len(c.supervisions[0].context_list.split()) for c in cuts]
+    print(f"{sum(has_context_list)}/{total_cuts} cuts have context list! ")
+    print(
+        f"Average length of non-empty context list is {sum(context_list_len)/sum(has_context_list)}"
+    )
+
+
+def write_error_stats(
+    f: TextIO,
+    test_set_name: str,
+    results: List[Tuple[str, str]],
+    enable_log: bool = True,
+    compute_CER: bool = False,
+    biasing_words: List[str] = None,
+) -> float:
+    """Write statistics based on predicted results and reference transcripts. It also calculates the
+    biasing word error rate as described in https://arxiv.org/pdf/2104.02194.pdf
+
+    It will write the following to the given file:
+
+        - WER
+        - number of insertions, deletions, substitutions, corrects and total
+          reference words. For example::
+
+              Errors: 23 insertions, 57 deletions, 212 substitutions, over 2606
+              reference words (2337 correct)
+
+        - The difference between the reference transcript and predicted result.
+          An instance is given below::
+
+            THE ASSOCIATION OF (EDISON->ADDISON) ILLUMINATING COMPANIES
+
+          The above example shows that the reference word is `EDISON`,
+          but it is predicted to `ADDISON` (a substitution error).
+
+          Another example is::
+
+            FOR THE FIRST DAY (SIR->*) I THINK
+
+          The reference word `SIR` is missing in the predicted
+          results (a deletion error).
+      results:
+        An iterable of tuples. The first element is the cut_id, the second is
+        the reference transcript and the third element is the predicted result.
+      enable_log:
+        If True, also print detailed WER to the console.
+        Otherwise, it is written only to the given file.
+      biasing_words:
+        All the words in the biasing list
+    Returns:
+      Return None.
+    """
+    subs: Dict[Tuple[str, str], int] = defaultdict(int)
+    ins: Dict[str, int] = defaultdict(int)
+    dels: Dict[str, int] = defaultdict(int)
+
+    # `words` stores counts per word, as follows:
+    #   corr, ref_sub, hyp_sub, ins, dels
+    words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
+    num_corr = 0
+    ERR = "*"
+
+    if compute_CER:
+        for i, res in enumerate(results):
+            cut_id, ref, hyp = res
+            ref = list("".join(ref))
+            hyp = list("".join(hyp))
+            results[i] = (cut_id, ref, hyp)
+
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        for ref_word, hyp_word in ali:
+            if ref_word == ERR:  # INSERTION
+                ins[hyp_word] += 1
+                words[hyp_word][3] += 1
+            elif hyp_word == ERR:  # DELETION
+                dels[ref_word] += 1
+                words[ref_word][4] += 1
+            elif hyp_word != ref_word:  # SUBSTITUTION
+                subs[(ref_word, hyp_word)] += 1
+                words[ref_word][1] += 1
+                words[hyp_word][2] += 1
+            else:
+                words[ref_word][0] += 1
+                num_corr += 1
+    ref_len = sum([len(r) for _, r, _ in results])
+    sub_errs = sum(subs.values())
+    ins_errs = sum(ins.values())
+    del_errs = sum(dels.values())
+    tot_errs = sub_errs + ins_errs + del_errs
+    tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len)
+
+    if enable_log:
+        logging.info(
+            f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} "
+            f"[{tot_errs} / {ref_len}, {ins_errs} ins, "
+            f"{del_errs} del, {sub_errs} sub ]"
+        )
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    print(
+        f"Errors: {ins_errs} insertions, {del_errs} deletions, "
+        f"{sub_errs} substitutions, over {ref_len} reference "
+        f"words ({num_corr} correct)",
+        file=f,
+    )
+    print(
+        "Search below for sections starting with PER-UTT DETAILS:, "
+        "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:",
+        file=f,
+    )
+
+    print("", file=f)
+    print("PER-UTT DETAILS: corr or (ref->hyp)  ", file=f)
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        combine_successive_errors = True
+        if combine_successive_errors:
+            ali = [[[x], [y]] for x, y in ali]
+            for i in range(len(ali) - 1):
+                if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]:
+                    ali[i + 1][0] = ali[i][0] + ali[i + 1][0]
+                    ali[i + 1][1] = ali[i][1] + ali[i + 1][1]
+                    ali[i] = [[], []]
+            ali = [
+                [
+                    list(filter(lambda a: a != ERR, x)),
+                    list(filter(lambda a: a != ERR, y)),
+                ]
+                for x, y in ali
+            ]
+            ali = list(filter(lambda x: x != [[], []], ali))
+            ali = [
+                [
+                    ERR if x == [] else " ".join(x),
+                    ERR if y == [] else " ".join(y),
+                ]
+                for x, y in ali
+            ]
+
+        print(
+            f"{cut_id}:\t"
+            + " ".join(
+                (
+                    ref_word if ref_word == hyp_word else f"({ref_word}->{hyp_word})"
+                    for ref_word, hyp_word in ali
+                )
+            ),
+            file=f,
+        )
+
+    print("", file=f)
+    print("SUBSTITUTIONS: count ref -> hyp", file=f)
+
+    for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True):
+        print(f"{count}   {ref} -> {hyp}", file=f)
+
+    print("", file=f)
+    print("DELETIONS: count ref", file=f)
+    for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True):
+        print(f"{count}   {ref}", file=f)
+
+    print("", file=f)
+    print("INSERTIONS: count hyp", file=f)
+    for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True):
+        print(f"{count}   {hyp}", file=f)
+
+    unbiased_word_counts = 0
+    unbiased_word_errs = 0
+    biased_word_counts = 0
+    biased_word_errs = 0
+
+    print("", file=f)
+    print("PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp", file=f)
+
+    for _, word, counts in sorted(
+        [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True
+    ):
+        (corr, ref_sub, hyp_sub, ins, dels) = counts
+        tot_errs = ref_sub + hyp_sub + ins + dels
+        # number of appearances of "word" in reference text
+        ref_count = (
+            corr + ref_sub + dels
+        )  # correct + in ref but got substituted + deleted
+        # number of appearances of "word" in hyp text
+        hyp_count = corr + hyp_sub + ins
+
+        if biasing_words is not None:
+            if word in biasing_words:
+                biased_word_counts += ref_count
+                biased_word_errs += ins + dels + ref_sub
+            else:
+                unbiased_word_counts += ref_count
+                unbiased_word_errs += ins + dels + hyp_sub
+
+        print(f"{word}   {corr} {tot_errs} {ref_count} {hyp_count}", file=f)
+
+    if biasing_words is not None:
+        B_WER = "%.2f" % (100 * biased_word_errs / biased_word_counts)
+        U_WER = "%.2f" % (100 * unbiased_word_errs / unbiased_word_counts)
+        logging.info(f"Biased WER: {B_WER} [{biased_word_errs}/{biased_word_counts}] ")
+        logging.info(
+            f"Un-biased WER: {U_WER} [{unbiased_word_errs}/{unbiased_word_counts}]"
+        )
+
+    return float(tot_err_rate)
+
+
+if __name__ == "__main__":
+    parser = get_parser()
+    args = parser.parse_args()
+    manifest_dir = args.manifest_dir
+    subset = args.subset
+    top_k = args.top_k
+    get_rare_words(subset=subset, top_k=top_k)
+    add_context_list_to_manifest(
+        manifest_dir=manifest_dir,
+        subset=subset,
+        top_k=top_k,
+    )
+    check(
+        manifest_dir=manifest_dir,
+        subset=subset,
+        top_k=top_k,
+    )
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py b/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py
new file mode 100644
index 000000000..d1cf90ffb
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py
@@ -0,0 +1,2310 @@
+#!/usr/bin/env python3
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Daniel Povey,
+#                                                       Zengwei Yao,
+#                                                       Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import logging
+import math
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+)
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationDropoutAndLinear,
+    Balancer,
+    BiasNorm,
+    ChunkCausalDepthwiseConv1d,
+    Dropout2,
+    FloatLike,
+    ScheduledFloat,
+    Whiten,
+    convert_num_channels,
+    limit_param_value,
+    penalize_abs_values_gt,
+    softmax,
+)
+from torch import Tensor, nn
+
+
+class Zipformer2(EncoderInterface):
+    """
+    Args:
+
+    Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length
+    as downsampling_factor if they are single ints or one-element tuples.  The length of
+    downsampling_factor defines the number of stacks.
+
+        output_downsampling_factor (int): how much to downsample at the output.  Note:
+            we also downsample by a factor of 2 in the Conv2dSubsampling encoder.
+            You should probably leave this at 2.
+        downsampling_factor (Tuple[int]): downsampling factor for each encoder stack.
+           Note: this is in addition to the downsampling factor of 2 that is applied in
+           the frontend (self.encoder_embed).
+        encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per
+           encoder stack.
+        num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack
+        encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of
+            the encoder stacks for purposes of per-frame dropout (recommend 256 for
+            now).
+        query_head_dim (int or Tuple[int]): dimension of query and key per attention
+           head: per stack, if a tuple..
+        value_head_dim (int or Tuple[int]): dimension of value in each attention head
+        pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per
+           attention head
+        num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism.
+              Must be at least 4.
+        feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules
+        cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module
+
+        pos_dim (int): the dimension of each positional-encoding vector prior to projection,
+            e.g. 128.
+
+        dropout (float): dropout rate
+        warmup_batches (float): number of batches to warm up over; this controls
+          dropout of encoder layers.
+        causal (bool): if True, support chunkwise causal convolution.  This should
+          not hurt WER as no modeling power is lost, but the convolution modules will be
+          slightly slower and use more memory.  Enables use of the chunk_size and
+          left_context_chunks options in forward(), which simulates streaming
+          decoding.
+        chunk_size: (list of int): only set this to other than [-1] if causal;
+           the chunk size will be randomly chosen from this list.  -1 means no chunking.
+        left_context_frames: (list of int): determines the number of left-
+           context chunks for causal training; will be rounded to a number of
+           chunks.  Must not be less than cnn_module_kernel (after factoring in
+           rounding and downsampling); an error will be thrown if this is violated.
+        memory_dim: if supplied and >0, will be the dimension of the memory embeddings
+            passed into the zipformer (e.g. this might be the output of another
+            Zipformer used to create embedding vectors.)
+        memory_dropout_rate: By this probability, do not use the provided memory for
+            cross-attention. This should give robustness to the model when evaluated
+            without memory.
+        memory_layer: if supplied and >0, only add cross-attention module starting from
+            the specified layer.
+    """
+
+    def __init__(
+        self,
+        output_downsampling_factor: int = 2,
+        downsampling_factor: Tuple[int] = (2, 4),
+        encoder_dim: Union[int, Tuple[int]] = 384,
+        num_encoder_layers: Union[int, Tuple[int]] = 4,
+        encoder_unmasked_dim: Union[int, Tuple[int]] = 256,
+        query_head_dim: Union[int, Tuple[int]] = 24,
+        pos_head_dim: Union[int, Tuple[int]] = 4,
+        value_head_dim: Union[int, Tuple[int]] = 12,
+        num_heads: Union[int, Tuple[int]] = 8,
+        feedforward_dim: Union[int, Tuple[int]] = 1536,
+        cnn_module_kernel: Union[int, Tuple[int]] = 31,
+        pos_dim: int = 192,
+        dropout: FloatLike = None,  # see code below for default
+        warmup_batches: float = 4000.0,
+        causal: bool = False,
+        chunk_size: Tuple[int] = [-1],
+        left_context_frames: Tuple[int] = [-1],
+        memory_dim: int = -1,
+        memory_dropout_rate: float = 0.05,
+        memory_layer: int = 0,
+    ) -> None:
+        super(Zipformer2, self).__init__()
+
+        if dropout is None:
+            dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1))
+
+        def _to_tuple(x):
+            """Converts a single int or a 1-tuple of an int to a tuple with the same length
+            as downsampling_factor"""
+            if isinstance(x, int):
+                x = (x,)
+            if len(x) == 1:
+                x = x * len(downsampling_factor)
+            else:
+                assert len(x) == len(downsampling_factor) and isinstance(x[0], int)
+            return x
+
+        self.output_downsampling_factor = output_downsampling_factor  # int
+        self.downsampling_factor = downsampling_factor  # tuple
+        self.encoder_dim = encoder_dim = _to_tuple(encoder_dim)  # tuple
+        self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple(
+            encoder_unmasked_dim
+        )  # tuple
+        num_encoder_layers = _to_tuple(num_encoder_layers)
+        self.num_encoder_layers = num_encoder_layers
+        self.query_head_dim = query_head_dim = _to_tuple(query_head_dim)
+        self.value_head_dim = value_head_dim = _to_tuple(value_head_dim)
+        pos_head_dim = _to_tuple(pos_head_dim)
+        self.num_heads = num_heads = _to_tuple(num_heads)
+        feedforward_dim = _to_tuple(feedforward_dim)
+        self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel)
+
+        self.causal = causal
+        self.chunk_size = chunk_size
+        self.left_context_frames = left_context_frames
+        self.memory_dropout_rate = memory_dropout_rate
+        self.memory_layer = memory_layer
+
+        for u, d in zip(encoder_unmasked_dim, encoder_dim):
+            assert u <= d
+
+        # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder
+        encoders = []
+
+        num_encoders = len(downsampling_factor)
+        for i in range(num_encoders):
+            encoder_layer = Zipformer2EncoderLayer(
+                embed_dim=encoder_dim[i],
+                pos_dim=pos_dim,
+                num_heads=num_heads[i],
+                query_head_dim=query_head_dim[i],
+                pos_head_dim=pos_head_dim[i],
+                value_head_dim=value_head_dim[i],
+                feedforward_dim=feedforward_dim[i],
+                memory_dim=memory_dim if i >= self.memory_layer else -1,
+                dropout=dropout,
+                cnn_module_kernel=cnn_module_kernel[i],
+                causal=causal,
+            )
+
+            # For the segment of the warmup period, we let the Conv2dSubsampling
+            # layer learn something.  Then we start to warm up the other encoders.
+            encoder = Zipformer2Encoder(
+                encoder_layer,
+                num_encoder_layers[i],
+                pos_dim=pos_dim,
+                dropout=dropout,
+                warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1),
+                warmup_end=warmup_batches * (i + 2) / (num_encoders + 1),
+                final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5),
+            )
+
+            if downsampling_factor[i] != 1:
+                encoder = DownsampledZipformer2Encoder(
+                    encoder,
+                    dim=encoder_dim[i],
+                    downsample=downsampling_factor[i],
+                    dropout=dropout,
+                )
+
+            encoders.append(encoder)
+
+        self.encoders = nn.ModuleList(encoders)
+
+        self.downsample_output = SimpleDownsample(
+            max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout
+        )
+
+    def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]:
+        """
+        In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of
+        randomized feature masks, one per encoder.
+        On e.g. 15% of frames, these masks will zero out all enocder dims larger than
+        some supplied number, e.g. >256, so in effect on those frames we are using
+        a smaller encoer dim.
+
+        We generate the random masks at this level because we want the 2 masks to 'agree'
+        all the way up the encoder stack. This will mean that the 1st mask will have
+        mask values repeated self.zipformer_subsampling_factor times.
+
+        Args:
+           x: the embeddings (needed for the shape and dtype and device), of shape
+             (1, batch_size, encoder_dims0)
+        """
+        num_encoders = len(self.encoder_dim)
+        if not self.training:
+            return [1.0] * num_encoders
+
+        (num_frames0, batch_size, _encoder_dims0) = x.shape
+
+        assert self.encoder_dim[0] == _encoder_dims0, (
+            self.encoder_dim[0],
+            _encoder_dims0,
+        )
+
+        feature_mask_dropout_prob = 0.125
+
+        # mask1 shape: (1, batch_size, 1)
+        mask1 = (
+            torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob
+        ).to(x.dtype)
+
+        # mask2 has additional sequences masked, about twice the number.
+        mask2 = torch.logical_and(
+            mask1,
+            (
+                torch.rand(1, batch_size, 1, device=x.device)
+                > feature_mask_dropout_prob
+            ).to(x.dtype),
+        )
+
+        # dim: (1, batch_size, 2)
+        mask = torch.cat((mask1, mask2), dim=-1)
+
+        feature_masks = []
+        for i in range(num_encoders):
+            channels = self.encoder_dim[i]
+            feature_mask = torch.ones(
+                1, batch_size, channels, dtype=x.dtype, device=x.device
+            )
+            u1 = self.encoder_unmasked_dim[i]
+            u2 = u1 + (channels - u1) // 2
+
+            feature_mask[:, :, u1:u2] *= mask[..., 0:1]
+            feature_mask[:, :, u2:] *= mask[..., 1:2]
+
+            feature_masks.append(feature_mask)
+
+        return feature_masks
+
+    def get_chunk_info(self) -> Tuple[int, int]:
+        """
+        Returns chunk_size and left_context_chunks.
+        """
+        if not self.causal:
+            return -1, -1
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            assert len(self.chunk_size) == 1, self.chunk_size
+            chunk_size = self.chunk_size[0]
+        else:
+            chunk_size = random.choice(self.chunk_size)
+        if chunk_size == -1:
+            left_context_chunks = -1
+        else:
+            if torch.jit.is_scripting() or torch.jit.is_tracing():
+                assert len(self.left_context_frames) == 1, self.left_context_frames
+                left_context_frames = self.left_context_frames[0]
+            else:
+                left_context_frames = random.choice(self.left_context_frames)
+            # Note: in Python, -1 // n == -1 for n > 0
+            left_context_chunks = left_context_frames // chunk_size
+            if left_context_chunks == 0:
+                left_context_chunks = 1
+
+        return chunk_size, left_context_chunks
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          src_key_padding_mask:
+            The mask for padding, of shape (batch_size, seq_len); True means
+             masked position. May be None.
+          memory:  optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim)
+          memory_key_padding_mask: optionally the mask for padding of memory input (for source-
+             attention), of shape  (batch_size, memory_len); True means
+              masked position.  May be None.
+
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, max(encoder_dim))
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        outputs = []
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            feature_masks = [1.0] * len(self.encoder_dim)
+        else:
+            feature_masks = self.get_feature_masks(x)
+
+        chunk_size, left_context_chunks = self.get_chunk_info()
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            # Not support exporting a model for simulating streaming decoding
+            attn_mask = None
+        else:
+            attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks)
+
+        if self.training and memory is not None:
+            batch_size = x.shape[1]
+            # setting memory to zero should be equivalent to not using the
+            # memory input at all, since the Attention module has no biases.
+            memory = memory * (
+                torch.rand(batch_size, 1, device=memory.device)
+                > self.memory_dropout_rate
+            )
+
+        for i, module in enumerate(self.encoders):
+            ds = self.downsampling_factor[i]
+            x = convert_num_channels(x, self.encoder_dim[i])
+
+            x = module(
+                x,
+                chunk_size=chunk_size,
+                feature_mask=feature_masks[i],
+                src_key_padding_mask=(
+                    None
+                    if src_key_padding_mask is None
+                    else src_key_padding_mask[..., ::ds]
+                ),
+                attn_mask=attn_mask,
+                memory=memory if i >= self.memory_layer else None,
+                memory_key_padding_mask=memory_key_padding_mask
+                if i >= self.memory_layer
+                else None,
+            )
+            outputs.append(x)
+
+        # if the last output has the largest dimension, x will be unchanged,
+        # it will be the same as outputs[-1].  Otherwise it will be concatenated
+        # from different pieces of 'outputs', taking each dimension from the
+        # most recent output that has it present.
+        x = self._get_full_dim_output(outputs)
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            lengths = (x_lens + 1) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                lengths = (x_lens + 1) // 2
+
+        return x, lengths
+
+    def _get_attn_mask(
+        self, x: Tensor, chunk_size: int, left_context_chunks: int
+    ) -> Optional[Tensor]:
+        """
+        Return None if chunk_size == -1, else return attention mask of shape
+          (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len).  True
+           means a masked position.
+        Args:
+           x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim).
+          chunk_size: chunk size, must divide
+        """
+        if chunk_size <= 0:
+            return None
+        assert all(chunk_size % d == 0 for d in self.downsampling_factor)
+        if left_context_chunks >= 0:
+            num_encoders = len(self.encoder_dim)
+            assert all(
+                chunk_size * left_context_chunks
+                >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i]
+                for i in range(num_encoders)
+            )
+        else:
+            left_context_chunks = 1000000
+
+        seq_len = x.shape[0]
+
+        # t is frame index, shape (seq_len,)
+        t = torch.arange(seq_len, dtype=torch.int32, device=x.device)
+        # c is chunk index for each frame, shape (seq_len,)
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            c = t // chunk_size
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                c = t // chunk_size
+        src_c = c
+        tgt_c = c.unsqueeze(-1)
+
+        attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks)
+        if __name__ == "__main__":
+            logging.info(f"attn_mask = {attn_mask}")
+        return attn_mask
+
+    def _get_full_dim_output(self, outputs: List[Tensor]):
+        num_encoders = len(self.encoder_dim)
+        assert len(outputs) == num_encoders
+        output_dim = max(self.encoder_dim)
+        output_pieces = [outputs[-1]]
+        cur_dim = self.encoder_dim[-1]
+        for i in range(num_encoders - 2, -1, -1):
+            d = self.encoder_dim[i]
+            if d > cur_dim:
+                this_output = outputs[i]
+                output_pieces.append(this_output[..., cur_dim:d])
+                cur_dim = d
+        assert cur_dim == output_dim
+        return torch.cat(output_pieces, dim=-1)
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        x_lens: Tensor,
+        states: List[Tensor],
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (seq_len, batch_size, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states: list of cached tensors of all encoder layers. For layer-i,
+            states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+            cached_conv1, cached_conv2).
+          src_key_padding_mask:
+            The mask for padding, of shape (batch_size, seq_len); True means
+            masked position. May be None.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim))
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+            - updated states
+        """
+        outputs = []
+        new_states = []
+        layer_offset = 0
+
+        for i, module in enumerate(self.encoders):
+            num_layers = module.num_layers
+            ds = self.downsampling_factor[i]
+            x = convert_num_channels(x, self.encoder_dim[i])
+
+            x, new_layer_states = module.streaming_forward(
+                x,
+                states=states[layer_offset * 6 : (layer_offset + num_layers) * 6],
+                left_context_len=self.left_context_frames[0] // ds,
+                src_key_padding_mask=src_key_padding_mask[..., ::ds],
+            )
+            layer_offset += num_layers
+            outputs.append(x)
+            new_states += new_layer_states
+
+        # if the last output has the largest dimension, x will be unchanged,
+        # it will be the same as outputs[-1].  Otherwise it will be concatenated
+        # from different pieces of 'outputs', taking each dimension from the
+        # most recent output that has it present.
+        x = self._get_full_dim_output(outputs)
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            lengths = (x_lens + 1) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                lengths = (x_lens + 1) // 2
+
+        return x, lengths, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[Tensor]:
+        """Get initial states.
+
+        A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        """
+        states = []
+        for i, module in enumerate(self.encoders):
+            num_layers = module.num_layers
+            embed_dim = self.encoder_dim[i]
+            ds = self.downsampling_factor[i]
+            num_heads = self.num_heads[i]
+            key_dim = self.query_head_dim[i] * num_heads
+            value_dim = self.value_head_dim[i] * num_heads
+            downsample_left = self.left_context_frames[0] // ds
+            nonlin_attn_head_dim = 3 * embed_dim // 4
+            conv_left_pad = self.cnn_module_kernel[i] // 2
+            for layer in range(num_layers):
+                cached_key = torch.zeros(downsample_left, batch_size, key_dim).to(
+                    device
+                )
+                cached_nonlin_attn = torch.zeros(
+                    1, batch_size, downsample_left, nonlin_attn_head_dim
+                ).to(device)
+                cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to(
+                    device
+                )
+                cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to(
+                    device
+                )
+                cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(
+                    device
+                )
+                cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(
+                    device
+                )
+                states += [
+                    cached_key,
+                    cached_nonlin_attn,
+                    cached_val1,
+                    cached_val2,
+                    cached_conv1,
+                    cached_conv2,
+                ]
+
+        return states
+
+
+def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat:
+    return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x)
+
+
+def _balancer_schedule(min_prob: float):
+    return ScheduledFloat((0.0, 0.4), (8000.0, min_prob))
+
+
+class Zipformer2EncoderLayer(nn.Module):
+    """
+    Args:
+        embed_dim: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        feedforward_dim: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        pos_dim: int,
+        num_heads: int,
+        query_head_dim: int,
+        pos_head_dim: int,
+        value_head_dim: int,
+        feedforward_dim: int,
+        dropout: FloatLike = 0.1,
+        cnn_module_kernel: int = 31,
+        causal: bool = False,
+        memory_dim: int = -1,
+        attention_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0
+        ),
+        conv_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0
+        ),
+        const_attention_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.25), (4000.0, 0.025), default=0
+        ),
+        ff2_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)
+        ),
+        ff3_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)
+        ),
+        bypass_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.5), (4000.0, 0.02), default=0
+        ),
+    ) -> None:
+        super(Zipformer2EncoderLayer, self).__init__()
+        self.embed_dim = embed_dim
+
+        # self.bypass implements layer skipping as well as bypass; see its default values.
+        self.bypass = BypassModule(
+            embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0
+        )
+        # bypass_mid is bypass used in the middle of the layer.
+        self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0)
+
+        # skip probability for dynamic modules (meaning: anything but feedforward).
+        self.attention_skip_rate = copy.deepcopy(attention_skip_rate)
+        # an additional skip probability that applies to ConvModule to stop it from
+        # contributing too much early on.
+        self.conv_skip_rate = copy.deepcopy(conv_skip_rate)
+
+        # ff2_skip_rate is to prevent the ff2 module from having output that's too big
+        # compared to its residual.
+        self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate)
+        self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate)
+
+        self.const_attention_rate = copy.deepcopy(const_attention_rate)
+
+        self.self_attn_weights = RelPositionMultiheadAttentionWeights(
+            embed_dim,
+            pos_dim=pos_dim,
+            num_heads=num_heads,
+            query_head_dim=query_head_dim,
+            pos_head_dim=pos_head_dim,
+            dropout=0.0,
+        )
+
+        self.self_attn1 = Attention(embed_dim, embed_dim, num_heads, value_head_dim)
+
+        self.self_attn2 = Attention(embed_dim, embed_dim, num_heads, value_head_dim)
+
+        if memory_dim > 0:
+            self.attn_weights = MultiheadAttentionWeights(
+                memory_dim,
+                embed_dim,
+                num_heads=num_heads,
+                head_dim=query_head_dim,
+                dropout=0.0,
+            )
+            self.src_attn1 = Attention(memory_dim, embed_dim, num_heads, value_head_dim)
+            self.src_attn2 = Attention(memory_dim, embed_dim, num_heads, value_head_dim)
+            self.memory_balancer = Balancer(
+                embed_dim,
+                channel_dim=-1,
+                min_abs=0.015,
+            )
+
+        self.feed_forward1 = FeedforwardModule(
+            embed_dim, (feedforward_dim * 3) // 4, dropout
+        )
+
+        self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout)
+
+        self.feed_forward3 = FeedforwardModule(
+            embed_dim, (feedforward_dim * 5) // 4, dropout
+        )
+
+        self.nonlin_attention = NonlinAttention(
+            embed_dim, hidden_channels=3 * embed_dim // 4
+        )
+
+        self.conv_module1 = ConvolutionModule(
+            embed_dim, cnn_module_kernel, causal=causal
+        )
+
+        self.conv_module2 = ConvolutionModule(
+            embed_dim, cnn_module_kernel, causal=causal
+        )
+
+        # self.attention_squeeze = AttentionSqueeze(embed_dim, embed_dim // 2)
+
+        self.norm = BiasNorm(embed_dim)
+
+        self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5))
+
+        self.balancer1 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            min_abs=0.2,
+            max_abs=4.0,
+        )
+
+        # balancer for output of NonlinAttentionModule
+        self.balancer_na = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)),
+            prob=0.05,  # out of concern for memory usage
+        )
+
+        # balancer for output of feedforward2, prevent it from staying too
+        # small.  give this a very small probability, even at the start of
+        # training, it's to fix a rare problem and it's OK to fix it slowly.
+        self.balancer_ff2 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0),
+            max_abs=2.0,
+            prob=0.05,
+        )
+
+        self.balancer_ff3 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0),
+            max_abs=4.0,
+            prob=0.05,
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(4.0, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.balancer2 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            min_abs=0.1,
+            max_abs=4.0,
+        )
+
+    def get_bypass_scale(self, batch_size: int):
+        # returns bypass-scale of shape (num_channels,),
+        # or (batch_size, num_channels,).  This is actually the
+        # scale on the non-residual term, so 0 correponds to bypassing
+        # this module.
+        if torch.jit.is_scripting() or not self.training:
+            return self.bypass_scale
+        else:
+            ans = limit_param_value(
+                self.bypass_scale,
+                min=float(self.bypass_min),
+                max=float(self.bypass_max),
+            )
+            layer_skip_rate = float(self.layer_skip_rate)
+            if layer_skip_rate != 0.0:
+                mask = torch.rand((batch_size, 1), device=ans.device) > layer_skip_rate
+                ans = ans * mask
+                # now ans is of shape (batch_size, num_channels), and is zero for sequences
+                # on which we have randomly chosen to do layer-skipping.
+            return ans
+
+    def get_sequence_dropout_mask(
+        self, x: Tensor, dropout_rate: float
+    ) -> Optional[Tensor]:
+        if dropout_rate == 0.0 or not self.training or torch.jit.is_scripting():
+            return None
+        batch_size = x.shape[1]
+        mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype)
+        return mask
+
+    def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor:
+        """
+        Apply sequence-level dropout to x.
+        x shape: (seq_len, batch_size, embed_dim)
+        """
+        dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate)
+        if dropout_mask is None:
+            return x
+        else:
+            return x * dropout_mask
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        chunk_size: int = -1,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+            Pass the input through the encoder layer.
+            Args:
+                src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+             pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim)
+             chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking.
+           feature_mask: something that broadcasts with src, that we'll multiply `src`
+                  by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+             attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                    interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                   True means masked position. May be None.
+        src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+
+            Returns:
+               A tensor which has the same shape as src
+        """
+        src_orig = src
+
+        # dropout rate for non-feedforward submodules
+        attention_skip_rate = float(self.attention_skip_rate) if self.training else 0.0
+
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        attn_weights = self.self_attn_weights(
+            src,
+            pos_emb=pos_emb,
+            attn_mask=attn_mask,
+            key_padding_mask=src_key_padding_mask,
+        )
+
+        if memory is not None and hasattr(self, "attn_weights"):
+            src_attn_weights = self.attn_weights(memory, src, memory_key_padding_mask)
+
+        src = src + self.feed_forward1(src)
+
+        attn_dropout_mask = self.get_sequence_dropout_mask(src, attention_skip_rate)
+
+        if True:
+            selected_attn_weights = attn_weights[0:2]
+            if random.random() < float(self.const_attention_rate):
+                # Make attention weights constant.  The intention is to
+                # encourage these modules to do something similar to an
+                # averaging-over-time operation.
+                # only need the mask, can just use the 1st one and expand later
+                selected_attn_weights = selected_attn_weights[0:1]
+                selected_attn_weights = (selected_attn_weights > 0.0).to(
+                    selected_attn_weights.dtype
+                )
+                selected_attn_weights = selected_attn_weights * (
+                    1.0 / selected_attn_weights.sum(dim=-1, keepdim=True)
+                )
+                selected_attn_weights = selected_attn_weights.expand(2, -1, -1, -1)
+
+        na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights[0:1]))
+
+        src = src + (na if attn_dropout_mask is None else na * attn_dropout_mask)
+
+        self_attn = self.self_attn1(src, attn_weights)
+
+        src = src + (
+            self_attn if attn_dropout_mask is None else self_attn * attn_dropout_mask
+        )
+
+        if memory is not None and hasattr(self, "attn_weights"):
+            src = src + self.sequence_dropout(
+                self.memory_balancer(self.src_attn1(memory, src_attn_weights)),
+                attention_skip_rate,
+            )
+
+        src = src + self.sequence_dropout(
+            self.conv_module1(
+                src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask
+            ),
+            float(self.conv_skip_rate),
+        )
+
+        src = src + self.sequence_dropout(
+            self.balancer_ff2(self.feed_forward2(src)), float(self.ff2_skip_rate)
+        )
+
+        # bypass in the middle of the layer.
+        src = self.bypass_mid(src_orig, src)
+
+        self_attn = self.self_attn2(src, attn_weights)
+
+        src = src + (
+            self_attn if attn_dropout_mask is None else self_attn * attn_dropout_mask
+        )
+
+        if memory is not None and hasattr(self, "attn_weights"):
+            src = src + self.sequence_dropout(
+                self.memory_balancer(self.src_attn2(memory, src_attn_weights)),
+                attention_skip_rate,
+            )
+
+        src = src + self.sequence_dropout(
+            self.conv_module2(
+                src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask
+            ),
+            float(self.conv_skip_rate),
+        )
+
+        src = src + self.sequence_dropout(
+            self.balancer_ff3(self.feed_forward3(src)), float(self.ff3_skip_rate)
+        )
+
+        src = self.balancer1(src)
+        src = self.norm(src)
+
+        src = self.bypass(src_orig, src)
+
+        src = self.balancer2(src)
+        src = self.whiten(src)
+
+        return src
+
+
+class Zipformer2Encoder(nn.Module):
+    r"""Zipformer2Encoder is a stack of N encoder layers
+
+    Args:
+     encoder_layer: an instance of the Zipformer2EncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+           pos_dim: the dimension for the relative positional encoding
+
+    Examples::
+        >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8)
+        >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = zipformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        pos_dim: int,
+        dropout: float,
+        warmup_begin: float,
+        warmup_end: float,
+        initial_layerdrop_rate: float = 0.5,
+        final_layerdrop_rate: float = 0.05,
+    ) -> None:
+        super().__init__()
+        self.encoder_pos = CompactRelPositionalEncoding(
+            pos_dim, dropout_rate=0.15, length_factor=1.0
+        )
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert 0 <= warmup_begin <= warmup_end
+
+        delta = (1.0 / num_layers) * (warmup_end - warmup_begin)
+        cur_begin = warmup_begin  # interpreted as a training batch index
+        for i in range(num_layers):
+            cur_end = cur_begin + delta
+            self.layers[i].bypass.skip_rate = ScheduledFloat(
+                (cur_begin, initial_layerdrop_rate),
+                (cur_end, final_layerdrop_rate),
+                default=0.0,
+            )
+            cur_begin = cur_end
+
+    def forward(
+        self,
+        src: Tensor,
+        chunk_size: int = -1,
+        feature_mask: Union[Tensor, float] = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking.
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+            attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                 interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                 True means masked position. May be None.
+            src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+            memory:  optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim)
+            memory_key_padding_mask: optionally the mask for padding of memory input (for source-
+                attention), of shape  (batch_size, memory_len); True means
+                 masked position.  May be None.
+
+        Returns: a Tensor with the same shape as src.
+        """
+        pos_emb = self.encoder_pos(src)
+        output = src
+
+        output = output * feature_mask
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                chunk_size=chunk_size,
+                attn_mask=attn_mask,
+                src_key_padding_mask=src_key_padding_mask,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+            )
+
+            output = output * feature_mask
+
+        return output
+
+
+class BypassModule(nn.Module):
+    """
+    An nn.Module that implements a learnable bypass scale, and also randomized per-sequence
+    layer-skipping.  The bypass is limited during early stages of training to be close to
+    "straight-through", i.e. to not do the bypass operation much initially, in order to
+    force all the modules to learn something.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        skip_rate: FloatLike = 0.0,
+        straight_through_rate: FloatLike = 0.0,
+        scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0),
+        scale_max: FloatLike = 1.0,
+    ):
+        super().__init__()
+        self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5))
+        self.skip_rate = copy.deepcopy(skip_rate)
+        self.straight_through_rate = copy.deepcopy(straight_through_rate)
+        self.scale_min = copy.deepcopy(scale_min)
+        self.scale_max = copy.deepcopy(scale_max)
+
+    def _get_bypass_scale(self, batch_size: int):
+        # returns bypass-scale of shape (num_channels,),
+        # or (batch_size, num_channels,).  This is actually the
+        # scale on the non-residual term, so 0 correponds to bypassing
+        # this module.
+        if torch.jit.is_scripting() or not self.training:
+            return self.bypass_scale
+        else:
+            ans = limit_param_value(
+                self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max)
+            )
+            skip_rate = float(self.skip_rate)
+            if skip_rate != 0.0:
+                mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate
+                ans = ans * mask
+                # now ans is of shape (batch_size, num_channels), and is zero for sequences
+                # on which we have randomly chosen to do layer-skipping.
+            straight_through_rate = float(self.straight_through_rate)
+            if straight_through_rate != 0.0:
+                mask = (
+                    torch.rand((batch_size, 1), device=ans.device)
+                    < straight_through_rate
+                )
+                ans = torch.maximum(ans, mask.to(ans.dtype))
+
+            return ans
+
+    def forward(self, src_orig: Tensor, src: Tensor):
+        """
+        Args: src_orig and src are both of shape (seq_len, batch_size, num_channels)
+        Returns: something with the same shape as src and src_orig
+        """
+        bypass_scale = self._get_bypass_scale(src.shape[1])
+        return src_orig + (src - src_orig) * bypass_scale
+
+
+class DownsampledZipformer2Encoder(nn.Module):
+    r"""
+    DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate,
+    after convolutional downsampling, and then upsampled again at the output, and combined
+    with the origin input, so that the output has the same shape as the input.
+    """
+
+    def __init__(
+        self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike
+    ):
+        super(DownsampledZipformer2Encoder, self).__init__()
+        self.downsample_factor = downsample
+        self.downsample = SimpleDownsample(dim, downsample, dropout)
+        self.encoder = encoder
+        self.upsample = SimpleUpsample(dim, downsample)
+        self.out_combiner = BypassModule(dim, straight_through_rate=0.025)
+
+    def forward(
+        self,
+        src: Tensor,
+        chunk_size: int = -1,
+        feature_mask: Union[Tensor, float] = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+            attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                 interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                 True means masked position. May be None.
+            src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+            memory:  optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim)
+            memory_key_padding_mask: optionally the mask for padding of memory input (for source-
+                attention), of shape  (batch_size, memory_len); True means
+                 masked position.  May be None.
+
+        Returns: a Tensor with the same shape as src.
+        """
+        src_orig = src
+        src = self.downsample(src)
+        ds = self.downsample_factor
+        if attn_mask is not None:
+            attn_mask = attn_mask[::ds, ::ds]
+
+        src = self.encoder(
+            src,
+            chunk_size=chunk_size // ds,
+            feature_mask=feature_mask,
+            attn_mask=attn_mask,
+            src_key_padding_mask=src_key_padding_mask,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src)
+
+
+class SimpleDownsample(torch.nn.Module):
+    """
+    Does downsampling with attention, by weighted sum, and a projection..
+    """
+
+    def __init__(self, channels: int, downsample: int, dropout: FloatLike):
+        super(SimpleDownsample, self).__init__()
+
+        self.bias = nn.Parameter(torch.zeros(downsample))
+
+        self.name = None  # will be set from training code
+        self.dropout = copy.deepcopy(dropout)
+
+        self.downsample = downsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, in_channels)
+        Returns a tensor of shape
+           ( (seq_len+downsample-1)//downsample, batch_size, channels)
+        """
+        (seq_len, batch_size, in_channels) = src.shape
+        ds = self.downsample
+        d_seq_len = (seq_len + ds - 1) // ds
+
+        # Pad to an exact multiple of self.downsample
+        if seq_len != d_seq_len * ds:
+            # right-pad src, repeating the last element.
+            pad = d_seq_len * ds - seq_len
+            src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2])
+            src = torch.cat((src, src_extra), dim=0)
+            assert src.shape[0] == d_seq_len * ds
+
+        src = src.reshape(d_seq_len, ds, batch_size, in_channels)
+
+        weights = self.bias.softmax(dim=0)
+        # weights: (downsample, 1, 1)
+        weights = weights.unsqueeze(-1).unsqueeze(-1)
+
+        # ans1 is the first `in_channels` channels of the output
+        ans = (src * weights).sum(dim=1)
+
+        return ans
+
+
+class SimpleUpsample(torch.nn.Module):
+    """
+    A very simple form of upsampling that mostly just repeats the input, but
+    also adds a position-specific bias.
+    """
+
+    def __init__(self, num_channels: int, upsample: int):
+        super(SimpleUpsample, self).__init__()
+        self.upsample = upsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, num_channels)
+        Returns a tensor of shape
+           ( (seq_len*upsample), batch_size, num_channels)
+        """
+        upsample = self.upsample
+        (seq_len, batch_size, num_channels) = src.shape
+        src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels)
+        src = src.reshape(seq_len * upsample, batch_size, num_channels)
+        return src
+
+
+class CompactRelPositionalEncoding(torch.nn.Module):
+    """
+    Relative positional encoding module.  This version is "compact" meaning it is able to encode
+    the important information about the relative position in a relatively small number of dimensions.
+    The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001)
+    make very little difference to the embedding.   Such differences were potentially important
+    when encoding absolute position, but not important when encoding relative position because there
+    is now no need to compare two large offsets with each other.
+
+    Our embedding works done by projecting the interval [-infinity,infinity] to a finite interval
+    using the atan() function, before doing the fourier transform of that fixed interval.  The
+    atan() function would compress the "long tails" too small,
+    making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic
+    function to compress large offsets to a smaller range before applying atan().
+    Scalings are chosen in such a way that the embedding can clearly distinguish invidual offsets as long
+    as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim)
+
+
+    Args:
+        embed_dim: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length: just a heuristic for initialization.
+        length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives
+           less weight to small differences of offset near the origin.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        dropout_rate: FloatLike,
+        max_len: int = 1000,
+        length_factor: float = 1.0,
+    ) -> None:
+        """Construct a CompactRelPositionalEncoding object."""
+        super(CompactRelPositionalEncoding, self).__init__()
+        self.embed_dim = embed_dim
+        assert embed_dim % 2 == 0
+        self.dropout = Dropout2(dropout_rate)
+        self.pe = None
+        assert length_factor >= 1.0
+        self.length_factor = length_factor
+        self.extend_pe(torch.tensor(0.0).expand(max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(0) >= x.size(0) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+
+        T = x.size(0)
+        # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ]
+        x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1)
+
+        freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device)
+
+        # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution
+        # for small time offsets but less resolution for large time offsets.
+        compression_length = self.embed_dim**0.5
+        # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity;
+        # but it does so more slowly than T for large absolute values of T.
+        # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which
+        # is important.
+        x_compressed = (
+            compression_length
+            * x.sign()
+            * ((x.abs() + compression_length).log() - math.log(compression_length))
+        )
+
+        # if self.length_factor == 1.0, then length_scale is chosen so that the
+        # FFT can exactly separate points close to the origin (T == 0).  So this
+        # part of the formulation is not really heuristic.
+        # But empirically, for ASR at least, length_factor > 1.0 seems to work better.
+        length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi)
+
+        # note for machine implementations: if atan is not available, we can use:
+        #   x.sign() * ((1 / (x.abs() + 1)) - 1)  * (-math.pi/2)
+        #  check on wolframalpha.com: plot(sign(x) *  (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x))
+        x_atan = (x_compressed / length_scale).atan()  # results between -pi and pi
+
+        cosines = (x_atan * freqs).cos()
+        sines = (x_atan * freqs).sin()
+
+        pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device)
+        pe[:, 0::2] = cosines
+        pe[:, 1::2] = sines
+        pe[:, -1] = 1.0  # for bias.
+
+        self.pe = pe.to(dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tensor:
+        """Create positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (time, batch, `*`).
+
+        Returns:
+            positional embedding, of shape (1, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            self.pe.size(0) // 2
+            - x.size(0)
+            + 1 : self.pe.size(0) // 2  # noqa E203
+            + x.size(0),
+            :,
+        ]
+        pos_emb = pos_emb.unsqueeze(0)
+        return self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttentionWeights(nn.Module):
+    r"""Module that computes multi-head attention weights with relative position encoding.
+    Various other modules consume the resulting attention weights: see, for example, the
+    SimpleAttention module which allows you to compute conventional attention.
+
+    This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
+    we have to write up the differences.
+
+
+    Args:
+           embed_dim: number of channels at the input to this module, e.g. 256
+             pos_dim: dimension of the positional encoding vectors, e.g. 128.
+           num_heads:  number of heads to compute weights for, e.g. 8
+     query_head_dim: dimension of the query (and key), per head.  e.g. 24.
+       pos_head_dim: dimension of the projected positional encoding per head, e.g. 4.
+            dropout: dropout probability for attn_output_weights. Default: 0.0.
+       pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on
+                     any given call to forward(), in training time.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        pos_dim: int,
+        num_heads: int,
+        query_head_dim: int,
+        pos_head_dim: int,
+        dropout: float = 0.0,
+        pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)),
+    ) -> None:
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.query_head_dim = query_head_dim
+        self.pos_head_dim = pos_head_dim
+        self.dropout = dropout
+        self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate)
+        self.name = None  # will be overwritten in training code; for diagnostics.
+
+        key_head_dim = query_head_dim
+        in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5 that has been used in previous forms of attention,
+        # dividing it between the query and key.   Note: this module is intended
+        # to be used with the ScaledAdam optimizer; with most other optimizers,
+        # it would be necessary to apply the scaling factor in the forward function.
+        self.in_proj = ScaledLinear(
+            embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25
+        )
+
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=_whitening_schedule(3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+        # add a balancer for the keys that runs with very small probability, and
+        # tries to enforce that all dimensions have mean around zero.  The
+        # weights produced by this module are invariant to adding a constant to
+        # the keys, so the derivative of the bias is mathematically zero; but
+        # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero
+        # bias because the small numerical roundoff tends to have a non-random
+        # sign.  This module is intended to prevent that.  Use a very small
+        # probability; that should be suffixient to fix the problem.
+        self.balance_keys = Balancer(
+            key_head_dim * num_heads,
+            channel_dim=-1,
+            min_positive=0.4,
+            max_positive=0.6,
+            min_abs=0.0,
+            max_abs=100.0,
+            prob=0.025,
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(
+            pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05
+        )
+
+        # the following are for diagnosics only, see --print-diagnostics option
+        self.copy_pos_query = Identity()
+        self.copy_query = Identity()
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        chunk_size: int = -1,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""
+        Args:
+            x: input of shape (seq_len, batch_size, embed_dim)
+            pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 2, pos_dim)
+           chunk_size
+            key_padding_mask: a bool tensor of shape (batch_size, seq_len).  Positions that
+               are True in this mask will be ignored as sources in the attention weighting.
+            attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len),
+               interpreted as ([batch_size,] tgt_seq_len, src_seq_len)
+               saying which positions are allowed to attend to which other positions.
+        Returns:
+           a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len)
+           interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len).
+        """
+        x = self.in_proj(x)
+        query_head_dim = self.query_head_dim
+        pos_head_dim = self.pos_head_dim
+        num_heads = self.num_heads
+
+        seq_len, batch_size, _ = x.shape
+
+        query_dim = query_head_dim * num_heads
+
+        q = x[..., 0:query_dim]
+        k = x[..., query_dim : 2 * query_dim]
+        # p is the position-encoding query
+        p = x[..., 2 * query_dim :]
+        assert p.shape[-1] == num_heads * pos_head_dim
+
+        q = self.copy_query(q)  # for diagnostics only, does nothing.
+        k = self.whiten_keys(self.balance_keys(k))  # does nothing in the forward pass.
+        p = self.copy_pos_query(p)  # for diagnostics only, does nothing.
+
+        q = q.reshape(seq_len, batch_size, num_heads, query_head_dim)
+        p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim)
+        k = k.reshape(seq_len, batch_size, num_heads, query_head_dim)
+
+        # time1 refers to target, time2 refers to source.
+        q = q.permute(2, 1, 0, 3)  # (head, batch, time1, query_head_dim)
+        p = p.permute(2, 1, 0, 3)  # (head, batch, time1, pos_head_dim)
+        k = k.permute(2, 1, 3, 0)  # (head, batch, d_k, time2)
+
+        attn_scores = torch.matmul(q, k)
+
+        if not self.training or random.random() >= float(self.pos_emb_skip_rate):
+            pos_emb = self.linear_pos(pos_emb)
+            seq_len2 = 2 * seq_len - 1
+            pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute(
+                2, 0, 3, 1
+            )
+            # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2)
+
+            # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2)
+            #  [where seq_len2 represents relative position.]
+            pos_scores = torch.matmul(p, pos_emb)
+            # the following .as_strided() expression converts the last axis of pos_scores from relative
+            # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+            # not, but let this code define which way round it is supposed to be.
+            pos_scores = pos_scores.as_strided(
+                (num_heads, batch_size, seq_len, seq_len),
+                (
+                    pos_scores.stride(0),
+                    pos_scores.stride(1),
+                    pos_scores.stride(2) - pos_scores.stride(3),
+                    pos_scores.stride(3),
+                ),
+                storage_offset=pos_scores.stride(3) * (seq_len - 1),
+            )
+
+            attn_scores = attn_scores + pos_scores
+
+        if self.training and random.random() < 0.1:
+            # This is away of limiting the attention scores to not be
+            # too large.  It incurs a penalty if any of them has an absolute
+            # value greater than 25.0.  this should be outside the normal range
+            # of the attention scores.  We use this mechanism instead of, say,
+            # something added to the loss function involving the entropy,
+            # because once the entropy gets very small gradients through the
+            # softmax can become very small, and we'd get zero derivatives.  The
+            # choices of 1.0e-04 as the scale on the penalty makes this
+            # mechanism vulnerable to the absolute scale of the loss function,
+            # but we view this as a failsafe to avoid "implausible" parameter
+            # values rather than a regularization method that should be active
+            # under normal circumstances.
+            attn_scores = penalize_abs_values_gt(
+                attn_scores, limit=25.0, penalty=1.0e-04, name=self.name
+            )
+
+        assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        if attn_mask is not None:
+            assert attn_mask.dtype == torch.bool
+            # use -1000 to avoid nan's where attn_mask and key_padding_mask make
+            # all scores zero.  It's important that this be large enough that exp(-1000)
+            # is exactly zero, for reasons related to const_attention_rate, it
+            # compares the final weights with zero.
+            attn_scores = attn_scores.masked_fill(attn_mask, -1000)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (
+                batch_size,
+                seq_len,
+            ), key_padding_mask.shape
+            attn_scores = attn_scores.masked_fill(
+                key_padding_mask.unsqueeze(1),
+                -1000,
+            )
+
+        # We use our own version of softmax, defined in scaling.py, which should
+        # save a little of the memory used in backprop by, if we are in
+        # automatic mixed precision mode (amp / autocast), by only storing the
+        # half-precision output for backprop purposes.
+        attn_weights = softmax(attn_scores, dim=-1)
+
+        if random.random() < 0.001:
+            self._print_attn_entropy(attn_weights)
+
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )
+
+        return attn_weights
+
+    def _print_attn_entropy(self, attn_weights: Tensor):
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .mean(dim=(1, 2))
+                )
+                logging.info(
+                    f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}"
+                )
+
+
+class Attention(nn.Module):
+    """
+    The simplest possible attention module.  This one works with already-computed attention
+    weights, e.g. as computed by RelPositionMultiheadAttentionWeights.
+
+    Args:
+          embed_dim_in: the input embedding dimension
+          embed_dim_out: the output embedding dimension (normally the same as input)
+          num_heads: the number of attention heads
+          value_head_dim: the value dimension per head
+    """
+
+    def __init__(
+        self,
+        embed_dim_in: int,
+        embed_dim_out: int,
+        num_heads: int,
+        value_head_dim: int,
+    ) -> None:
+        super().__init__()
+        self.in_proj = nn.Linear(embed_dim_in, num_heads * value_head_dim, bias=False)
+
+        # Note we set bias to False so that input of 0 will have no effect
+        self.out_proj = ScaledLinear(
+            num_heads * value_head_dim, embed_dim_out, bias=False, initial_scale=0.05
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """
+        Args:
+          x: input tensor, of shape (seq_len, batch_size, embed_dim)
+         attn_weights: a tensor of shape (num_heads, batch_size, query_len, key_len),
+          Expect attn_weights.sum(dim=-1) == 1. The input here is the value in the
+          original attention mechanism.
+        Returns:
+           a tensor with the same shape as x.
+        """
+        (num_heads, batch_size, query_len, key_len) = attn_weights.shape
+
+        x = self.in_proj(x)  # (key_len, batch_size, num_heads * value_head_dim)
+        x = x.reshape(key_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, key_len, value_head_dim)
+        value_head_dim = x.shape[-1]
+
+        # todo: see whether there is benefit in overriding matmul
+        x = torch.matmul(attn_weights, x)
+        # v: (num_heads, batch_size, query_len, value_head_dim)
+
+        x = (
+            x.permute(2, 1, 0, 3)
+            .contiguous()
+            .view(query_len, batch_size, num_heads * value_head_dim)
+        )
+
+        # returned value is of shape (query_len, batch_size, embed_dim), like the input.
+        x = self.out_proj(x)
+        x = self.whiten(x)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_val: Tensor,
+        left_context_len: int,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            x: input tensor, of shape (seq_len, batch_size, embed_dim)
+            attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len),
+              with seq_len being interpreted as (tgt_seq_len, src_seq_len).  Expect
+              attn_weights.sum(dim=-1) == 1.
+            cached_val: cached attention value tensor of left context,
+              of shape (left_context_len, batch_size, value_dim)
+            left_context_len: number of left context frames.
+
+        Returns:
+           - attention weighted output, a tensor with the same shape as x.
+           - updated cached attention value tensor of left context.
+        """
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        seq_len2 = seq_len + left_context_len
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2)
+
+        x = self.in_proj(x)  # (seq_len, batch_size, num_heads * value_head_dim)
+
+        # Pad cached left contexts
+        assert cached_val.shape[0] == left_context_len, (
+            cached_val.shape[0],
+            left_context_len,
+        )
+        x = torch.cat([cached_val, x], dim=0)
+        # Update cached left contexts
+        cached_val = x[-left_context_len:, ...]
+
+        x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, value_head_dim)
+        value_head_dim = x.shape[-1]
+
+        # todo: see whether there is benefit in overriding matmul
+        x = torch.matmul(attn_weights, x)
+        # v: (num_heads, batch_size, seq_len, value_head_dim)
+
+        x = (
+            x.permute(2, 1, 0, 3)
+            .contiguous()
+            .view(seq_len, batch_size, num_heads * value_head_dim)
+        )
+
+        # returned value is of shape (seq_len, batch_size, embed_dim), like the input.
+        x = self.out_proj(x)
+
+        return x, cached_val
+
+
+class MultiheadAttentionWeights(nn.Module):
+    r"""Module that computes multi-head cross-attention weights.  Allows src and target
+    to have different dims.
+
+    Args:
+          key_embed_dim: number of channels of the thing that we'll project to
+              make the query (corresponds to source).  e.g. 256
+          query_embed_dim: number of channels of the thing that we'll project to
+              make the query (corresponds to target).  e.g. 256
+          num_heads:  number of heads to compute weights for, e.g. 8
+           head_dim: dimension of the query and key, per head.  e.g. 24.
+             dropout: dropout probability for attn_output_weights. Default: 0.0.
+    """
+
+    def __init__(
+        self,
+        key_embed_dim: int,
+        query_embed_dim: int,
+        num_heads: int,
+        head_dim: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super().__init__()
+        self.key_embed_dim = key_embed_dim
+        self.query_embed_dim = query_embed_dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+        self.dropout = dropout
+        self.name = None  # will be overwritten in training code; for diagnostics.
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5 that has been used in previous forms of attention,
+        # dividing it between the query and key.   Note: this module is intended
+        # to be used with the ScaledAdam optimizer; with most other optimizers,
+        # it would be necessary to apply the scaling factor in the forward function.
+        self.query_in_proj = ScaledLinear(
+            query_embed_dim,
+            head_dim * num_heads,
+            bias=True,
+            initial_scale=head_dim**-0.25,
+        )
+
+        # weights produced by this module are invariant to adding a constant to
+        # the keys, so we don't need a bias for the keys.
+        self.key_in_proj = ScaledLinear(
+            key_embed_dim,
+            head_dim * num_heads,
+            bias=False,
+            initial_scale=head_dim**-0.25,
+        )
+
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=_whitening_schedule(3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+    def forward(
+        self,
+        key: Tensor,
+        query: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""
+        Args:
+              key: input of shape (key_len, batch_size, key_embed_dim)
+            query: input of shape (query_len, batch_size, query_embed_dim)
+          key_padding_mask: an optional bool tensor of shape (batch_size, key_len).  Positions that
+               are True in this mask will be ignored as sources in the attention weighting.
+        Returns:
+           a tensor of attention weights, of shape (hum_heads, batch_size, query_len, key_len)
+        """
+        q = self.query_in_proj(query)
+        k = self.key_in_proj(key)
+
+        head_dim = self.head_dim
+        num_heads = self.num_heads
+
+        query_len, batch_size, _ = q.shape
+        key_len, _batch_size, _ = k.shape
+        assert _batch_size == batch_size
+
+        k = self.whiten_keys(k)  # does nothing in the forward pass.
+
+        q = q.reshape(query_len, batch_size, num_heads, head_dim)
+        k = k.reshape(key_len, batch_size, num_heads, head_dim)
+
+        # time1 refers to target, time2 refers to source.
+        q = q.permute(2, 1, 0, 3)  # (head, batch, time1, query_head_dim)
+        k = k.permute(2, 1, 3, 0)  # (head, batch, d_k, time2)
+
+        attn_scores = torch.matmul(q, k)
+
+        if self.training and random.random() < 0.1:
+            # This is a way of limiting the attention scores to not be
+            # too large.  It incurs a penalty if any of them has an absolute
+            # value greater than 25.0.  this should be outside the normal range
+            # of the attention scores.  We use this mechanism instead of, say,
+            # something added to the loss function involving the entropy,
+            # because once the entropy gets very small gradients through the
+            # softmax can become very small, and we'd get zero derivatives.  The
+            # choices of 1.0e-04 as the scale on the penalty makes this
+            # mechanism vulnerable to the absolute scale of the loss function,
+            # but we view this as a failsafe to avoid "implausible" parameter
+            # values rather than a regularization method that should be active
+            # under normal circumstances.
+            attn_scores = penalize_abs_values_gt(
+                attn_scores, limit=25.0, penalty=1.0e-04, name=self.name
+            )
+
+        assert attn_scores.shape == (num_heads, batch_size, query_len, key_len)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (
+                batch_size,
+                key_len,
+            ), key_padding_mask.shape
+            attn_scores = attn_scores.masked_fill(
+                key_padding_mask.unsqueeze(1),
+                -1000,
+            )
+
+        # We use our own version of softmax, defined in scaling.py, which should
+        # save a little of the memory used in backprop by, if we are in
+        # automatic mixed precision mode (amp / autocast), by only storing the
+        # half-precision output for backprop purposes.
+        attn_weights = softmax(attn_scores, dim=-1)
+
+        if random.random() < 0.001:
+            self._print_attn_entropy(attn_weights)
+
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )
+
+        return attn_weights
+
+    def _print_attn_entropy(self, attn_weights: Tensor):
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .mean(dim=(1, 2))
+                )
+                logging.info(
+                    f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}"
+                )
+
+
+class FeedforwardModule(nn.Module):
+    """Feedforward module in Zipformer2 model."""
+
+    def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike):
+        super(FeedforwardModule, self).__init__()
+        self.in_proj = nn.Linear(embed_dim, feedforward_dim)
+
+        self.hidden_balancer = Balancer(
+            feedforward_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=1.0,
+            min_abs=0.75,
+            max_abs=5.0,
+        )
+
+        # shared_dim=0 means we share the dropout mask along the time axis
+        self.out_proj = ActivationDropoutAndLinear(
+            feedforward_dim,
+            embed_dim,
+            activation="SwooshL",
+            dropout_p=dropout,
+            dropout_shared_dim=0,
+            bias=True,
+            initial_scale=0.1,
+        )
+
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(self, x: Tensor):
+        x = self.in_proj(x)
+        x = self.hidden_balancer(x)
+        # out_proj contains SwooshL activation, then dropout, then linear.
+        x = self.out_proj(x)
+        x = self.out_whiten(x)
+        return x
+
+
+class NonlinAttention(nn.Module):
+    """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed
+       from the attention module) in place of actual convolution.  We also took out the second nonlinearity, the
+       one after the attention mechanism.
+
+    Args:
+        channels (int): The number of channels of conv layers.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        hidden_channels: int,
+    ) -> None:
+        super().__init__()
+
+        self.hidden_channels = hidden_channels
+
+        self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True)
+
+        # balancer that goes before the sigmoid.  Have quite a large min_abs value, at 2.0,
+        # because we noticed that well-trained instances of this module have abs-value before the sigmoid
+        # starting from about 3, and poorly-trained instances of the module have smaller abs values
+        # before the sigmoid.
+        self.balancer = Balancer(
+            hidden_channels,
+            channel_dim=-1,
+            min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)),
+            max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)),
+            min_abs=0.5,
+            max_abs=5.0,
+        )
+        self.tanh = nn.Tanh()
+
+        self.identity1 = Identity()  # for diagnostics.
+        self.identity2 = Identity()  # for diagnostics.
+        self.identity3 = Identity()  # for diagnostics.
+
+        self.out_proj = ScaledLinear(
+            hidden_channels, channels, bias=True, initial_scale=0.05
+        )
+
+        self.whiten1 = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(5.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.whiten2 = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(5.0, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """.
+                Args:
+                   x: a Tensor of shape (seq_len, batch_size, num_channels)
+        attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
+                Returns:
+                   a Tensor with the same shape as x
+        """
+        x = self.in_proj(x)
+
+        (seq_len, batch_size, _) = x.shape
+        hidden_channels = self.hidden_channels
+
+        s, x, y = x.chunk(3, dim=-1)
+
+        # s will go through tanh.
+
+        s = self.balancer(s)
+        s = self.tanh(s)
+
+        s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels)
+        x = self.whiten1(x)
+        x = x * s
+        x = self.identity1(x)  # diagnostics only, it's the identity.
+
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = torch.matmul(attn_weights, x)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1)
+
+        y = self.identity2(y)
+        x = x * y
+        x = self.identity3(x)
+
+        x = self.out_proj(x)
+        x = self.whiten2(x)
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_x: Tensor,
+        left_context_len: int,
+    ) -> Tuple[Tensor, Tensor]:
+        """.
+        Args:
+            x: a Tensor of shape (seq_len, batch_size, num_channels)
+            attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
+            cached_x: left context, a Tensor of shape
+              (num_heads, batch_size, left_context_len, head_dim)
+            left_context_len: number of left context frames.
+        Returns:
+            - a Tensor with the same shape as x
+            - updated left context with same shape as cached_x
+        """
+        x = self.in_proj(x)
+
+        (seq_len, batch_size, _) = x.shape
+        hidden_channels = self.hidden_channels
+
+        s, x, y = x.chunk(3, dim=-1)
+
+        # s will go through tanh.
+        s = self.tanh(s)
+
+        s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels)
+        x = x * s
+
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (
+            num_heads,
+            batch_size,
+            seq_len,
+            left_context_len + seq_len,
+        )
+
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+
+        # Pad cached tensor
+        assert cached_x.shape[2] == left_context_len, (
+            cached_x.shape[2],
+            left_context_len,
+        )
+        x_pad = torch.cat([cached_x, x], dim=2)
+        # Update cached tensor
+        cached_x = x_pad[:, :, -left_context_len:, :]
+
+        x = torch.matmul(attn_weights, x_pad)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1)
+
+        x = x * y
+
+        x = self.out_proj(x)
+        return x, cached_x
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Zipformer2 model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        causal: bool,
+    ) -> None:
+        """Construct a ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        bottleneck_dim = channels
+        self.causal = causal
+
+        self.in_proj = nn.Linear(
+            channels,
+            2 * bottleneck_dim,
+        )
+        # the gradients on in_proj are a little noisy, likely to do with the
+        # sigmoid in glu.
+
+        # after in_proj we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in
+        # the range 1 to 4, but sometimes, for some reason, for layer 0 the
+        # rms ends up being very large, between 50 and 100 for different channels.
+        # This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.balancer1 = Balancer(
+            bottleneck_dim,
+            channel_dim=-1,
+            min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)),
+            max_positive=1.0,
+            min_abs=1.5,
+            max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0),
+        )
+
+        self.activation1 = Identity()  # for diagnostics
+
+        self.sigmoid = nn.Sigmoid()
+
+        self.activation2 = Identity()  # for diagnostics
+
+        assert kernel_size % 2 == 1
+
+        self.depthwise_conv = (
+            ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size)
+            if causal
+            else nn.Conv1d(
+                in_channels=bottleneck_dim,
+                out_channels=bottleneck_dim,
+                groups=bottleneck_dim,
+                kernel_size=kernel_size,
+                padding=kernel_size // 2,
+            )
+        )
+
+        self.balancer2 = Balancer(
+            bottleneck_dim,
+            channel_dim=1,
+            min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)),
+            max_positive=1.0,
+            min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)),
+            max_abs=10.0,
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.out_proj = ActivationDropoutAndLinear(
+            bottleneck_dim,
+            channels,
+            activation="SwooshR",
+            dropout_p=0.0,
+            initial_scale=0.05,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+        chunk_size: int = -1,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+           src_key_padding_mask: the mask for the src keys per batch (optional):
+               (batch, #time), contains True in masked positions.
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+
+        x = self.in_proj(x)  # (time, batch, 2*channels)
+
+        x, s = x.chunk(2, dim=-1)
+        s = self.balancer1(s)
+        s = self.sigmoid(s)
+        x = self.activation1(x)  # identity.
+        x = x * s
+        x = self.activation2(x)  # identity
+
+        # (time, batch, channels)
+
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        if (
+            not torch.jit.is_scripting()
+            and not torch.jit.is_tracing()
+            and chunk_size >= 0
+        ):
+            # Not support exporting a model for simulated streaming decoding
+            assert (
+                self.causal
+            ), "Must initialize model with causal=True if you use chunk_size"
+            x = self.depthwise_conv(x, chunk_size=chunk_size)
+        else:
+            x = self.depthwise_conv(x)
+
+        x = self.balancer2(x)
+        x = x.permute(2, 0, 1)  # (time, batch, channels)
+
+        x = self.whiten(x)  # (time, batch, channels)
+        x = self.out_proj(x)  # (time, batch, channels)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cache: Tensor,
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module in streaming forward mode.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: cached left context for depthwise_conv of shape
+              (#batch, channels, left_pad)
+            src_key_padding_mask: the mask for the src keys per batch (optional):
+              (batch, #time), contains True in masked positions.
+
+        Returns:
+            - Output tensor (#time, batch, channels).
+            - Updated cache (#batch, channels, left_pad)
+        """
+
+        x = self.in_proj(x)  # (time, batch, 2*channels)
+
+        x, s = x.chunk(2, dim=2)
+        s = self.sigmoid(s)
+        x = x * s
+        # (time, batch, channels)
+
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        x, cache = self.depthwise_conv.streaming_forward(x, cache=cache)
+
+        x = x.permute(2, 0, 1)  # (time, batch, channels)
+
+        x = self.out_proj(x)  # (time, batch, channels)
+
+        return x, cache
+
+
+class ScalarMultiply(nn.Module):
+    def __init__(self, scale: float):
+        super().__init__()
+        self.scale = scale
+
+    def forward(self, x):
+        return x * self.scale
+
+
+def _test_zipformer_main(causal: bool = False):
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    memory_dim = 100
+
+    c = Zipformer2(
+        encoder_dim=(64, 96),
+        encoder_unmasked_dim=(48, 64),
+        num_heads=(4, 4),
+        causal=causal,
+        chunk_size=(4,) if causal else (-1,),
+        left_context_frames=(64,),
+        memory_dim=memory_dim,
+    )
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(seq_len, batch_size, 64),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        memory=torch.randn(101, batch_size, memory_dim),
+    )
+    f[0].sum().backward()
+    c.eval()
+    f = c(
+        torch.randn(seq_len, batch_size, 64),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_zipformer_main(False)
+    _test_zipformer_main(True)
diff --git a/egs/librispeech/ASR/.gitignore b/egs/librispeech/ASR/.gitignore
new file mode 100644
index 000000000..1c26f7978
--- /dev/null
+++ b/egs/librispeech/ASR/.gitignore
@@ -0,0 +1,3 @@
+log-*
+.DS_Store
+run*.sh
diff --git a/egs/librispeech/ASR/README.md b/egs/librispeech/ASR/README.md
new file mode 100644
index 000000000..1c8930818
--- /dev/null
+++ b/egs/librispeech/ASR/README.md
@@ -0,0 +1,58 @@
+# Introduction
+
+Please refer to <https://icefall.readthedocs.io/en/latest/recipes/Non-streaming-ASR/librispeech/index.html> for how to run models in this recipe.
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                                           |
+|---------------------------------------|---------------------|--------------------|---------------------------------------------------|
+| `transducer`                          | Conformer           | LSTM               |                                                   |
+| `transducer_stateless`                | Conformer           | Embedding + Conv1d | Using optimized_transducer from computing RNN-T loss  |
+| `transducer_stateless2`               | Conformer           | Embedding + Conv1d | Using torchaudio for computing RNN-T loss             |
+| `transducer_lstm`                     | LSTM                | LSTM               |                                                   |
+| `transducer_stateless_multi_datasets` | Conformer           | Embedding + Conv1d | Using data from GigaSpeech as extra training data |
+| `pruned_transducer_stateless`         | Conformer           | Embedding + Conv1d | Using k2 pruned RNN-T loss                        |
+| `pruned_transducer_stateless2`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss                        |
+| `pruned_transducer_stateless3`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data |
+| `pruned_transducer_stateless4`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless2 + save averaged models periodically during training + delay penalty |
+| `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
+| `pruned_transducer_stateless6`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert|
+| `pruned_transducer_stateless7`        | Zipformer | Embedding + Conv1d | First experiment with Zipformer from Dan|
+| `pruned_transducer_stateless7_ctc`    | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but with extra CTC head|
+| `pruned_transducer_stateless7_ctc_bs` | Zipformer | Embedding + Conv1d | pruned_transducer_stateless7_ctc + blank skip |
+| `pruned_transducer_stateless7_streaming` | Streaming Zipformer | Embedding + Conv1d | streaming version of pruned_transducer_stateless7 |
+| `pruned_transducer_stateless7_streaming_multi` | Streaming Zipformer | Embedding + Conv1d | same as pruned_transducer_stateless7_streaming, trained on LibriSpeech + GigaSpeech  |
+| `pruned_transducer_stateless8`        | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but using extra data from GigaSpeech|
+| `pruned_stateless_emformer_rnnt2`     | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
+| `conv_emformer_transducer_stateless`  | ConvEmformer | Embedding + Conv1d | Using ConvEmformer for streaming ASR + mechanisms in reworked model |
+| `conv_emformer_transducer_stateless2` | ConvEmformer | Embedding + Conv1d | Using ConvEmformer with simplified memory for streaming ASR + mechanisms in reworked model |
+| `lstm_transducer_stateless`           | LSTM | Embedding + Conv1d | Using LSTM with mechanisms in reworked model |
+| `lstm_transducer_stateless2`          | LSTM | Embedding + Conv1d | Using LSTM with mechanisms in reworked model + gigaspeech (multi-dataset setup) |
+| `lstm_transducer_stateless3`          | LSTM | Embedding + Conv1d | Using LSTM with mechanisms in reworked model + gradient filter + delay penalty |
+| `zipformer`                           | Upgraded Zipformer | Embedding + Conv1d | The latest recipe |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
+
+# CTC
+
+|                              | Encoder            | Comment                      |
+|------------------------------|--------------------|------------------------------|
+| `conformer-ctc`              | Conformer          | Use auxiliary attention head |
+| `conformer-ctc2`             | Reworked Conformer | Use auxiliary attention head |
+| `conformer-ctc3`             | Reworked Conformer | Streaming version + delay penalty |
+| `zipformer-ctc`              | Zipformer          | Use auxiliary attention head |
+| `zipformer`                  | Upgraded Zipformer | Use auxiliary transducer head | The latest recipe |
+
+# MMI
+
+|                              | Encoder   | Comment                                           |
+|------------------------------|-----------|---------------------------------------------------|
+| `conformer-mmi`              | Conformer |                                                   |
+| `zipformer-mmi`              | Zipformer | CTC warmup + use HP as decoding graph for decoding |
diff --git a/egs/librispeech/ASR/RESULTS-100hours.md b/egs/librispeech/ASR/RESULTS-100hours.md
new file mode 100644
index 000000000..3a064e69d
--- /dev/null
+++ b/egs/librispeech/ASR/RESULTS-100hours.md
@@ -0,0 +1,102 @@
+# Results for train-clean-100
+
+This page shows the WERs for test-clean/test-other using only
+train-clean-100 subset as training data.
+
+## Distillation with hubert
+### 2022-05-27
+Related models/log/tensorboard:
+https://huggingface.co/GuoLiyong/stateless6_baseline_vs_disstillation
+
+Following results are obtained by ./distillation_with_hubert.sh
+
+The only differences is in pruned_transducer_stateless6/train.py.
+
+For baseline: set enable_distillation=False
+
+For distillation: set enable_distillation=True (the default)
+
+Decoding method is modified beam search.
+|                                     | test-clean | test-other | comment                                  |
+|-------------------------------------|------------|------------|------------------------------------------|
+| baseline no vq distillation         | 7.09       | 18.88      | --epoch 20, --avg 10, --max-duration 200 |
+| baseline no vq distillation         | 6.83       | 18.19      | --epoch 30, --avg 10, --max-duration 200 |
+| baseline no vq distillation         | 6.73       | 17.79      | --epoch 40, --avg 10, --max-duration 200 |
+| baseline no vq distillation         | 6.75       | 17.68      | --epoch 50, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.82       | 15.98      | --epoch 20, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.52       | 15.15      | --epoch 30, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.45       | 14.94      | --epoch 40, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.50       | 14.77      | --epoch 50, --avg 10, --max-duration 200 |
+
+## Conformer encoder + embedding decoder
+
+### 2022-02-21
+
+Using commit `2332ba312d7ce72f08c7bac1e3312f7e3dd722dc`.
+
+|                                     | test-clean | test-other | comment                                  |
+|-------------------------------------|------------|------------|------------------------------------------|
+| greedy search (max sym per frame 1) | 6.34       | 16.7       | --epoch 57, --avg 17, --max-duration 100 |
+| greedy search (max sym per frame 2) | 6.34       | 16.7       | --epoch 57, --avg 17, --max-duration 100 |
+| greedy search (max sym per frame 3) | 6.34       | 16.7       | --epoch 57, --avg 17, --max-duration 100 |
+| modified beam search (beam size 4)  | 6.31       | 16.3       | --epoch 57, --avg 17, --max-duration 100 |
+
+
+The training command for reproducing is given below:
+
+```bash
+cd egs/librispeech/ASR/
+./prepare.sh
+./prepare_giga_speech.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./transducer_stateless_multi_datasets/train.py \
+  --world-size 2 \
+  --num-epochs 60 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_multi_datasets/exp-100-2 \
+  --full-libri 0 \
+  --max-duration 300 \
+  --lr-factor 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --modified-transducer-prob 0.25
+  --giga-prob 0.2
+```
+
+The decoding command is given below:
+
+```bash
+for epoch in 57; do
+  for avg in 17; do
+    for sym in 1 2 3; do
+    ./transducer_stateless_multi_datasets/decode.py \
+      --epoch $epoch \
+      --avg $avg \
+      --exp-dir transducer_stateless_multi_datasets/exp-100-2 \
+      --bpe-model ./data/lang_bpe_500/bpe.model \
+      --max-duration 100 \
+      --context-size 2 \
+      --max-sym-per-frame $sym
+    done
+  done
+done
+
+epoch=57
+avg=17
+./transducer_stateless_multi_datasets/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless_multi_datasets/exp-100-2 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --context-size 2 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+```
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/qUEKzMnrTZmOz1EXPda9RA/>
+
+A pre-trained model and decoding logs can be found at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-100h-transducer-stateless-multi-datasets-bpe-500-2022-02-21>
diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md
new file mode 100644
index 000000000..ebf5e89c4
--- /dev/null
+++ b/egs/librispeech/ASR/RESULTS.md
@@ -0,0 +1,4086 @@
+## Results
+
+### zipformer (zipformer + pruned stateless transducer + CTC)
+
+See <https://github.com/k2-fsa/icefall/pull/1111> for more details.
+
+[zipformer](./zipformer)
+
+#### Non-streaming
+
+##### normal-scaled model, number of model parameters: 65805511, i.e., 65.81 M
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/Lo3Qlad7TP68ulM2K0ixgQ/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Results of the CTC head:
+
+| decoding method         | test-clean | test-other | comment            |
+|-------------------------|------------|------------|--------------------|
+| ctc-decoding            | 2.40       | 5.66       | --epoch 40 --avg 16 |
+| 1best                   | 2.46       | 5.11       | --epoch 40 --avg 16 |
+| nbest                   | 2.46       | 5.11       | --epoch 40 --avg 16 |
+| nbest-rescoring         | 2.37       | 4.93       | --epoch 40 --avg 16 |
+| whole-lattice-rescoring | 2.37       | 4.88       | --epoch 40 --avg 16 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-ctc-rnnt \
+  --causal 0 \
+  --use-transducer 1 \
+  --use-ctc 1 \
+  --ctc-loss-scale 0.2 \
+  --full-libri 1 \
+  --max-duration 1000
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in ctc-decoding 1best nbest nbest-rescoring whole-lattice-rescoring; do
+  ./zipformer/ctc_decode.py \
+      --epoch 40 \
+      --avg 16 \
+      --exp-dir zipformer/exp-ctc-rnnt \
+      --use-transducer 1 \
+      --use-ctc 1 \
+      --max-duration 300 \
+      --causal 0 \
+      --num-paths 100 \
+      --nbest-scale 1.0 \
+      --hlg-scale 0.6 \
+      --decoding-method $m
+done
+```
+
+### zipformer (zipformer + pruned stateless transducer)
+
+See <https://github.com/k2-fsa/icefall/pull/1058> for more details.
+
+[zipformer](./zipformer)
+
+#### Non-streaming
+
+##### normal-scaled model, number of model parameters: 65549011, i.e., 65.55 M
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/R2DT9Ju4QiadC4e2ioKh5A/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 2.27       | 5.1        | --epoch 30 --avg 9 |
+| modified_beam_search | 2.25       | 5.06       | --epoch 30 --avg 9 |
+| fast_beam_search     | 2.25       | 5.04       | --epoch 30 --avg 9 |
+| greedy_search        | 2.23       | 4.96       | --epoch 40 --avg 16 |
+| modified_beam_search | 2.21       | 4.91       | --epoch 40 --avg 16 |
+| fast_beam_search     | 2.24       | 4.93       | --epoch 40 --avg 16 |
+| greedy_search        | 2.22       | 4.87       | --epoch 50 --avg 25 |
+| modified_beam_search | 2.21       | 4.79       | --epoch 50 --avg 25 |
+| fast_beam_search     | 2.21       | 4.82       | --epoch 50 --avg 25 |
+| modified_beam_search_shallow_fusion | 2.01 | 4.37 | --epoch 40 --avg 16 --beam-size 12 --lm-scale 0.3 |
+| modified_beam_search_LODR | 1.94 | 4.17 | --epoch 40 --avg 16 --beam-size 12 --lm-scale 0.52 --LODR-scale -0.26 |
+| modified_beam_search_rescore | 2.04 | 4.39 | --epoch 40 --avg 16 --beam-size 12 |
+| modified_beam_search_rescore_LODR | 2.01  | 4.33 | --epoch 40 --avg 16 --beam-size 12  |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 50 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 0 \
+  --full-libri 1 \
+  --max-duration 1000
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./zipformer/decode.py \
+    --epoch 50 \
+    --avg 25 \
+    --use-averaged-model 1 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method $m
+done
+```
+
+To decode with external language models, please refer to the documentation [here](https://k2-fsa.github.io/icefall/decoding-with-langugage-models/index.html).
+
+##### small-scaled model, number of model parameters: 23285615, i.e., 23.3 M
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/M9C8cYPWSN2MVBYaBIX3EQ/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-small-2023-05-16>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 2.64       | 6.14       | --epoch 30 --avg 8 |
+| modified_beam_search | 2.6        | 6.01       | --epoch 30 --avg 8 |
+| fast_beam_search     | 2.62       | 6.06       | --epoch 30 --avg 8 |
+| greedy_search        | 2.49       | 5.91       | --epoch 40 --avg 13 |
+| modified_beam_search | 2.46       | 5.83       | --epoch 40 --avg 13 |
+| fast_beam_search     | 2.46       | 5.87       | --epoch 40 --avg 13 |
+| greedy_search        | 2.46       | 5.86       | --epoch 50 --avg 23 |
+| modified_beam_search | 2.42       | 5.73       | --epoch 50 --avg 23 |
+| fast_beam_search     | 2.46       | 5.78       | --epoch 50 --avg 23 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1"
+./zipformer/train.py \
+  --world-size 2 \
+  --num-epochs 50 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-small \
+  --causal 0 \
+  --num-encoder-layers 2,2,2,2,2,2 \
+  --feedforward-dim 512,768,768,768,768,768 \
+  --encoder-dim 192,256,256,256,256,256 \
+  --encoder-unmasked-dim 192,192,192,192,192,192 \
+  --base-lr 0.04 \
+  --full-libri 1 \
+  --max-duration 1500
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./zipformer/decode.py \
+    --epoch 50 \
+    --avg 23 \
+    --exp-dir zipformer/exp-small \
+    --max-duration 600 \
+    --causal 0 \
+    --decoding-method $m \
+    --num-encoder-layers 2,2,2,2,2,2 \
+    --feedforward-dim 512,768,768,768,768,768 \
+    --encoder-dim 192,256,256,256,256,256 \
+    --encoder-unmasked-dim 192,192,192,192,192,192
+done
+```
+
+##### large-scaled model, number of model parameters: 148439574, i.e., 148.4 M
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/C5ZPE5u1So2ZwhYLKW0FVg/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-large-2023-05-16>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 2.12       | 4.91       | --epoch 30 --avg 9 |
+| modified_beam_search | 2.11       | 4.9        | --epoch 30 --avg 9 |
+| fast_beam_search     | 2.13       | 4.93       | --epoch 30 --avg 9 |
+| greedy_search        | 2.12       | 4.8        | --epoch 40 --avg 13 |
+| modified_beam_search | 2.11       | 4.7        | --epoch 40 --avg 13 |
+| fast_beam_search     | 2.13       | 4.78       | --epoch 40 --avg 13 |
+| greedy_search        | 2.08       | 4.69       | --epoch 50 --avg 30 |
+| modified_beam_search | 2.06       | 4.63       | --epoch 50 --avg 30 |
+| fast_beam_search     | 2.09       | 4.68       | --epoch 50 --avg 30 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 50 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-large \
+  --causal 0 \
+  --num-encoder-layers 2,2,4,5,4,2 \
+  --feedforward-dim 512,768,1536,2048,1536,768 \
+  --encoder-dim 192,256,512,768,512,256 \
+  --encoder-unmasked-dim 192,192,256,320,256,192 \
+  --full-libri 1 \
+  --max-duration 1000
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./zipformer/decode.py \
+    --epoch 50 \
+    --avg 30 \
+    --exp-dir zipformer/exp-large \
+    --max-duration 600 \
+    --causal 0 \
+    --decoding-method $m \
+    --num-encoder-layers 2,2,4,5,4,2 \
+    --feedforward-dim 512,768,1536,2048,1536,768 \
+    --encoder-dim 192,256,512,768,512,256 \
+    --encoder-unmasked-dim 192,192,256,320,256,192
+done
+```
+
+##### large-scaled model, number of model parameters: 148439574, i.e., 148.4 M, trained on 8 80G-A100 GPUs
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/95TdNyEuQXaWK2PzFpD9yg/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-large-2023-10-26-8-a100>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | test-clean | test-other | comment               |
+|----------------------|------------|------------|-----------------------|
+| greedy_search        | 2.00       | 4.47       | --epoch 174 --avg 172 |
+| modified_beam_search | 2.00       | 4.38       | --epoch 174 --avg 172 |
+| fast_beam_search     | 2.00       | 4.42       | --epoch 174 --avg 172 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 174 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-large \
+  --causal 0 \
+  --num-encoder-layers 2,2,4,5,4,2 \
+  --feedforward-dim 512,768,1536,2048,1536,768 \
+  --encoder-dim 192,256,512,768,512,256 \
+  --encoder-unmasked-dim 192,192,256,320,256,192 \
+  --full-libri 1 \
+  --max-duration 2200
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./zipformer/decode.py \
+    --epoch 174 \
+    --avg 172 \
+    --exp-dir zipformer/exp-large \
+    --max-duration 600 \
+    --causal 0 \
+    --decoding-method $m \
+    --num-encoder-layers 2,2,4,5,4,2 \
+    --feedforward-dim 512,768,1536,2048,1536,768 \
+    --encoder-dim 192,256,512,768,512,256 \
+    --encoder-unmasked-dim 192,192,256,320,256,192
+done
+```
+
+#### streaming
+
+##### normal-scaled model, number of model parameters: 66110931, i.e., 66.11 M
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/9rD0i6rMSWq1O61poWi71A>
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | chunk size | test-clean | test-other | decoding mode       | comment            |
+|----------------------|------------|------------|------------|---------------------|--------------------|
+| greedy_search        | 320ms      | 3.06       | 7.81       | simulated streaming | --epoch 30 --avg 8 --chunk-size 16 --left-context-frames 128 |
+| greedy_search        | 320ms      | 3.06       | 7.79       | chunk-wise          | --epoch 30 --avg 8 --chunk-size 16 --left-context-frames 128 |
+| modified_beam_search | 320ms      | 3.01       | 7.69       | simulated streaming | --epoch 30 --avg 8 --chunk-size 16 --left-context-frames 128 |
+| modified_beam_search | 320ms      | 3.05       | 7.69       | chunk-wise          | --epoch 30 --avg 8 --chunk-size 16 --left-context-frames 128 |
+| fast_beam_search     | 320ms      | 3.04       | 7.68       | simulated streaming | --epoch 30 --avg 8 --chunk-size 16 --left-context-frames 128 |
+| fast_beam_search     | 320ms      | 3.07       | 7.69       | chunk-wise          | --epoch 30 --avg 8 --chunk-size 16 --left-context-frames 128 |
+| greedy_search        | 640ms      | 2.81       | 7.15       | simulated streaming | --epoch 30 --avg 8 --chunk-size 32 --left-context-frames 256 |
+| greedy_search        | 640ms      | 2.84       | 7.16       | chunk-wise          | --epoch 30 --avg 8 --chunk-size 32 --left-context-frames 256 |
+| modified_beam_search | 640ms      | 2.79       | 7.05       | simulated streaming | --epoch 30 --avg 8 --chunk-size 32 --left-context-frames 256 |
+| modified_beam_search | 640ms      | 2.81       | 7.11       | chunk-wise          | --epoch 30 --avg 8 --chunk-size 32 --left-context-frames 256 |
+| fast_beam_search     | 640ms      | 2.84       | 7.04       | simulated streaming | --epoch 30 --avg 8 --chunk-size 32 --left-context-frames 256 |
+| fast_beam_search     | 640ms      | 2.83       | 7.1        | chunk-wise          | --epoch 30 --avg 8 --chunk-size 32 --left-context-frames 256 |
+
+Note: For decoding mode, `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
+  while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-causal \
+  --causal 1 \
+  --full-libri 1 \
+  --max-duration 1000
+```
+
+The simulated streaming decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./zipformer/decode.py \
+    --epoch 30 \
+    --avg 8 \
+    --use-averaged-model 1 \
+    --exp-dir ./zipformer/exp-causal \
+    --causal 1 \
+    --chunk-size 16 \
+    --left-context-frames 128 \
+    --max-duration 600 \
+    --decoding-method $m
+done
+```
+
+The chunk-wise streaming decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./zipformer/streaming_decode.py \
+    --epoch 30 \
+    --avg 8 \
+    --use-averaged-model 1 \
+    --exp-dir ./zipformer/exp-causal \
+    --causal 1 \
+    --chunk-size 16 \
+    --left-context-frames 128 \
+    --num-decode-streams 2000 \
+    --decoding-method $m
+done
+```
+
+### Zipformer CTC
+
+#### [zipformer_ctc](./zipformer_ctc)
+
+See <https://github.com/k2-fsa/icefall/pull/941> for more details.
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/desh2608/icefall-asr-librispeech-zipformer-ctc>
+
+Number of model parameters: 86083707, i.e., 86.08 M
+
+| decoding method         | test-clean | test-other | comment             |
+|-------------------------|------------|------------|---------------------|
+| ctc-decoding            | 2.50       | 5.86       | --epoch 30 --avg 9  |
+| whole-lattice-rescoring | 2.44       | 5.38       | --epoch 30 --avg 9  |
+| attention-rescoring     | 2.35       | 5.16       | --epoch 30 --avg 9  |
+| 1best                   | 2.01       | 4.61       | --epoch 30 --avg 9  |
+
+The training commands are:
+```bash
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_ctc/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_ctc/exp \
+  --full-libri 1 \
+  --max-duration 1000 \
+  --master-port 12345
+```
+
+The tensorboard log can be found at:
+<https://tensorboard.dev/experiment/IjPSJjHOQFKPYA5Z0Vf8wg>
+
+The decoding command is:
+
+```bash
+./zipformer_ctc/decode.py \
+  --epoch 30 --avg 9 --use-averaged-model True \
+  --exp-dir zipformer_ctc/exp \
+  --lang-dir data/lang_bpe_500 \
+  --lm-dir data/lm \
+  --method ctc-decoding
+```
+
+### pruned_transducer_stateless7 (Fine-tune with mux)
+
+See <https://github.com/k2-fsa/icefall/pull/1059> for more details.
+
+[pruned_transducer_stateless7](./pruned_transducer_stateless7)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/MaNDZfO7RzW2Czzf3R2ZRA/>
+
+You can find the pretrained model and bpe model needed for fine-tuning at:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11>
+
+You can find a fine-tuned model, fine-tuning logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/yfyeung/icefall-asr-finetune-mux-pruned_transducer_stateless7-2023-05-19>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+| decoding method      |    dev     |    test    | test-clean | test-other |      comment       |
+|----------------------|------------|------------|------------|------------|--------------------|
+| greedy_search        |   14.27    |   14.22    |    2.08    |    4.79    | --epoch 20 --avg 5 |
+| modified_beam_search |   14.22    |   14.08    |    2.06    |    4.72    | --epoch 20 --avg 5 |
+| fast_beam_search     |   14.23    |   14.17    |    2.08    |    4.09    | --epoch 20 --avg 5 |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless7/finetune.py \
+  --world-size 2 \
+  --num-epochs 20 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp_giga_finetune \
+  --subset S \
+  --use-fp16 1 \
+  --base-lr 0.005 \
+  --lr-epochs 100 \
+  --lr-batches 100000 \
+  --bpe-model icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/data/lang_bpe_500/bpe.model \
+  --do-finetune True \
+  --use-mux True \
+  --finetune-ckpt icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp/pretrain.pt \
+  --max-duration 500
+```
+
+The decoding commands are:
+```bash
+# greedy_search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 20 \
+    --avg 5 \
+    --use-averaged-model 1 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+# modified_beam_search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 20 \
+    --avg 5 \
+    --use-averaged-model 1 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+# fast_beam_search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 20 \
+    --avg 5 \
+    --use-averaged-model 1 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+```
+
+### pruned_transducer_stateless7 (zipformer + multidataset(LibriSpeech + GigaSpeech + CommonVoice 13.0))
+
+See <https://github.com/k2-fsa/icefall/pull/1010> for more details.
+
+[pruned_transducer_stateless7](./pruned_transducer_stateless7)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/SwdJoHgZSZWn8ph9aJLb8g/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/yfyeung/icefall-asr-multidataset-pruned_transducer_stateless7-2023-05-04>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 1.91       | 4.06       | --epoch 30 --avg 7 |
+| modified_beam_search | 1.90       | 3.99       | --epoch 30 --avg 7 |
+| fast_beam_search     | 1.90       | 3.98       | --epoch 30 --avg 7 |
+
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --use-multidataset 1 \
+  --use-fp16 1 \
+  --max-duration 750 \
+  --exp-dir pruned_transducer_stateless7/exp
+```
+
+The decoding commands are:
+```bash
+# greedy_search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 30 \
+    --avg 7 \
+    --use-averaged-model 1 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+# modified_beam_search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 30 \
+    --avg 7 \
+    --use-averaged-model 1 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+# fast_beam_search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 30 \
+    --avg 7 \
+    --use-averaged-model 1 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+```
+
+### Streaming Zipformer-Transducer (Pruned Stateless Transducer + Streaming Zipformer + Multi-Dataset)
+
+#### [pruned_transducer_stateless7_streaming_multi](./pruned_transducer_stateless7_streaming_multi)
+
+See <https://github.com/k2-fsa/icefall/pull/984> for more details.
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at: <https://huggingface.co/marcoyang/icefall-libri-giga-pruned-transducer-stateless7-streaming-2023-04-04>
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+##### training on full librispeech + full gigaspeech (with giga_prob=0.9)
+
+The WERs are:
+
+
+| decoding method      | chunk size | test-clean | test-other | comment             | decoding mode        |
+|----------------------|------------|------------|------------|---------------------|----------------------|
+| greedy search        | 320ms      | 2.43       | 6.0        | --epoch 20 --avg 4  | simulated streaming  |
+| greedy search        | 320ms      | 2.47       | 6.13       | --epoch 20 --avg 4  | chunk-wise           |
+| fast beam search     | 320ms      | 2.43       | 5.99       | --epoch 20 --avg 4  | simulated streaming  |
+| fast beam search     | 320ms      | 2.8        | 6.46       | --epoch 20 --avg 4  | chunk-wise           |
+| modified beam search | 320ms      | 2.4        | 5.96       | --epoch 20 --avg 4  | simulated streaming  |
+| modified beam search | 320ms      | 2.42       | 6.03       | --epoch 20 --avg 4  | chunk-size           |
+| greedy search        | 640ms      | 2.26       | 5.58       | --epoch 20 --avg 4  | simulated streaming  |
+| greedy search        | 640ms      | 2.33       | 5.76       | --epoch 20 --avg 4  | chunk-wise           |
+| fast beam search     | 640ms      | 2.27       | 5.54       | --epoch 20 --avg 4  | simulated streaming  |
+| fast beam search     | 640ms      | 2.37       | 5.75       | --epoch 20 --avg 4  | chunk-wise           |
+| modified beam search | 640ms      | 2.22       | 5.5        | --epoch 20 --avg 4  | simulated streaming  |
+| modified beam search | 640ms      | 2.25       | 5.69       | --epoch 20 --avg 4  | chunk-size           |
+
+The model also has good WERs on GigaSpeech. The following WERs are achieved on GigaSpeech test and dev sets:
+
+| decoding method      | chunk size | dev        | test | comment    | decoding mode       |
+|----------------------|------------|-----|------|------------|---------------------|
+| greedy search        | 320ms      | 12.08       | 11.98       | --epoch 20 --avg 4  | simulated streaming  |
+| greedy search        | 640ms      | 11.66       | 11.71       | --epoch 20 --avg 4  | simulated streaming  |
+| modified beam search | 320ms      | 11.95       | 11.83       | --epoch 20 --avg 4  | simulated streaming  |
+| modified beam search | 320ms      | 11.65       | 11.56       | --epoch 20 --avg 4  | simulated streaming  |
+
+
+Note: `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
+while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless7_streaming_multi/train.py \
+  --world-size 4 \
+  --num-epochs 20 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming_multi/exp \
+  --full-libri 1 \
+  --giga-prob 0.9 \
+  --max-duration 750 \
+  --master-port 12345
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/G4yDMLXGQXexf41i4MA2Tg/#scalars>
+
+The simulated streaming decoding command (e.g., chunk-size=320ms) is:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  ./pruned_transducer_stateless7_streaming_multi/decode.py \
+    --epoch 20 \
+    --avg 4 \
+    --exp-dir ./pruned_transducer_stateless7_streaming_multi/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --right-padding 64 \
+    --decoding-method $m
+done
+```
+
+The streaming chunk-size decoding command (e.g., chunk-size=320ms) is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless7_streaming_multi/streaming_decode.py \
+    --epoch 20 \
+    --avg 4 \
+    --exp-dir ./pruned_transducer_stateless7_streaming_multi/exp \
+    --decoding-method $m \
+    --decode-chunk-len 32 \
+    --num-decode-streams 2000
+done
+```
+
+#### Smaller model
+
+We also provide a very small version (only 6.1M parameters) of this setup. The training command for the small model is:
+
+```bash
+./pruned_transducer_stateless7_streaming_multi/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming_multi/exp \
+  --full-libri 1 \
+  --giga-prob 0.9 \
+  --num-encoder-layers "2,2,2,2,2" \
+  --feedforward-dims "256,256,512,512,256" \
+  --nhead "4,4,4,4,4" \
+  --encoder-dims "128,128,128,128,128" \
+  --attention-dims "96,96,96,96,96" \
+  --encoder-unmasked-dims "96,96,96,96,96" \
+  --max-duration 1200 \
+  --master-port 12345
+```
+
+You can find this pretrained small model and its training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/marcoyang/icefall-libri-giga-pruned-transducer-stateless7-streaming-6M-2023-04-03>
+
+
+| decoding method      | chunk size | test-clean | test-other | comment             | decoding mode        |
+|----------------------|------------|------------|------------|---------------------|----------------------|
+| greedy search        | 320ms      | 5.95       | 15.03       | --epoch 30 --avg 1  | simulated streaming  |
+| greedy search        | 640ms      | 5.61       | 13.86       | --epoch 30 --avg 1  | simulated streaming  |
+| modified beam search | 320ms      | 5.72       | 14.34      | --epoch 30 --avg 1  | simulated streaming  |
+| modified beam search | 640ms      | 5.43       | 13.16      | --epoch 30 --avg 1  | simulated streaming  |
+| fast beam search | 320ms      | 5.88       | 14.45      | --epoch 30 --avg 1  | simulated streaming  |
+| fast beam search | 640ms      | 5.48       | 13.31      | --epoch 30 --avg 1  | simulated streaming  |
+
+This small model achieves the following WERs on GigaSpeech test and dev sets:
+
+| decoding method      | chunk size | dev | test | comment    | decoding mode       |
+|----------------------|------------|------------|------------|---------------------|----------------------|
+| greedy search        | 320ms      | 17.57       | 17.2     | --epoch 30 --avg 1  | simulated streaming  |
+| modified beam search | 320ms      | 16.98       | 11.98       | --epoch 30 --avg 1  | simulated streaming  |
+
+You can find the tensorboard logs at <https://tensorboard.dev/experiment/tAc5iXxTQrCQxky5O5OLyw/#scalars>.
+
+
+### Streaming Zipformer-Transducer (Pruned Stateless Transducer + Streaming Zipformer)
+
+#### [pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
+
+See <https://github.com/k2-fsa/icefall/pull/787> for more details.
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+##### training on full librispeech
+
+The WERs are:
+
+| decoding method      | chunk size | test-clean | test-other | comment             | decoding mode        |
+|----------------------|------------|------------|------------|---------------------|----------------------|
+| greedy search        | 320ms      | 3.15       | 8.09       | --epoch 30 --avg 9  | simulated streaming  |
+| greedy search        | 320ms      | 3.17       | 8.24       | --epoch 30 --avg 9  | chunk-wise           |
+| fast beam search     | 320ms      | 3.2        | 8.04       | --epoch 30 --avg 9  | simulated streaming  |
+| fast beam search     | 320ms      | 3.36       | 8.19       | --epoch 30 --avg 9  | chunk-wise           |
+| modified beam search | 320ms      | 3.11       | 7.93       | --epoch 30 --avg 9  | simulated streaming  |
+| modified beam search | 320ms      | 3.12       | 8.11       | --epoch 30 --avg 9  | chunk-size           |
+| greedy search        | 640ms      | 2.97       | 7.5        | --epoch 30 --avg 9  | simulated streaming  |
+| greedy search        | 640ms      | 2.98       | 7.67       | --epoch 30 --avg 9  | chunk-wise           |
+| fast beam search     | 640ms      | 3.02       | 7.47       | --epoch 30 --avg 9  | simulated streaming  |
+| fast beam search     | 640ms      | 2.96       | 7.61       | --epoch 30 --avg 9  | chunk-wise           |
+| modified beam search | 640ms      | 2.94       | 7.36       | --epoch 30 --avg 9  | simulated streaming  |
+| modified beam search | 640ms      | 2.95       | 7.53       | --epoch 30 --avg 9  | chunk-size           |
+
+Note: `simulated streaming` indicates feeding full utterance during decoding using `decode.py`,
+while `chunk-size` indicates feeding certain number of frames at each time using `streaming_decode.py`.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 750 \
+  --master-port 12345
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/A46UpqEWQWS7oDi5VcQ8rg/>
+
+The simulated streaming decoding command (e.g., chunk-size=320ms) is:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  ./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method $m
+done
+```
+
+The streaming chunk-size decoding command (e.g., chunk-size=320ms) is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless7_streaming/streaming_decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --decoding-method $m \
+    --decode-chunk-len 32 \
+    --num-decode-streams 2000
+done
+```
+We also support decoding with neural network LMs. After combining with language models, the WERs are
+| decoding method      | chunk size | test-clean | test-other | comment             | decoding mode        |
+|----------------------|------------|------------|------------|---------------------|----------------------|
+| `modified_beam_search` | 320ms      | 3.11       | 7.93       | --epoch 30 --avg 9  | simulated streaming  |
+| `modified_beam_search_lm_shallow_fusion` | 320ms      | 2.58       | 6.65       | --epoch 30 --avg 9  | simulated streaming  |
+| `modified_beam_search_lm_rescore` | 320ms      | 2.59       | 6.86       | --epoch 30 --avg 9  | simulated streaming  |
+| `modified_beam_search_lm_rescore_LODR` | 320ms      | 2.52       | 6.73       | --epoch 30 --avg 9  | simulated streaming  |
+
+Please use the following command for `modified_beam_search_lm_shallow_fusion`:
+```bash
+for lm_scale in $(seq 0.15 0.01 0.38); do
+    for beam_size in 4 8 12; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+            --epoch 99 \
+            --avg 1 \
+            --use-averaged-model False \
+            --beam-size $beam_size \
+            --exp-dir ./pruned_transducer_stateless7_streaming/exp-large-LM \
+            --max-duration 600 \
+            --decode-chunk-len 32 \
+            --decoding-method modified_beam_search_lm_shallow_fusion \
+            --use-shallow-fusion 1 \
+            --lm-type rnn \
+            --lm-exp-dir rnn_lm/exp \
+            --lm-epoch 99 \
+            --lm-scale $lm_scale \
+            --lm-avg 1 \
+            --rnn-lm-embedding-dim 2048 \
+            --rnn-lm-hidden-dim 2048 \
+            --rnn-lm-num-layers 3 \
+            --lm-vocab-size 500
+    done
+done
+```
+
+Please use the following command for `modified_beam_search_lm_rescore`:
+```bash
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --use-averaged-model True \
+    --beam-size 8 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search_lm_rescore \
+    --use-shallow-fusion 0 \
+    --lm-type rnn \
+    --lm-exp-dir rnn_lm/exp \
+    --lm-epoch 99 \
+    --lm-avg 1 \
+    --rnn-lm-embedding-dim 2048 \
+    --rnn-lm-hidden-dim 2048 \
+    --rnn-lm-num-layers 3 \
+    --lm-vocab-size 500
+```
+
+Please use the following command for `modified_beam_search_lm_rescore_LODR`:
+```bash
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --use-averaged-model True \
+    --beam-size 8 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search_lm_rescore_LODR \
+    --use-shallow-fusion 0 \
+    --lm-type rnn \
+    --lm-exp-dir rnn_lm/exp \
+    --lm-epoch 99 \
+    --lm-avg 1 \
+    --rnn-lm-embedding-dim 2048 \
+    --rnn-lm-hidden-dim 2048 \
+    --rnn-lm-num-layers 3 \
+    --lm-vocab-size 500 \
+    --tokens-ngram 2 \
+    --backoff-id 500
+```
+
+A well-trained RNNLM can be found here: <https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>. The bi-gram used in LODR decoding
+can be found here: <https://huggingface.co/marcoyang/librispeech_bigram>.
+
+
+#### Smaller model
+
+A smaller model (~20M params) is also available with configuration based on [this comment](https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740). The WERs are:
+
+| decoding method      | chunk size | test-clean | test-other | comment             | decoding mode        |
+|----------------------|------------|------------|------------|---------------------|----------------------|
+| greedy search        | 320ms      | 3.94       | 9.79       | --epoch 30 --avg 9  | simulated streaming  |
+| modified beam search | 320ms      | 3.88       | 9.53       | --epoch 30 --avg 9  | simulated streaming  |
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at: <https://huggingface.co/desh2608/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-small>
+
+
+### zipformer_mmi (zipformer with mmi loss)
+
+See <https://github.com/k2-fsa/icefall/pull/746> for more details.
+
+[zipformer_mmi](./zipformer_mmi)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08>
+
+Number of model parameters: 69136519, i.e., 69.14 M
+
+|                        | test-clean | test-other | comment             |
+| ---------------------- | ---------- | ---------- | ------------------- |
+| 1best                  | 2.54       | 5.65       | --epoch 30 --avg 10 |
+| nbest                  | 2.54       | 5.66       | --epoch 30 --avg 10 |
+| nbest-rescoring-LG     | 2.49       | 5.42       | --epoch 30 --avg 10 |
+| nbest-rescoring-3-gram | 2.52       | 5.62       | --epoch 30 --avg 10 |
+| nbest-rescoring-4-gram | 2.5        | 5.51       | --epoch 30 --avg 10 |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_mmi/train.py \
+  --world-size 4 \
+  --master-port 12345 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 500 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_mmi/exp
+```
+
+The decoding commands for the transducer branch are:
+```bash
+export CUDA_VISIBLE_DEVICES="5"
+
+for m in nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+  ./zipformer_mmi/decode.py \
+    --epoch 30 \
+    --avg 10 \
+    --exp-dir ./zipformer_mmi/exp/ \
+    --max-duration 100 \
+    --lang-dir data/lang_bpe_500 \
+    --nbest-scale 1.2 \
+    --hp-scale 1.0 \
+    --decoding-method $m
+done
+```
+
+### pruned_transducer_stateless7_ctc_bs (zipformer with transducer loss and ctc loss using blank skip)
+
+See https://github.com/k2-fsa/icefall/pull/730 for more details.
+
+[pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/rrNZ7l83Qu6RKoD7y49wiA/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29>
+
+Number of model parameters: 76804822, i.e., 76.80 M
+
+Test on 8-card V100 cluster, with 4-card busy and 4-card idle.
+
+#### greedy_search
+
+| model                                                        | test-clean | test-other | decoding time(s) | comment             |
+| ------------------------------------------------------------ | ---------- | ---------- | ---------------- | ------------------- |
+| [pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs) | 2.28       | 5.53       | 48.939           | --epoch 30 --avg 13 |
+| [pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc) | 2.24       | 5.18       | 91.900           | --epoch 30 --avg 8  |
+
+- [pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs) applies blank skip both on training and decoding, and [pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc) doesn`t apply blank skip.
+- Applying blank skip both on training and decoding is **1.88 times** faster than the model that doesn't apply blank skip without obvious performance loss.
+
+#### modified_beam_search
+
+| model                                                        | test-clean | test-other | decoding time(s) | comment             |
+| ------------------------------------------------------------ | ---------- | ---------- | ---------------- | ------------------- |
+| [pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs) | 2.26       | 5.44       | 80.446           | --epoch 30 --avg 13 |
+| [pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc) | 2.20       | 5.12       | 283.676          | --epoch 30 --avg 8  |
+
+- [pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs) applies blank skip both on training and decoding, and [pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc) doesn`t apply blank skip.
+- Applying blank skip both on training and decoding is **3.53 times**  faster than the model that doesn't apply blank skip without obvious performance loss.
+
+The training commands for the model using blank skip ([pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs)) are:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_ctc_bs/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --max-duration 750 \
+  --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --ctc-loss-scale 0.2 \
+  --master-port 12535
+```
+
+The decoding commands for the transducer branch of the model using blank skip ([pruned_transducer_stateless7_ctc_bs](./pruned_transducer_stateless7_ctc_bs)) are:
+
+```bash
+for m in greedy_search modified_beam_search fast_beam_search; do
+  for epoch in 30; do
+    for avg in 15; do
+      ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+          --epoch $epoch \
+          --avg $avg \
+          --use-averaged-model 1 \
+          --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+          --feedforward-dims  "1024,1024,2048,2048,1024" \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
+
+The decoding commands for the transducer branch of the model without blank skip ([pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc)) are:
+
+```bash
+for m in greedy_search modified_beam_search fast_beam_search; do
+  for epoch in 30; do
+    for avg in 15; do
+      ./pruned_transducer_stateless7_ctc/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --use-averaged-model 1 \
+          --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+          --feedforward-dims  "1024,1024,2048,2048,1024" \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
+
+### pruned_transducer_stateless7_ctc (zipformer with transducer loss and ctc loss)
+
+See <https://github.com/k2-fsa/icefall/pull/683> for more details.
+
+[pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/hxlGAhOPToGmRLZFnAzPWw/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01>
+
+Number of model parameters: 70561891, i.e., 70.56 M
+
+|                          | test-clean | test-other  | comment            |
+|--------------------------|------------|-------------|--------------------|
+| greedy search            | 2.23       | 5.19        | --epoch 30 --avg 8 |
+| modified beam search     | 2.21       | 5.12        | --epoch 30 --avg 8 |
+| fast beam search         | 2.23       | 5.18        | --epoch 30 --avg 8 |
+| ctc decoding             | 2.48       | 5.82        | --epoch 30 --avg 9 |
+| 1best                    | 2.43       | 5.22        | --epoch 30 --avg 9 |
+| nbest                    | 2.43       | 5.22        | --epoch 30 --avg 9 |
+| nbest rescoring          | 2.34       | 5.05        | --epoch 30 --avg 9 |
+| whole lattice rescoring  | 2.34       | 5.04        | --epoch 30 --avg 9 |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_ctc/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --max-duration 750 \
+  --exp-dir pruned_transducer_stateless7_ctc/exp \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --ctc-loss-scale 0.2 \
+  --master-port 12535
+```
+
+The decoding commands for the transducer branch are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search ; do
+  for epoch in 30; do
+    for avg in 8; do
+      ./pruned_transducer_stateless7_ctc/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --use-averaged-model 1 \
+          --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+          --feedforward-dims  "1024,1024,2048,2048,1024" \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
+
+The decoding commands for the ctc branch are:
+```bash
+for m in ctc-decoding nbest nbest-rescoring whole-lattice-rescoring; do
+  for epoch in 30; do
+    for avg in 9; do
+      ./pruned_transducer_stateless7_ctc/ctc_decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+          --max-duration 100 \
+          --decoding-method $m \
+          --hlg-scale 0.6 \
+          --lm-dir data/lm
+    done
+  done
+done
+```
+
+
+### LibriSpeech BPE training results (Conformer CTC, supporting delay penalty)
+
+#### [conformer_ctc3](./conformer_ctc3)
+
+It implements Conformer model training with CTC loss.
+For streaming mode, it supports symbol delay penalty.
+
+See <https://github.com/k2-fsa/icefall/pull/669> for more details.
+
+##### training on full librispeech
+
+This model contains 12 encoder layers. The number of model parameters is 77352694.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment              |
+|-------------------------------------|------------|------------|----------------------|
+| ctc-decoding                        | 3.09       | 7.62       | --epoch 25 --avg 7   |
+| 1best                               | 2.87       | 6.44       | --epoch 25 --avg 7   |
+| nbest                               | 2.88       | 6.5        | --epoch 25 --avg 7   |
+| nbest-rescoring                     | 2.71       | 6.1        | --epoch 25 --avg 7   |
+| whole-lattice-rescoring             | 2.71       | 6.04       | --epoch 25 --avg 7   |
+
+The training command is:
+
+```bash
+./conformer_ctc3/train.py \
+  --world-size 4 \
+  --num-epochs 25 \
+  --start-epoch 1 \
+  --exp-dir conformer_ctc3/full \
+  --full-libri 1 \
+  --max-duration 300 \
+  --master-port 12345
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/4jbxIQ2SQIaQeRqsR6bOSA>
+
+The decoding command using different methods is:
+```bash
+for method in ctc-decoding 1best nbest nbest-rescoring whole-lattice-rescoring; do
+  ./conformer_ctc3/decode.py \
+    --epoch 25 \
+    --avg 7 \
+    --exp-dir conformer_ctc3/exp \
+    --max-duration 300 \
+    --decoding-method $method \
+    --manifest-dir data/fbank \
+    --lm-dir data/lm \
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-conformer-ctc3-2022-11-27>
+
+The command to train a streaming model with symbol delay penalty is:
+```bash
+./conformer_ctc3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conformer_ctc3/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 25 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --delay-penalty 0.1
+```
+To evaluate symbol delay, you should:
+(1) Generate cuts with word-time alignments:
+```bash
+./local/add_alignment_librispeech.py \
+  --alignments-dir data/alignment \
+  --cuts-in-dir data/fbank \
+  --cuts-out-dir data/fbank_ali
+```
+(2) Set the argument "--manifest-dir data/fbank_ali" while decoding.
+For example:
+```bash
+./conformer_ctc3/decode.py \
+  --epoch 25 \
+  --avg 7 \
+  --exp-dir ./conformer_ctc3/exp \
+  --max-duration 300 \
+  --decoding-method ctc-decoding \
+  --simulate-streaming 1 \
+  --causal-convolution 1 \
+  --decode-chunk-size 16 \
+  --left-context 64 \
+  --manifest-dir data/fbank_ali
+```
+Note: It supports to calculate symbol delay with following decoding methods:
+  - ctc-greedy-search
+  - ctc-decoding
+  - 1best
+
+
+### pruned_transducer_stateless8 (zipformer + multidataset)
+
+See <https://github.com/k2-fsa/icefall/pull/675> for more details.
+
+[pruned_transducer_stateless8](./pruned_transducer_stateless8)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/3e9AfOcgRwOXpLQlZvHZrQ>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 1.81       | 4.18       | --epoch 20 --avg 4 |
+| fast_beam_search     | 1.82       | 4.15       | --epoch 20 --avg 4 |
+| modified_beam_search | 1.78       | **4.08**   | --epoch 20 --avg 4 |
+| greedy_search        | 1.84       | 4.3        | --epoch 19 --avg 8 |
+| fast_beam_search     |**1.77**    | 4.25       | --epoch 19 --avg 8 |
+| modified_beam_search | 1.81       | 4.16       | --epoch 19 --avg 8 |
+
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless8/train.py \
+  --world-size 8 \
+  --num-epochs 20 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --max-duration 750 \
+  --exp-dir pruned_transducer_stateless8/exp \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --master-port 12535 \
+  --giga-prob 0.9
+```
+
+The decoding commands are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  for epoch in $(seq 20 -1 10); do
+    for avg in $(seq 9 -1 1); do
+      ./pruned_transducer_stateless8/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --use-averaged-model 1 \
+          --exp-dir ./pruned_transducer_stateless8/exp \
+          --feedforward-dims "1024,1024,2048,2048,1024" \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
+
+
+### pruned_transducer_stateless7 (zipformer)
+
+See <https://github.com/k2-fsa/icefall/pull/672> for more details.
+
+[pruned_transducer_stateless7](./pruned_transducer_stateless7)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/P7vXWqK7QVu1mU9Ene1gGg/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+|                      | test-clean | test-other  | comment                                |
+|----------------------|------------|-------------|----------------------------------------|
+| greedy search        | 2.17       | 5.23        | --epoch 30 --avg 9 --max-duration 600  |
+| modified beam search | 2.15       | 5.20        | --epoch 30 --avg 9 --max-duration 600  |
+| modified beam search + RNNLM shallow fusion | 1.99       | 4.73        | --epoch 30 --avg 9 --max-duration 600  |
+| modified beam search + TransformerLM shallow fusion | 1.94       | 4.73        | --epoch 30 --avg 9 --max-duration 600  |
+| modified beam search + RNNLM + LODR | 1.91       | 4.57        | --epoch 30 --avg 9 --max-duration 600  |
+| modified beam search + TransformerLM + LODR | 1.91       | 4.51        | --epoch 30 --avg 9 --max-duration 600  |
+| fast beam search     | 2.15       | 5.22        | --epoch 30 --avg 9 --max-duration 600  |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,3,6,7"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --max-duration 750 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --feedforward-dims  "1024,1024,2048,2048,1024" \
+  --master-port 12535
+```
+
+The decoding commands are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search ; do
+  for epoch in 30; do
+    for avg in 9; do
+      ./pruned_transducer_stateless7/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --use-averaged-model 1 \
+          --exp-dir ./pruned_transducer_stateless7/exp \
+          --feedforward-dims  "1024,1024,2048,2048,1024" \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
+
+Note that a small change is made to the `pruned_transducer_stateless7/decoder.py` in
+this [PR](https://github.com/k2-fsa/icefall/pull/942) to address the
+problem of emitting the first symbol at the very beginning. If you need a
+model without this issue, please download the model from here: <https://huggingface.co/marcoyang/icefall-asr-librispeech-pruned-transducer-stateless7-2023-03-10>
+
+### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T + gradient filter)
+
+#### [lstm_transducer_stateless3](./lstm_transducer_stateless3)
+
+It implements LSTM model with mechanisms in reworked model for streaming ASR.
+Gradient filter is applied inside each lstm module to stabilize the training.
+
+See <https://github.com/k2-fsa/icefall/pull/564> for more details.
+
+##### training on full librispeech
+
+This model contains 12 encoder layers (LSTM module + Feedforward module). The number of model parameters is 84689496.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment              | decoding mode        |
+|-------------------------------------|------------|------------|----------------------|----------------------|
+| greedy search (max sym per frame 1) | 3.66       | 9.51       | --epoch 40 --avg 15  | simulated streaming  |
+| greedy search (max sym per frame 1) | 3.66       | 9.48       | --epoch 40 --avg 15  | streaming            |
+| fast beam search                    | 3.55       | 9.33       | --epoch 40 --avg 15  | simulated streaming  |
+| fast beam search                    | 3.57       | 9.25       | --epoch 40 --avg 15  | streaming            |
+| modified beam search                | 3.55       | 9.28       | --epoch 40 --avg 15  | simulated streaming  |
+| modified beam search                | 3.54       | 9.25       | --epoch 40 --avg 15  | streaming            |
+
+Note: `simulated streaming` indicates feeding full utterance during decoding, while `streaming` indicates feeding certain number of frames at each time.
+
+
+The training command is:
+
+```bash
+./lstm_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir lstm_transducer_stateless3/exp \
+  --full-libri 1 \
+  --max-duration 500 \
+  --master-port 12325 \
+  --num-encoder-layers 12 \
+  --grad-norm-threshold 25.0 \
+  --rnn-hidden-size 1024
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/caNPyr5lT8qAl9qKsXEeEQ/>
+
+The simulated streaming decoding command using greedy search, fast beam search, and modified beam search is:
+```bash
+for decoding_method in greedy_search fast_beam_search modified_beam_search; do
+  ./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 15 \
+    --exp-dir lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --num-encoder-layers 12 \
+    --rnn-hidden-size 1024 \
+    --decoding-method $decoding_method \
+    --use-averaged-model True \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8 \
+    --beam-size 4
+done
+```
+
+The streaming decoding command using greedy search, fast beam search, and modified beam search is:
+```bash
+for decoding_method in greedy_search fast_beam_search modified_beam_search; do
+  ./lstm_transducer_stateless3/streaming_decode.py \
+    --epoch 40 \
+    --avg 15 \
+    --exp-dir lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --num-encoder-layers 12 \
+    --rnn-hidden-size 1024 \
+    --decoding-method $decoding_method \
+    --use-averaged-model True \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8 \
+    --beam-size 4
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless3-2022-09-28>
+
+
+### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T + multi-dataset)
+
+#### [lstm_transducer_stateless2](./lstm_transducer_stateless2)
+
+See <https://github.com/k2-fsa/icefall/pull/558> for more details.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                 |
+|-------------------------------------|------------|------------|-------------------------|
+| greedy search (max sym per frame 1) | 2.78       | 7.36       | --iter 468000 --avg 16  |
+| modified_beam_search                | 2.73       | 7.15       | --iter 468000 --avg 16  |
+| modified_beam_search + RNNLM shallow fusion   | 2.42     |  6.46      | --iter 468000 --avg 16  |
+| modified_beam_search + TransformerLM shallow fusion   | 2.37     |  6.48      | --iter 468000 --avg 16  |
+| modified_beam_search + RNNLM + LODR   | 2.24     |  5.89      | --iter 468000 --avg 16  |
+| modified_beam_search + TransformerLM + LODR   | 2.19     |  5.90      | --iter 468000 --avg 16  |
+| fast_beam_search                    | 2.76       | 7.31       | --iter 468000 --avg 16  |
+| greedy search (max sym per frame 1) | 2.77       | 7.35       | --iter 472000 --avg 18  |
+| modified_beam_search                | 2.75       | 7.08       | --iter 472000 --avg 18  |
+| fast_beam_search                    | 2.77       | 7.29       | --iter 472000 --avg 18  |
+
+
+The training command is:
+
+```bash
+#!/usr/bin/env bash
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./lstm_transducer_stateless2/train.py \
+  --world-size 8 \
+  --num-epochs 35 \
+  --start-epoch 1 \
+  --full-libri 1 \
+  --exp-dir lstm_transducer_stateless2/exp \
+  --max-duration 500 \
+  --use-fp16 0 \
+  --lr-epochs 10 \
+  --num-workers 2 \
+  --giga-prob 0.9
+```
+**Note**: It was killed manually after getting `epoch-18.pt`. Also, we resumed
+training after getting `epoch-9.pt`.
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/1ziQ2LFmQY2mt4dlUr5dyA/>
+
+The decoding command is
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  for iter in 472000; do
+    for avg in 8 10 12 14 16 18; do
+      ./lstm_transducer_stateless2/decode.py \
+        --iter $iter \
+        --avg $avg \
+        --exp-dir lstm_transducer_stateless2/exp \
+        --max-duration 600 \
+        --num-encoder-layers 12 \
+        --rnn-hidden-size 1024 \
+        --decoding-method $m \
+        --use-averaged-model True \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        --beam-size 4
+    done
+  done
+done
+```
+
+You may also decode using shallow fusion with external neural network LM. To do so you need to
+download a well-trained NN LM:
+RNN LM: <https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>
+Transformer LM: <https://huggingface.co/marcoyang/icefall-librispeech-transformer-lm/tree/main>
+
+```bash
+for iter in 472000; do
+    for avg in 8 10 12 14 16 18; do
+        ./lstm_transducer_stateless2/decode.py \
+                --iter $iter \
+                --avg $avg \
+                --exp-dir ./lstm_transducer_stateless2/exp \
+                --max-duration 600 \
+                --decoding-method modified_beam_search_lm_shallow_fusion \
+                --use-shallow-fusion 1 \
+                --lm-type rnn \
+                --lm-exp-dir /ceph-data4/yangxiaoyu/pretrained_models/LM/icefall-librispeech-rnn-lm/exp \
+                --lm-epoch 99 \
+                --lm-scale $lm_scale \
+                --lm-avg 1 \
+    done
+done
+```
+
+You may also decode using LODR + LM shallow fusion. This decoding method is proposed in <https://arxiv.org/pdf/2203.16776.pdf>.
+It subtracts the internal language model score during shallow fusion, which is approximated by a bi-gram model. The bi-gram can be
+generated by `generate-lm.sh`, or you may download it from <https://huggingface.co/marcoyang/librispeech_bigram>.
+
+The decoding command is as follows:
+
+```bash
+for iter in 472000; do
+    for avg in 8 10 12 14 16 18; do
+        ./lstm_transducer_stateless2/decode.py \
+                --iter $iter \
+                --avg $avg \
+                --exp-dir ./lstm_transducer_stateless2/exp \
+                --max-duration 600 \
+                --decoding-method modified_beam_search_LODR \
+                --beam 4 \
+                --max-contexts 4 \
+                --use-shallow-fusion 1 \
+                --lm-type rnn \
+                --lm-exp-dir /ceph-data4/yangxiaoyu/pretrained_models/LM/icefall-librispeech-rnn-lm/exp \
+                --lm-epoch 99 \
+                --lm-scale 0.4 \
+                --lm-avg 1 \
+                --tokens-ngram 2 \
+                --ngram-lm-scale -0.16
+    done
+done
+```
+Note that you can also set `--lm-type transformer` to use transformer LM during LODR. But it will be slower
+because it has not been optimized. The pre-trained transformer LM is available at <https://huggingface.co/marcoyang/icefall-librispeech-transformer-lm/tree/main>
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03>
+
+
+### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T)
+
+#### [lstm_transducer_stateless](./lstm_transducer_stateless)
+
+It implements LSTM model with mechanisms in reworked model for streaming ASR.
+
+See <https://github.com/k2-fsa/icefall/pull/479> for more details.
+
+##### training on full librispeech
+
+This model contains 12 encoder layers (LSTM module + Feedforward module). The number of model parameters is 84689496.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment              | decoding mode        |
+|-------------------------------------|------------|------------|----------------------|----------------------|
+| greedy search (max sym per frame 1) | 3.81       | 9.73       | --epoch 35 --avg 15  | simulated streaming  |
+| greedy search (max sym per frame 1) | 3.78       | 9.79       | --epoch 35 --avg 15  | streaming            |
+| fast beam search                    | 3.74       | 9.59       | --epoch 35 --avg 15  | simulated streaming  |
+| fast beam search                    | 3.73       | 9.61       | --epoch 35 --avg 15  | streaming            |
+| modified beam search                | 3.64       | 9.55       | --epoch 35 --avg 15  | simulated streaming  |
+| modified beam search                | 3.65       | 9.51       | --epoch 35 --avg 15  | streaming            |
+
+Note: `simulated streaming` indicates feeding full utterance during decoding, while `streaming` indicates feeding certain number of frames at each time.
+
+The training command is:
+
+```bash
+./lstm_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 35 \
+  --start-epoch 1 \
+  --exp-dir lstm_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 500 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --rnn-hidden-size 1024
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/FWrM20mjTeWo6dTpFYOsYQ/>
+
+The simulated streaming decoding command using greedy search, fast beam search, and modified beam search is:
+```bash
+for decoding_method in greedy_search fast_beam_search modified_beam_search; do
+  ./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --num-encoder-layers 12 \
+    --rnn-hidden-size 1024 \
+    --decoding-method $decoding_method \
+    --use-averaged-model True \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8 \
+    --beam-size 4
+done
+```
+
+The streaming decoding command using greedy search, fast beam search, and modified beam search is:
+```bash
+for decoding_method in greedy_search fast_beam_search modified_beam_search; do
+  ./lstm_transducer_stateless/streaming_decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --num-encoder-layers 12 \
+    --rnn-hidden-size 1024 \
+    --decoding-method $decoding_method \
+    --use-averaged-model True \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8 \
+    --beam-size 4
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18>
+
+
+#### LibriSpeech BPE training results (Pruned Stateless Conv-Emformer RNN-T 2)
+
+[conv_emformer_transducer_stateless2](./conv_emformer_transducer_stateless2)
+
+It implements [Emformer](https://arxiv.org/abs/2010.10759) augmented with convolution module and simplified memory bank for streaming ASR.
+It is modified from [torchaudio](https://github.com/pytorch/audio).
+
+See <https://github.com/k2-fsa/icefall/pull/440> for more details.
+
+##### With lower latency setup, training on full librispeech
+
+In this model, the lengths of chunk and right context are 32 frames (i.e., 0.32s) and 8 frames (i.e., 0.08s), respectively.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment              | decoding mode        |
+|-------------------------------------|------------|------------|----------------------|----------------------|
+| greedy search (max sym per frame 1) | 3.5        | 9.09       | --epoch 30 --avg 10  | simulated streaming  |
+| greedy search (max sym per frame 1) | 3.57       | 9.1        | --epoch 30 --avg 10  | streaming            |
+| fast beam search                    | 3.5        | 8.91       | --epoch 30 --avg 10  | simulated streaming  |
+| fast beam search                    | 3.54       | 8.91       | --epoch 30 --avg 10  | streaming            |
+| modified beam search                | 3.43       | 8.86       | --epoch 30 --avg 10  | simulated streaming  |
+| modified beam search                | 3.48       | 8.88       | --epoch 30 --avg 10  | streaming            |
+
+The training command is:
+
+```bash
+./conv_emformer_transducer_stateless2/train.py \
+  --world-size 6 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 280 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/W5MpxekiQLSPyM4fe5hbKg/>
+
+The simulated streaming decoding command using greedy search is:
+```bash
+./conv_emformer_transducer_stateless2/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The simulated streaming decoding command using fast beam search is:
+```bash
+./conv_emformer_transducer_stateless2/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The simulated streaming decoding command using modified beam search is:
+```bash
+./conv_emformer_transducer_stateless2/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+The streaming decoding command using greedy search is:
+```bash
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The streaming decoding command using fast beam search is:
+```bash
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The streaming decoding command using modified beam search is:
+```bash
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>
+
+##### With higher latency setup, training on full librispeech
+
+In this model, the lengths of chunk and right context are 64 frames (i.e., 0.64s) and 16 frames (i.e., 0.16s), respectively.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment              | decoding mode        |
+|-------------------------------------|------------|------------|----------------------|----------------------|
+| greedy search (max sym per frame 1) | 3.3        | 8.71       | --epoch 30 --avg 10  | simulated streaming  |
+| greedy search (max sym per frame 1) | 3.35       | 8.65       | --epoch 30 --avg 10  | streaming            |
+| fast beam search                    | 3.27       | 8.58       | --epoch 30 --avg 10  | simulated streaming  |
+| fast beam search                    | 3.31       | 8.48       | --epoch 30 --avg 10  | streaming            |
+| modified beam search                | 3.26       | 8.56       | --epoch 30 --avg 10  | simulated streaming  |
+| modified beam search                | 3.29       | 8.47       | --epoch 30 --avg 10  | streaming            |
+
+The training command is:
+
+```bash
+./conv_emformer_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 280 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/eRx6XwbOQhGlywgD8lWBjw/>
+
+The simulated streaming decoding command using greedy search is:
+```bash
+./conv_emformer_transducer_stateless2/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The simulated streaming decoding command using fast beam search is:
+```bash
+./conv_emformer_transducer_stateless2/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The simulated streaming decoding command using modified beam search is:
+```bash
+./conv_emformer_transducer_stateless2/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+The streaming decoding command using greedy search is:
+```bash
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The streaming decoding command using fast beam search is:
+```bash
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The streaming decoding command using modified beam search is:
+```bash
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 64 \
+  --cnn-module-kernel 31 \
+  --left-context-length 64 \
+  --right-context-length 16 \
+  --memory-size 32 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-larger-latency-2022-07-06>
+
+
+### LibriSpeech BPE training results (Pruned Stateless Streaming Conformer RNN-T)
+
+#### [pruned_transducer_stateless](./pruned_transducer_stateless)
+
+See <https://github.com/k2-fsa/icefall/pull/380> for more details.
+
+##### Training on full librispeech
+The WERs are (the number in the table formatted as test-clean & test-other):
+
+We only trained 25 epochs for saving time, if you want to get better results you can train more epochs.
+
+| decoding method      | left context | chunk size = 2 | chunk size = 4 | chunk size = 8 | chunk size = 16|
+|----------------------|--------------|----------------|----------------|----------------|----------------|
+| greedy search        | 32           | 4.74 & 11.38   | 4.57 & 10.86   | 4.18 & 10.37   | 3.87 & 9.85    |
+| greedy search        | 64           | 4.74 & 11.25   | 4.48 & 10.72   | 4.1 & 10.24    | 3.85 & 9.73    |
+| fast beam search     | 32           | 4.75 & 11.1    | 4.48 & 10.65   | 4.12 & 10.18   | 3.95 & 9.67    |
+| fast beam search     | 64           | 4.7 & 11       | 4.37 & 10.49   | 4.07 & 10.04   | 3.89 & 9.53    |
+| modified beam search | 32           | 4.64 & 10.94   | 4.38 & 10.51   | 4.11 & 10.14   | 3.87 & 9.61    |
+| modified beam search | 64           | 4.59 & 10.81   | 4.29 & 10.39   | 4.02 & 10.02   | 3.84 & 9.43    |
+
+**NOTE:** The WERs in table above were decoded with simulate streaming method (i.e. using masking strategy), see commands below. We also have [real streaming decoding](./pruned_transducer_stateless/streaming_decode.py) script which should produce almost the same results. We tried adding right context in the real streaming decoding, but it seemed not to benefit the performance for all the models, the reasons might be the training and decoding mismatching.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless/train.py \
+  --exp-dir pruned_transducer_stateless/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 20 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --world-size 4 \
+  --start-epoch 0 \
+  --num-epochs 25
+```
+
+You can find the tensorboard log here <https://tensorboard.dev/experiment/ofxRakE6R7WHB1AoB8Bweg/>
+
+The decoding command is:
+```bash
+decoding_method="greedy_search"  # "fast_beam_search", "modified_beam_search"
+
+for chunk in 2 4 8 16; do
+  for left in 32 64; do
+    ./pruned_transducer_stateless/decode.py \
+            --simulate-streaming 1 \
+            --decode-chunk-size ${chunk} \
+            --left-context ${left} \
+            --causal-convolution 1 \
+            --epoch 24 \
+            --avg 10 \
+            --exp-dir ./pruned_transducer_stateless/exp \
+            --max-sym-per-frame 1 \
+            --max-duration 1000 \
+            --decoding-method ${decoding_method}
+  done
+done
+```
+
+Pre-trained models, training and decoding logs, and decoding results are available at <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless_20220625>
+
+#### [pruned_transducer_stateless2](./pruned_transducer_stateless2)
+
+See <https://github.com/k2-fsa/icefall/pull/380> for more details.
+
+##### Training on full librispeech
+The WERs are (the number in the table formatted as test-clean & test-other):
+
+We only trained 25 epochs for saving time, if you want to get better results you can train more epochs.
+
+| decoding method      | left context | chunk size = 2 | chunk size = 4 | chunk size = 8 | chunk size = 16|
+|----------------------|--------------|----------------|----------------|----------------|----------------|
+| greedy search        | 32           | 4.2 & 10.64    | 3.97 & 10.03   | 3.83 & 9.58    | 3.7 & 9.11     |
+| greedy search        | 64           | 4.16 & 10.5    | 3.93 & 9.99    | 3.73 & 9.45    | 3.63 & 9.04    |
+| fast beam search     | 32           | 4.13 & 10.3    | 3.93 & 9.82    | 3.8 & 9.35     | 3.62 & 8.93    |
+| fast beam search     | 64           | 4.13 & 10.22   | 3.89 & 9.68    | 3.73 & 9.27    | 3.52 & 8.82    |
+| modified beam search | 32           | 4.02 & 10.22   | 3.9 & 9.71     | 3.74 & 9.33    | 3.59 & 8.87    |
+| modified beam search | 64           | 4.05 & 10.08   | 3.81 & 9.67    | 3.68 & 9.21    | 3.56 & 8.77    |
+
+**NOTE:** The WERs in table above were decoded with simulate streaming method (i.e. using masking strategy), see commands below. We also have [real streaming decoding](./pruned_transducer_stateless2/streaming_decode.py) script which should produce almost the same results. We tried adding right context in the real streaming decoding, but it seemed not to benefit the performance for all the models, the reasons might be the training and decoding mismatching.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless2/train.py \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 20 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --world-size 4 \
+  --start-epoch 0 \
+  --num-epochs 25
+```
+
+You can find the tensorboard log here <https://tensorboard.dev/experiment/hbltNS5TQ1Kiw0D1vcoakw/>
+
+The decoding command is:
+```bash
+decoding_method="greedy_search"  # "fast_beam_search", "modified_beam_search"
+
+for chunk in 2 4 8 16; do
+  for left in 32 64; do
+    ./pruned_transducer_stateless2/decode.py \
+            --simulate-streaming 1 \
+            --decode-chunk-size ${chunk} \
+            --left-context ${left} \
+            --causal-convolution 1 \
+            --epoch 24 \
+            --avg 10 \
+            --exp-dir ./pruned_transducer_stateless2/exp \
+            --max-sym-per-frame 1 \
+            --max-duration 1000 \
+            --decoding-method ${decoding_method}
+  done
+done
+```
+
+Pre-trained models, training and decoding logs, and decoding results are available at <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless2_20220625>
+
+#### [pruned_transducer_stateless3](./pruned_transducer_stateless3)
+
+See <https://github.com/k2-fsa/icefall/pull/380> for more details.
+
+##### Training on full librispeech (**Use giga_prob = 0.5**)
+
+The WERs are (the number in the table formatted as test-clean & test-other):
+
+| decoding method      | left context | chunk size = 2 | chunk size = 4 | chunk size = 8 | chunk size = 16|
+|----------------------|--------------|----------------|----------------|----------------|----------------|
+| greedy search        | 32           | 3.7 & 9.53     | 3.45 & 8.88    | 3.28 & 8.45    | 3.13 & 7.93    |
+| greedy search        | 64           | 3.69 & 9.36    | 3.39 & 8.68    | 3.28 & 8.19    | 3.08 & 7.83    |
+| fast beam search     | 32           | 3.71 & 9.18    | 3.36 & 8.65    | 3.23 & 8.23    | 3.17 & 7.78    |
+| fast beam search     | 64           | 3.61 & 9.03    | 3.46 & 8.43    | 3.2 & 8.0      | 3.11 & 7.63    |
+| modified beam search | 32           | 3.56 & 9.08    | 3.34 & 8.58    | 3.21 & 8.14    | 3.06 & 7.73    |
+| modified beam search | 64           | 3.55 & 8.86    | 3.29 & 8.34    | 3.16 & 8.01    | 3.05 & 7.57    |
+
+**NOTE:** The WERs in table above were decoded with simulate streaming method (i.e. using masking strategy), see commands below. We also have [real streaming decoding](./pruned_transducer_stateless3/streaming_decode.py) script which should produce almost the same results. We tried adding right context in the real streaming decoding, but it seemed not to benefit the performance for all the models, the reasons might be the training and decoding mismatching.
+
+The training command is (Note: this model was trained with mix-precision training):
+
+```bash
+./pruned_transducer_stateless3/train.py \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 32 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --world-size 4 \
+  --use-fp16 1 \
+  --start-epoch 0 \
+  --num-epochs 37 \
+  --num-workers 2 \
+  --giga-prob 0.5
+```
+
+You can find the tensorboard log here <https://tensorboard.dev/experiment/vL7dWVZqTYaSeoOED4rtow/>
+
+The decoding command is:
+```bash
+decoding_method="greedy_search"  # "fast_beam_search", "modified_beam_search"
+
+for chunk in 2 4 8 16; do
+  for left in 32 64; do
+    ./pruned_transducer_stateless3/decode.py \
+            --simulate-streaming 1 \
+            --decode-chunk-size ${chunk} \
+            --left-context ${left} \
+            --causal-convolution 1 \
+            --epoch 36 \
+            --avg 8 \
+            --exp-dir ./pruned_transducer_stateless3/exp \
+            --max-sym-per-frame 1 \
+            --max-duration 1000 \
+            --decoding-method ${decoding_method}
+  done
+done
+```
+
+Pre-trained models, training and decoding logs, and decoding results are available at <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless3_giga_0.5_20220625>
+
+##### Training on full librispeech (**Use giga_prob = 0.9**)
+
+The WERs are (the number in the table formatted as test-clean & test-other):
+
+| decoding method      | left context | chunk size = 2 | chunk size = 4 | chunk size = 8 | chunk size = 16|
+|----------------------|--------------|----------------|----------------|----------------|----------------|
+| greedy search        | 32           | 3.25 & 8.2     | 3.07 & 7.67    | 2.91 & 7.28    | 2.8 & 6.89     |
+| greedy search        | 64           | 3.22 & 8.12    | 3.05 & 7.59    | 2.91 & 7.07    | 2.78 & 6.81    |
+| fast beam search     | 32           | 3.26 & 8.2     | 3.06 & 7.56    | 2.98 & 7.08    | 2.77 & 6.75    |
+| fast beam search     | 64           | 3.24 & 8.09    | 3.06 & 7.43    | 2.88 & 7.03    | 2.73 & 6.68    |
+| modified beam search | 32           | 3.13 & 7.91    | 2.99 & 7.45    | 2.83 & 6.98    | 2.68 & 6.75    |
+| modified beam search | 64           | 3.08 & 7.8     | 2.97 & 7.37    | 2.81 & 6.82    | 2.66 & 6.67    |
+
+**NOTE:** The WERs in table above were decoded with simulate streaming method (i.e. using masking strategy), see commands below. We also have [real streaming decoding](./pruned_transducer_stateless3/streaming_decode.py) script which should produce almost the same results. We tried adding right context in the real streaming decoding, but it seemed not to benefit the performance for all the models, the reasons might be the training and decoding mismatching.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless3/train.py \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 25 \
+  --num-left-chunks 8 \
+  --max-duration 300 \
+  --world-size 8 \
+  --start-epoch 0 \
+  --num-epochs 26 \
+  --num-workers 2 \
+  --giga-prob 0.9
+```
+
+You can find the tensorboard log here <https://tensorboard.dev/experiment/WBGBDzt7SByRnvCBEfQpGQ/>
+
+The decoding command is:
+```bash
+decoding_method="greedy_search"  # "fast_beam_search", "modified_beam_search"
+
+for chunk in 2 4 8 16; do
+  for left in 32 64; do
+    ./pruned_transducer_stateless3/decode.py \
+            --simulate-streaming 1 \
+            --decode-chunk-size ${chunk} \
+            --left-context ${left} \
+            --causal-convolution 1 \
+            --epoch 25 \
+            --avg 12 \
+            --exp-dir ./pruned_transducer_stateless3/exp \
+            --max-sym-per-frame 1 \
+            --max-duration 1000 \
+            --decoding-method ${decoding_method}
+  done
+done
+```
+
+Pre-trained models, training and decoding logs, and decoding results are available at <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless3_giga_0.9_20220625>
+
+#### [pruned_transducer_stateless4](./pruned_transducer_stateless4)
+
+See <https://github.com/k2-fsa/icefall/pull/380> for more details.
+
+##### Training on full librispeech
+The WERs are (the number in the table formatted as test-clean & test-other):
+
+We only trained 25 epochs for saving time, if you want to get better results you can train more epochs.
+
+| decoding method      | left context | chunk size = 2 | chunk size = 4 | chunk size = 8 | chunk size = 16|
+|----------------------|--------------|----------------|----------------|----------------|----------------|
+| greedy search        | 32           | 3.96 & 10.45   | 3.73 & 9.97    | 3.54 & 9.56    | 3.45 & 9.08    |
+| greedy search        | 64           | 3.9 & 10.34    | 3.7 & 9.9      | 3.53 & 9.41    | 3.39 & 9.03    |
+| fast beam search     | 32           | 3.9 & 10.09    | 3.69 & 9.65    | 3.58 & 9.28    | 3.46 & 8.91    |
+| fast beam search     | 64           | 3.82 & 10.03   | 3.67 & 9.56    | 3.51 & 9.18    | 3.43 & 8.78    |
+| modified beam search | 32           | 3.78 & 10.0    | 3.63 & 9.54    | 3.43 & 9.29    | 3.39 & 8.84    |
+| modified beam search | 64           | 3.76 & 9.95    | 3.54 & 9.48    | 3.4 & 9.13     | 3.33 & 8.74    |
+
+**NOTE:** The WERs in table above were decoded with simulate streaming method (i.e. using masking strategy), see commands below. We also have [real streaming decoding](./pruned_transducer_stateless4/streaming_decode.py) script which should produce almost the same results. We tried adding right context in the real streaming decoding, but it seemed not to benefit the performance for all the models, the reasons might be the training and decoding mismatching.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless4/train.py \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 20 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --world-size 4 \
+  --start-epoch 1 \
+  --num-epochs 25
+```
+
+You can find the tensorboard log here <https://tensorboard.dev/experiment/97VKXf80Ru61CnP2ALWZZg/>
+
+The decoding command is:
+```bash
+decoding_method="greedy_search"  # "fast_beam_search", "modified_beam_search"
+
+for chunk in 2 4 8 16; do
+  for left in 32 64; do
+    ./pruned_transducer_stateless4/decode.py \
+            --simulate-streaming 1 \
+            --decode-chunk-size ${chunk} \
+            --left-context ${left} \
+            --causal-convolution 1 \
+            --epoch 25 \
+            --avg 3 \
+            --exp-dir ./pruned_transducer_stateless4/exp \
+            --max-sym-per-frame 1 \
+            --max-duration 1000 \
+            --decoding-method ${decoding_method}
+  done
+done
+```
+
+Pre-trained models, training and decoding logs, and decoding results are available at <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless4_20220625>
+
+#### [pruned_transducer_stateless5](./pruned_transducer_stateless5)
+
+See <https://github.com/k2-fsa/icefall/pull/454> for more details.
+
+##### Training on full librispeech
+The WERs are (the number in the table formatted as test-clean & test-other):
+
+We only trained 25 epochs for saving time, if you want to get better results you can train more epochs.
+
+| decoding method      | left context | chunk size = 2 | chunk size = 4 | chunk size = 8 | chunk size = 16|
+|----------------------|--------------|----------------|----------------|----------------|----------------|
+| greedy search        | 32           | 3.93 & 9.88    | 3.64 & 9.43    | 3.51 & 8.92    | 3.26 & 8.37    |
+| greedy search        | 64           | 4.84 & 9.81    | 3.59 & 9.27    | 3.44 & 8.83    | 3.23 & 8.33    |
+| fast beam search     | 32           | 3.86 & 9.77    | 3.67 & 9.3     | 3.5 & 8.83     | 3.27 & 8.33    |
+| fast beam search     | 64           | 3.79 & 9.68    | 3.57 & 9.21    | 3.41 & 8.72    | 3.25 & 8.27    |
+| modified beam search | 32           | 3.84 & 9.71    | 3.66 & 9.38    | 3.47 & 8.86    | 3.26 & 8.42    |
+| modified beam search | 64           | 3.81 & 9.59    | 3.58 & 9.2     | 3.44 & 8.74    | 3.23 & 8.35    |
+
+
+**NOTE:** The WERs in table above were decoded with simulate streaming method (i.e. using masking strategy), see commands below. We also have [real streaming decoding](./pruned_transducer_stateless5/streaming_decode.py) script which should produce almost the same results. We tried adding right context in the real streaming decoding, but it seemed not to benefit the performance for all the models, the reasons might be the training and decoding mismatching.
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless5/train.py \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 20 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --world-size 4 \
+  --start-epoch 1 \
+  --num-epochs 25
+```
+
+You can find the tensorboard log here <https://tensorboard.dev/experiment/rO04h6vjTLyw0qSxjp4m4Q>
+
+The decoding command is:
+```bash
+decoding_method="greedy_search"  # "fast_beam_search", "modified_beam_search"
+
+for chunk in 2 4 8 16; do
+  for left in 32 64; do
+    ./pruned_transducer_stateless5/decode.py \
+            --num-encoder-layers 18 \
+            --dim-feedforward 2048 \
+            --nhead 8 \
+            --encoder-dim 512 \
+            --decoder-dim 512 \
+            --joiner-dim 512 \
+            --simulate-streaming 1 \
+            --decode-chunk-size ${chunk} \
+            --left-context ${left} \
+            --causal-convolution 1 \
+            --epoch 25 \
+            --avg 3 \
+            --exp-dir ./pruned_transducer_stateless5/exp \
+            --max-sym-per-frame 1 \
+            --max-duration 1000 \
+            --decoding-method ${decoding_method}
+  done
+done
+```
+
+Pre-trained models, training and decoding logs, and decoding results are available at <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729>
+
+
+### LibriSpeech BPE training results (Pruned Stateless Conv-Emformer RNN-T)
+
+#### [conv_emformer_transducer_stateless](./conv_emformer_transducer_stateless)
+
+It implements [Emformer](https://arxiv.org/abs/2010.10759) augmented with convolution module for streaming ASR.
+It is modified from [torchaudio](https://github.com/pytorch/audio).
+
+See <https://github.com/k2-fsa/icefall/pull/389> for more details.
+
+##### Training on full librispeech
+
+In this model, the lengths of chunk and right context are 32 frames (i.e., 0.32s) and 8 frames (i.e., 0.08s), respectively.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment              | decoding mode        |
+|-------------------------------------|------------|------------|----------------------|----------------------|
+| greedy search (max sym per frame 1) | 3.63       | 9.61       | --epoch 30 --avg 10  | simulated streaming  |
+| greedy search (max sym per frame 1) | 3.64       | 9.65       | --epoch 30 --avg 10  | streaming            |
+| fast beam search                    | 3.61       | 9.4        | --epoch 30 --avg 10  | simulated streaming  |
+| fast beam search                    | 3.58       | 9.5        | --epoch 30 --avg 10  | streaming            |
+| modified beam search                | 3.56       | 9.41       | --epoch 30 --avg 10  | simulated streaming  |
+| modified beam search                | 3.54       | 9.46       | --epoch 30 --avg 10  | streaming            |
+
+The training command is:
+
+```bash
+./conv_emformer_transducer_stateless/train.py \
+  --world-size 6 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/4em2FLsxRwGhmoCRQUEoDw/>
+
+The simulated streaming decoding command using greedy search is:
+```bash
+./conv_emformer_transducer_stateless/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The simulated streaming decoding command using fast beam search is:
+```bash
+./conv_emformer_transducer_stateless/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The simulated streaming decoding command using modified beam search is:
+```bash
+./conv_emformer_transducer_stateless/decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --max-duration 300 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+The streaming decoding command using greedy search is:
+```bash
+./conv_emformer_transducer_stateless/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The streaming decoding command using fast beam search is:
+```bash
+./conv_emformer_transducer_stateless/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The streaming decoding command using modified beam search is:
+```bash
+./conv_emformer_transducer_stateless/streaming_decode.py \
+  --epoch 30 \
+  --avg 10 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --num-decode-streams 2000 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless-2022-06-11>
+
+### LibriSpeech BPE training results (Pruned Stateless Emformer RNN-T)
+
+#### [pruned_stateless_emformer_rnnt2](./pruned_stateless_emformer_rnnt2)
+
+Use <https://github.com/k2-fsa/icefall/pull/390>.
+
+Use [Emformer](https://arxiv.org/abs/2010.10759) from [torchaudio](https://github.com/pytorch/audio)
+for streaming ASR. The Emformer model is imported from torchaudio without modifications.
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+|                                     | test-clean | test-other | comment                                |
+|-------------------------------------|------------|------------|----------------------------------------|
+| greedy search (max sym per frame 1) | 4.28       | 11.42       | --epoch 39 --avg 6  --max-duration 600 |
+| modified beam search                | 4.22       | 11.16       | --epoch 39 --avg 6  --max-duration 600 |
+| fast beam search                    | 4.29       | 11.26       | --epoch 39 --avg 6 --max-duration 600  |
+
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_stateless_emformer_rnnt2/train.py \
+  --world-size 8 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir pruned_stateless_emformer_rnnt2/exp-full \
+  --full-libri 1 \
+  --use-fp16 0 \
+  --max-duration 200 \
+  --prune-range 5 \
+  --lm-scale 0.25 \
+  --master-port 12358 \
+  --num-encoder-layers 18 \
+  --left-context-length 128 \
+  --segment-length 8 \
+  --right-context-length 4
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/ZyiqhAhmRjmr49xml4ofLw/>
+
+The decoding commands are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  for epoch in 39; do
+    for avg in 6; do
+      ./pruned_stateless_emformer_rnnt2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --use-averaged-model 1 \
+        --exp-dir pruned_stateless_emformer_rnnt2/exp-full \
+        --max-duration 50 \
+        --decoding-method $m \
+        --num-encoder-layers 18 \
+        --left-context-length 128 \
+        --segment-length 8 \
+        --right-context-length 4
+    done
+  done
+done
+```
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-stateless-emformer-rnnt2-2022-06-01>
+
+
+### LibriSpeech BPE training results (Pruned Stateless Transducer 5)
+
+#### [pruned_transducer_stateless5](./pruned_transducer_stateless5)
+
+Same as `Pruned Stateless Transducer 2` but with more layers.
+
+See <https://github.com/k2-fsa/icefall/pull/330>
+
+Note that models in `pruned_transducer_stateless` and `pruned_transducer_stateless2`
+have about 80 M parameters.
+
+The notations `large` and `medium` below are from the [Conformer](https://arxiv.org/pdf/2005.08100.pdf)
+paper, where the large model has about 118 M parameters and the medium model
+has 30.8 M parameters.
+
+##### Large
+
+Number of model parameters 118129516 (i.e, 118.13 M).
+
+|                                     | test-clean | test-other | comment                                |
+|-------------------------------------|------------|------------|----------------------------------------|
+| greedy search (max sym per frame 1) | 2.43       | 5.72       | --epoch 30 --avg 10 --max-duration 600 |
+| modified beam search                | 2.43       | 5.69       | --epoch 30 --avg 10 --max-duration 600 |
+| fast beam search                    | 2.43       | 5.67       | --epoch 30 --avg 10 --max-duration 600 |
+
+The training commands are:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --full-libri 1 \
+  --exp-dir pruned_transducer_stateless5/exp-L \
+  --max-duration 300 \
+  --use-fp16 0 \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/aWzDj5swSE2VmcOYgoe3vQ>
+
+The decoding commands are:
+
+```bash
+for method in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless5/decode.py \
+    --epoch 30 \
+    --avg 10 \
+    --exp-dir ./pruned_transducer_stateless5/exp-L \
+    --max-duration 600 \
+    --decoding-method $method \
+    --max-sym-per-frame 1 \
+    --num-encoder-layers 18 \
+    --dim-feedforward 2048 \
+    --nhead 8 \
+    --encoder-dim 512 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    --use-averaged-model True
+done
+```
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-2022-07-07>
+
+
+##### Medium
+
+Number of model parameters 30896748 (i.e, 30.9 M).
+
+|                                     | test-clean | test-other | comment                                 |
+|-------------------------------------|------------|------------|-----------------------------------------|
+| greedy search (max sym per frame 1) | 2.87       | 6.92       | --epoch 30 --avg 10  --max-duration 600 |
+| modified beam search                | 2.83       | 6.75       | --epoch 30 --avg 10  --max-duration 600 |
+| fast beam search                    | 2.81       | 6.76       | --epoch 30 --avg 10  --max-duration 600  |
+
+The training commands are:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --full-libri 1 \
+  --exp-dir pruned_transducer_stateless5/exp-M \
+  --max-duration 300 \
+  --use-fp16 0 \
+  --num-encoder-layers 18 \
+  --dim-feedforward 1024 \
+  --nhead 4 \
+  --encoder-dim 256 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/04xtWUKPRmebSnpzN1GMHQ>
+
+The decoding commands are:
+
+```bash
+for method in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless5/decode.py \
+    --epoch 30 \
+    --avg 10 \
+    --exp-dir ./pruned_transducer_stateless5/exp-M \
+    --max-duration 600 \
+    --decoding-method $method \
+    --max-sym-per-frame 1 \
+    --num-encoder-layers 18 \
+    --dim-feedforward 1024 \
+    --nhead 4 \
+    --encoder-dim 256 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    --use-averaged-model True
+done
+```
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-M-2022-07-07>
+
+
+##### Baseline-2
+
+It has 87.8 M parameters. Compared to the model in pruned_transducer_stateless2, its has more
+layers (24 v.s 12) but a narrower model (1536 feedforward dim and 384 encoder dim vs 2048 feed forward dim and 512 encoder dim).
+
+|                                     | test-clean | test-other | comment                                 |
+|-------------------------------------|------------|------------|-----------------------------------------|
+| greedy search (max sym per frame 1) | 2.54       | 5.72       | --epoch 30 --avg 10  --max-duration 600 |
+| modified beam search                | 2.47       | 5.71       | --epoch 30 --avg 10  --max-duration 600 |
+| modified beam search + RNNLM shallow fusion     | 2.27       | 5.24      | --epoch 30 --avg 10  --max-duration 600 |
+| modified beam search + RNNLM + LODR     | 2.23       | 5.17      | --epoch 30 --avg 10  --max-duration 600 |
+| modified beam search + TransformerLM shallow fusion     | 2.27       | 5.26      | --epoch 30 --avg 10  --max-duration 600 |
+| modified beam search + TransformerLM + LODR     | 2.22       | 5.11      | --epoch 30 --avg 10  --max-duration 600 |
+| fast beam search                    | 2.5        | 5.72       | --epoch 30 --avg 10  --max-duration 600 |
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --full-libri 1 \
+  --exp-dir pruned_transducer_stateless5/exp-B \
+  --max-duration 300 \
+  --use-fp16 0 \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/foVHNyqiRi2LhybmRUOAyg>
+
+The decoding commands are:
+
+```bash
+for method in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless5/decode.py \
+    --epoch 30 \
+    --avg 10 \
+    --exp-dir ./pruned_transducer_stateless5/exp-B \
+    --max-duration 600 \
+    --decoding-method $method \
+    --max-sym-per-frame 1 \
+    --num-encoder-layers 24 \
+    --dim-feedforward 1536 \
+    --nhead 8 \
+    --encoder-dim 384 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    --use-averaged-model True
+done
+```
+
+To decode with RNNLM shallow fusion, use the following decoding command. A well-trained RNNLM
+can be found here: <https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>
+
+```bash
+for method in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless5/decode.py \
+    --epoch 30 \
+    --avg 10 \
+    --exp-dir ./pruned_transducer_stateless5/exp-B \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_rnnlm_shallow_fusion \
+    --max-sym-per-frame 1 \
+    --num-encoder-layers 24 \
+    --dim-feedforward 1536 \
+    --nhead 8 \
+    --encoder-dim 384 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    --use-averaged-model True
+    --beam 4 \
+    --max-contexts 4 \
+    --rnn-lm-scale 0.4 \
+    --rnn-lm-exp-dir /path/to/RNNLM/exp \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+done
+```
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-B-2022-07-07>
+
+
+### LibriSpeech BPE training results (Pruned Stateless Transducer 4)
+
+#### [pruned_transducer_stateless4](./pruned_transducer_stateless4)
+
+This version saves averaged model during training, and decodes with averaged model.
+
+See <https://github.com/k2-fsa/icefall/issues/337> for details about the idea of model averaging.
+
+##### Training on full librispeech
+
+See <https://github.com/k2-fsa/icefall/pull/344>
+
+Using commit `ec0b0e92297cc03fdb09f48cd235e84d2c04156b`.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|-------------------------------------------------------------------------------|
+| greedy search (max sym per frame 1) | 2.75       | 6.74       | --epoch 30 --avg 6  --use-averaged-model False                                |
+| greedy search (max sym per frame 1) | 2.69       | 6.64       | --epoch 30 --avg 6  --use-averaged-model True                                 |
+| fast beam search                    | 2.72       | 6.67       | --epoch 30 --avg 6  --use-averaged-model False                                |
+| fast beam search                    | 2.66       | 6.6        | --epoch 30 --avg 6  --use-averaged-model True                                 |
+| modified beam search                | 2.67       | 6.68       | --epoch 30 --avg 6  --use-averaged-model False                                |
+| modified beam search                | 2.62       | 6.57       | --epoch 30 --avg 6  --use-averaged-model True                                 |
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless4/train.py \
+  --world-size 6 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --save-every-n 8000 \
+  --keep-last-k 20 \
+  --average-period 100
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/QOGSPBgsR8KzcRMmie9JGw/>
+
+The decoding command using greedy search is:
+```bash
+./pruned_transducer_stateless4/decode.py \
+  --epoch 30 \
+  --avg 6 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --max-duration 300 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The decoding command using fast beam search is:
+```bash
+./pruned_transducer_stateless4/decode.py \
+  --epoch 30 \
+  --avg 6 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --max-duration 300 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The decoding command using modified beam search is:
+```bash
+./pruned_transducer_stateless4/decode.py \
+  --epoch 30 \
+  --avg 6 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --max-duration 300 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless4-2022-06-03>
+
+##### Training on train-clean-100
+
+See <https://github.com/k2-fsa/icefall/pull/344>
+
+Using commit `ec0b0e92297cc03fdb09f48cd235e84d2c04156b`.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|-------------------------------------------------------------------------------|
+| greedy search (max sym per frame 1) | 7.0        | 18.95      | --epoch 30 --avg 10 --use_averaged_model False                                |
+| greedy search (max sym per frame 1) | 6.92       | 18.65      | --epoch 30 --avg 10 --use_averaged_model True                                 |
+| fast beam search                    | 6.82       | 18.47      | --epoch 30 --avg 10 --use_averaged_model False                                |
+| fast beam search                    | 6.74       | 18.2       | --epoch 30 --avg 10 --use_averaged_model True                                 |
+| modified beam search                | 6.74       | 18.39      | --epoch 30 --avg 10 --use_averaged_model False                                |
+| modified beam search                | 6.74       | 18.12      | --epoch 30 --avg 10 --use_averaged_model True                                 |
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless4/train.py \
+  --world-size 3 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 0 \
+  --max-duration 300 \
+  --save-every-n 8000 \
+  --keep-last-k 20 \
+  --average-period 100
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/YVYHq1irQS69s9bW1vQ06Q/>
+
+### LibriSpeech BPE training results (Pruned Stateless Transducer 3, 2022-04-29)
+
+#### [pruned_transducer_stateless3](./pruned_transducer_stateless3)
+Same as `Pruned Stateless Transducer 2` but using the XL subset from
+[GigaSpeech](https://github.com/SpeechColab/GigaSpeech) as extra training data.
+
+During training, it selects either a batch from GigaSpeech with prob `giga_prob`
+or a batch from LibriSpeech with prob `1 - giga_prob`. All utterances within
+a batch come from the same dataset.
+
+Using commit `ac84220de91dee10c00e8f4223287f937b1930b6`.
+
+See <https://github.com/k2-fsa/icefall/pull/312>.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|----------------------------------------|
+| greedy search (max sym per frame 1) | 2.21       | 5.09       | --epoch 27 --avg 2  --max-duration 600 |
+| greedy search (max sym per frame 1) | 2.25       | 5.02       | --epoch 27 --avg 12 --max-duration 600 |
+| modified beam search                | 2.19       | 5.03       | --epoch 25 --avg 6  --max-duration 600 |
+| modified beam search                | 2.23       | 4.94       | --epoch 27 --avg 10 --max-duration 600 |
+| beam search                         | 2.16       | 4.95       | --epoch 25 --avg 7  --max-duration 600 |
+| fast beam search                    | 2.21       | 4.96       | --epoch 27 --avg 10 --max-duration 600 |
+| fast beam search                    | 2.19       | 4.97       | --epoch 27 --avg 12 --max-duration 600 |
+
+The training commands are:
+
+```bash
+./prepare.sh
+./prepare_giga_speech.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --full-libri 1 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --max-duration 300 \
+  --use-fp16 1 \
+  --lr-epochs 4 \
+  --num-workers 2 \
+  --giga-prob 0.8
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/gaD34WeYSMCOkzoo3dZXGg/>
+(Note: The training process is killed manually after saving `epoch-28.pt`.)
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-04-29>
+
+The decoding commands are:
+
+```bash
+
+# greedy search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 27 \
+    --avg 2 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search \
+    --max-sym-per-frame 1
+
+# modified beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 25 \
+    --avg 6 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --max-sym-per-frame 1
+
+# beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 25 \
+    --avg 7 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --max-sym-per-frame 1
+
+# fast beam search
+for epoch in 27; do
+  for avg in 10 12; do
+    ./pruned_transducer_stateless3/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --max-duration 600 \
+        --decoding-method fast_beam_search \
+        --max-states 32 \
+        --beam 8
+  done
+done
+```
+
+The following table shows the
+[Nbest oracle WER](http://kaldi-asr.org/doc/lattices.html#lattices_operations_oracle)
+for fast beam search.
+
+| epoch | avg | num_paths | nbest_scale | test-clean | test-other |
+|-------|-----|-----------|-------------|------------|------------|
+|  27   | 10  |   50      | 0.5         |  0.91      |  2.74      |
+|  27   | 10  |   50      | 0.8         |  0.94      |  2.82      |
+|  27   | 10  |   50      | 1.0         |  1.06      |  2.88      |
+|  27   | 10  |   100     | 0.5         |  0.82      |  2.58      |
+|  27   | 10  |   100     | 0.8         |  0.92      |  2.65      |
+|  27   | 10  |   100     | 1.0         |  0.95      |  2.77      |
+|  27   | 10  |   200     | 0.5         |  0.81      |  2.50      |
+|  27   | 10  |   200     | 0.8         |  0.85      |  2.56      |
+|  27   | 10  |   200     | 1.0         |  0.91      |  2.64      |
+|  27   | 10  |   400     | 0.5         |  N/A       |  N/A       |
+|  27   | 10  |   400     | 0.8         |  0.81      |  2.49      |
+|  27   | 10  |   400     | 1.0         |  0.85      |  2.54      |
+
+The Nbest oracle WER is computed using the following steps:
+
+  - 1. Use `fast_beam_search` to produce a lattice.
+  - 2. Extract `N` paths from the lattice using [k2.random_path](https://k2-fsa.github.io/k2/python_api/api.html#random-paths)
+  - 3. [Unique](https://k2-fsa.github.io/k2/python_api/api.html#unique) paths so that each path
+       has a distinct sequence of tokens
+  - 4. Compute the edit distance of each path with the ground truth
+  - 5. The path with the lowest edit distance is the final output and is used to
+       compute the WER
+
+The command to compute the Nbest oracle WER is:
+
+```bash
+for epoch in 27; do
+  for avg in 10 ; do
+    for num_paths in 50 100 200 400; do
+      for nbest_scale in 0.5 0.8 1.0; do
+        ./pruned_transducer_stateless3/decode.py \
+            --epoch $epoch \
+            --avg $avg \
+            --exp-dir ./pruned_transducer_stateless3/exp \
+            --max-duration 600 \
+            --decoding-method fast_beam_search_nbest_oracle \
+            --num-paths $num_paths \
+            --max-states 32 \
+            --beam 8 \
+            --nbest-scale $nbest_scale
+      done
+    done
+  done
+done
+```
+
+### LibriSpeech BPE training results (Pruned Transducer 3, 2022-05-13)
+
+Same setup as [pruned_transducer_stateless3](./pruned_transducer_stateless3) (2022-04-29)
+but change `--giga-prob` from 0.8 to 0.9. Also use `repeat` on gigaspeech XL
+subset so that the gigaspeech dataloader never exhausts.
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|---------------------------------------------|
+| greedy search (max sym per frame 1) | 2.03       | 4.70       | --iter 1224000 --avg 14  --max-duration 600 |
+| modified beam search                | 2.00       | 4.63       | --iter 1224000 --avg 14  --max-duration 600 |
+| modified beam search + rnnlm shallow fusion  | 1.94     |  4.2    | --iter 1224000 --avg 14  --max-duration 600 |
+| modified beam search + rnnlm + LODR         | 1.77       | 3.99       | --iter 1224000 --avg 14  --max-duration 600 |
+| modified beam search + TransformerLM + LODR    | 1.75       | 3.94       | --iter 1224000 --avg 14  --max-duration 600 |
+| fast beam search                    | 2.10       | 4.68       | --iter 1224000 --avg 14 --max-duration 600 |
+
+The training commands are:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./prepare.sh
+./prepare_giga_speech.sh
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --full-libri 1 \
+  --exp-dir pruned_transducer_stateless3/exp-0.9 \
+  --max-duration 300 \
+  --use-fp16 1 \
+  --lr-epochs 4 \
+  --num-workers 2 \
+  --giga-prob 0.9
+```
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/HpocR7dKS9KCQkJeYxfXug/>
+
+Decoding commands:
+
+```bash
+for iter in 1224000; do
+  for avg in 14; do
+    for method in greedy_search modified_beam_search fast_beam_search ; do
+      ./pruned_transducer_stateless3/decode.py \
+        --iter $iter \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless3/exp-0.9/ \
+        --max-duration 600 \
+        --decoding-method $method \
+        --max-sym-per-frame 1 \
+        --beam 4 \
+        --max-contexts 32
+    done
+  done
+done
+```
+You may also decode using shallow fusion with external neural network LM. To do so you need to
+download a well-trained NN LM:
+RNN LM: <https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>
+Transformer LM: <https://huggingface.co/marcoyang/icefall-librispeech-transformer-lm/tree/main>
+
+```bash
+rnn_lm_scale=0.3
+
+for iter in 1224000; do
+  for avg in 14; do
+    for method in modified_beam_search_rnnlm_shallow_fusion ; do
+      ./pruned_transducer_stateless3/decode.py \
+        --iter $iter \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless3/exp-0.9/ \
+        --max-duration 600 \
+        --decoding-method $method \
+        --max-sym-per-frame 1 \
+        --beam 4 \
+        --max-contexts 32 \
+        --rnn-lm-scale $rnn_lm_scale \
+        --rnn-lm-exp-dir /path/to/RNNLM \
+        --rnn-lm-epoch 99 \
+        --rnn-lm-avg 1 \
+        --rnn-lm-num-layers 3 \
+        --rnn-lm-tie-weights 1
+    done
+  done
+done
+```
+
+If you want to try out with LODR decoding, use the following command. This assums you have a bi-gram LM trained on LibriSpeech text. You can also download the bi-gram LM from here <https://huggingface.co/marcoyang/librispeech_bigram/tree/main> and put it under the directory `data/lang_bpe_500`.
+
+```bash
+rnn_lm_scale=0.4
+
+for iter in 1224000; do
+  for avg in 14; do
+    for method in modified_beam_search_rnnlm_LODR ; do
+      ./pruned_transducer_stateless3/decode.py \
+        --iter $iter \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless3/exp-0.9/ \
+        --max-duration 600 \
+        --decoding-method $method \
+        --max-sym-per-frame 1 \
+        --beam 4 \
+        --max-contexts 32 \
+        --rnn-lm-scale $rnn_lm_scale \
+        --rnn-lm-exp-dir /path/to/RNNLM \
+        --rnn-lm-epoch 99 \
+        --rnn-lm-avg 1 \
+        --rnn-lm-num-layers 3 \
+        --rnn-lm-tie-weights 1 \
+        --tokens-ngram 2 \
+        --ngram-lm-scale -0.14
+    done
+  done
+done
+```
+
+The pretrained models, training logs, decoding logs, and decoding results
+can be found at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13>
+
+
+### LibriSpeech BPE training results (Pruned Transducer 2)
+
+#### [pruned_transducer_stateless2](./pruned_transducer_stateless2)
+This is with a reworked version of the conformer encoder, with many changes.
+
+##### Training on full librispeech
+
+Using commit `34aad74a2c849542dd5f6359c9e6b527e8782fd6`.
+See <https://github.com/k2-fsa/icefall/pull/288>
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|-------------------------------------------------------------------------------|
+| greedy search (max sym per frame 1) | 2.62       | 6.37       | --epoch 25 --avg 8  --max-duration 600                                        |
+| fast beam search                    | 2.61       | 6.17       | --epoch 25 --avg 8  --max-duration 600 --decoding-method fast_beam_search     |
+| modified beam search                | 2.59       | 6.19       | --epoch 25 --avg 8  --max-duration 600 --decoding-method modified_beam_search |
+| greedy search (max sym per frame 1) | 2.70       | 6.04       | --epoch 34 --avg 10 --max-duration 600                                        |
+| fast beam search                    | 2.66       | 6.00       | --epoch 34 --avg 10  --max-duration 600 --decoding-method fast_beam_search    |
+| greedy search (max sym per frame 1) | 2.62       | 6.03       | --epoch 38 --avg 10 --max-duration 600                                        |
+| fast beam search                    | 2.57       | 5.95       | --epoch 38 --avg 10  --max-duration 600 --decoding-method fast_beam_search    |
+
+
+
+
+The train and decode commands are:
+```bash
+python3 ./pruned_transducer_stateless2/train.py \
+  --exp-dir=pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 26 \
+  --full-libri 1 \
+  --max-duration 300
+```
+
+and:
+
+```bash
+python3 ./pruned_transducer_stateless2/decode.py \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --epoch 25 \
+  --avg 8 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 600
+```
+
+The Tensorboard log is at <https://tensorboard.dev/experiment/Xoz0oABMTWewo1slNFXkyA> (apologies, log starts
+only from epoch 3).
+
+The pretrained models, training logs, decoding logs, and decoding results
+can be found at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless2-2022-04-29>
+
+
+##### Training on train-clean-100:
+
+Trained with 1 job:
+```
+python3 ./pruned_transducer_stateless2/train.py \
+  --exp-dir=pruned_transducer_stateless2/exp_100h_ws1 \
+  --world-size 1 \
+  --num-epochs 40  \
+  --full-libri 0 \
+  --max-duration 300
+```
+
+and decoded with:
+
+```
+python3 ./pruned_transducer_stateless2/decode.py \
+  --exp-dir pruned_transducer_stateless2/exp_100h_ws1 \
+  --epoch 19 \
+  --avg 8 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 600
+```
+
+
+The Tensorboard log is at <https://tensorboard.dev/experiment/AhnhooUBRPqTnaggoqo7lg> (learning rate
+schedule is not visible due to a since-fixed bug).
+
+|                                     | test-clean | test-other | comment                                               |
+|-------------------------------------|------------|------------|-------------------------------------------------------|
+| greedy search (max sym per frame 1) | 7.12       | 18.42      | --epoch 19 --avg 8                                    |
+| greedy search (max sym per frame 1) | 6.71       | 17.77      | --epoch 29 --avg 8                                    |
+| greedy search (max sym per frame 1) | 6.64       | 17.19      | --epoch 39 --avg 10                                    |
+| fast beam search                    | 6.58       | 17.27      | --epoch 29 --avg 8 --decoding-method fast_beam_search |
+| fast beam search                    | 6.53       | 16.82      | --epoch 39 --avg 10 --decoding-method fast_beam_search |
+
+Trained with 2 jobs:
+
+```bash
+python3 ./pruned_transducer_stateless2/train.py \
+  --exp-dir=pruned_transducer_stateless2/exp_100h_ws2 \
+  --world-size 2 \
+  --num-epochs 40  \
+  --full-libri 0 \
+  --max-duration 300
+```
+
+and decoded with:
+
+```
+python3 ./pruned_transducer_stateless2/decode.py \
+  --exp-dir pruned_transducer_stateless2/exp_100h_ws2 \
+  --epoch 19 \
+  --avg 8 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 600
+```
+
+The Tensorboard log is at <https://tensorboard.dev/experiment/dvOC9wsrSdWrAIdsebJILg/>
+(learning rate schedule is not visible due to a since-fixed bug).
+
+|                                     | test-clean | test-other | comment               |
+|-------------------------------------|------------|------------|-----------------------|
+| greedy search (max sym per frame 1) | 7.05       | 18.77      | --epoch 19  --avg 8   |
+| greedy search (max sym per frame 1) | 6.82       | 18.14      | --epoch 29  --avg 8   |
+| greedy search (max sym per frame 1) | 6.81       | 17.66      | --epoch 30  --avg 10  |
+
+
+Trained with 4 jobs:
+
+```
+python3 ./pruned_transducer_stateless2/train.py \
+  --exp-dir=pruned_transducer_stateless2/exp_100h_ws4 \
+  --world-size 4 \
+  --num-epochs 40  \
+  --full-libri 0 \
+  --max-duration 300
+```
+
+and decoded with:
+
+```
+python3 ./pruned_transducer_stateless2/decode.py \
+  --exp-dir pruned_transducer_stateless2/exp_100h_ws4 \
+  --epoch 19 \
+  --avg 8 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 600
+```
+
+
+The Tensorboard log is at <https://tensorboard.dev/experiment/a3T0TyC0R5aLj5bmFbRErA/>
+(learning rate schedule is not visible due to a since-fixed bug).
+
+|                                     | test-clean | test-other | comment               |
+|-------------------------------------|------------|------------|-----------------------|
+| greedy search (max sym per frame 1) | 7.31       | 19.55      | --epoch 19  --avg 8   |
+| greedy search (max sym per frame 1) | 7.08       | 18.59      | --epoch 29  --avg 8   |
+| greedy search (max sym per frame 1) | 6.86       | 18.29      | --epoch 30  --avg 10  |
+
+
+
+Trained with 1 job, with  --use-fp16=True --max-duration=300 i.e. with half-precision
+floats (but without increasing max-duration), after merging <https://github.com/k2-fsa/icefall/pull/305>.
+Train command was
+
+```
+python3 ./pruned_transducer_stateless2/train.py \
+  --exp-dir=pruned_transducer_stateless2/exp_100h_fp16 \
+  --world-size 1 \
+  --num-epochs 40  \
+  --full-libri 0 \
+  --max-duration 300 \
+  --use-fp16 True
+```
+
+The Tensorboard log is at <https://tensorboard.dev/experiment/DAtGG9lpQJCROUDwPNxwpA>
+
+|                                     | test-clean | test-other | comment               |
+|-------------------------------------|------------|------------|-----------------------|
+| greedy search (max sym per frame 1) | 7.10       | 18.57      | --epoch 19  --avg 8   |
+| greedy search (max sym per frame 1) | 6.81       | 17.84      | --epoch 29  --avg 8   |
+| greedy search (max sym per frame 1) | 6.63       | 17.39      | --epoch 30  --avg 10  |
+
+
+Trained with 1 job, with  --use-fp16=True --max-duration=500, i.e. with half-precision
+floats and max-duration increased from 300 to 500, after merging <https://github.com/k2-fsa/icefall/pull/305>.
+Train command was
+
+```
+python3 ./pruned_transducer_stateless2/train.py \
+  --exp-dir=pruned_transducer_stateless2/exp_100h_fp16 \
+  --world-size 1 \
+  --num-epochs 40  \
+  --full-libri 0 \
+  --max-duration 500 \
+  --use-fp16 True
+```
+
+The Tensorboard log is at <https://tensorboard.dev/experiment/Km7QBHYnSLWs4qQnAJWsaA>
+
+|                                     | test-clean | test-other | comment               |
+|-------------------------------------|------------|------------|-----------------------|
+| greedy search (max sym per frame 1) | 7.10       | 18.79      | --epoch 19  --avg 8   |
+| greedy search (max sym per frame 1) | 6.92       | 18.16      | --epoch 29  --avg 8   |
+| greedy search (max sym per frame 1) | 6.89       | 17.75      | --epoch 30  --avg 10  |
+
+
+
+### LibriSpeech BPE training results (Pruned Transducer)
+
+Conformer encoder + non-current decoder. The decoder
+contains only an embedding layer, a Conv1d (with kernel size 2) and a linear
+layer (to transform tensor dim).
+
+#### 2022-03-12
+
+[pruned_transducer_stateless](./pruned_transducer_stateless)
+
+Using commit `1603744469d167d848e074f2ea98c587153205fa`.
+See <https://github.com/k2-fsa/icefall/pull/248>
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                  |
+|-------------------------------------|------------|------------|------------------------------------------|
+| greedy search (max sym per frame 1) | 2.62       | 6.37       | --epoch 42  --avg 11  --max-duration 100 |
+| greedy search (max sym per frame 2) | 2.62       | 6.37       | --epoch 42  --avg 11  --max-duration 100 |
+| greedy search (max sym per frame 3) | 2.62       | 6.37       | --epoch 42  --avg 11  --max-duration 100 |
+| modified beam search (beam size 4)  | 2.56       | 6.27       | --epoch 42  --avg 11  --max-duration 100 |
+| beam search (beam size 4)           | 2.57       | 6.27       | --epoch 42  --avg 11  --max-duration 100 |
+
+
+
+
+
+The decoding time for `test-clean` and `test-other` is given below:
+(A V100 GPU with 32 GB RAM is used for decoding. Note: Not all GPU RAM is used during decoding.)
+
+| decoding method | test-clean (seconds) | test-other (seconds)|
+|---|---:|---:|
+| greedy search (--max-sym-per-frame=1) | 160 | 159 |
+| greedy search (--max-sym-per-frame=2) | 184 | 177 |
+| greedy search (--max-sym-per-frame=3) | 210 | 213 |
+| modified beam search (--beam-size 4)| 273 | 269 |
+|beam search (--beam-size 4) | 2741 | 2221 |
+
+We recommend you to use `modified_beam_search`.
+
+Training command:
+
+```bash
+cd egs/librispeech/ASR/
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+. path.sh
+
+./pruned_transducer_stateless/train.py \
+  --world-size 8 \
+  --num-epochs 60 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --prune-range 5 \
+  --lr-factor 5 \
+  --lm-scale 0.25
+```
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/WKRFY5fYSzaVBHahenpNlA/>
+
+The command for decoding is:
+
+```bash
+epoch=42
+avg=11
+sym=1
+
+# greedy search
+
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --max-duration 100 \
+  --decoding-method greedy_search \
+  --beam-size 4 \
+  --max-sym-per-frame $sym
+
+# modified beam search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+# beam search
+# (not recommended)
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --max-duration 100 \
+  --decoding-method beam_search \
+  --beam-size 4
+```
+
+You can find a pre-trained model, decoding logs, and decoding results at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12>
+
+#### 2022-02-18
+
+[pruned_transducer_stateless](./pruned_transducer_stateless)
+
+
+The WERs are
+
+|                           | test-clean | test-other | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             | 2.85       | 6.98       | --epoch 28  --avg 15  --max-duration 100 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --prune-range 5 \
+  --lr-factor 5 \
+  --lm-scale 0.25 \
+```
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/ejG7VpakRYePNNj6AbDEUw/#scalars>
+
+The decoding command is:
+```
+epoch=28
+avg=15
+
+## greedy search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --max-duration 100
+```
+
+
+### LibriSpeech BPE training results (Transducer)
+
+#### Conformer encoder + embedding decoder
+
+Conformer encoder + non-recurrent decoder. The decoder
+contains only an embedding layer and a Conv1d (with kernel size 2).
+
+See
+
+- [./transducer_stateless](./transducer_stateless)
+- [./transducer_stateless_multi_datasets](./transducer_stateless_multi_datasets)
+
+##### 2022-03-01
+
+Using commit `2332ba312d7ce72f08c7bac1e3312f7e3dd722dc`.
+
+It uses [GigaSpeech](https://github.com/SpeechColab/GigaSpeech)
+as extra training data. 20% of the time it selects a batch from L subset of
+GigaSpeech and 80% of the time it selects a batch from LibriSpeech.
+
+The WERs are
+
+|                                     | test-clean | test-other | comment                                  |
+|-------------------------------------|------------|------------|------------------------------------------|
+| greedy search (max sym per frame 1) | 2.64       | 6.55       | --epoch 39  --avg 15  --max-duration 100 |
+| modified beam search (beam size 4)  | 2.61       | 6.46       | --epoch 39  --avg 15  --max-duration 100 |
+
+The training command for reproducing is given below:
+
+```bash
+cd egs/librispeech/ASR/
+./prepare.sh
+./prepare_giga_speech.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless_multi_datasets/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_multi_datasets/exp-full-2 \
+  --full-libri 1 \
+  --max-duration 300 \
+  --lr-factor 5 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --modified-transducer-prob 0.25 \
+  --giga-prob 0.2
+```
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/xmo5oCgrRVelH9dCeOkYBg/>
+
+The decoding command is:
+
+```bash
+epoch=39
+avg=15
+sym=1
+
+# greedy search
+./transducer_stateless_multi_datasets/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless_multi_datasets/exp-full-2 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --context-size 2 \
+  --max-sym-per-frame $sym
+
+# modified beam search
+./transducer_stateless_multi_datasets/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless_multi_datasets/exp-full-2 \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --context-size 2 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+```
+
+You can find a pretrained model by visiting
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-multi-datasets-bpe-500-2022-03-01>
+
+
+##### 2022-04-19
+
+[transducer_stateless2](./transducer_stateless2)
+
+This version uses torchaudio's RNN-T loss.
+
+Using commit `fce7f3cd9a486405ee008bcbe4999264f27774a3`.
+See <https://github.com/k2-fsa/icefall/pull/316>
+
+|                                     | test-clean | test-other | comment                                                                        |
+|-------------------------------------|------------|------------|--------------------------------------------------------------------------------|
+| greedy search (max sym per frame 1) | 2.65       | 6.30       | --epoch 59 --avg 10  --max-duration 600                                        |
+| greedy search (max sym per frame 2) | 2.62       | 6.23       | --epoch 59 --avg 10  --max-duration 100                                        |
+| greedy search (max sym per frame 3) | 2.62       | 6.23       | --epoch 59 --avg 10  --max-duration 100                                        |
+| modified beam search                | 2.63       | 6.15       | --epoch 59 --avg 10  --max-duration 100 --decoding-method modified_beam_search |
+| beam search                         | 2.59       | 6.15       | --epoch 59 --avg 10  --max-duration 100 --decoding-method beam_search          |
+
+**Note**: This model is trained with standard RNN-T loss. Neither modified transducer nor pruned RNN-T is used.
+You can see that there is a performance degradation in WER when we limit the max symbol per frame to 1.
+
+The number of active paths in `modified_beam_search` and `beam_search` is 4.
+
+The training and decoding commands are:
+
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./transducer_stateless2/train.py \
+  --world-size 8 \
+  --num-epochs 60 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless2/exp-2 \
+  --full-libri 1 \
+  --max-duration 300 \
+  --lr-factor 5
+
+epoch=59
+avg=10
+# greedy search
+./transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./transducer_stateless2/exp-2 \
+  --max-duration 600 \
+  --decoding-method greedy_search \
+  --max-sym-per-frame 1
+
+# modified beam search
+./transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./transducer_stateless2/exp-2 \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+
+# beam search
+./transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./transducer_stateless2/exp-2 \
+  --max-duration 100 \
+  --decoding-method beam_search \
+```
+
+The tensorboard log is at <https://tensorboard.dev/experiment/oAlle3dxQD2EY8ePwjIGuw/>.
+
+
+You can find a pre-trained model, decoding logs, and decoding results at
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless2-torchaudio-2022-04-19>
+
+
+
+##### 2022-02-07
+
+Using commit `a8150021e01d34ecbd6198fe03a57eacf47a16f2`.
+
+
+The WERs are
+
+|                                     | test-clean | test-other | comment                                  |
+|-------------------------------------|------------|------------|------------------------------------------|
+| greedy search (max sym per frame 1) | 2.67       | 6.67       | --epoch 63  --avg 19  --max-duration 100 |
+| greedy search (max sym per frame 2) | 2.67       | 6.67       | --epoch 63  --avg 19  --max-duration 100 |
+| greedy search (max sym per frame 3) | 2.67       | 6.67       | --epoch 63  --avg 19  --max-duration 100 |
+| modified beam search (beam size 4)  | 2.67       | 6.57       | --epoch 63  --avg 19  --max-duration 100 |
+
+
+The training command for reproducing is given below:
+
+```
+cd egs/librispeech/ASR/
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 76 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp-full \
+  --full-libri 1 \
+  --max-duration 300 \
+  --lr-factor 5 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --modified-transducer-prob 0.25
+```
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/qgvWkbF2R46FYA6ZMNmOjA/#scalars>
+
+The decoding command is:
+```
+epoch=63
+avg=19
+
+## greedy search
+for sym in 1 2 3; do
+  ./transducer_stateless/decode.py \
+    --epoch $epoch \
+    --avg $avg \
+    --exp-dir transducer_stateless/exp-full \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --max-duration 100 \
+    --max-sym-per-frame $sym
+done
+
+## modified beam search
+
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp-full \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --context-size 2 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+```
+
+You can find a pretrained model by visiting
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07>
+
+
+#### Conformer encoder + LSTM decoder
+Using commit `8187d6236c2926500da5ee854f758e621df803cc`.
+
+Conformer encoder + LSTM decoder.
+
+The best WER is
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 3.07       | 7.51       |
+
+using `--epoch 34 --avg 11` with **greedy search**.
+
+The training command to reproduce the above WER is:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer/train.py \
+  --world-size 4 \
+  --num-epochs 35 \
+  --start-epoch 0 \
+  --exp-dir transducer/exp-lr-2.5-full \
+  --full-libri 1 \
+  --max-duration 180 \
+  --lr-factor 2.5
+```
+
+The decoding command is:
+
+```
+epoch=34
+avg=11
+
+./transducer/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer/exp-lr-2.5-full \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100
+```
+
+You can find the tensorboard log at: <https://tensorboard.dev/experiment/D7NQc3xqTpyVmWi5FnWjrA>
+
+
+### LibriSpeech BPE training results (Conformer-CTC 2)
+
+#### [conformer_ctc2](./conformer_ctc2)
+
+#### 2022-07-21
+
+It implements a 'reworked' version of CTC attention model.
+As demenstrated by pruned_transducer_stateless2, reworked Conformer model has superior performance compared to the original Conformer.
+So in this modified version of CTC attention model, it has the reworked Conformer as the encoder and the reworked Transformer as the decoder.
+conformer_ctc2 also integrates with the idea of the 'averaging models' in pruned_transducer_stateless4.
+
+The WERs on comparisons with a baseline model, for the librispeech test dataset, are listed as below.
+
+The baseline model is the original conformer CTC attention model trained with icefall/egs/librispeech/ASR/conformer_ctc.
+The model is downloaded from  <https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09>.
+This model has 12 layers of Conformer encoder layers and 6 Transformer decoder layers.
+Number of model parameters is 109,226,120.
+It has been trained with 90 epochs with full Librispeech dataset.
+
+For this reworked CTC attention model, it has 12 layers of reworked Conformer encoder layers and 6 reworked Transformer decoder layers.
+Number of model parameters is 103,071,035.
+With full Librispeech data set, it was trained for **only** 30 epochs because the reworked model would converge much faster.
+Please refer to <https://tensorboard.dev/experiment/GR1s6VrJRTW5rtB50jakew/#scalars> to see the loss convergence curve.
+Please find the above trained model at <https://huggingface.co/WayneWiser/icefall-asr-librispeech-conformer-ctc2-jit-bpe-500-2022-07-21> in huggingface.
+
+The decoding configuration for the reworked model is --epoch 30, --avg 8, --use-averaged-model True, which is the best after searching.
+
+| WER | reworked ctc attention | with --epoch 30 --avg 8 --use-averaged-model True | | ctc attention| with --epoch 77 --avg 55 | |
+|------------------------|-------|------|------|------|------|-----|
+| test sets | test-clean | test-other | Avg | test-clean | test-other | Avg |
+| ctc-greedy-search      | 2.98% | 7.14%| 5.06%| 2.90%| 7.47%| 5.19%|
+| ctc-decoding           | 2.98% | 7.14%| 5.06%| 2.90%| 7.47%| 5.19%|
+| 1best                  | 2.93% | 6.37%| 4.65%| 2.70%| 6.49%| 4.60%|
+| nbest                  | 2.94% | 6.39%| 4.67%| 2.70%| 6.48%| 4.59%|
+| nbest-rescoring        | 2.68% | 5.77%| 4.23%| 2.55%| 6.07%| 4.31%|
+| whole-lattice-rescoring| 2.66% | 5.76%| 4.21%| 2.56%| 6.04%| 4.30%|
+| attention-decoder      | 2.59% | 5.54%| 4.07%| 2.41%| 5.77%| 4.09%|
+| nbest-oracle           | 1.53% | 3.47%| 2.50%| 1.69%| 4.02%| 2.86%|
+| rnn-lm                 | 2.37% | 4.98%| 3.68%| 2.31%| 5.35%| 3.83%|
+
+
+
+conformer_ctc2 also implements the CTC greedy search decoding, it has the identical WERs with the CTC-decoding method.
+For other decoding methods, the average WER of the two test sets with the two models is similar.
+Except for the 1best and nbest methods, the overall performance of reworked model is better than the baseline model.
+
+
+To reproduce the above result, use the following commands.
+
+The training commands are:
+
+```bash
+    WORLD_SIZE=8
+    export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+    ./conformer_ctc2/train.py \
+    --manifest-dir data/fbank \
+    --exp-dir conformer_ctc2/exp \
+    --full-libri 1 \
+    --spec-aug-time-warp-factor 80 \
+    --max-duration 300 \
+    --world-size ${WORLD_SIZE} \
+    --start-epoch 1 \
+    --num-epochs 30 \
+    --att-rate 0.7 \
+    --num-decoder-layers 6
+```
+
+
+And the following commands are for decoding:
+
+```bash
+
+
+for method in ctc-greedy-search ctc-decoding 1best nbest-oracle; do
+  python3 ./conformer_ctc2/decode.py \
+  --exp-dir conformer_ctc2/exp \
+  --use-averaged-model True --epoch 30 --avg 8 --max-duration 200 --method $method
+done
+
+for method in nbest nbest-rescoring whole-lattice-rescoring attention-decoder ; do
+  python3 ./conformer_ctc2/decode.py \
+  --exp-dir conformer_ctc2/exp \
+  --use-averaged-model True --epoch 30 --avg 8 --max-duration 20 --method $method
+done
+
+rnn_dir=$(git rev-parse --show-toplevel)/icefall/rnn_lm
+./conformer_ctc2/decode.py \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --lm-dir data/lm \
+  --max-duration 30 \
+  --concatenate-cuts 0 \
+  --bucketing-sampler 1 \
+  --num-paths 1000 \
+  --use-averaged-model True \
+  --epoch 30 \
+  --avg 8 \
+  --nbest-scale 0.5 \
+  --rnn-lm-exp-dir ${rnn_dir}/exp \
+  --rnn-lm-epoch 29 \
+  --rnn-lm-avg 3 \
+  --rnn-lm-embedding-dim 2048 \
+  --rnn-lm-hidden-dim 2048 \
+  --rnn-lm-num-layers 3 \
+  --rnn-lm-tie-weights true \
+  --method rnn-lm
+```
+
+You can find the RNN-LM pre-trained model at
+<https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>
+
+
+### LibriSpeech BPE training results (Conformer-CTC)
+
+#### 2021-11-09
+
+The best WER, as of 2022-06-20, for the librispeech test dataset is below
+(using HLG decoding + n-gram LM rescoring + attention decoder rescoring + rnn lm rescoring):
+
+|     | test-clean | test-other |
+|-----|------------|------------|
+| WER | 2.32       | 5.39       |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+
+| ngram_lm_scale | attention_scale | rnn_lm_scale |
+|----------------|-----------------|--------------|
+| 0.3            | 2.1             | 2.2          |
+
+
+To reproduce the above result, use the following commands for training:
+
+```
+cd egs/librispeech/ASR/conformer_ctc
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./conformer_ctc/train.py \
+  --exp-dir conformer_ctc/exp_500_att0.8 \
+  --lang-dir data/lang_bpe_500 \
+  --att-rate 0.8 \
+  --full-libri 1 \
+  --max-duration 200 \
+  --concatenate-cuts 0 \
+  --world-size 4 \
+  --bucketing-sampler 1 \
+  --start-epoch 0 \
+  --num-epochs 90
+# Note: It trains for 90 epochs, but the best WER is at epoch-77.pt
+
+# Train the RNN-LM
+cd icefall
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./rnn_lm/train.py \
+  --exp-dir rnn_lm/exp_2048_3_tied \
+  --start-epoch 0 \
+  --world-size 4 \
+  --num-epochs 30 \
+  --use-fp16 1 \
+  --embedding-dim 2048 \
+  --hidden-dim 2048 \
+  --num-layers 3 \
+  --batch-size 500 \
+  --tie-weights true
+```
+
+and the following command for decoding
+
+```
+rnn_dir=$(git rev-parse --show-toplevel)/icefall/rnn_lm
+./conformer_ctc/decode.py \
+  --exp-dir conformer_ctc/exp_500_att0.8 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 30 \
+  --concatenate-cuts 0 \
+  --bucketing-sampler 1 \
+  --num-paths 1000 \
+  --epoch 77 \
+  --avg 55 \
+  --nbest-scale 0.5 \
+  --rnn-lm-exp-dir ${rnn_dir}/exp_2048_3_tied \
+  --rnn-lm-epoch 29 \
+  --rnn-lm-avg 3 \
+  --rnn-lm-embedding-dim 2048 \
+  --rnn-lm-hidden-dim 2048 \
+  --rnn-lm-num-layers 3 \
+  --rnn-lm-tie-weights true \
+  --method rnn-lm
+```
+
+You can find the Conformer-CTC pre-trained model by visiting
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09>
+
+and the RNN-LM pre-trained model:
+<https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>
+
+The tensorboard log for training is available at
+<https://tensorboard.dev/experiment/hZDWrZfaSqOMqtW0NEfXKg/#scalars>
+
+
+#### 2021-08-19
+(Wei Kang): Result of https://github.com/k2-fsa/icefall/pull/13
+
+TensorBoard log is available at https://tensorboard.dev/experiment/GnRzq8WWQW62dK4bklXBTg/#scalars
+
+Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_librispeech_conformer_ctc
+
+The best decoding results (WER) are listed below, we got this results by averaging models from epoch 15 to 34, and using `attention-decoder` decoder with num_paths equals to 100.
+
+||test-clean|test-other|
+|--|--|--|
+|WER| 2.57% | 5.94% |
+
+To get more unique paths, we scaled the lattice.scores with 0.5 (see https://github.com/k2-fsa/icefall/pull/10#discussion_r690951662 for more details), we searched the lm_score_scale and attention_score_scale for best results, the scales that produced the WER above are also listed below.
+
+||lm_scale|attention_scale|
+|--|--|--|
+|test-clean|1.3|1.2|
+|test-other|1.2|1.1|
+
+You can use the following commands to reproduce our results:
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+
+# It was using ef233486, you may not need to switch to it
+# git checkout ef233486
+
+cd egs/librispeech/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+python conformer_ctc/train.py --bucketing-sampler True \
+                              --concatenate-cuts False \
+                              --max-duration 200 \
+                              --full-libri True \
+                              --world-size 4 \
+                              --lang-dir data/lang_bpe_5000
+
+python conformer_ctc/decode.py --nbest-scale 0.5 \
+                               --epoch 34 \
+                               --avg 20 \
+                               --method attention-decoder \
+                               --max-duration 20 \
+                               --num-paths 100 \
+                               --lang-dir data/lang_bpe_5000
+```
+
+### LibriSpeech training results (Tdnn-Lstm)
+#### 2021-08-24
+
+(Wei Kang): Result of phone based Tdnn-Lstm model.
+
+Icefall version: https://github.com/k2-fsa/icefall/commit/caa0b9e9425af27e0c6211048acb55a76ed5d315
+
+Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_librispeech_tdnn-lstm_ctc
+
+The best decoding results (WER) are listed below, we got this results by averaging models from epoch 19 to 14, and using `whole-lattice-rescoring` decoding method.
+
+||test-clean|test-other|
+|--|--|--|
+|WER| 6.59% | 17.69% |
+
+We searched the lm_score_scale for best results, the scales that produced the WER above are also listed below.
+
+||lm_scale|
+|--|--|
+|test-clean|0.8|
+|test-other|0.9|
diff --git a/egs/librispeech/ASR/add_alignments.sh b/egs/librispeech/ASR/add_alignments.sh
new file mode 100755
index 000000000..6c47d25a2
--- /dev/null
+++ b/egs/librispeech/ASR/add_alignments.sh
@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+# align could be in ("mfa", "torchaudio")
+# We recommend "torchaudio"
+align="torchaudio"
+
+# It adds alignments to the existing fbank features dir (e.g., data/fbank)
+# and save cuts to a new dir (e.g., data/fbank_ali).
+cuts_in_dir=data/fbank
+cuts_out_dir=data/fbank_ali
+
+if [ $align == "mfa" ]; then
+  # It add alignments from https://github.com/CorentinJ/librispeech-alignments,
+  # generated using the Montreal Forced Aligner (https://montreal-forced-aligner.readthedocs.io).
+  alignments_dir=data/alignment
+
+  python3 ./local/add_alignment_librispeech.py \
+    --alignments-dir $alignments_dir \
+    --cuts-in-dir $cuts_in_dir \
+    --cuts-out-dir $cuts_out_dir
+elif [ $align == "torchaudio" ]; then
+  # See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/bin/modes/workflows.py for details.
+  #
+  # It use a pretrained ASR model from torchaudio to generate alignments.
+  # It will attach word-level alignment information (start, end, and score) to the
+  # supervisions in each cut.
+  mkdir -p $cuts_out_dir
+
+  parts=(
+    train-clean-100
+    train-clean-360
+    train-other-500
+    test-clean
+    test-other
+    dev-clean
+    dev-other
+  )
+
+  echo "The alignments will be saved to $cuts_out_dir"
+  for part in ${parts[@]}; do
+    echo "Start to align $part"
+    lhotse workflows align-with-torchaudio --dont-normalize-text \
+      $cuts_in_dir/librispeech_cuts_${part}.jsonl.gz \
+      $cuts_out_dir/librispeech_cuts_${part}.jsonl.gz
+  done
+  echo "Finished"
+else
+  echo "align is expected to be in ('mfa', 'torchaudio'), but got $align"
+  exit 1
+fi
diff --git a/egs/librispeech/ASR/conformer_ctc/README.md b/egs/librispeech/ASR/conformer_ctc/README.md
new file mode 100644
index 000000000..37ace4204
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/README.md
@@ -0,0 +1,75 @@
+## Introduction
+
+Please visit
+<https://icefall.readthedocs.io/en/latest/recipes/librispeech/conformer_ctc.html>
+for how to run this recipe.
+
+## How to compute framewise alignment information
+
+### Step 1: Train a model
+
+Please use `conformer_ctc/train.py` to train a model.
+See <https://icefall.readthedocs.io/en/latest/recipes/librispeech/conformer_ctc.html>
+for how to do it.
+
+### Step 2: Compute framewise alignment
+
+Run
+
+```
+# Choose a checkpoint and determine the number of checkpoints to average
+epoch=30
+avg=15
+./conformer_ctc/ali.py \
+  --epoch $epoch \
+  --avg $avg \
+  --max-duration 500 \
+  --bucketing-sampler 0 \
+  --full-libri 1 \
+  --exp-dir conformer_ctc/exp \
+  --lang-dir data/lang_bpe_500 \
+  --ali-dir data/ali_500
+```
+and  you will get four files inside the folder `data/ali_500`:
+
+```
+$ ls -lh data/ali_500
+total 546M
+-rw-r--r-- 1 kuangfangjun root 1.1M Sep 28 08:06 test_clean.pt
+-rw-r--r-- 1 kuangfangjun root 1.1M Sep 28 08:07 test_other.pt
+-rw-r--r-- 1 kuangfangjun root 542M Sep 28 11:36 train-960.pt
+-rw-r--r-- 1 kuangfangjun root 2.1M Sep 28 11:38 valid.pt
+```
+
+**Note**: It can take more than 3 hours to compute the alignment
+for the training dataset, which contains 960 * 3 = 2880 hours of data.
+
+**Caution**: The model parameters in `conformer_ctc/ali.py` have to match those
+in `conformer_ctc/train.py`.
+
+**Caution**: You have to set the parameter `preserve_id` to `True` for `CutMix`.
+Search `./conformer_ctc/asr_datamodule.py` for `preserve_id`.
+
+### Step 3: Check your extracted alignments
+
+There is a file `test_ali.py` in `icefall/test` that can be used to test your
+alignments. It uses pre-computed alignments to modify a randomly generated
+`nnet_output` and it checks that we can decode the correct transcripts
+from the resulting `nnet_output`.
+
+You should get something like the following if you run that script:
+
+```
+$ ./test/test_ali.py
+['THE GOOD NATURED AUDIENCE IN PITY TO FALLEN MAJESTY SHOWED FOR ONCE GREATER DEFERENCE TO THE KING THAN TO THE MINISTER AND SUNG THE PSALM WHICH THE FORMER HAD CALLED FOR', 'THE OLD SERVANT TOLD HIM QUIETLY AS THEY CREPT BACK TO DWELL THAT THIS PASSAGE THAT LED FROM THE HUT IN THE PLEASANCE TO SHERWOOD AND THAT GEOFFREY FOR THE TIME WAS HIDING WITH THE OUTLAWS IN THE FOREST', 'FOR A WHILE SHE LAY IN HER CHAIR IN HAPPY DREAMY PLEASURE AT SUN AND BIRD AND TREE', "BUT THE ESSENCE OF LUTHER'S LECTURES IS THERE"]
+['THE GOOD NATURED AUDIENCE IN PITY TO FALLEN MAJESTY SHOWED FOR ONCE GREATER DEFERENCE TO THE KING THAN TO THE MINISTER AND SUNG THE PSALM WHICH THE FORMER HAD CALLED FOR', 'THE OLD SERVANT TOLD HIM QUIETLY AS THEY CREPT BACK TO GAMEWELL THAT THIS PASSAGE WAY LED FROM THE HUT IN THE PLEASANCE TO SHERWOOD AND THAT GEOFFREY FOR THE TIME WAS HIDING WITH THE OUTLAWS IN THE FOREST', 'FOR A WHILE SHE LAY IN HER CHAIR IN HAPPY DREAMY PLEASURE AT SUN AND BIRD AND TREE', "BUT THE ESSENCE OF LUTHER'S LECTURES IS THERE"]
+```
+
+### Step 4: Use your alignments in training
+
+Please refer to `conformer_mmi/train.py` for usage. Some useful
+functions are:
+
+- `load_alignments()`, it loads alignment saved by `conformer_ctc/ali.py`
+- `convert_alignments_to_tensor()`, it converts alignments to PyTorch tensors
+- `lookup_alignments()`, it returns the alignments of utterances by giving the cut ID of the utterances.
diff --git a/egs/librispeech/ASR/conformer_ctc/__init__.py b/egs/librispeech/ASR/conformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/conformer_ctc/ali.py b/egs/librispeech/ASR/conformer_ctc/ali.py
new file mode 100755
index 000000000..42e14abac
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/ali.py
@@ -0,0 +1,395 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+    ./conformer_ctc/ali.py \
+            --exp-dir ./conformer_ctc/exp \
+            --lang-dir ./data/lang_bpe_500 \
+            --epoch 20 \
+            --avg 10 \
+            --max-duration 300 \
+            --dataset train-clean-100 \
+            --out-dir data/ali
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import numpy as np
+import torch
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse import CutSet
+from lhotse.features.io import FeaturesWriter, NumpyHdf5Writer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import one_best_decoding
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    encode_supervisions,
+    get_alignments,
+    setup_logger,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--out-dir",
+        type=str,
+        required=True,
+        help="""Output directory.
+        It contains 3 generated files:
+
+        - labels_xxx.h5
+        - aux_labels_xxx.h5
+        - librispeech_cuts_xxx.jsonl.gz
+
+        where xxx is the value of `--dataset`. For instance, if
+        `--dataset` is `train-clean-100`, it will contain 3 files:
+
+        - `labels_train-clean-100.h5`
+        - `aux_labels_train-clean-100.h5`
+        - `librispeech_cuts_train-clean-100.jsonl.gz`
+
+        Note: Both labels_xxx.h5 and aux_labels_xxx.h5 contain framewise
+        alignment. The difference is that labels_xxx.h5 contains repeats.
+        """,
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="""The name of the dataset to compute alignments for.
+        Possible values are:
+            - test-clean.
+            - test-other
+            - train-clean-100
+            - train-clean-360
+            - train-other-500
+            - dev-clean
+            - dev-other
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "lm_dir": Path("data/lm"),
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "subsampling_factor": 4,
+            # Set it to 0 since attention decoder
+            # is not used for computing alignments
+            "num_decoder_layers": 0,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "output_beam": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def compute_alignments(
+    model: torch.nn.Module,
+    dl: torch.utils.data.DataLoader,
+    labels_writer: FeaturesWriter,
+    aux_labels_writer: FeaturesWriter,
+    params: AttributeDict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+) -> CutSet:
+    """Compute the framewise alignments of a dataset.
+
+    Args:
+      model:
+        The neural network model.
+      dl:
+        Dataloader containing the dataset.
+      params:
+        Parameters for computing alignments.
+      graph_compiler:
+        It converts token IDs to decoding graphs.
+    Returns:
+      Return a CutSet. Each cut has two custom fields: labels_alignment
+      and aux_labels_alignment, containing framewise alignments information.
+      Both are of type `lhotse.array.TemporalArray`. The difference between
+      the two alignments is that `labels_alignment` contain repeats.
+    """
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+    num_cuts = 0
+
+    device = graph_compiler.device
+    cuts = []
+    for batch_idx, batch in enumerate(dl):
+        feature = batch["inputs"]
+
+        # at entry, feature is [N, T, C]
+        assert feature.ndim == 3
+        feature = feature.to(device)
+
+        supervisions = batch["supervisions"]
+        cut_list = supervisions["cut"]
+
+        for cut in cut_list:
+            assert len(cut.supervisions) == 1, f"{len(cut.supervisions)}"
+
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is [N, T, C]
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+        # we need also to sort cut_ids as encode_supervisions()
+        # reorders "texts".
+        # In general, new2old is an identity map since lhotse sorts the returned
+        # cuts by duration in descending order
+        new2old = supervision_segments[:, 0].tolist()
+
+        cut_list = [cut_list[i] for i in new2old]
+
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            nnet_output,
+            supervision_segments,
+            allow_truncate=params.subsampling_factor - 1,
+        )
+
+        lattice = k2.intersect_dense(
+            decoding_graph,
+            dense_fsa_vec,
+            params.output_beam,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice,
+            use_double_scores=params.use_double_scores,
+        )
+
+        labels_ali = get_alignments(best_path, kind="labels")
+        aux_labels_ali = get_alignments(best_path, kind="aux_labels")
+        assert len(labels_ali) == len(aux_labels_ali) == len(cut_list)
+        for cut, labels, aux_labels in zip(cut_list, labels_ali, aux_labels_ali):
+            cut.labels_alignment = labels_writer.store_array(
+                key=cut.id,
+                value=np.asarray(labels, dtype=np.int32),
+                # frame shift is 0.01s, subsampling_factor is 4
+                frame_shift=0.04,
+                temporal_dim=0,
+                start=0,
+            )
+            cut.aux_labels_alignment = aux_labels_writer.store_array(
+                key=cut.id,
+                value=np.asarray(aux_labels, dtype=np.int32),
+                # frame shift is 0.01s, subsampling_factor is 4
+                frame_shift=0.04,
+                temporal_dim=0,
+                start=0,
+            )
+
+        cuts += cut_list
+
+        num_cuts += len(cut_list)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return CutSet.from_cuts(cuts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    args.enable_spec_aug = False
+    args.enable_musan = False
+    args.return_cuts = True
+    args.concatenate_cuts = False
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-ali")
+
+    logging.info(f"Computing alignments for {params.dataset} - started")
+    logging.info(params)
+
+    out_dir = Path(params.out_dir)
+    out_dir.mkdir(exist_ok=True)
+
+    out_labels_ali_filename = out_dir / f"labels_{params.dataset}.h5"
+    out_aux_labels_ali_filename = out_dir / f"aux_labels_{params.dataset}.h5"
+    out_manifest_filename = out_dir / f"librispeech_cuts_{params.dataset}.jsonl.gz"
+
+    for f in (
+        out_labels_ali_filename,
+        out_aux_labels_ali_filename,
+        out_manifest_filename,
+    ):
+        if f.exists():
+            logging.info(f"{f} exists - skipping")
+            return
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(
+            f"{params.exp_dir}/epoch-{params.epoch}.pt", model, strict=False
+        )
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.eval()
+
+    librispeech = LibriSpeechAsrDataModule(args)
+    if params.dataset == "test-clean":
+        test_clean_cuts = librispeech.test_clean_cuts()
+        dl = librispeech.test_dataloaders(test_clean_cuts)
+    elif params.dataset == "test-other":
+        test_other_cuts = librispeech.test_other_cuts()
+        dl = librispeech.test_dataloaders(test_other_cuts)
+    elif params.dataset == "train-clean-100":
+        train_clean_100_cuts = librispeech.train_clean_100_cuts()
+        dl = librispeech.train_dataloaders(train_clean_100_cuts)
+    elif params.dataset == "train-clean-360":
+        train_clean_360_cuts = librispeech.train_clean_360_cuts()
+        dl = librispeech.train_dataloaders(train_clean_360_cuts)
+    elif params.dataset == "train-other-500":
+        train_other_500_cuts = librispeech.train_other_500_cuts()
+        dl = librispeech.train_dataloaders(train_other_500_cuts)
+    elif params.dataset == "dev-clean":
+        dev_clean_cuts = librispeech.dev_clean_cuts()
+        dl = librispeech.valid_dataloaders(dev_clean_cuts)
+    else:
+        assert params.dataset == "dev-other", f"{params.dataset}"
+        dev_other_cuts = librispeech.dev_other_cuts()
+        dl = librispeech.valid_dataloaders(dev_other_cuts)
+
+    logging.info(f"Processing {params.dataset}")
+    with NumpyHdf5Writer(out_labels_ali_filename) as labels_writer:
+        with NumpyHdf5Writer(out_aux_labels_ali_filename) as aux_labels_writer:
+            cut_set = compute_alignments(
+                model=model,
+                dl=dl,
+                labels_writer=labels_writer,
+                aux_labels_writer=aux_labels_writer,
+                params=params,
+                graph_compiler=graph_compiler,
+            )
+
+    cut_set.to_file(out_manifest_filename)
+
+    logging.info(
+        f"For dataset {params.dataset}, its alignments with repeats are "
+        f"saved to {out_labels_ali_filename}, the alignments without repeats "
+        f"are saved to {out_aux_labels_ali_filename}, and the cut manifest "
+        f"file is {out_manifest_filename}. Number of cuts: {len(cut_set)}"
+    )
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/asr_datamodule.py b/egs/librispeech/ASR/conformer_ctc/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc/conformer.py b/egs/librispeech/ASR/conformer_ctc/conformer.py
new file mode 100644
index 000000000..a1cfe6e75
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/conformer.py
@@ -0,0 +1,910 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: Union[float, bool] = 0.1,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        use_conv_batchnorm = True
+        if isinstance(use_feat_batchnorm, float):
+            use_conv_batchnorm = False
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+            use_conv_batchnorm,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self, x: Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        use_conv_batchnorm: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(
+            d_model, cnn_module_kernel, use_batchnorm=use_conv_batchnorm
+        )
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        use_batchnorm: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+        self.use_batchnorm = use_batchnorm
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        if self.use_batchnorm:
+            self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        if self.use_batchnorm:
+            x = self.norm(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/librispeech/ASR/conformer_ctc/decode.py b/egs/librispeech/ASR/conformer_ctc/decode.py
new file mode 100755
index 000000000..7e0bf5b7b
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/decode.py
@@ -0,0 +1,815 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=77,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[int], List[int]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        model = load_averaged_model(
+            params.exp_dir, model, params.epoch, params.avg, device
+        )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        if params.rnn_lm_avg == 1:
+            load_checkpoint(
+                f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                rnn_lm_model,
+            )
+            rnn_lm_model.to(device)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/export.py b/egs/librispeech/ASR/conformer_ctc/export.py
new file mode 100755
index 000000000..f0bb97560
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/export.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from conformer import Conformer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            "num_decoder_layers": 6,
+        }
+    )
+    return params
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    num_classes = num_tokens(token_table) + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
new file mode 100755
index 000000000..4bdec9e11
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with H
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+(1) LibriSpeech conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_H.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --H ./data/lang_bpe_500/H.fst \
+      --tokens ./data/lang_bpe_500/tokens.txt \
+      ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
+      ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
+
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_H.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --H ./data/lang_char/H.fst \
+      --tokens ./data/lang_char/tokens.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./conformer_ctc/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument("--H", type=str, required=True, help="Path to H.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_tokens(tokens_txt: str) -> Dict[int, str]:
+    id2token = dict()
+    with open(tokens_txt, encoding="utf-8") as f:
+        for line in f:
+            token, idx = line.strip().split()
+            id2token[int(idx)] = token
+
+    return id2token
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    H: kaldifst,
+    id2token: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      nnet_output:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      H:
+        The H graph.
+      id2token:
+        A map mapping token ID to token string.
+    Returns:
+      Return a list of decoded tokens.
+    """
+    logging.info(f"{filename}, {nnet_output.shape}")
+    decodable = DecodableCtc(nnet_output.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(H, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # tokens are incremented during graph construction
+    # so they need to be decremented
+    hyps = [id2token[i - 1] for i in osymbols_out]
+    #  hyps = "".join(hyps).split("▁")
+    hyps = "".join(hyps).split("\u2581")  # unicode codepoint of ▁
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading H from {args.H}")
+    H = kaldifst.StdVectorFst.read(args.H)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.shape[0] for f in features]
+    feature_lengths = torch.tensor(feature_lengths)
+
+    supervisions = dict()
+    supervisions["sequence_idx"] = torch.arange(len(features))
+    supervisions["start_frame"] = torch.zeros(len(features))
+    supervisions["num_frames"] = feature_lengths
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output, _, _ = model(features, supervisions)
+    feature_lengths = ((feature_lengths - 1) // 2 - 1) // 2
+
+    id2token = read_tokens(args.tokens)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            nnet_output=nnet_output[i, : feature_lengths[i]],
+            H=H,
+            id2token=id2token,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
new file mode 100755
index 000000000..d5a1dba3c
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with HL
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+(1) LibriSpeech conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HL.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HL ./data/lang_bpe_500/HL.fst \
+      --words ./data/lang_bpe_500/words.txt \
+      ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
+      ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HL.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HL ./data/lang_char/HL.fst \
+      --words ./data/lang_char/words.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+
+Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./conformer_ctc/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HL", type=str, required=True, help="Path to HL.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_words(words_txt: str) -> Dict[int, str]:
+    id2word = dict()
+    with open(words_txt, encoding="utf-8") as f:
+        for line in f:
+            word, idx = line.strip().split()
+            id2word[int(idx)] = word
+
+    return id2word
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    HL: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      nnet_output:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HL:
+        The HL graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {nnet_output.shape}")
+    decodable = DecodableCtc(nnet_output.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HL, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HL from {args.HL}")
+    HL = kaldifst.StdVectorFst.read(args.HL)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.shape[0] for f in features]
+    feature_lengths = torch.tensor(feature_lengths)
+
+    supervisions = dict()
+    supervisions["sequence_idx"] = torch.arange(len(features))
+    supervisions["start_frame"] = torch.zeros(len(features))
+    supervisions["num_frames"] = feature_lengths
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output, _, _ = model(features, supervisions)
+    feature_lengths = ((feature_lengths - 1) // 2 - 1) // 2
+
+    id2word = read_words(args.words)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            nnet_output=nnet_output[i, : feature_lengths[i]],
+            HL=HL,
+            id2word=id2word,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
new file mode 100755
index 000000000..216677a23
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
@@ -0,0 +1,243 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with HLG
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+(1) LibriSpeech conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HLG ./data/lang_bpe_500/HLG.fst \
+      --words ./data/lang_bpe_500/words.txt \
+      ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
+      ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HLG ./data/lang_char/HLG.fst \
+      --words ./data/lang_char/words.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./conformer_ctc/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_words(words_txt: str) -> Dict[int, str]:
+    id2word = dict()
+    with open(words_txt, encoding="utf-8") as f:
+        for line in f:
+            word, idx = line.strip().split()
+            id2word[int(idx)] = word
+
+    return id2word
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    HLG: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      nnet_output:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HLG:
+        The HLG graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {nnet_output.shape}")
+    decodable = DecodableCtc(nnet_output.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HLG, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HLG from {args.HLG}")
+    HLG = kaldifst.StdVectorFst.read(args.HLG)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.shape[0] for f in features]
+    feature_lengths = torch.tensor(feature_lengths)
+
+    supervisions = dict()
+    supervisions["sequence_idx"] = torch.arange(len(features))
+    supervisions["start_frame"] = torch.zeros(len(features))
+    supervisions["num_frames"] = feature_lengths
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output, _, _ = model(features, supervisions)
+    feature_lengths = ((feature_lengths - 1) // 2 - 1) // 2
+
+    id2word = read_words(args.words)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            nnet_output=nnet_output[i, : feature_lengths[i]],
+            HLG=HLG,
+            id2word=id2word,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/label_smoothing.py b/egs/librispeech/ASR/conformer_ctc/label_smoothing.py
new file mode 100644
index 000000000..52d2eda3b
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/label_smoothing.py
@@ -0,0 +1,109 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+class LabelSmoothingLoss(torch.nn.Module):
+    """
+    Implement the LabelSmoothingLoss proposed in the following paper
+    https://arxiv.org/pdf/1512.00567.pdf
+    (Rethinking the Inception Architecture for Computer Vision)
+
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -1,
+        label_smoothing: float = 0.1,
+        reduction: str = "sum",
+    ) -> None:
+        """
+        Args:
+          ignore_index:
+            ignored class id
+          label_smoothing:
+            smoothing rate (0.0 means the conventional cross entropy loss)
+          reduction:
+            It has the same meaning as the reduction in
+            `torch.nn.CrossEntropyLoss`. It can be one of the following three
+            values: (1) "none": No reduction will be applied. (2) "mean": the
+            mean of the output is taken. (3) "sum": the output will be summed.
+        """
+        super().__init__()
+        assert 0.0 <= label_smoothing < 1.0, f"{label_smoothing}"
+        assert reduction in ("none", "sum", "mean"), reduction
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.reduction = reduction
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute loss between x and target.
+
+        Args:
+          x:
+            prediction of dimension
+            (batch_size, input_length, number_of_classes).
+          target:
+            target masked with self.ignore_index of
+            dimension (batch_size, input_length).
+
+        Returns:
+          A scalar tensor containing the loss without normalization.
+        """
+        assert x.ndim == 3
+        assert target.ndim == 2
+        assert x.shape[:2] == target.shape
+        num_classes = x.size(-1)
+        x = x.reshape(-1, num_classes)
+        # Now x is of shape (N*T, C)
+
+        # We don't want to change target in-place below,
+        # so we make a copy of it here
+        target = target.clone().reshape(-1)
+
+        ignored = target == self.ignore_index
+
+        # See https://github.com/k2-fsa/icefall/issues/240
+        # and https://github.com/k2-fsa/icefall/issues/297
+        # for why we don't use target[ignored] = 0 here
+        target = torch.where(ignored, torch.zeros_like(target), target)
+
+        true_dist = torch.nn.functional.one_hot(target, num_classes=num_classes).to(x)
+
+        true_dist = (
+            true_dist * (1 - self.label_smoothing) + self.label_smoothing / num_classes
+        )
+
+        # Set the value of ignored indexes to 0
+        #
+        # See https://github.com/k2-fsa/icefall/issues/240
+        # and https://github.com/k2-fsa/icefall/issues/297
+        # for why we don't use true_dist[ignored] = 0 here
+        true_dist = torch.where(
+            ignored.unsqueeze(1).repeat(1, true_dist.shape[1]),
+            torch.zeros_like(true_dist),
+            true_dist,
+        )
+
+        loss = -1 * (torch.log_softmax(x, dim=1) * true_dist)
+        if self.reduction == "sum":
+            return loss.sum()
+        elif self.reduction == "mean":
+            return loss.sum() / (~ignored).sum()
+        else:
+            return loss.sum(dim=-1)
diff --git a/egs/librispeech/ASR/conformer_ctc/pretrained.py b/egs/librispeech/ASR/conformer_ctc/pretrained.py
new file mode 100755
index 000000000..df3e4d819
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/pretrained.py
@@ -0,0 +1,434 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from conformer import Conformer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a tokens.txt file 
+            to convert tokens to actual words or characters. It needs 
+            neither a lexicon nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) attention-decoder - Extract n paths from the rescored
+            lattice and use the transformer attention decoder for
+            rescoring.
+            We call it HLG decoding + n-gram LM rescoring + attention
+            decoder rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or attention-decoder.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and attention-decoder.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--attention-decoder-scale",
+        type=float,
+        default=1.2,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for attention decoder scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sos-id",
+        type=int,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the SOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--eos-id",
+        type=int,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the EOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    if args.method != "attention-decoder":
+        # to save memory as the attention decoder
+        # will not be used
+        params.num_decoder_layers = 0
+
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    hyps = []
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    # Note: We don't use key padding mask for attention during decoding
+    with torch.no_grad():
+        nnet_output, memory, memory_key_padding_mask = model(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        max_token_id = params.num_classes - 1
+
+        # Load tokens.txt here
+        token_table = k2.SymbolTable.from_file(params.tokens)
+
+        def token_ids_to_words(token_ids: List[int]) -> str:
+            text = ""
+            for i in token_ids:
+                text += token_table[i]
+            return text.replace("▁", " ").strip()
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=params.num_classes > 500,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        hyp_tokens = get_texts(best_path)
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method in [
+        "1best",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = G.to(device)
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "attention-decoder":
+            logging.info("Use HLG + LM rescoring + attention decoder rescoring")
+            rescored_lattice = rescore_with_whole_lattice(
+                lattice=lattice, G_with_epsilon_loops=G, lm_scale_list=None
+            )
+            best_path_dict = rescore_with_attention_decoder(
+                lattice=rescored_lattice,
+                num_paths=params.num_paths,
+                model=model,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                sos_id=params.sos_id,
+                eos_id=params.eos_id,
+                nbest_scale=params.nbest_scale,
+                ngram_lm_scale=params.ngram_lm_scale,
+                attention_scale=params.attention_decoder_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyp_tokens = get_texts(best_path)
+        for hyp in hyp_tokens:
+            hyps.append(" ".join([word_sym_table[i] for i in hyp]))
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/subsampling.py b/egs/librispeech/ASR/conformer_ctc/subsampling.py
new file mode 100644
index 000000000..8e0f73d05
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/subsampling.py
@@ -0,0 +1,153 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x
diff --git a/egs/librispeech/ASR/conformer_ctc/test_label_smoothing.py b/egs/librispeech/ASR/conformer_ctc/test_label_smoothing.py
new file mode 100755
index 000000000..5d4438fd1
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/test_label_smoothing.py
@@ -0,0 +1,52 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from distutils.version import LooseVersion
+
+import torch
+from label_smoothing import LabelSmoothingLoss
+
+torch_ver = LooseVersion(torch.__version__)
+
+
+def test_with_torch_label_smoothing_loss():
+    if torch_ver < LooseVersion("1.10.0"):
+        print(f"Current torch version: {torch_ver}")
+        print("Please use torch >= 1.10 to run this test - skipping")
+        return
+    torch.manual_seed(20211105)
+    x = torch.rand(20, 30, 5000)
+    tgt = torch.randint(low=-1, high=x.size(-1), size=x.shape[:2])
+    for reduction in ["none", "sum", "mean"]:
+        custom_loss_func = LabelSmoothingLoss(
+            ignore_index=-1, label_smoothing=0.1, reduction=reduction
+        )
+        custom_loss = custom_loss_func(x, tgt)
+
+        torch_loss_func = torch.nn.CrossEntropyLoss(
+            ignore_index=-1, reduction=reduction, label_smoothing=0.1
+        )
+        torch_loss = torch_loss_func(x.reshape(-1, x.size(-1)), tgt.reshape(-1))
+        assert torch.allclose(custom_loss, torch_loss)
+
+
+def main():
+    test_with_torch_label_smoothing_loss()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/test_subsampling.py b/egs/librispeech/ASR/conformer_ctc/test_subsampling.py
new file mode 100755
index 000000000..81fa234dd
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/test_subsampling.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+
+def test_conv2d_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = Conv2dSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
+
+
+def test_vgg_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = VggSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
diff --git a/egs/librispeech/ASR/conformer_ctc/test_transformer.py b/egs/librispeech/ASR/conformer_ctc/test_transformer.py
new file mode 100644
index 000000000..667057c51
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/test_transformer.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from torch.nn.utils.rnn import pad_sequence
+from transformer import (
+    Transformer,
+    add_eos,
+    add_sos,
+    decoder_padding_mask,
+    encoder_padding_mask,
+    generate_square_subsequent_mask,
+)
+
+
+def test_encoder_padding_mask():
+    supervisions = {
+        "sequence_idx": torch.tensor([0, 1, 2]),
+        "start_frame": torch.tensor([0, 0, 0]),
+        "num_frames": torch.tensor([18, 7, 13]),
+    }
+
+    max_len = ((18 - 1) // 2 - 1) // 2
+    mask = encoder_padding_mask(max_len, supervisions)
+    expected_mask = torch.tensor(
+        [
+            [False, False, False],  # ((18 - 1)//2 - 1)//2 = 3,
+            [False, True, True],  # ((7 - 1)//2 - 1)//2 = 1,
+            [False, False, True],  # ((13 - 1)//2 - 1)//2 = 2,
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_transformer():
+    num_features = 40
+    num_classes = 87
+    model = Transformer(num_features=num_features, num_classes=num_classes)
+
+    N = 31
+
+    for T in range(7, 30):
+        x = torch.rand(N, T, num_features)
+        y, _, _ = model(x)
+        assert y.shape == (N, (((T - 1) // 2) - 1) // 2, num_classes)
+
+
+def test_generate_square_subsequent_mask():
+    s = 5
+    mask = generate_square_subsequent_mask(s)
+    inf = float("inf")
+    expected_mask = torch.tensor(
+        [
+            [0.0, -inf, -inf, -inf, -inf],
+            [0.0, 0.0, -inf, -inf, -inf],
+            [0.0, 0.0, 0.0, -inf, -inf],
+            [0.0, 0.0, 0.0, 0.0, -inf],
+            [0.0, 0.0, 0.0, 0.0, 0.0],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_decoder_padding_mask():
+    x = [torch.tensor([1, 2]), torch.tensor([3]), torch.tensor([2, 5, 8])]
+    y = pad_sequence(x, batch_first=True, padding_value=-1)
+    mask = decoder_padding_mask(y, ignore_id=-1)
+    expected_mask = torch.tensor(
+        [
+            [False, False, True],
+            [False, True, True],
+            [False, False, False],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_add_sos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_sos(x, sos_id=0)
+    expected_y = [[0, 1, 2], [0, 3], [0, 2, 5, 8]]
+    assert y == expected_y
+
+
+def test_add_eos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_eos(x, eos_id=0)
+    expected_y = [[1, 2, 0], [3, 0], [2, 5, 8, 0]]
+    assert y == expected_y
diff --git a/egs/librispeech/ASR/conformer_ctc/train.py b/egs/librispeech/ASR/conformer_ctc/train.py
new file mode 100755
index 000000000..828106f41
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/train.py
@@ -0,0 +1,818 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./conformer_ctc/train.py \
+     --exp-dir ./conformer_ctc/exp \
+     --world-size 4 \
+     --full-libri 1 \
+     --max-duration 200 \
+     --num-epochs 20
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=78,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Normalization for the input features, can be a
+                              boolean indicating whether to do batch
+                              normalization, or a float which means just scaling
+                              the input features with this float value.
+                              If given a float value, we will remove batchnorm
+                              layer in `ConvolutionModule` as well.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    if isinstance(graph_compiler, BpeCtcTrainingGraphCompiler):
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+    elif isinstance(graph_compiler, CtcTrainingGraphCompiler):
+        # Works with a phone lexicon
+        decoding_graph = graph_compiler.compile(texts)
+    else:
+        raise ValueError(f"Unsupported type of graph compiler: {type(graph_compiler)}")
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    if "lang_bpe" in str(params.lang_dir):
+        graph_compiler = BpeCtcTrainingGraphCompiler(
+            params.lang_dir,
+            device=device,
+            sos_token="<sos/eos>",
+            eos_token="<sos/eos>",
+        )
+    elif "lang_phone" in str(params.lang_dir):
+        assert params.att_rate == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. Set --att-rate=0 "
+            "for pure CTC training when using a phone-based lang dir."
+        )
+        assert params.num_decoder_layers == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. "
+            "Set --num-decoder-layers=0 for pure CTC training when using "
+            "a phone-based lang dir."
+        )
+        graph_compiler = CtcTrainingGraphCompiler(
+            lexicon,
+            device=device,
+        )
+        # Manually add the sos/eos ID with their default values
+        # from the BPE recipe which we're adapting here.
+        graph_compiler.sos_id = 1
+        graph_compiler.eos_id = 1
+    else:
+        raise ValueError(
+            f"Unsupported type of lang dir (we expected it to have "
+            f"'lang_bpe' or 'lang_phone' in its name): {params.lang_dir}"
+        )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/transformer.py b/egs/librispeech/ASR/conformer_ctc/transformer.py
new file mode 100644
index 000000000..0566cfc81
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/transformer.py
@@ -0,0 +1,928 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from label_smoothing import LabelSmoothingLoss
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: Union[float, bool] = 0.1,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+            Float value to scale the input layer.
+            False to do nothing.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        assert isinstance(use_feat_batchnorm, (float, bool))
+        if isinstance(use_feat_batchnorm, bool) and use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if isinstance(self.use_feat_batchnorm, bool) and self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        if isinstance(self.use_feat_batchnorm, float):
+            x *= self.use_feat_batchnorm
+        encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision)
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/librispeech/ASR/conformer_ctc2/__init__.py b/egs/librispeech/ASR/conformer_ctc2/__init__.py
new file mode 120000
index 000000000..b24e5e357
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/__init__.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc2/asr_datamodule.py b/egs/librispeech/ASR/conformer_ctc2/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc2/attention.py b/egs/librispeech/ASR/conformer_ctc2/attention.py
new file mode 100644
index 000000000..356d3f21b
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/attention.py
@@ -0,0 +1,243 @@
+# Copyright    2022  Xiaomi Corp.        (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+from torch import Tensor
+from torch.nn.init import xavier_normal_
+
+
+class MultiheadAttention(nn.Module):
+    r"""Allows the model to jointly attend to information
+    from different representation subspaces.
+    See `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_.
+
+    .. math::
+        \text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
+
+    where :math:`head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)`.
+
+    Args:
+        embed_dim: Total dimension of the model.
+        num_heads: Number of parallel attention heads. Note that ``embed_dim`` will be split
+            across ``num_heads`` (i.e. each head will have dimension ``embed_dim // num_heads``).
+        dropout: Dropout probability on ``attn_output_weights``. Default: ``0.0`` (no dropout).
+        bias: If specified, adds bias to input / output projection layers. Default: ``True``.
+        add_bias_kv: If specified, adds bias to the key and value sequences at dim=0. Default: ``False``.
+        add_zero_attn: If specified, adds a new batch of zeros to the key and value sequences at dim=1.
+            Default: ``False``.
+        kdim: Total number of features for keys. Default: ``None`` (uses ``kdim=embed_dim``).
+        vdim: Total number of features for values. Default: ``None`` (uses ``vdim=embed_dim``).
+        batch_first: If ``True``, then the input and output tensors are provided
+            as (batch, seq, feature). Default: ``False`` (seq, batch, feature).
+
+    Examples::
+
+        >>> multihead_attn = nn.MultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value)
+    """
+    __constants__ = ["batch_first"]
+    bias_k: Optional[torch.Tensor]
+    bias_v: Optional[torch.Tensor]
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        dropout=0.0,
+        bias=True,
+        add_bias_kv=False,
+        add_zero_attn=False,
+        kdim=None,
+        vdim=None,
+        batch_first=False,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(MultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.batch_first = batch_first
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        if self._qkv_same_embed_dim is False:
+            self.q_proj_weight = ScaledLinear(embed_dim, embed_dim, bias=bias)
+            self.k_proj_weight = ScaledLinear(self.kdim, embed_dim, bias=bias)
+            self.v_proj_weight = ScaledLinear(self.vdim, embed_dim, bias=bias)
+            self.register_parameter("in_proj_weight", None)
+        else:
+            self.in_proj_weight = ScaledLinear(embed_dim, 3 * embed_dim, bias=bias)
+            self.register_parameter("q_proj_weight", None)
+            self.register_parameter("k_proj_weight", None)
+            self.register_parameter("v_proj_weight", None)
+
+        if not bias:
+            self.register_parameter("in_proj_bias", None)
+
+        self.out_proj = ScaledLinear(embed_dim, embed_dim, bias=bias)
+
+        if add_bias_kv:
+            self.bias_k = nn.Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
+            self.bias_v = nn.Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
+        else:
+            self.bias_k = self.bias_v = None
+
+        self.add_zero_attn = add_zero_attn
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        if self.bias_k is not None:
+            xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            xavier_normal_(self.bias_v)
+
+    def __setstate__(self, state):
+        # Support loading old MultiheadAttention checkpoints generated by v1.1.0
+        if "_qkv_same_embed_dim" not in state:
+            state["_qkv_same_embed_dim"] = True
+
+        super(MultiheadAttention, self).__setstate__(state)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query: Query embeddings of shape :math:`(L, N, E_q)` when ``batch_first=False`` or :math:`(N, L, E_q)`
+                when ``batch_first=True``, where :math:`L` is the target sequence length, :math:`N` is the batch size,
+                and :math:`E_q` is the query embedding dimension ``embed_dim``. Queries are compared against
+                key-value pairs to produce the output. See "Attention Is All You Need" for more details.
+            key: Key embeddings of shape :math:`(S, N, E_k)` when ``batch_first=False`` or :math:`(N, S, E_k)` when
+                ``batch_first=True``, where :math:`S` is the source sequence length, :math:`N` is the batch size, and
+                :math:`E_k` is the key embedding dimension ``kdim``. See "Attention Is All You Need" for more details.
+            value: Value embeddings of shape :math:`(S, N, E_v)` when ``batch_first=False`` or :math:`(N, S, E_v)` when
+                ``batch_first=True``, where :math:`S` is the source sequence length, :math:`N` is the batch size, and
+                :math:`E_v` is the value embedding dimension ``vdim``. See "Attention Is All You Need" for more details.
+            key_padding_mask: If specified, a mask of shape :math:`(N, S)` indicating which elements within ``key``
+                to ignore for the purpose of attention (i.e. treat as "padding"). Binary and byte masks are supported.
+                For a binary mask, a ``True`` value indicates that the corresponding ``key`` value will be ignored for
+                the purpose of attention. For a byte mask, a non-zero value indicates that the corresponding ``key``
+                value will be ignored.
+            need_weights: If specified, returns ``attn_output_weights`` in addition to ``attn_outputs``.
+                Default: ``True``.
+            attn_mask: If specified, a 2D or 3D mask preventing attention to certain positions. Must be of shape
+                :math:`(L, S)` or :math:`(N\cdot\text{num\_heads}, L, S)`, where :math:`N` is the batch size,
+                :math:`L` is the target sequence length, and :math:`S` is the source sequence length. A 2D mask will be
+                broadcasted across the batch while a 3D mask allows for a different mask for each entry in the batch.
+                Binary, byte, and float masks are supported. For a binary mask, a ``True`` value indicates that the
+                corresponding position is not allowed to attend. For a byte mask, a non-zero value indicates that the
+                corresponding position is not allowed to attend. For a float mask, the mask values will be added to
+                the attention weight.
+
+        Outputs:
+            - **attn_output** - Attention outputs of shape :math:`(L, N, E)` when ``batch_first=False`` or
+              :math:`(N, L, E)` when ``batch_first=True``, where :math:`L` is the target sequence length, :math:`N` is
+              the batch size, and :math:`E` is the embedding dimension ``embed_dim``.
+            - **attn_output_weights** - Attention output weights of shape :math:`(N, L, S)`, where :math:`N` is the batch
+              size, :math:`L` is the target sequence length, and :math:`S` is the source sequence length. Only returned
+              when ``need_weights=True``.
+        """
+        if self.batch_first:
+            query, key, value = [x.transpose(1, 0) for x in (query, key, value)]
+
+        if not self._qkv_same_embed_dim:
+            q_proj_weight = (
+                self.q_proj_weight.get_weight()
+                if self.q_proj_weight is not None
+                else None
+            )
+            k_proj_weight = (
+                self.k_proj_weight.get_weight()
+                if self.k_proj_weight is not None
+                else None
+            )
+            v_proj_weight = (
+                self.v_proj_weight.get_weight()
+                if self.v_proj_weight is not None
+                else None
+            )
+            (
+                attn_output,
+                attn_output_weights,
+            ) = nn.functional.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                self.in_proj_weight.get_weight(),
+                self.in_proj_weight.get_bias(),
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout,
+                self.out_proj.get_weight(),
+                self.out_proj.get_bias(),
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                use_separate_proj_weight=True,
+                q_proj_weight=q_proj_weight,
+                k_proj_weight=k_proj_weight,
+                v_proj_weight=v_proj_weight,
+            )
+        else:
+            (
+                attn_output,
+                attn_output_weights,
+            ) = nn.functional.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                self.in_proj_weight.get_weight(),
+                self.in_proj_weight.get_bias(),
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout,
+                self.out_proj.get_weight(),
+                self.out_proj.get_bias(),
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+            )
+        if self.batch_first:
+            return attn_output.transpose(1, 0), attn_output_weights
+        else:
+            return attn_output, attn_output_weights
diff --git a/egs/librispeech/ASR/conformer_ctc2/conformer.py b/egs/librispeech/ASR/conformer_ctc2/conformer.py
new file mode 100644
index 000000000..09f1eb000
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/conformer.py
@@ -0,0 +1,945 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corp.                              (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        x = self.encoder(
+            x, pos_emb, src_key_padding_mask=mask, warmup=warmup
+        )  # (T, N, C)
+
+        # x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        # return x, lengths
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src = src + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/librispeech/ASR/conformer_ctc2/decode.py b/egs/librispeech/ASR/conformer_ctc2/decode.py
new file mode 100755
index 000000000..0b271a51c
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/decode.py
@@ -0,0 +1,991 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=77,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) ctc-greedy-search. It only use CTC output and a sentence piece
+              model for decoding. It produces the same results with ctc-decoding.
+            - (2) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (3) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (6) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (7) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (8) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "encoder_dim": 512,
+            "num_encoder_layers": 12,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def ctc_greedy_search(
+    nnet_output: torch.Tensor,
+    memory: torch.Tensor,
+    memory_key_padding_mask: torch.Tensor,
+) -> List[List[int]]:
+    """Apply CTC greedy search
+
+     Args:
+         speech (torch.Tensor): (batch, max_len, feat_dim)
+         speech_length (torch.Tensor): (batch, )
+    Returns:
+         List[List[int]]: best path result
+    """
+    batch_size = memory.shape[1]
+    # Let's assume B = batch_size
+    encoder_out = memory
+    encoder_mask = memory_key_padding_mask
+    maxlen = encoder_out.size(0)
+
+    ctc_probs = nnet_output  # (B, maxlen, vocab_size)
+    topk_prob, topk_index = ctc_probs.topk(1, dim=2)  # (B, maxlen, 1)
+    topk_index = topk_index.view(batch_size, maxlen)  # (B, maxlen)
+    topk_index = topk_index.masked_fill_(encoder_mask, 0)  # (B, maxlen)
+    hyps = [hyp.tolist() for hyp in topk_index]
+    scores = topk_prob.max(1)
+    hyps = [remove_duplicates_and_blank(hyp) for hyp in hyps]
+    return hyps, scores
+
+
+def remove_duplicates_and_blank(hyp: List[int]) -> List[int]:
+    # from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/common.py
+    new_hyp: List[int] = []
+    cur = 0
+    while cur < len(hyp):
+        if hyp[cur] != 0:
+            new_hyp.append(hyp[cur])
+        prev = cur
+        while cur < len(hyp) and hyp[cur] == hyp[prev]:
+            cur += 1
+    return new_hyp
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "ctc-greedy-search":
+        hyps, _ = ctc_greedy_search(
+            nnet_output,
+            memory,
+            memory_key_padding_mask,
+        )
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(hyps)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-greedy-search"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+        # TODO: pass `lattice` instead of `rescored_lattice` to
+        # `rescore_with_attention_decoder`
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding" or params.method == "ctc-greedy-search":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        if params.rnn_lm_avg == 1:
+            load_checkpoint(
+                f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                rnn_lm_model,
+            )
+            rnn_lm_model.to(device)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc2/export.py b/egs/librispeech/ASR/conformer_ctc2/export.py
new file mode 100755
index 000000000..26a95dbfa
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/export.py
@@ -0,0 +1,279 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./conformer_ctc2/export.py \
+  --exp-dir ./conformer_ctc2/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `conformer_ctc2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./conformer_ctc2/decode.py \
+        --exp-dir ./conformer_ctc2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from conformer import Conformer
+from decode import get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    num_classes = num_tokens(token_table) + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc2/label_smoothing.py b/egs/librispeech/ASR/conformer_ctc2/label_smoothing.py
new file mode 120000
index 000000000..08734abd7
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/label_smoothing.py
@@ -0,0 +1 @@
+../conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc2/optim.py b/egs/librispeech/ASR/conformer_ctc2/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc2/scaling.py b/egs/librispeech/ASR/conformer_ctc2/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc2/subsampling.py b/egs/librispeech/ASR/conformer_ctc2/subsampling.py
new file mode 100644
index 000000000..85a4dc8df
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/subsampling.py
@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corporation                        (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+
+class Conv2dSubsampling(torch.nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = torch.nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
diff --git a/egs/librispeech/ASR/conformer_ctc2/train.py b/egs/librispeech/ASR/conformer_ctc2/train.py
new file mode 100755
index 000000000..c4a13b101
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/train.py
@@ -0,0 +1,1081 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,
+#                                                  Zengwei Yao,
+#                                                  Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conformer_ctc2/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./conformer_ctc2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir conformer_ctc2/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for ctc loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(
+            feature, supervisions, warmup=warmup
+        )
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    if isinstance(graph_compiler, BpeCtcTrainingGraphCompiler):
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+    elif isinstance(graph_compiler, CtcTrainingGraphCompiler):
+        # Works with a phone lexicon
+        decoding_graph = graph_compiler.compile(texts)
+    else:
+        raise ValueError(f"Unsupported type of graph compiler: {type(graph_compiler)}")
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    scheduler: LRSchedulerType,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        # scaler.scale(loss).backward()
+
+        try:
+            # loss.backward()
+            scaler.scale(loss).backward()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(f"failing batch size:{batch_size} ")
+            raise
+
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+            if loss_info["ctc_loss"] == float("inf") or loss_info["att_loss"] == float(
+                "inf"
+            ):
+                logging.error("Your loss contains inf, something goes wrong")
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    if "lang_bpe" in str(params.lang_dir):
+        graph_compiler = BpeCtcTrainingGraphCompiler(
+            params.lang_dir,
+            device=device,
+            sos_token="<sos/eos>",
+            eos_token="<sos/eos>",
+        )
+    elif "lang_phone" in str(params.lang_dir):
+        assert params.att_rate == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. Set --att-rate=0 "
+            "for pure CTC training when using a phone-based lang dir."
+        )
+        assert params.num_decoder_layers == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. "
+            "Set --num-decoder-layers=0 for pure CTC training when using "
+            "a phone-based lang dir."
+        )
+        graph_compiler = CtcTrainingGraphCompiler(
+            lexicon,
+            device=device,
+        )
+        # Manually add the sos/eos ID with their default values
+        # from the BPE recipe which we're adapting here.
+        graph_compiler.sos_id = 1
+        graph_compiler.eos_id = 1
+    else:
+        raise ValueError(
+            f"Unsupported type of lang dir (we expected it to have "
+            f"'lang_bpe' or 'lang_phone' in its name): {params.lang_dir}"
+        )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    print(model)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    def remove_invalid_utt_ctc(c: Cut):
+        # Caution: We assume the subsampling factor is 4!
+        # num_tokens = len(sp.encode(c.supervisions[0].text, out_type=int))
+        num_tokens = len(graph_compiler.texts_to_ids(c.supervisions[0].text))
+        min_output_input_ratio = 0.0005
+        max_output_input_ratio = 0.1
+        return (
+            min_output_input_ratio
+            < num_tokens / float(c.features.num_frames)
+            < max_output_input_ratio
+        )
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.filter(remove_invalid_utt_ctc)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            scheduler=scheduler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    batch=batch,
+                    graph_compiler=graph_compiler,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc2/transformer.py b/egs/librispeech/ASR/conformer_ctc2/transformer.py
new file mode 100644
index 000000000..d3443dc94
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc2/transformer.py
@@ -0,0 +1,1049 @@
+# Copyright    2021  University of Chinese Academy of Sciences (author: Han Zhu)
+# Copyright    2022  Xiaomi Corp.                              (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from attention import MultiheadAttention
+from label_smoothing import LabelSmoothingLoss
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledEmbedding,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          layer_dropout (float): layer-dropout rate.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.encoder = TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), ScaledLinear(d_model, num_classes, bias=True)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = ScaledEmbedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+            )
+
+            self.decoder = TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+            )
+
+            self.decoder_output_layer = ScaledLinear(
+                d_model, self.decoder_num_class, bias=True
+            )
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        supervision: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+
+        encoder_memory, memory_key_padding_mask = self.run_encoder(
+            x, supervision, warmup
+        )
+
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask, warmup=warmup)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.self_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # src_att = self.self_attn(src, src, src, src_mask)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.layer_dropout = layer_dropout
+        self.self_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+          warmup: controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+
+
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        tgt_orig = tgt
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # tgt_att = self.self_attn(tgt, tgt, tgt, tgt_mask)
+        tgt_att = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(tgt_att)
+
+        # src_att = self.src_attn(tgt, memory, memory, memory_mask)
+        src_att = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(src_att)
+
+        tgt = tgt + self.dropout(self.feed_forward(tgt))
+
+        tgt = self.norm_final(self.balancer(tgt))
+
+        if alpha != 1.0:
+            tgt = alpha * tgt + (1 - alpha) * tgt_orig
+
+        return tgt
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class TransformerDecoder(nn.Module):
+    r"""TransformerDecoder is a stack of N decoder layers
+
+    Args:
+        decoder_layer: an instance of the TransformerDecoderLayer() class (required).
+        num_layers: the number of sub-decoder-layers in the decoder (required).
+
+    Examples::
+        >>> decoder_layer = TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> transformer_decoder = TransformerDecoder(decoder_layer, num_layers=6)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(10, 32, 512)
+        >>> out = transformer_decoder(tgt, memory)
+    """
+
+    def __init__(self, decoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(decoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        r"""Pass the input through the decoder layers in turn.
+
+        Args:
+            tgt: the sequence to the decoder (required).
+            memory: the sequence from the last layer of the encoder (required).
+            tgt_mask: the mask for the tgt sequence (optional).
+            memory_mask: the mask for the memory sequence (optional).
+            tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
+            memory_key_padding_mask: the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (S, N, E).
+            tgt_mask: (S, S).
+            tgt_key_padding_mask: (N, S).
+
+        """
+        output = tgt
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                memory,
+                tgt_mask=tgt_mask,
+                memory_mask=memory_mask,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                memory_key_padding_mask=memory_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/librispeech/ASR/conformer_ctc3/__init__.py b/egs/librispeech/ASR/conformer_ctc3/__init__.py
new file mode 120000
index 000000000..b24e5e357
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/__init__.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/asr_datamodule.py b/egs/librispeech/ASR/conformer_ctc3/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/conformer.py b/egs/librispeech/ASR/conformer_ctc3/conformer.py
new file mode 120000
index 000000000..3b84b9573
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/conformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/decode.py b/egs/librispeech/ASR/conformer_ctc3/decode.py
new file mode 100755
index 000000000..e6327bb5e
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/decode.py
@@ -0,0 +1,1049 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) decode in non-streaming mode (take ctc-decoding as an example)
+./conformer_ctc3/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./conformer_ctc3/exp \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) decode in streaming mode (take ctc-decoding as an example)
+./conformer_ctc3/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --decode-chunk-size 16 \
+    --left-context 64 \
+    --exp-dir ./conformer_ctc3/exp \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+To evaluate symbol delay, you should:
+(1) Generate cuts with word-time alignments:
+./add_alignments.sh
+(2) Set the argument "--manifest-dir data/fbank_ali" while decoding.
+For example:
+./conformer_ctc3/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./conformer_ctc3/exp \
+    --max-duration 600 \
+    --decoding-method ctc-decoding \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --decode-chunk-size 16 \
+    --left-context 64 \
+    --manifest-dir data/fbank_ali
+Note: It supports calculating symbol delay with following decoding methods:
+    - ctc-decoding
+    - 1best
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    convert_timestamp,
+    get_texts,
+    make_pad_mask,
+    parse_bpe_start_end_pairs,
+    parse_fsa_timestamps_and_texts,
+    setup_logger,
+    store_transcripts_and_timestamps,
+    str2bool,
+    write_error_stats_with_timestamps,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (0) ctc-greedy-search. It uses a sentence piece model,
+          i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.8,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def ctc_greedy_search(
+    ctc_probs: torch.Tensor,
+    nnet_output_lens: torch.Tensor,
+    sp: spm.SentencePieceProcessor,
+    subsampling_factor: int = 4,
+    frame_shift_ms: float = 10,
+) -> Tuple[List[Tuple[float, float]], List[List[str]]]:
+    """Apply CTC greedy search
+    Args:
+      ctc_probs (torch.Tensor):
+        (batch, max_len, feat_dim)
+      nnet_output_lens (torch.Tensor):
+        (batch, )
+      sp:
+        The BPE model.
+      subsampling_factor:
+        The subsampling factor of the model.
+      frame_shift_ms:
+        Frame shift in milliseconds between two contiguous frames.
+
+    Returns:
+      utt_time_pairs:
+        A list of pair list. utt_time_pairs[i] is a list of
+        (start-time, end-time) pairs for each word in
+        utterance-i.
+      utt_words:
+        A list of str list. utt_words[i] is a word list of utterence-i.
+    """
+    topk_prob, topk_index = ctc_probs.topk(1, dim=2)  # (B, maxlen, 1)
+    topk_index = topk_index.squeeze(2)  # (B, maxlen)
+    mask = make_pad_mask(nnet_output_lens)
+    topk_index = topk_index.masked_fill_(mask, 0)  # (B, maxlen)
+    hyps = [hyp.tolist() for hyp in topk_index]
+
+    def get_first_tokens(tokens: List[int]) -> List[bool]:
+        is_first_token = []
+        first_tokens = []
+        for t in range(len(tokens)):
+            if tokens[t] != 0 and (t == 0 or tokens[t - 1] != tokens[t]):
+                is_first_token.append(True)
+                first_tokens.append(tokens[t])
+            else:
+                is_first_token.append(False)
+        return first_tokens, is_first_token
+
+    utt_time_pairs = []
+    utt_words = []
+    for utt in range(len(hyps)):
+        first_tokens, is_first_token = get_first_tokens(hyps[utt])
+        all_tokens = sp.id_to_piece(hyps[utt])
+        index_pairs = parse_bpe_start_end_pairs(all_tokens, is_first_token)
+        words = sp.decode(first_tokens).split()
+        assert len(index_pairs) == len(words), (
+            len(index_pairs),
+            len(words),
+            all_tokens,
+        )
+        start = convert_timestamp(
+            frames=[i[0] for i in index_pairs],
+            subsampling_factor=subsampling_factor,
+            frame_shift_ms=frame_shift_ms,
+        )
+        end = convert_timestamp(
+            # The duration in frames is (end_frame_index - start_frame_index + 1)
+            frames=[i[1] + 1 for i in index_pairs],
+            subsampling_factor=subsampling_factor,
+            frame_shift_ms=frame_shift_ms,
+        )
+        utt_time_pairs.append(list(zip(start, end)))
+        utt_words.append(words)
+
+    return utt_time_pairs, utt_words
+
+
+def remove_duplicates_and_blank(hyp: List[int]) -> Tuple[List[int], List[int]]:
+    # modified from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/common.py
+    new_hyp: List[int] = []
+    time: List[Tuple[int, int]] = []
+    cur = 0
+    start, end = -1, -1
+    while cur < len(hyp):
+        if hyp[cur] != 0:
+            new_hyp.append(hyp[cur])
+            start = cur
+        prev = cur
+        while cur < len(hyp) and hyp[cur] == hyp[prev]:
+            if start != -1:
+                end = cur
+            cur += 1
+        if start != -1 and end != -1:
+            time.append((start, end))
+            start, end = -1, -1
+    return new_hyp, time
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, Tuple[List[List[str]], List[List[float]]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(feature, feature_lens)
+
+    nnet_output = model.get_ctc_output(encoder_out)
+    # nnet_output is (N, T, C)
+
+    if params.decoding_method == "ctc-greedy-search":
+        timestamps, hyps = ctc_greedy_search(
+            ctc_probs=nnet_output,
+            nnet_output_lens=encoder_out_lens,
+            sp=bpe_model,
+            subsampling_factor=params.subsampling_factor,
+            frame_shift_ms=params.frame_shift_ms,
+        )
+        key = "ctc-greedy-search"
+        return {key: (hyps, timestamps)}
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            encoder_out_lens.cpu(),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        timestamps, hyps = parse_fsa_timestamps_and_texts(
+            best_paths=best_path,
+            sp=bpe_model,
+            subsampling_factor=params.subsampling_factor,
+            frame_shift_ms=params.frame_shift_ms,
+        )
+        key = "ctc-decoding"
+        return {key: (hyps, timestamps)}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        timestamps = [[] for _ in range(len(hyps))]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}_hlg_scale_{params.hlg_scale}"  # noqa
+        return {key: (hyps, timestamps)}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = f"no_rescore_hlg_scale_{params.hlg_scale}"
+            timestamps, hyps = parse_fsa_timestamps_and_texts(
+                best_paths=best_path,
+                word_table=word_table,
+                subsampling_factor=params.subsampling_factor,
+                frame_shift_ms=params.frame_shift_ms,
+            )
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}-hlg-scale-{params.hlg_scale}"  # noqa
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            timestamps = [[] for _ in range(len(hyps))]
+        return {key: (hyps, timestamps)}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            timestamps = [[] for _ in range(len(hyps))]
+            ans[lm_scale_str] = (hyps, timestamps)
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[
+    str,
+    List[
+        Tuple[
+            str,
+            List[str],
+            List[str],
+            List[Tuple[float, float]],
+            List[Tuple[float, float]],
+        ]
+    ],
+]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        timestamps_ref = []
+        for cut in batch["supervisions"]["cut"]:
+            for s in cut.supervisions:
+                time = []
+                if s.alignment is not None and "word" in s.alignment:
+                    time = [
+                        (aliword.start, aliword.end)
+                        for aliword in s.alignment["word"]
+                        if aliword.symbol != ""
+                    ]
+                timestamps_ref.append(time)
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        for name, (hyps, timestamps_hyp) in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts) and len(timestamps_hyp) == len(
+                timestamps_ref
+            )
+            for cut_id, hyp_words, ref_text, time_hyp, time_ref in zip(
+                cut_ids, hyps, texts, timestamps_hyp, timestamps_ref
+            ):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words, time_ref, time_hyp))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[
+        str,
+        List[
+            Tuple[
+                List[str],
+                List[str],
+                List[str],
+                List[Tuple[float, float]],
+                List[Tuple[float, float]],
+            ]
+        ],
+    ],
+):
+    test_set_wers = dict()
+    test_set_delays = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts_and_timestamps(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer, mean_delay, var_delay = write_error_stats_with_timestamps(
+                f,
+                f"{test_set_name}-{key}",
+                results,
+                enable_log=True,
+                with_end_time=True,
+            )
+            test_set_wers[key] = wer
+            test_set_delays[key] = (mean_delay, var_delay)
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    # sort according to the mean start symbol delay
+    test_set_delays = sorted(test_set_delays.items(), key=lambda x: x[1][0][0])
+    delays_info = (
+        params.res_dir / f"symbol-delay-summary-{test_set_name}-{params.suffix}.txt"
+    )
+    with open(delays_info, "w") as f:
+        print("settings\t(start, end) symbol-delay (s) (start, end)", file=f)
+        for key, val in test_set_delays:
+            print(
+                "{}\tmean: {}, variance: {}".format(key, val[0], val[1]),
+                file=f,
+            )
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    s = "\nFor {}, (start, end) symbol-delay (s) of different settings are:\n".format(
+        test_set_name
+    )
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_delays:
+        s += "{}\tmean: {}, variance: {}{}\n".format(key, val[0], val[1], note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-greedy-search",
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.vocab_size = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.decoding_method in ["ctc-decoding", "ctc-greedy-search"]:
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc3/encoder_interface.py b/egs/librispeech/ASR/conformer_ctc3/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/export.py b/egs/librispeech/ASR/conformer_ctc3/export.py
new file mode 100755
index 000000000..5cb9b4b6d
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/export.py
@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+
+(1) Export to torchscript model using torch.jit.trace()
+
+./conformer_ctc3/export.py \
+  --exp-dir ./conformer_ctc3/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10 \
+  --jit-trace 1
+
+It will generates the file: `jit_trace.pt`.
+
+(2) Export `model.state_dict()`
+
+./conformer_ctc3/export.py \
+  --exp-dir ./conformer_ctc3/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `conformer_ctc3/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./conformer_ctc3/decode.py \
+        --exp-dir ./conformer_ctc3/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_bpe_500
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    num_classes = num_tokens(token_table) + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit_trace:
+        # TODO: will support streaming mode
+        assert not params.streaming_model
+        convert_scaled_to_non_scaled(model, inplace=True)
+
+        logging.info("Using torch.jit.trace()")
+
+        x = torch.zeros(1, 100, 80, dtype=torch.float32)
+        x_lens = torch.tensor([100], dtype=torch.int64)
+        traced_model = torch.jit.trace(model, (x, x_lens))
+
+        filename = params.exp_dir / "jit_trace.pt"
+        traced_model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.trace()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc3/jit_pretrained.py b/egs/librispeech/ASR/conformer_ctc3/jit_pretrained.py
new file mode 100755
index 000000000..76db46cc8
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/jit_pretrained.py
@@ -0,0 +1,413 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Mingshuang Luo,)
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+Usage (for non-streaming mode):
+
+(1) ctc-decoding
+./conformer_ctc3/pretrained.py \
+  --nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./conformer_ctc3/pretrained.py \
+  --nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./conformer_ctc3/pretrained.py \
+  --nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) whole-lattice-rescoring
+./conformer_ctc3/pretrained.py \
+  --nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+    params.vocab_size = params.num_classes
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.model_filename)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    nnet_output, _ = model(features, feature_lengths)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc3/lstmp.py b/egs/librispeech/ASR/conformer_ctc3/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/model.py b/egs/librispeech/ASR/conformer_ctc3/model.py
new file mode 100644
index 000000000..f56df2006
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/model.py
@@ -0,0 +1,122 @@
+# Copyright  2021-2022  Xiaomi Corp.     (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+
+class CTCModel(nn.Module):
+    """It implements https://www.cs.toronto.edu/~graves/icml_2006.pdf
+    "Connectionist Temporal Classification: Labelling Unsegmented
+    Sequence Data with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        encoder_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          encoder_dim:
+            The feature embedding dimension.
+          vocab_size:
+            The vocabulary size.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder = encoder
+        self.ctc_output_module = nn.Sequential(
+            nn.Dropout(p=0.1),
+            ScaledLinear(encoder_dim, vocab_size),
+        )
+
+    def get_ctc_output(
+        self,
+        encoder_out: torch.Tensor,
+        delay_penalty: float = 0.0,
+        blank_threshold: float = 0.99,
+    ):
+        """Compute ctc log-prob and optionally (delay_penalty > 0) apply delay penalty.
+        We first split utterance into sub-utterances according to the
+        blank probs, and then add sawtooth-like "blank-bonus" values to
+        the blank probs.
+        See https://github.com/k2-fsa/icefall/pull/669 for details.
+
+        Args:
+          encoder_out:
+            A tensor with shape of (N, T, C).
+          delay_penalty:
+            A constant used to scale the delay penalty score.
+          blank_threshold:
+            The threshold used to split utterance into sub-utterances.
+        """
+        output = self.ctc_output_module(encoder_out)
+        log_prob = nn.functional.log_softmax(output, dim=-1)
+
+        if self.training and delay_penalty > 0:
+            T_arange = torch.arange(encoder_out.shape[1]).to(device=encoder_out.device)
+            # split into sub-utterances using the blank-id
+            mask = log_prob[:, :, 0] >= math.log(blank_threshold)  # (B, T)
+            mask[:, 0] = True
+            cummax_out = (T_arange * mask).cummax(dim=-1)[0]  # (B, T)
+            # the sawtooth "blank-bonus" value
+            penalty = T_arange - cummax_out  # (B, T)
+            penalty_all = torch.zeros_like(log_prob)
+            penalty_all[:, :, 0] = delay_penalty * penalty
+            # apply latency penalty on probs
+            log_prob = log_prob + penalty_all
+
+        return log_prob
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        warmup: float = 1.0,
+        delay_penalty: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          warmup: a floating point value which increases throughout training;
+            values >= 1.0 are fully warmed up and have all modules present.
+          delay_penalty:
+            A constant used to scale the delay penalty score.
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(encoder_out_lens > 0)
+        nnet_output = self.get_ctc_output(encoder_out, delay_penalty=delay_penalty)
+        return nnet_output, encoder_out_lens
diff --git a/egs/librispeech/ASR/conformer_ctc3/optim.py b/egs/librispeech/ASR/conformer_ctc3/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/pretrained.py b/egs/librispeech/ASR/conformer_ctc3/pretrained.py
new file mode 100755
index 000000000..c37b99cce
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/pretrained.py
@@ -0,0 +1,465 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Mingshuang Luo,)
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+Usage (for non-streaming mode):
+
+(1) ctc-decoding
+./conformer_ctc3/pretrained.py \
+  --checkpoint conformer_ctc3/exp/pretrained.pt \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  test_wavs/1089-134686-0001.wav
+
+(2) 1best
+./conformer_ctc3/pretrained.py \
+  --checkpoint conformer_ctc3/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  test_wavs/1089-134686-0001.wav
+
+(3) nbest-rescoring
+./conformer_ctc3/pretrained.py \
+  --checkpoint conformer_ctc3/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  test_wavs/1089-134686-0001.wav
+
+(4) whole-lattice-rescoring
+./conformer_ctc3/pretrained.py \
+  --checkpoint conformer_ctc3/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  test_wavs/1089-134686-0001.wav
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a tokens.txt file 
+            to convert tokens to actual words or characters. It needs 
+            neither a lexicon nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+    params.vocab_size = params.num_classes
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    hyps = []
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    # model forward
+    if params.simulate_streaming:
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+    nnet_output = model.get_ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        max_token_id = params.num_classes - 1
+
+        # Load tokens.txt here
+        token_table = k2.SymbolTable.from_file(params.tokens)
+
+        def token_ids_to_words(token_ids: List[int]) -> str:
+            text = ""
+            for i in token_ids:
+                text += token_table[i]
+            return text.replace("▁", " ").strip()
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        hyp_tokens = get_texts(best_path)
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyp_tokens = get_texts(best_path)
+        for hyp in hyp_tokens:
+            hyps.append(" ".join([word_sym_table[i] for i in hyp]))
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc3/scaling.py b/egs/librispeech/ASR/conformer_ctc3/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/scaling_converter.py b/egs/librispeech/ASR/conformer_ctc3/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_ctc3/test_model.py b/egs/librispeech/ASR/conformer_ctc3/test_model.py
new file mode 100755
index 000000000..b97b7eed8
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/test_model.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./conformer_ctc3/test_model.py
+"""
+
+import torch
+
+from train import get_params, get_ctc_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.dynamic_chunk_training = False
+    params.short_chunk_size = 25
+    params.num_left_chunks = 4
+    params.causal_convolution = False
+
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    features = torch.randn(2, 100, 80)
+    feature_lengths = torch.full((2,), 100)
+    model(x=features, x_lens=feature_lengths)
+
+
+def test_model_streaming():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.dynamic_chunk_training = True
+    params.short_chunk_size = 25
+    params.num_left_chunks = 4
+    params.causal_convolution = True
+
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    features = torch.randn(2, 100, 80)
+    feature_lengths = torch.full((2,), 100)
+    encoder_out, _ = model.encoder(x=features, x_lens=feature_lengths)
+    model.get_ctc_output(encoder_out)
+
+
+def main():
+    test_model()
+    test_model_streaming()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc3/train.py b/egs/librispeech/ASR/conformer_ctc3/train.py
new file mode 100755
index 000000000..2cd223945
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc3/train.py
@@ -0,0 +1,1109 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conformer_ctc3/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./conformer_ctc3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir conformer_ctc3/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+# train a streaming model
+./conformer_ctc3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conformer_ctc3/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 25 \
+  --num-left-chunks 4 \
+  --max-duration 300 \
+  --delay-penalty 0.0
+"""
+
+import argparse
+import copy
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import CTCModel
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc3/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant used to scale the symbol delay penalty,
+        to encourage symbol emit earlier for streaming models.
+        It is almost the same as the `delay_penalty` in our `rnnt_loss`, See
+        https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    parser.add_argument(
+        "--nnet-delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant to penalize symbol delay, which is applied on
+        the nnet_output after log-softmax.
+        We recommend using --delay-penalty instead.
+        See https://github.com/k2-fsa/icefall/pull/669 for details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "none",
+            "use_double_scores": True,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_ctc_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    model = CTCModel(
+        encoder=encoder,
+        encoder_dim=params.encoder_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: Union[BpeCtcTrainingGraphCompiler, CtcTrainingGraphCompiler],
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_out_lens = model(
+            feature,
+            feature_lens,
+            warmup=warmup,
+            delay_penalty=params.nnet_delay_penalty if warmup >= 1.0 else 0,
+        )
+        assert torch.all(encoder_out_lens > 0)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    if isinstance(graph_compiler, BpeCtcTrainingGraphCompiler):
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+    elif isinstance(graph_compiler, CtcTrainingGraphCompiler):
+        # Works with a phone lexicon
+        decoding_graph = graph_compiler.compile(texts)
+    else:
+        raise ValueError(f"Unsupported type of graph compiler: {type(graph_compiler)}")
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        delay_penalty=params.delay_penalty if warmup >= 1.0 else 0.0,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+    ctc_loss_is_finite = torch.isfinite(ctc_loss)
+    if not torch.all(ctc_loss_is_finite):
+        logging.info("Not all losses are finite!\n" f"ctc_loss: {ctc_loss}")
+        ctc_loss = ctc_loss[ctc_loss_is_finite]
+
+        # If either all simple_loss or pruned_loss is inf or nan,
+        # we stop the training process by raising an exception
+        if torch.all(~ctc_loss_is_finite):
+            raise ValueError(
+                "There are too many utterances in this batch "
+                "leading to inf or nan losses."
+            )
+    loss = ctc_loss.sum()
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    # info["frames"] is an approximate number for two reasons:
+    # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+    # (2) If some utterances in the batch lead to inf/nan loss, they
+    #     are filtered out.
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: Union[BpeCtcTrainingGraphCompiler, CtcTrainingGraphCompiler],
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: Union[BpeCtcTrainingGraphCompiler, CtcTrainingGraphCompiler],
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    params.vocab_size = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    if "lang_bpe" in str(params.lang_dir):
+        graph_compiler = BpeCtcTrainingGraphCompiler(
+            params.lang_dir,
+            device=device,
+            sos_token="<sos/eos>",
+            eos_token="<sos/eos>",
+        )
+    elif "lang_phone" in str(params.lang_dir):
+        graph_compiler = CtcTrainingGraphCompiler(
+            lexicon,
+            device=device,
+            need_repeat_flag=params.delay_penalty > 0,
+        )
+        # Manually add the sos/eos ID with their default values
+        # from the BPE recipe which we're adapting here.
+        graph_compiler.sos_id = 1
+        graph_compiler.eos_id = 1
+    else:
+        raise ValueError(
+            f"Unsupported type of lang dir (we expected it to have "
+            f"'lang_bpe' or 'lang_phone' in its name): {params.lang_dir}"
+        )
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: Union[BpeCtcTrainingGraphCompiler, CtcTrainingGraphCompiler],
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_mmi/__init__.py b/egs/librispeech/ASR/conformer_mmi/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/conformer_mmi/asr_datamodule.py b/egs/librispeech/ASR/conformer_mmi/asr_datamodule.py
new file mode 120000
index 000000000..a73848de9
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/asr_datamodule.py
@@ -0,0 +1 @@
+../conformer_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conformer_mmi/conformer.py b/egs/librispeech/ASR/conformer_mmi/conformer.py
new file mode 100644
index 000000000..53e48eb13
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/conformer.py
@@ -0,0 +1,894 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self, x: Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        x = self.activation(self.norm(x))
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/librispeech/ASR/conformer_mmi/decode.py b/egs/librispeech/ASR/conformer_mmi/decode.py
new file mode 100755
index 000000000..74f6e73fa
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/decode.py
@@ -0,0 +1,701 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        conformer_ctc/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp_500",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="Number of attention decoder layers",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "lm_dir": Path("data/lm"),
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+        # TODO: pass `lattice` instead of `rescored_lattice` to
+        # `rescore_with_attention_decoder`
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        for lm_scale in lm_scale_list:
+            ans["empty"] = [[] * lattice.shape[0]]
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu")
+        )
+        HLG = HLG.to(device)
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location="cpu")
+            G = k2.Fsa.from_dict(d).to(device)
+
+        if params.method in ["whole-lattice-rescoring", "attention-decoder"]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    # CAUTION: `test_sets` is for displaying only.
+    # If you want to skip test-clean, you have to skip
+    # it inside the for loop. That is, use
+    #
+    #   if test_set == 'test-clean': continue
+    test_sets = ["test-clean", "test-other"]
+    test_dls = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_mmi/subsampling.py b/egs/librispeech/ASR/conformer_mmi/subsampling.py
new file mode 100644
index 000000000..ad9415987
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/subsampling.py
@@ -0,0 +1,136 @@
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape [N, T, idim] to an output
+    with shape [N, T', odim], where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is [N, T, idim].
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim]
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is [N, T, idim].
+
+        Returns:
+          Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
+        """
+        # On entry, x is [N, T, idim]
+        x = x.unsqueeze(1)  # [N, T, idim] -> [N, 1, T, idim] i.e., [N, C, H, W]
+        x = self.conv(x)
+        # Now x is of shape [N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2]
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape [N, ((T-1)//2 - 1))//2, odim]
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape [N, T, idim] to an output
+    with shape [N, T', odim], where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is [N, T, idim].
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim]
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is [N, T, idim].
+
+        Returns:
+          Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x
diff --git a/egs/librispeech/ASR/conformer_mmi/test_subsampling.py b/egs/librispeech/ASR/conformer_mmi/test_subsampling.py
new file mode 100755
index 000000000..d0bb017dd
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/test_subsampling.py
@@ -0,0 +1,32 @@
+#!/usr/bin/env python3
+
+import torch
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+
+def test_conv2d_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = Conv2dSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
+
+
+def test_vgg_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = VggSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
diff --git a/egs/librispeech/ASR/conformer_mmi/test_transformer.py b/egs/librispeech/ASR/conformer_mmi/test_transformer.py
new file mode 100644
index 000000000..25d18076d
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/test_transformer.py
@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+
+import torch
+from torch.nn.utils.rnn import pad_sequence
+from transformer import (
+    Transformer,
+    add_eos,
+    add_sos,
+    decoder_padding_mask,
+    encoder_padding_mask,
+    generate_square_subsequent_mask,
+)
+
+
+def test_encoder_padding_mask():
+    supervisions = {
+        "sequence_idx": torch.tensor([0, 1, 2]),
+        "start_frame": torch.tensor([0, 0, 0]),
+        "num_frames": torch.tensor([18, 7, 13]),
+    }
+
+    max_len = ((18 - 1) // 2 - 1) // 2
+    mask = encoder_padding_mask(max_len, supervisions)
+    expected_mask = torch.tensor(
+        [
+            [False, False, False],  # ((18 - 1)//2 - 1)//2 = 3,
+            [False, True, True],  # ((7 - 1)//2 - 1)//2 = 1,
+            [False, False, True],  # ((13 - 1)//2 - 1)//2 = 2,
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_transformer():
+    num_features = 40
+    num_classes = 87
+    model = Transformer(num_features=num_features, num_classes=num_classes)
+
+    N = 31
+
+    for T in range(7, 30):
+        x = torch.rand(N, T, num_features)
+        y, _, _ = model(x)
+        assert y.shape == (N, (((T - 1) // 2) - 1) // 2, num_classes)
+
+
+def test_generate_square_subsequent_mask():
+    s = 5
+    mask = generate_square_subsequent_mask(s)
+    inf = float("inf")
+    expected_mask = torch.tensor(
+        [
+            [0.0, -inf, -inf, -inf, -inf],
+            [0.0, 0.0, -inf, -inf, -inf],
+            [0.0, 0.0, 0.0, -inf, -inf],
+            [0.0, 0.0, 0.0, 0.0, -inf],
+            [0.0, 0.0, 0.0, 0.0, 0.0],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_decoder_padding_mask():
+    x = [torch.tensor([1, 2]), torch.tensor([3]), torch.tensor([2, 5, 8])]
+    y = pad_sequence(x, batch_first=True, padding_value=-1)
+    mask = decoder_padding_mask(y, ignore_id=-1)
+    expected_mask = torch.tensor(
+        [
+            [False, False, True],
+            [False, True, True],
+            [False, False, False],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_add_sos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_sos(x, sos_id=0)
+    expected_y = [[0, 1, 2], [0, 3], [0, 2, 5, 8]]
+    assert y == expected_y
+
+
+def test_add_eos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_eos(x, eos_id=0)
+    expected_y = [[1, 2, 0], [3, 0], [2, 5, 8, 0]]
+    assert y == expected_y
diff --git a/egs/librispeech/ASR/conformer_mmi/train-with-attention.py b/egs/librispeech/ASR/conformer_mmi/train-with-attention.py
new file mode 100755
index 000000000..100bc846a
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/train-with-attention.py
@@ -0,0 +1,860 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Dict, Optional
+
+import k2
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.ali import convert_alignments_to_tensor, load_alignments, lookup_alignments
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.lexicon import Lexicon
+from icefall.mmi import LFMMILoss
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import AttributeDict, encode_supervisions, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=50,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_mmi/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--ali-dir",
+        type=str,
+        default="data/ali_500",
+        help="""This folder is expected to contain
+        two files, train-960.pt and valid.pt, which
+        contain framewise alignment information for
+        the training set and validation set.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp-attn",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--use-pruned-intersect",
+        type=str2bool,
+        default=False,
+        help="""Whether to use `intersect_dense_pruned` to get denominator
+        lattice.""",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Whether to do batch normalization for the
+                              input features.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - lr_factor: The lr_factor for Noam optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            # parameters for loss
+            "beam_size": 6,  # will change it to 8 after some batches (see code)
+            "reduction": "sum",
+            "use_double_scores": True,
+            "att_rate": 0.7,
+            "num_decoder_layers": 6,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "lr_factor": 5.0,
+            "warm_step": 80000,
+            "den_scale": 1.0,
+            # use alignments before this number of batches
+            "use_ali_until": 13000,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: MmiTrainingGraphCompiler,
+    is_training: bool,
+    ali: Optional[Dict[str, torch.Tensor]],
+):
+    """
+    Compute LF-MMI loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+      ali:
+        Precomputed alignments.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+        # NOTE: We need `encode_supervisions` to sort sequences with
+        # different duration in decreasing order, required by
+        # `k2.intersect_dense` called in `LFMMILoss.forward()`
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+
+        if ali is not None and params.batch_idx_train < params.use_ali_until:
+            cut_ids = [cut.id for cut in supervisions["cut"]]
+
+            # As encode_supervisions reorders cuts, we need
+            # also to reorder cut IDs here
+            new2old = supervision_segments[:, 0].tolist()
+            cut_ids = [cut_ids[i] for i in new2old]
+
+            # Check that new2old is just a permutation,
+            # i.e., each cut contains only one utterance
+            new2old.sort()
+            assert new2old == torch.arange(len(new2old)).tolist()
+            mask = lookup_alignments(
+                cut_ids=cut_ids,
+                alignments=ali,
+                num_classes=nnet_output.shape[2],
+            ).to(nnet_output)
+
+            min_len = min(nnet_output.shape[1], mask.shape[1])
+            ali_scale = 500.0 / (params.batch_idx_train + 500)
+
+            nnet_output = nnet_output.clone()
+            nnet_output[:, :min_len, :] += ali_scale * mask[:, :min_len, :]
+
+        if params.batch_idx_train > params.use_ali_until and params.beam_size < 8:
+            #  logging.info("Change beam size to 8")
+            params.beam_size = 8
+        else:
+            params.beam_size = 6
+
+        loss_fn = LFMMILoss(
+            graph_compiler=graph_compiler,
+            use_pruned_intersect=params.use_pruned_intersect,
+            den_scale=params.den_scale,
+            beam_size=params.beam_size,
+        )
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            nnet_output,
+            supervision_segments,
+            allow_truncate=params.subsampling_factor - 1,
+        )
+        mmi_loss = loss_fn(dense_fsa_vec=dense_fsa_vec, texts=texts)
+
+    if params.att_rate != 0.0:
+        token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * mmi_loss + params.att_rate * att_loss
+    else:
+        loss = mmi_loss
+        att_loss = torch.tensor([0])
+
+    # train_frames and valid_frames are used for printing.
+    if is_training:
+        params.train_frames = supervision_segments[:, 2].sum().item()
+    else:
+        params.valid_frames = supervision_segments[:, 2].sum().item()
+
+    assert loss.requires_grad == is_training
+
+    return loss, mmi_loss.detach(), att_loss.detach()
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: MmiTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+    ali: Optional[Dict[str, torch.Tensor]] = None,
+) -> None:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = 0.0
+    tot_mmi_loss = 0.0
+    tot_att_loss = 0.0
+    tot_frames = 0.0
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, mmi_loss, att_loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+            ali=ali,
+        )
+        assert loss.requires_grad is False
+        assert mmi_loss.requires_grad is False
+        assert att_loss.requires_grad is False
+
+        loss_cpu = loss.detach().cpu().item()
+        tot_loss += loss_cpu
+
+        tot_mmi_loss += mmi_loss.detach().cpu().item()
+        tot_att_loss += att_loss.detach().cpu().item()
+
+        tot_frames += params.valid_frames
+
+    if world_size > 1:
+        s = torch.tensor(
+            [tot_loss, tot_mmi_loss, tot_att_loss, tot_frames],
+            device=loss.device,
+        )
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        s = s.cpu().tolist()
+        tot_loss = s[0]
+        tot_mmi_loss = s[1]
+        tot_att_loss = s[2]
+        tot_frames = s[3]
+
+    params.valid_loss = tot_loss / tot_frames
+    params.valid_mmi_loss = tot_mmi_loss / tot_frames
+    params.valid_att_loss = tot_att_loss / tot_frames
+
+    if params.valid_loss < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = params.valid_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: MmiTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    train_ali: Optional[Dict[str, torch.Tensor]],
+    valid_ali: Optional[Dict[str, torch.Tensor]],
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      train_ali:
+        Precomputed alignments for the training set.
+      valid_ali:
+        Precomputed alignments for the validation set.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = 0.0  # sum of losses over all batches
+    tot_mmi_loss = 0.0
+    tot_att_loss = 0.0
+
+    tot_frames = 0.0  # sum of frames over all batches
+    params.tot_loss = 0.0
+    params.tot_frames = 0.0
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, mmi_loss, att_loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+            ali=train_ali,
+        )
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        loss_cpu = loss.detach().cpu().item()
+        mmi_loss_cpu = mmi_loss.detach().cpu().item()
+        att_loss_cpu = att_loss.detach().cpu().item()
+
+        tot_frames += params.train_frames
+        tot_loss += loss_cpu
+        tot_mmi_loss += mmi_loss_cpu
+        tot_att_loss += att_loss_cpu
+
+        params.tot_frames += params.train_frames
+        params.tot_loss += loss_cpu
+
+        tot_avg_loss = tot_loss / tot_frames
+        tot_avg_mmi_loss = tot_mmi_loss / tot_frames
+        tot_avg_att_loss = tot_att_loss / tot_frames
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"batch avg mmi loss {mmi_loss_cpu/params.train_frames:.4f}, "
+                f"batch avg att loss {att_loss_cpu/params.train_frames:.4f}, "
+                f"batch avg loss {loss_cpu/params.train_frames:.4f}, "
+                f"total avg mmiloss: {tot_avg_mmi_loss:.4f}, "
+                f"total avg att loss: {tot_avg_att_loss:.4f}, "
+                f"total avg loss: {tot_avg_loss:.4f}, "
+                f"batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/current_mmi_loss",
+                    mmi_loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/current_att_loss",
+                    att_loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/current_loss",
+                    loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/tot_avg_mmi_loss",
+                    tot_avg_mmi_loss,
+                    params.batch_idx_train,
+                )
+
+                tb_writer.add_scalar(
+                    "train/tot_avg_att_loss",
+                    tot_avg_att_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/tot_avg_loss",
+                    tot_avg_loss,
+                    params.batch_idx_train,
+                )
+        if batch_idx > 0 and batch_idx % params.reset_interval == 0:
+            tot_loss = 0.0  # sum of losses over all batches
+            tot_mmi_loss = 0.0
+            tot_att_loss = 0.0
+
+            tot_frames = 0.0  # sum of frames over all batches
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+                ali=valid_ali,
+            )
+            model.train()
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"valid mmi loss {params.valid_mmi_loss:.4f},"
+                f"valid att loss {params.valid_att_loss:.4f},"
+                f"valid loss {params.valid_loss:.4f},"
+                f" best valid loss: {params.best_valid_loss:.4f} "
+                f"best valid epoch: {params.best_valid_epoch}"
+            )
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/valid_mmi_loss",
+                    params.valid_mmi_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/valid_att_loss",
+                    params.valid_att_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/valid_loss",
+                    params.valid_loss,
+                    params.batch_idx_train,
+                )
+
+    params.train_loss = params.tot_loss / params.tot_frames
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+
+    logging.info("About to create model")
+    if params.att_rate == 0:
+        assert params.num_decoder_layers == 0, f"{params.num_decoder_layers}"
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    train_960_ali_filename = Path(params.ali_dir) / "train-960.pt"
+    if (
+        params.batch_idx_train < params.use_ali_until
+        and train_960_ali_filename.is_file()
+    ):
+        logging.info("Use pre-computed alignments")
+        subsampling_factor, train_ali = load_alignments(train_960_ali_filename)
+        assert subsampling_factor == params.subsampling_factor
+        assert len(train_ali) == 843723, f"{len(train_ali)} vs 843723"
+
+        valid_ali_filename = Path(params.ali_dir) / "valid.pt"
+        subsampling_factor, valid_ali = load_alignments(valid_ali_filename)
+        assert subsampling_factor == params.subsampling_factor
+
+        train_ali = convert_alignments_to_tensor(train_ali, device=device)
+        valid_ali = convert_alignments_to_tensor(valid_ali, device=device)
+    else:
+        logging.info("Not using alignments")
+        train_ali = None
+        valid_ali = None
+
+    librispeech = LibriSpeechAsrDataModule(args)
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+        if params.batch_idx_train >= params.use_ali_until and train_ali is not None:
+            # Delete the alignments to save memory
+            train_ali = None
+            valid_ali = None
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            train_ali=train_ali,
+            valid_ali=valid_ali,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_mmi/train.py b/egs/librispeech/ASR/conformer_mmi/train.py
new file mode 100755
index 000000000..f9f80632e
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/train.py
@@ -0,0 +1,861 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Dict, Optional
+
+import k2
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.ali import convert_alignments_to_tensor, load_alignments, lookup_alignments
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.lexicon import Lexicon
+from icefall.mmi import LFMMILoss
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import AttributeDict, encode_supervisions, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=50,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_mmi/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--ali-dir",
+        type=str,
+        default="data/ali_500",
+        help="""This folder is expected to contain
+        two files, train-960.pt and valid.pt, which
+        contain framewise alignment information for
+        the training set and validation set.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--use-pruned-intersect",
+        type=str2bool,
+        default=False,
+        help="""Whether to use `intersect_dense_pruned` to get denominator
+        lattice.""",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Whether to do batch normalization for the
+                              input features.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - lr_factor: The lr_factor for Noam optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            # parameters for loss
+            "beam_size": 6,  # will change it to 8 after some batches (see code)
+            "reduction": "sum",
+            "use_double_scores": True,
+            "att_rate": 0.0,
+            "num_decoder_layers": 0,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "lr_factor": 5.0,
+            "warm_step": 80000,
+            "den_scale": 1.0,
+            # use alignments before this number of batches
+            "use_ali_until": 13000,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: MmiTrainingGraphCompiler,
+    is_training: bool,
+    ali: Optional[Dict[str, torch.Tensor]],
+):
+    """
+    Compute LF-MMI loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+      ali:
+        Precomputed alignments.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+        # NOTE: We need `encode_supervisions` to sort sequences with
+        # different duration in decreasing order, required by
+        # `k2.intersect_dense` called in `LFMMILoss.forward()`
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+
+        if ali is not None and params.batch_idx_train < params.use_ali_until:
+            cut_ids = [cut.id for cut in supervisions["cut"]]
+
+            # As encode_supervisions reorders cuts, we need
+            # also to reorder cut IDs here
+            new2old = supervision_segments[:, 0].tolist()
+            cut_ids = [cut_ids[i] for i in new2old]
+
+            # Check that new2old is just a permutation,
+            # i.e., each cut contains only one utterance
+            new2old.sort()
+            assert new2old == torch.arange(len(new2old)).tolist()
+            mask = lookup_alignments(
+                cut_ids=cut_ids,
+                alignments=ali,
+                num_classes=nnet_output.shape[2],
+            ).to(nnet_output)
+
+            min_len = min(nnet_output.shape[1], mask.shape[1])
+            ali_scale = 500.0 / (params.batch_idx_train + 500)
+
+            nnet_output = nnet_output.clone()
+            nnet_output[:, :min_len, :] += ali_scale * mask[:, :min_len, :]
+
+        if params.batch_idx_train > params.use_ali_until and params.beam_size < 8:
+            logging.info("Change beam size to 8")
+            params.beam_size = 8
+        else:
+            params.beam_size = 6
+
+        loss_fn = LFMMILoss(
+            graph_compiler=graph_compiler,
+            use_pruned_intersect=params.use_pruned_intersect,
+            den_scale=params.den_scale,
+            beam_size=params.beam_size,
+        )
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            nnet_output,
+            supervision_segments,
+            allow_truncate=params.subsampling_factor - 1,
+        )
+        mmi_loss = loss_fn(dense_fsa_vec=dense_fsa_vec, texts=texts)
+
+    if params.att_rate != 0.0:
+        token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * mmi_loss + params.att_rate * att_loss
+    else:
+        loss = mmi_loss
+        att_loss = torch.tensor([0])
+
+    # train_frames and valid_frames are used for printing.
+    if is_training:
+        params.train_frames = supervision_segments[:, 2].sum().item()
+    else:
+        params.valid_frames = supervision_segments[:, 2].sum().item()
+
+    assert loss.requires_grad == is_training
+
+    return loss, mmi_loss.detach(), att_loss.detach()
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: MmiTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+    ali: Optional[Dict[str, torch.Tensor]] = None,
+) -> None:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = 0.0
+    tot_mmi_loss = 0.0
+    tot_att_loss = 0.0
+    tot_frames = 0.0
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, mmi_loss, att_loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+            ali=ali,
+        )
+        assert loss.requires_grad is False
+        assert mmi_loss.requires_grad is False
+        assert att_loss.requires_grad is False
+
+        loss_cpu = loss.detach().cpu().item()
+        tot_loss += loss_cpu
+
+        tot_mmi_loss += mmi_loss.detach().cpu().item()
+        tot_att_loss += att_loss.detach().cpu().item()
+
+        tot_frames += params.valid_frames
+
+    if world_size > 1:
+        s = torch.tensor(
+            [tot_loss, tot_mmi_loss, tot_att_loss, tot_frames],
+            device=loss.device,
+        )
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        s = s.cpu().tolist()
+        tot_loss = s[0]
+        tot_mmi_loss = s[1]
+        tot_att_loss = s[2]
+        tot_frames = s[3]
+
+    params.valid_loss = tot_loss / tot_frames
+    params.valid_mmi_loss = tot_mmi_loss / tot_frames
+    params.valid_att_loss = tot_att_loss / tot_frames
+
+    if params.valid_loss < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = params.valid_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: MmiTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    train_ali: Optional[Dict[str, torch.Tensor]],
+    valid_ali: Optional[Dict[str, torch.Tensor]],
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      train_ali:
+        Precomputed alignments for the training set.
+      valid_ali:
+        Precomputed alignments for the validation set.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = 0.0  # sum of losses over all batches
+    tot_mmi_loss = 0.0
+    tot_att_loss = 0.0
+
+    tot_frames = 0.0  # sum of frames over all batches
+    params.tot_loss = 0.0
+    params.tot_frames = 0.0
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, mmi_loss, att_loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+            ali=train_ali,
+        )
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        loss_cpu = loss.detach().cpu().item()
+        mmi_loss_cpu = mmi_loss.detach().cpu().item()
+        att_loss_cpu = att_loss.detach().cpu().item()
+
+        tot_frames += params.train_frames
+        tot_loss += loss_cpu
+        tot_mmi_loss += mmi_loss_cpu
+        tot_att_loss += att_loss_cpu
+
+        params.tot_frames += params.train_frames
+        params.tot_loss += loss_cpu
+
+        tot_avg_loss = tot_loss / tot_frames
+        tot_avg_mmi_loss = tot_mmi_loss / tot_frames
+        tot_avg_att_loss = tot_att_loss / tot_frames
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"batch avg mmi loss {mmi_loss_cpu/params.train_frames:.4f}, "
+                f"batch avg att loss {att_loss_cpu/params.train_frames:.4f}, "
+                f"batch avg loss {loss_cpu/params.train_frames:.4f}, "
+                f"total avg mmiloss: {tot_avg_mmi_loss:.4f}, "
+                f"total avg att loss: {tot_avg_att_loss:.4f}, "
+                f"total avg loss: {tot_avg_loss:.4f}, "
+                f"batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/current_mmi_loss",
+                    mmi_loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/current_att_loss",
+                    att_loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/current_loss",
+                    loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/tot_avg_mmi_loss",
+                    tot_avg_mmi_loss,
+                    params.batch_idx_train,
+                )
+
+                tb_writer.add_scalar(
+                    "train/tot_avg_att_loss",
+                    tot_avg_att_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/tot_avg_loss",
+                    tot_avg_loss,
+                    params.batch_idx_train,
+                )
+        if batch_idx > 0 and batch_idx % params.reset_interval == 0:
+            tot_loss = 0.0  # sum of losses over all batches
+            tot_mmi_loss = 0.0
+            tot_att_loss = 0.0
+
+            tot_frames = 0.0  # sum of frames over all batches
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+                ali=valid_ali,
+            )
+            model.train()
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"valid mmi loss {params.valid_mmi_loss:.4f},"
+                f"valid att loss {params.valid_att_loss:.4f},"
+                f"valid loss {params.valid_loss:.4f},"
+                f" best valid loss: {params.best_valid_loss:.4f} "
+                f"best valid epoch: {params.best_valid_epoch}"
+            )
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/valid_mmi_loss",
+                    params.valid_mmi_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/valid_att_loss",
+                    params.valid_att_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/valid_loss",
+                    params.valid_loss,
+                    params.batch_idx_train,
+                )
+
+    params.train_loss = params.tot_loss / params.tot_frames
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+
+    logging.info("About to create model")
+    if params.att_rate == 0:
+        assert params.num_decoder_layers == 0, f"{params.num_decoder_layers}"
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    train_960_ali_filename = Path(params.ali_dir) / "train-960.pt"
+    if (
+        params.batch_idx_train < params.use_ali_until
+        and train_960_ali_filename.is_file()
+    ):
+        logging.info("Use pre-computed alignments")
+        subsampling_factor, train_ali = load_alignments(train_960_ali_filename)
+        assert subsampling_factor == params.subsampling_factor
+        assert len(train_ali) == 843723, f"{len(train_ali)} vs 843723"
+
+        valid_ali_filename = Path(params.ali_dir) / "valid.pt"
+        subsampling_factor, valid_ali = load_alignments(valid_ali_filename)
+        assert subsampling_factor == params.subsampling_factor
+
+        train_ali = convert_alignments_to_tensor(train_ali, device=device)
+        valid_ali = convert_alignments_to_tensor(valid_ali, device=device)
+    else:
+        logging.info("Not using alignments")
+        train_ali = None
+        valid_ali = None
+
+    librispeech = LibriSpeechAsrDataModule(args)
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+        if params.batch_idx_train >= params.use_ali_until and train_ali is not None:
+            # Delete the alignments to save memory
+            train_ali = None
+            valid_ali = None
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            train_ali=train_ali,
+            valid_ali=valid_ali,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conformer_mmi/transformer.py b/egs/librispeech/ASR/conformer_mmi/transformer.py
new file mode 100644
index 000000000..2542d9abe
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_mmi/transformer.py
@@ -0,0 +1,984 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        if use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss(self.decoder_num_class)
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision)
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=eos_id)
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=-1)
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=eos_id)
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=-1)
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        self.pe = None
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+class LabelSmoothingLoss(nn.Module):
+    """
+    Label-smoothing loss. KL-divergence between
+    q_{smoothed ground truth prob.}(w)
+    and p_{prob. computed by model}(w) is minimized.
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/label_smoothing_loss.py  # noqa
+
+    Args:
+        size: the number of class
+        padding_idx: padding_idx: ignored class id
+        smoothing: smoothing rate (0.0 means the conventional CE)
+        normalize_length: normalize loss by sequence length if True
+        criterion: loss function to be smoothed
+    """
+
+    def __init__(
+        self,
+        size: int,
+        padding_idx: int = -1,
+        smoothing: float = 0.1,
+        normalize_length: bool = False,
+        criterion: nn.Module = nn.KLDivLoss(reduction="none"),
+    ) -> None:
+        """Construct an LabelSmoothingLoss object."""
+        super(LabelSmoothingLoss, self).__init__()
+        self.criterion = criterion
+        self.padding_idx = padding_idx
+        assert 0.0 < smoothing <= 1.0
+        self.confidence = 1.0 - smoothing
+        self.smoothing = smoothing
+        self.size = size
+        self.true_dist = None
+        self.normalize_length = normalize_length
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute loss between x and target.
+
+        Args:
+          x:
+            prediction of dimension
+            (batch_size, input_length, number_of_classes).
+          target:
+            target masked with self.padding_id of
+            dimension (batch_size, input_length).
+
+        Returns:
+          A scalar tensor containing the loss without normalization.
+        """
+        assert x.size(2) == self.size
+        #  batch_size = x.size(0)
+        x = x.view(-1, self.size)
+        target = target.view(-1)
+        with torch.no_grad():
+            true_dist = x.clone()
+            true_dist.fill_(self.smoothing / (self.size - 1))
+            ignore = target == self.padding_idx  # (B,)
+            total = len(target) - ignore.sum().item()
+            target = target.masked_fill(ignore, 0)  # avoid -1 index
+            true_dist.scatter_(1, target.unsqueeze(1), self.confidence)
+        kl = self.criterion(torch.log_softmax(x, dim=1), true_dist)
+        #  denom = total if self.normalize_length else batch_size
+        denom = total if self.normalize_length else 1
+        return kl.masked_fill(ignore.unsqueeze(1), 0).sum() / denom
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    ans = []
+    for utt in token_ids:
+        ans.append([sos_id] + utt)
+    return ans
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    ans = []
+    for utt in token_ids:
+        ans.append(utt + [eos_id])
+    return ans
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py
new file mode 100755
index 000000000..7be3299f3
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py
@@ -0,0 +1,644 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./conv_emformer_transducer_stateless/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless/exp \
+      --max-duration 300 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method greedy_search \
+      --use-averaged-model True
+
+(2) modified beam search
+./conv_emformer_transducer_stateless/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless/exp \
+      --max-duration 300 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method modified_beam_search \
+      --use-averaged-model True \
+      --beam-size 4
+
+(3) fast beam search
+./conv_emformer_transducer_stateless/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless/exp \
+      --max-duration 300 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method fast_beam_search \
+      --use-averaged-model True \
+      --beam 4 \
+      --max-contexts 4 \
+      --max-states 8
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += params.chunk_length
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.chunk_length),
+        value=LOG_EPS,
+    )
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0
+            start = params.epoch - params.avg
+            assert start >= 1
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py
new file mode 100644
index 000000000..91f50cf67
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py
@@ -0,0 +1,1872 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# It is modified based on
+# 1) https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py  # noqa
+# 2) https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py  # noqa
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+from icefall.utils import make_pad_mask
+
+LOG_EPSILON = math.log(1e-10)
+
+
+def unstack_states(
+    states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]]
+) -> List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]:
+    """Unstack the emformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Args:
+      states:
+        A tuple of 2 elements.
+        ``states[0]`` is the attention caches of a batch of utterance.
+        ``states[1]`` is the convolution caches of a batch of utterance.
+        ``len(states[0])`` and ``len(states[1])`` both eqaul to number of layers.  # noqa
+
+    Returns:
+      A list of states.
+      ``states[i]`` is a tuple of 2 elements of i-th utterance.
+      ``states[i][0]`` is the attention caches of i-th utterance.
+      ``states[i][1]`` is the convolution caches of i-th utterance.
+      ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers.  # noqa
+    """
+
+    attn_caches, conv_caches = states
+    batch_size = conv_caches[0].size(0)
+    num_layers = len(attn_caches)
+
+    list_attn_caches = [None] * batch_size
+    for i in range(batch_size):
+        list_attn_caches[i] = [[] for _ in range(num_layers)]
+    for li, layer in enumerate(attn_caches):
+        for s in layer:
+            s_list = s.unbind(dim=1)
+            for bi, b in enumerate(list_attn_caches):
+                b[li].append(s_list[bi])
+
+    list_conv_caches = [None] * batch_size
+    for i in range(batch_size):
+        list_conv_caches[i] = [None] * num_layers
+    for li, layer in enumerate(conv_caches):
+        c_list = layer.unbind(dim=0)
+        for bi, b in enumerate(list_conv_caches):
+            b[li] = c_list[bi]
+
+    ans = [None] * batch_size
+    for i in range(batch_size):
+        ans[i] = [list_attn_caches[i], list_conv_caches[i]]
+
+    return ans
+
+
+def stack_states(
+    state_list: List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]
+) -> Tuple[List[List[torch.Tensor]], List[torch.Tensor]]:
+    """Stack list of emformer states that correspond to separate utterances
+    into a single emformer state so that it can be used as an input for
+    emformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the emformer model for a single utterance.
+        ``states[i]`` is a tuple of 2 elements of i-th utterance.
+        ``states[i][0]`` is the attention caches of i-th utterance.
+        ``states[i][1]`` is the convolution caches of i-th utterance.
+        ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers.  # noqa
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+
+    attn_caches = []
+    for layer in state_list[0][0]:
+        if batch_size > 1:
+            # Note: We will stack attn_caches[layer][s][] later to get attn_caches[layer][s]  # noqa
+            attn_caches.append([[s] for s in layer])
+        else:
+            attn_caches.append([s.unsqueeze(1) for s in layer])
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states[0]):
+            for si, s in enumerate(layer):
+                attn_caches[li][si].append(s)
+                if b == batch_size - 1:
+                    attn_caches[li][si] = torch.stack(attn_caches[li][si], dim=1)
+
+    conv_caches = []
+    for layer in state_list[0][1]:
+        if batch_size > 1:
+            # Note: We will stack conv_caches[layer][] later to get conv_caches[layer]  # noqa
+            conv_caches.append([layer])
+        else:
+            conv_caches.append(layer.unsqueeze(0))
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states[1]):
+            conv_caches[li].append(layer)
+            if b == batch_size - 1:
+                conv_caches[li] = torch.stack(conv_caches[li], dim=0)
+
+    return [attn_caches, conv_caches]
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule.
+
+    Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa
+
+    Args:
+      chunk_length (int):
+        Length of each chunk.
+      right_context_length (int):
+        Length of right context.
+      channels (int):
+        The number of input channels and output channels of conv layers.
+      kernel_size (int):
+        Kernerl size of conv layers.
+      bias (bool):
+        Whether to use bias in conv layers (default=True).
+    """
+
+    def __init__(
+        self,
+        chunk_length: int,
+        right_context_length: int,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super().__init__()
+        # kernerl_size should be an odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0, kernel_size
+
+        self.chunk_length = chunk_length
+        self.right_context_length = right_context_length
+        self.channels = channels
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        # After pointwise_conv1 we put x through a gated linear unit
+        # (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in
+        # the range 1 to 4, but sometimes, for some reason, for layer 0 the rms
+        # ends up being very large, between 50 and 100 for different channels.
+        # This will cause very peaky and sparse derivatives for the sigmoid
+        # gating function, which will tend to make the loss function not learn
+        # effectively.  (for most layers the average absolute values are in the
+        # range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for
+        # different layers, which likely breaks down as 0.5 for the "linear"
+        # half and 0.2 to 0.3 for the part that goes into the sigmoid.
+        # The idea is that if we constrain the rms values to a reasonable range
+        # via a constraint of max_abs=10.0, it will be in a better position to
+        # start learning something, i.e. to latch onto the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        # make it causal by padding cached (kernel_size - 1) frames on the left
+        self.cache_size = kernel_size - 1
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def _split_right_context(
+        self,
+        pad_utterance: torch.Tensor,
+        right_context: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          pad_utterance:
+            Its shape is (cache_size + U, B, D).
+          right_context:
+            Its shape is (R, B, D).
+
+        Returns:
+          Right context segments padding with corresponding context.
+          Its shape is (num_segs * B, D, cache_size + right_context_length).
+        """
+        U_, B, D = pad_utterance.size()
+        R = right_context.size(0)
+        assert self.right_context_length != 0
+        assert R % self.right_context_length == 0
+        num_chunks = R // self.right_context_length
+        right_context = right_context.reshape(
+            num_chunks, self.right_context_length, B, D
+        )
+        right_context = right_context.permute(0, 2, 1, 3).reshape(
+            num_chunks * B, self.right_context_length, D
+        )
+
+        intervals = torch.arange(
+            0, self.chunk_length * (num_chunks - 1), self.chunk_length
+        )
+        first = torch.arange(self.chunk_length, self.chunk_length + self.cache_size)
+        indexes = intervals.unsqueeze(1) + first.unsqueeze(0)
+        indexes = torch.cat(
+            [indexes, torch.arange(U_ - self.cache_size, U_).unsqueeze(0)]
+        )
+        padding = pad_utterance[indexes]  # (num_chunks, cache_size, B, D)
+        padding = padding.permute(0, 2, 1, 3).reshape(
+            num_chunks * B, self.cache_size, D
+        )
+
+        pad_right_context = torch.cat([padding, right_context], dim=1)
+        # (num_chunks * B, cache_size + right_context_length, D)
+        return pad_right_context.permute(0, 2, 1)
+
+    def _merge_right_context(self, right_context: torch.Tensor, B: int) -> torch.Tensor:
+        """
+        Args:
+          right_context:
+            Right context segments.
+            It shape is (num_segs * B, D, right_context_length).
+          B:
+            Batch size.
+
+        Returns:
+          A tensor of shape (B, D, R), where
+          R = num_segs * right_context_length.
+        """
+        right_context = right_context.reshape(
+            -1, B, self.channels, self.right_context_length
+        )
+        right_context = right_context.permute(1, 2, 0, 3)
+        right_context = right_context.reshape(B, self.channels, -1)
+        return right_context
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Causal convolution module.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance tensor of shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context tensor of shape (R, B, D).
+
+        Returns:
+          A tuple of 2 tensors:
+          - output utterance of shape (U, B, D).
+          - output right_context of shape (R, B, D).
+        """
+        U, B, D = utterance.size()
+        R, _, _ = right_context.size()
+
+        # point-wise conv and GLU mechanism
+        x = torch.cat([right_context, utterance], dim=0)  # (R + U, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, R + U)
+        x = self.pointwise_conv1(x)  # (B, 2 * D, R + U)
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (B, D, R + U)
+        utterance = x[:, :, R:]  # (B, D, U)
+        right_context = x[:, :, :R]  # (B, D, R)
+
+        # make causal convolution
+        cache = torch.zeros(B, D, self.cache_size, device=x.device, dtype=x.dtype)
+        pad_utterance = torch.cat([cache, utterance], dim=2)  # (B, D, cache + U)
+
+        # depth-wise conv on utterance
+        utterance = self.depthwise_conv(pad_utterance)  # (B, D, U)
+
+        if self.right_context_length > 0:
+            # depth-wise conv on right_context
+            pad_right_context = self._split_right_context(
+                pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1)
+            )  # (num_segs * B, D, cache_size + right_context_length)
+            right_context = self.depthwise_conv(
+                pad_right_context
+            )  # (num_segs * B, D, right_context_length)
+            right_context = self._merge_right_context(right_context, B)  # (B, D, R)
+
+        x = torch.cat([right_context, utterance], dim=2)  # (B, D, R + U)
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        # point-wise conv
+        x = self.pointwise_conv2(x)  # (B, D, R + U)
+
+        right_context = x[:, :, :R]  # (B, D, R)
+        utterance = x[:, :, R:]  # (B, D, U)
+        return (
+            utterance.permute(2, 0, 1),
+            right_context.permute(2, 0, 1),
+        )
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        cache: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Causal convolution module applied on both utterance and right_context.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance tensor of shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context tensor of shape (R, B, D).
+          cache (torch.Tensor, optional):
+            Cached tensor for left padding of shape (B, D, cache_size).
+
+        Returns:
+          A tuple of 3 tensors:
+            - output utterance of shape (U, B, D).
+            - output right_context of shape (R, B, D).
+            - updated cache tensor of shape (B, D, cache_size).
+        """
+        U, B, D = utterance.size()
+        R, _, _ = right_context.size()
+
+        # point-wise conv
+        x = torch.cat([utterance, right_context], dim=0)  # (U + R, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, U + R)
+        x = self.pointwise_conv1(x)  # (B, 2 * D, U + R)
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (B, D, U + R)
+
+        # make causal convolution
+        assert cache.shape == (B, D, self.cache_size), cache.shape
+        x = torch.cat([cache, x], dim=2)  # (B, D, cache_size + U + R)
+        # update cache
+        new_cache = x[:, :, -R - self.cache_size : -R]
+
+        # 1-D depth-wise conv
+        x = self.depthwise_conv(x)  # (B, D, U + R)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        # point-wise conv
+        x = self.pointwise_conv2(x)  # (B, D, U + R)
+
+        utterance = x[:, :, :U]  # (B, D, U)
+        right_context = x[:, :, U:]  # (B, D, R)
+        return (
+            utterance.permute(2, 0, 1),
+            right_context.permute(2, 0, 1),
+            new_cache,
+        )
+
+
+class EmformerAttention(nn.Module):
+    r"""Emformer layer attention module.
+
+    Args:
+      embed_dim (int):
+        Embedding dimension.
+      nhead (int):
+        Number of attention heads in each Emformer layer.
+      dropout (float, optional):
+        Dropout probability. (Default: 0.0)
+      tanh_on_mem (bool, optional):
+        If ``True``, applies tanh to memory elements. (Default: ``False``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (Default: -1e8)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        nhead: int,
+        dropout: float = 0.0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        if embed_dim % nhead != 0:
+            raise ValueError(
+                f"embed_dim ({embed_dim}) is not a multiple of nhead ({nhead})."
+            )
+
+        self.embed_dim = embed_dim
+        self.nhead = nhead
+        self.tanh_on_mem = tanh_on_mem
+        self.negative_inf = negative_inf
+        self.head_dim = embed_dim // nhead
+        self.dropout = dropout
+
+        self.emb_to_key_value = ScaledLinear(embed_dim, 2 * embed_dim, bias=True)
+        self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+    def _gen_attention_probs(
+        self,
+        attention_weights: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Given the entire attention weights, mask out unecessary connections
+        and optionally with padding positions, to obtain underlying chunk-wise
+        attention probabilities.
+
+        B: batch size;
+        Q: length of query;
+        KV: length of key and value.
+
+        Args:
+          attention_weights (torch.Tensor):
+            Attention weights computed on the entire concatenated tensor
+            with shape (B * nhead, Q, KV).
+          attention_mask (torch.Tensor):
+            Mask tensor where chunk-wise connections are filled with `False`,
+            and other unnecessary connections are filled with `True`,
+            with shape (Q, KV).
+          padding_mask (torch.Tensor, optional):
+            Mask tensor where the padding positions are fill with `True`,
+            and other positions are filled with `False`, with shapa `(B, KV)`.
+
+        Returns:
+          A tensor of shape (B * nhead, Q, KV).
+        """
+        attention_weights_float = attention_weights.float()
+        attention_weights_float = attention_weights_float.masked_fill(
+            attention_mask.unsqueeze(0), self.negative_inf
+        )
+        if padding_mask is not None:
+            Q = attention_weights.size(1)
+            B = attention_weights.size(0) // self.nhead
+            attention_weights_float = attention_weights_float.view(B, self.nhead, Q, -1)
+            attention_weights_float = attention_weights_float.masked_fill(
+                padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+                self.negative_inf,
+            )
+            attention_weights_float = attention_weights_float.view(
+                B * self.nhead, Q, -1
+            )
+
+        attention_probs = nn.functional.softmax(
+            attention_weights_float, dim=-1
+        ).type_as(attention_weights)
+
+        attention_probs = nn.functional.dropout(
+            attention_probs, p=self.dropout, training=self.training
+        )
+        return attention_probs
+
+    def _forward_impl(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        summary: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        left_context_key: Optional[torch.Tensor] = None,
+        left_context_val: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Underlying chunk-wise attention implementation."""
+        U, B, _ = utterance.size()
+        R = right_context.size(0)
+        M = memory.size(0)
+        scaling = float(self.head_dim) ** -0.5
+
+        # compute query with [right_context, utterance, summary].
+        query = self.emb_to_query(torch.cat([right_context, utterance, summary]))
+        # compute key and value with [memory, right_context, utterance].
+        key, value = self.emb_to_key_value(
+            torch.cat([memory, right_context, utterance])
+        ).chunk(chunks=2, dim=2)
+
+        if left_context_key is not None and left_context_val is not None:
+            # now compute key and value with
+            #   [memory, right context, left context, uttrance]
+            # this is used in inference mode
+            key = torch.cat([key[: M + R], left_context_key, key[M + R :]])
+            value = torch.cat([value[: M + R], left_context_val, value[M + R :]])
+        Q = query.size(0)
+        # KV = key.size(0)
+
+        reshaped_query, reshaped_key, reshaped_value = [
+            tensor.contiguous().view(-1, B * self.nhead, self.head_dim).transpose(0, 1)
+            for tensor in [query, key, value]
+        ]  # (B * nhead, Q or KV, head_dim)
+        attention_weights = torch.bmm(
+            reshaped_query * scaling, reshaped_key.transpose(1, 2)
+        )  # (B * nhead, Q, KV)
+
+        # compute attention probabilities
+        attention_probs = self._gen_attention_probs(
+            attention_weights, attention_mask, padding_mask
+        )
+
+        # compute attention outputs
+        attention = torch.bmm(attention_probs, reshaped_value)
+        assert attention.shape == (B * self.nhead, Q, self.head_dim)
+        attention = attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim)
+
+        # apply output projection
+        outputs = self.out_proj(attention)
+
+        output_right_context_utterance = outputs[: R + U]
+        output_memory = outputs[R + U :]
+        if self.tanh_on_mem:
+            output_memory = torch.tanh(output_memory)
+        else:
+            output_memory = torch.clamp(output_memory, min=-10, max=10)
+
+        return output_right_context_utterance, output_memory, key, value
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        summary: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # TODO: Modify docs.
+        """Forward pass for training and validation mode.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of the hard-copied right contexts;
+        U: length of full utterance;
+        S: length of summary vectors;
+        M: length of memory vectors.
+
+        It computes a `big` attention matrix on full utterance and
+        then utilizes a pre-computed mask to simulate chunk-wise attention.
+
+        It concatenates three blocks: hard-copied right contexts,
+        full utterance, and summary vectors, as a `big` block,
+        to compute the query tensor:
+        query = [right_context, utterance, summary],
+        with length Q = R + U + S.
+        It concatenates the three blocks: memory vectors,
+        hard-copied right contexts, and full utterance as another `big` block,
+        to compute the key and value tensors:
+        key & value = [memory, right_context, utterance],
+        with length KV = M + R + U.
+        Attention scores is computed with above `big` query and key.
+
+        Then the underlying chunk-wise attention is obtained by applying
+        the attention mask. Suppose
+        c_i: chunk at index i;
+        r_i: right context that c_i can use;
+        l_i: left context that c_i can use;
+        m_i: past memory vectors from previous layer that c_i can use;
+        s_i: summary vector of c_i;
+        The target chunk-wise attention is:
+        c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key);
+        s_i (in query) -> l_i, c_i, r_i (in key).
+
+        Args:
+          utterance (torch.Tensor):
+            Full utterance frames, with shape (U, B, D).
+          right_context (torch.Tensor):
+            Hard-copied right context frames, with shape (R, B, D),
+            where R = num_chunks * right_context_length
+          summary (torch.Tensor):
+            Summary elements with shape (S, B, D), where S = num_chunks.
+            It is an empty tensor without using memory.
+          memory (torch.Tensor):
+            Memory elements, with shape (M, B, D), where M = num_chunks - 1.
+            It is an empty tensor without using memory.
+          attention_mask (torch.Tensor):
+            Pre-computed attention mask to simulate underlying chunk-wise
+            attention, with shape (Q, KV).
+          padding_mask (torch.Tensor):
+            Padding mask of key tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 2 tensors:
+            - output of right context and utterance, with shape (R + U, B, D).
+            - memory output, with shape (M, B, D), where M = S - 1 or M = 0.
+        """
+        (output_right_context_utterance, output_memory, _, _,) = self._forward_impl(
+            utterance,
+            right_context,
+            summary,
+            memory,
+            attention_mask,
+            padding_mask=padding_mask,
+        )
+        return output_right_context_utterance, output_memory[:-1]
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        summary: torch.Tensor,
+        memory: torch.Tensor,
+        left_context_key: torch.Tensor,
+        left_context_val: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward pass for inference.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of right context;
+        U: length of utterance, i.e., current chunk;
+        L: length of cached left context;
+        S: length of summary vectors, S = 1;
+        M: length of cached memory vectors.
+
+        It concatenates the right context, utterance (i.e., current chunk)
+        and summary vector of current chunk, to compute the query tensor:
+        query = [right_context, utterance, summary],
+        with length Q = R + U + S.
+        It concatenates the memory vectors, right context, left context, and
+        current chunk, to compute the key and value tensors:
+        key & value = [memory, right_context, left_context, utterance],
+        with length KV = M + R + L + U.
+
+        The chunk-wise attention is:
+        chunk, right context (in query) ->
+          left context, chunk, right context, memory vectors (in key);
+        summary (in query) -> left context, chunk, right context (in key).
+
+        Args:
+          utterance (torch.Tensor):
+            Current chunk frames, with shape (U, B, D), where U = chunk_length.
+          right_context (torch.Tensor):
+            Right context frames, with shape (R, B, D),
+            where R = right_context_length.
+          summary (torch.Tensor):
+            Summary vector with shape (1, B, D), or empty tensor.
+          memory (torch.Tensor):
+            Memory vectors, with shape (M, B, D), or empty tensor.
+          left_context_key (torch,Tensor):
+            Cached attention key of left context from preceding computation,
+            with shape (L, B, D).
+          left_context_val (torch.Tensor):
+            Cached attention value of left context from preceding computation,
+            with shape (L, B, D).
+          padding_mask (torch.Tensor):
+            Padding mask of key tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 4 tensors:
+            - output of right context and utterance, with shape (R + U, B, D).
+            - memory output, with shape (1, B, D) or (0, B, D).
+            - attention key of left context and utterance, which would be cached
+              for next computation, with shape (L + U, B, D).
+            - attention value of left context and utterance, which would be
+              cached for next computation, with shape (L + U, B, D).
+        """
+        U = utterance.size(0)
+        R = right_context.size(0)
+        L = left_context_key.size(0)
+        S = summary.size(0)
+        M = memory.size(0)
+
+        # TODO: move it outside
+        # query = [right context, utterance, summary]
+        Q = R + U + S
+        # key, value = [memory, right context, left context, uttrance]
+        KV = M + R + L + U
+        attention_mask = torch.zeros(Q, KV).to(
+            dtype=torch.bool, device=utterance.device
+        )
+        # disallow attention bettween the summary vector with the memory bank
+        attention_mask[-1, :M] = True
+        (
+            output_right_context_utterance,
+            output_memory,
+            key,
+            value,
+        ) = self._forward_impl(
+            utterance,
+            right_context,
+            summary,
+            memory,
+            attention_mask,
+            padding_mask=padding_mask,
+            left_context_key=left_context_key,
+            left_context_val=left_context_val,
+        )
+        return (
+            output_right_context_utterance,
+            output_memory,
+            key[M + R :],
+            value[M + R :],
+        )
+
+
+class EmformerEncoderLayer(nn.Module):
+    """Emformer layer that constitutes Emformer.
+
+    Args:
+      d_model (int):
+        Input dimension.
+      nhead (int):
+        Number of attention heads.
+      dim_feedforward (int):
+        Hidden layer dimension of feedforward network.
+      chunk_length (int):
+        Length of each input segment.
+      dropout (float, optional):
+        Dropout probability. (Default: 0.0)
+      layer_dropout (float, optional):
+        Layer dropout probability. (Default: 0.0)
+      cnn_module_kernel (int):
+        Kernel size of convolution module.
+      left_context_length (int, optional):
+        Length of left context. (Default: 0)
+      right_context_length (int, optional):
+        Length of right context. (Default: 0)
+      memory_size (int, optional):
+        Number of memory elements to use. (Default: 0)
+      tanh_on_mem (bool, optional):
+        If ``True``, applies tanh to memory elements. (Default: ``False``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (Default: -1e8)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        chunk_length: int,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        self.attention = EmformerAttention(
+            embed_dim=d_model,
+            nhead=nhead,
+            dropout=dropout,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+        )
+        self.summary_op = nn.AvgPool1d(
+            kernel_size=chunk_length, stride=chunk_length, ceil_mode=True
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(
+            chunk_length,
+            right_context_length,
+            d_model,
+            cnn_module_kernel,
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean
+        # (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.layer_dropout = layer_dropout
+        self.left_context_length = left_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+        self.d_model = d_model
+        self.use_memory = memory_size > 0
+
+    def _update_attn_cache(
+        self,
+        next_key: torch.Tensor,
+        next_val: torch.Tensor,
+        memory: torch.Tensor,
+        attn_cache: List[torch.Tensor],
+    ) -> List[torch.Tensor]:
+        """Update cached attention state:
+        1) output memory of current chunk in the lower layer;
+        2) attention key and value in current chunk's computation, which would
+        be resued in next chunk's computation.
+        """
+        new_memory = torch.cat([attn_cache[0], memory])
+        new_key = torch.cat([attn_cache[1], next_key])
+        new_val = torch.cat([attn_cache[2], next_val])
+        attn_cache[0] = new_memory[new_memory.size(0) - self.memory_size :]
+        attn_cache[1] = new_key[new_key.size(0) - self.left_context_length :]
+        attn_cache[2] = new_val[new_val.size(0) - self.left_context_length :]
+        return attn_cache
+
+    def _apply_conv_module_forward(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+    ) -> torch.Tensor:
+        """Apply convolution module in training and validation mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+        utterance, right_context = self.conv_module(utterance, right_context)
+        right_context_utterance = torch.cat([right_context, utterance])
+        return right_context_utterance
+
+    def _apply_conv_module_infer(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        conv_cache: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply convolution module on utterance in inference mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+        utterance, right_context, conv_cache = self.conv_module.infer(
+            utterance, right_context, conv_cache
+        )
+        right_context_utterance = torch.cat([right_context, utterance])
+        return right_context_utterance, conv_cache
+
+    def _apply_attention_module_forward(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply attention module in training and validation mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+
+        if self.use_memory:
+            summary = self.summary_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)
+        else:
+            summary = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+        output_right_context_utterance, output_memory = self.attention(
+            utterance=utterance,
+            right_context=right_context,
+            summary=summary,
+            memory=memory,
+            attention_mask=attention_mask,
+            padding_mask=padding_mask,
+        )
+
+        return output_right_context_utterance, output_memory
+
+    def _apply_attention_module_infer(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        memory: torch.Tensor,
+        attn_cache: List[torch.Tensor],
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """Apply attention module in inference mode.
+        1) Unpack cached states including:
+           - memory from previous chunks in the lower layer;
+           - attention key and value of left context from preceding
+             chunk's compuation;
+        2) Apply attention computation;
+        3) Update cached attention states including:
+           - output memory of current chunk in the lower layer;
+           - attention key and value in current chunk's computation, which would
+             be resued in next chunk's computation.
+        """
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+
+        pre_memory = attn_cache[0]
+        left_context_key = attn_cache[1]
+        left_context_val = attn_cache[2]
+
+        if self.use_memory:
+            summary = self.summary_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)
+            summary = summary[:1]
+        else:
+            summary = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+        (
+            output_right_context_utterance,
+            output_memory,
+            next_key,
+            next_val,
+        ) = self.attention.infer(
+            utterance=utterance,
+            right_context=right_context,
+            summary=summary,
+            memory=pre_memory,
+            left_context_key=left_context_key,
+            left_context_val=left_context_val,
+            padding_mask=padding_mask,
+        )
+        attn_cache = self._update_attn_cache(next_key, next_val, memory, attn_cache)
+        return output_right_context_utterance, output_memory, attn_cache
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        r"""Forward pass for training and validation mode.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance frames, with shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context frames, with shape (R, B, D).
+          memory (torch.Tensor):
+            Memory elements, with shape (M, B, D).
+            It is an empty tensor without using memory.
+          attention_mask (torch.Tensor):
+            Attention mask for underlying attention module,
+            with shape (Q, KV), where Q = R + U + S, KV = M + R + U.
+          padding_mask (torch.Tensor):
+            Padding mask of ker tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 3 tensors:
+            - output utterance, with shape (U, B, D).
+            - output right context, with shape (R, B, D).
+            - output memory, with shape (M, B, D).
+        """
+        R = right_context.size(0)
+        src = torch.cat([right_context, utterance])
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # emformer attention module
+        src_att, output_memory = self._apply_attention_module_forward(
+            src, R, memory, attention_mask, padding_mask=padding_mask
+        )
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src_conv = self._apply_conv_module_forward(src, R)
+        src = src + self.dropout(src_conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        output_utterance = src[R:]
+        output_right_context = src[:R]
+        return output_utterance, output_right_context, output_memory
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        attn_cache: List[torch.Tensor],
+        conv_cache: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        List[torch.Tensor],
+        torch.Tensor,
+    ]:
+        """Forward pass for inference.
+
+         B: batch size;
+         D: embedding dimension;
+         R: length of right_context;
+         U: length of utterance;
+         M: length of memory.
+
+        Args:
+           utterance (torch.Tensor):
+             Utterance frames, with shape (U, B, D).
+           right_context (torch.Tensor):
+             Right context frames, with shape (R, B, D).
+           memory (torch.Tensor):
+             Memory elements, with shape (M, B, D).
+           attn_cache (List[torch.Tensor]):
+             Cached attention tensors generated in preceding computation,
+             including memory, key and value of left context.
+           conv_cache (torch.Tensor, optional):
+             Cache tensor of left context for causal convolution.
+           padding_mask (torch.Tensor):
+             Padding mask of ker tensor.
+
+         Returns:
+           (Tensor, Tensor, List[torch.Tensor], Tensor):
+             - output utterance, with shape (U, B, D);
+             - output right_context, with shape (R, B, D);
+             - output memory, with shape (1, B, D) or (0, B, D).
+             - updated attention cache.
+             - updated convolution cache.
+        """
+        R = right_context.size(0)
+        src = torch.cat([right_context, utterance])
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # emformer attention module
+        (src_att, output_memory, attn_cache,) = self._apply_attention_module_infer(
+            src, R, memory, attn_cache, padding_mask=padding_mask
+        )
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache)
+        src = src + self.dropout(src_conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        output_utterance = src[R:]
+        output_right_context = src[:R]
+        return (
+            output_utterance,
+            output_right_context,
+            output_memory,
+            attn_cache,
+            conv_cache,
+        )
+
+
+def _gen_attention_mask_block(
+    col_widths: List[int],
+    col_mask: List[bool],
+    num_rows: int,
+    device: torch.device,
+) -> torch.Tensor:
+    assert len(col_widths) == len(
+        col_mask
+    ), "Length of col_widths must match that of col_mask"
+
+    mask_block = [
+        torch.ones(num_rows, col_width, device=device)
+        if is_ones_col
+        else torch.zeros(num_rows, col_width, device=device)
+        for col_width, is_ones_col in zip(col_widths, col_mask)
+    ]
+    return torch.cat(mask_block, dim=1)
+
+
+class EmformerEncoder(nn.Module):
+    """Implements the Emformer architecture introduced in
+    *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency
+    Streaming Speech Recognition*
+    [:footcite:`shi2021emformer`].
+
+    Args:
+      d_model (int):
+        Input dimension.
+      nhead (int):
+        Number of attention heads in each emformer layer.
+      dim_feedforward (int):
+        Hidden layer dimension of each emformer layer's feedforward network.
+      num_encoder_layers (int):
+        Number of emformer layers to instantiate.
+      chunk_length (int):
+        Length of each input segment.
+      dropout (float, optional):
+        Dropout probability. (default: 0.0)
+      layer_dropout (float, optional):
+        Layer dropout probability. (default: 0.0)
+      cnn_module_kernel (int):
+        Kernel size of convolution module.
+      left_context_length (int, optional):
+        Length of left context. (default: 0)
+      right_context_length (int, optional):
+        Length of right context. (default: 0)
+      memory_size (int, optional):
+        Number of memory elements to use. (default: 0)
+      tanh_on_mem (bool, optional):
+        If ``true``, applies tanh to memory elements. (default: ``false``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (default: -1e8)
+    """
+
+    def __init__(
+        self,
+        chunk_length: int,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        assert (
+            chunk_length - 1
+        ) & chunk_length == 0, "chunk_length should be a power of 2."
+        self.shift = int(math.log(chunk_length, 2))
+
+        self.use_memory = memory_size > 0
+        self.init_memory_op = nn.AvgPool1d(
+            kernel_size=chunk_length,
+            stride=chunk_length,
+            ceil_mode=True,
+        )
+
+        self.emformer_layers = nn.ModuleList(
+            [
+                EmformerEncoderLayer(
+                    d_model=d_model,
+                    nhead=nhead,
+                    dim_feedforward=dim_feedforward,
+                    chunk_length=chunk_length,
+                    dropout=dropout,
+                    layer_dropout=layer_dropout,
+                    cnn_module_kernel=cnn_module_kernel,
+                    left_context_length=left_context_length,
+                    right_context_length=right_context_length,
+                    memory_size=memory_size,
+                    tanh_on_mem=tanh_on_mem,
+                    negative_inf=negative_inf,
+                )
+                for layer_idx in range(num_encoder_layers)
+            ]
+        )
+
+        self.num_encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.left_context_length = left_context_length
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+        self.cnn_module_kernel = cnn_module_kernel
+
+    def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor:
+        """Hard copy each chunk's right context and concat them."""
+        T = x.shape[0]
+        num_chunks = math.ceil((T - self.right_context_length) / self.chunk_length)
+        # first (num_chunks - 1) right context block
+        intervals = torch.arange(
+            0, self.chunk_length * (num_chunks - 1), self.chunk_length
+        )
+        first = torch.arange(
+            self.chunk_length, self.chunk_length + self.right_context_length
+        )
+        indexes = intervals.unsqueeze(1) + first.unsqueeze(0)
+        # cat last right context block
+        indexes = torch.cat(
+            [
+                indexes,
+                torch.arange(T - self.right_context_length, T).unsqueeze(0),
+            ]
+        )
+        right_context_blocks = x[indexes.reshape(-1)]
+        return right_context_blocks
+
+    def _gen_attention_mask_col_widths(self, chunk_idx: int, U: int) -> List[int]:
+        """Calculate column widths (key, value) in attention mask for the
+        chunk_idx chunk."""
+        num_chunks = math.ceil(U / self.chunk_length)
+        rc = self.right_context_length
+        lc = self.left_context_length
+        rc_start = chunk_idx * rc
+        rc_end = rc_start + rc
+        chunk_start = max(chunk_idx * self.chunk_length - lc, 0)
+        chunk_end = min((chunk_idx + 1) * self.chunk_length, U)
+        R = rc * num_chunks
+
+        if self.use_memory:
+            m_start = max(chunk_idx - self.memory_size, 0)
+            M = num_chunks - 1
+            col_widths = [
+                m_start,  # before memory
+                chunk_idx - m_start,  # memory
+                M - chunk_idx,  # after memory
+                rc_start,  # before right context
+                rc,  # right context
+                R - rc_end,  # after right context
+                chunk_start,  # before chunk
+                chunk_end - chunk_start,  # chunk
+                U - chunk_end,  # after chunk
+            ]
+        else:
+            col_widths = [
+                rc_start,  # before right context
+                rc,  # right context
+                R - rc_end,  # after right context
+                chunk_start,  # before chunk
+                chunk_end - chunk_start,  # chunk
+                U - chunk_end,  # after chunk
+            ]
+
+        return col_widths
+
+    def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor:
+        """Generate attention mask to simulate underlying chunk-wise attention
+        computation, where chunk-wise connections are filled with `False`,
+        and other unnecessary connections beyond chunk are filled with `True`.
+
+        R: length of hard-copied right contexts;
+        U: length of full utterance;
+        S: length of summary vectors;
+        M: length of memory vectors;
+        Q: length of attention query;
+        KV: length of attention key and value.
+
+        The shape of attention mask is (Q, KV).
+        If self.use_memory is `True`:
+          query = [right_context, utterance, summary];
+          key, value = [memory, right_context, utterance];
+          Q = R + U + S, KV = M + R + U.
+        Otherwise:
+          query = [right_context, utterance]
+          key, value = [right_context, utterance]
+          Q = R + U, KV = R + U.
+
+        Suppose:
+          c_i: chunk at index i;
+          r_i: right context that c_i can use;
+          l_i: left context that c_i can use;
+          m_i: past memory vectors from previous layer that c_i can use;
+          s_i: summary vector of c_i.
+        The target chunk-wise attention is:
+          c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key);
+          s_i (in query) -> l_i, c_i, r_i (in key).
+        """
+        U = utterance.size(0)
+        num_chunks = math.ceil(U / self.chunk_length)
+
+        right_context_mask = []
+        utterance_mask = []
+        summary_mask = []
+
+        if self.use_memory:
+            num_cols = 9
+            # right context and utterance both attend to memory, right context,
+            # utterance
+            right_context_utterance_cols_mask = [
+                idx in [1, 4, 7] for idx in range(num_cols)
+            ]
+            # summary attends to right context, utterance
+            summary_cols_mask = [idx in [4, 7] for idx in range(num_cols)]
+            masks_to_concat = [right_context_mask, utterance_mask, summary_mask]
+        else:
+            num_cols = 6
+            # right context and utterance both attend to right context and
+            # utterance
+            right_context_utterance_cols_mask = [
+                idx in [1, 4] for idx in range(num_cols)
+            ]
+            summary_cols_mask = None
+            masks_to_concat = [right_context_mask, utterance_mask]
+
+        for chunk_idx in range(num_chunks):
+            col_widths = self._gen_attention_mask_col_widths(chunk_idx, U)
+
+            right_context_mask_block = _gen_attention_mask_block(
+                col_widths,
+                right_context_utterance_cols_mask,
+                self.right_context_length,
+                utterance.device,
+            )
+            right_context_mask.append(right_context_mask_block)
+
+            utterance_mask_block = _gen_attention_mask_block(
+                col_widths,
+                right_context_utterance_cols_mask,
+                min(
+                    self.chunk_length,
+                    U - chunk_idx * self.chunk_length,
+                ),
+                utterance.device,
+            )
+            utterance_mask.append(utterance_mask_block)
+
+            if summary_cols_mask is not None:
+                summary_mask_block = _gen_attention_mask_block(
+                    col_widths, summary_cols_mask, 1, utterance.device
+                )
+                summary_mask.append(summary_mask_block)
+
+        attention_mask = (
+            1 - torch.cat([torch.cat(mask) for mask in masks_to_concat])
+        ).to(torch.bool)
+        return attention_mask
+
+    def forward(
+        self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass for training and validation mode.
+
+        B: batch size;
+        D: input dimension;
+        U: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (U + right_context_length, B, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, which contains the
+            right_context at the end.
+
+        Returns:
+          A tuple of 2 tensors:
+            - output utterance frames, with shape (U, B, D).
+            - output_lengths, with shape (B,), without containing the
+              right_context at the end.
+        """
+        U = x.size(0) - self.right_context_length
+
+        right_context = self._gen_right_context(x)
+        utterance = x[:U]
+        output_lengths = torch.clamp(lengths - self.right_context_length, min=0)
+        attention_mask = self._gen_attention_mask(utterance)
+        memory = (
+            self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[:-1]
+            if self.use_memory
+            else torch.empty(0).to(dtype=x.dtype, device=x.device)
+        )
+        padding_mask = make_pad_mask(
+            memory.size(0) + right_context.size(0) + output_lengths
+        )
+
+        output = utterance
+        for layer in self.emformer_layers:
+            output, right_context, memory = layer(
+                output,
+                right_context,
+                memory,
+                attention_mask,
+                padding_mask=padding_mask,
+                warmup=warmup,
+            )
+
+        return output, output_lengths
+
+    @torch.jit.export
+    def infer(
+        self,
+        x: torch.Tensor,
+        lengths: torch.Tensor,
+        num_processed_frames: torch.Tensor,
+        states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ]:
+        """Forward pass for streaming inference.
+
+        B: batch size;
+        D: input dimension;
+        U: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (U + right_context_length, B, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, which contains the
+            right_context at the end.
+          states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa
+            Cached states containing:
+            - attn_caches: attention states from preceding chunk's computation,
+              where each element corresponds to each emformer layer
+            - conv_caches: left context for causal convolution, where each
+              element corresponds to each layer.
+
+        Returns:
+          (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]):
+            - output utterance frames, with shape (U, B, D).
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+            - updated states from current chunk's computation.
+        """
+        assert num_processed_frames.shape == (x.size(1),)
+
+        attn_caches = states[0]
+        assert len(attn_caches) == self.num_encoder_layers, len(attn_caches)
+        for i in range(len(attn_caches)):
+            assert attn_caches[i][0].shape == (
+                self.memory_size,
+                x.size(1),
+                self.d_model,
+            ), attn_caches[i][0].shape
+            assert attn_caches[i][1].shape == (
+                self.left_context_length,
+                x.size(1),
+                self.d_model,
+            ), attn_caches[i][1].shape
+            assert attn_caches[i][2].shape == (
+                self.left_context_length,
+                x.size(1),
+                self.d_model,
+            ), attn_caches[i][2].shape
+
+        conv_caches = states[1]
+        assert len(conv_caches) == self.num_encoder_layers, len(conv_caches)
+        for i in range(len(conv_caches)):
+            assert conv_caches[i].shape == (
+                x.size(1),
+                self.d_model,
+                self.cnn_module_kernel - 1,
+            ), conv_caches[i].shape
+
+        right_context = x[-self.right_context_length :]
+        utterance = x[: -self.right_context_length]
+        output_lengths = torch.clamp(lengths - self.right_context_length, min=0)
+        memory = (
+            self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)
+            if self.use_memory
+            else torch.empty(0).to(dtype=x.dtype, device=x.device)
+        )
+
+        # calcualte padding mask to mask out initial zero caches
+        chunk_mask = make_pad_mask(output_lengths).to(x.device)
+        memory_mask = (
+            (
+                (num_processed_frames >> self.shift).view(x.size(1), 1)
+                <= torch.arange(self.memory_size, device=x.device).expand(
+                    x.size(1), self.memory_size
+                )
+            ).flip(1)
+            if self.use_memory
+            else torch.empty(0).to(dtype=torch.bool, device=x.device)
+        )
+        left_context_mask = (
+            num_processed_frames.view(x.size(1), 1)
+            <= torch.arange(self.left_context_length, device=x.device).expand(
+                x.size(1), self.left_context_length
+            )
+        ).flip(1)
+        right_context_mask = torch.zeros(
+            x.size(1),
+            self.right_context_length,
+            dtype=torch.bool,
+            device=x.device,
+        )
+        padding_mask = torch.cat(
+            [memory_mask, right_context_mask, left_context_mask, chunk_mask],
+            dim=1,
+        )
+
+        output = utterance
+        output_attn_caches: List[List[torch.Tensor]] = []
+        output_conv_caches: List[torch.Tensor] = []
+        for layer_idx, layer in enumerate(self.emformer_layers):
+            (
+                output,
+                right_context,
+                memory,
+                output_attn_cache,
+                output_conv_cache,
+            ) = layer.infer(
+                output,
+                right_context,
+                memory,
+                padding_mask=padding_mask,
+                attn_cache=attn_caches[layer_idx],
+                conv_cache=conv_caches[layer_idx],
+            )
+            output_attn_caches.append(output_attn_cache)
+            output_conv_caches.append(output_conv_cache)
+
+        output_states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = (
+            output_attn_caches,
+            output_conv_caches,
+        )
+        return output, output_lengths, output_states
+
+    @torch.jit.export
+    def init_states(self, device: torch.device = torch.device("cpu")):
+        """Create initial states."""
+        attn_caches = [
+            [
+                torch.zeros(self.memory_size, self.d_model, device=device),
+                torch.zeros(self.left_context_length, self.d_model, device=device),
+                torch.zeros(self.left_context_length, self.d_model, device=device),
+            ]
+            for _ in range(self.num_encoder_layers)
+        ]
+        conv_caches = [
+            torch.zeros(self.d_model, self.cnn_module_kernel - 1, device=device)
+            for _ in range(self.num_encoder_layers)
+        ]
+        states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = (
+            attn_caches,
+            conv_caches,
+        )
+        return states
+
+
+class Emformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        chunk_length: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 3,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+        if chunk_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "chunk_length must be a mutiple of subsampling_factor."
+            )
+        if left_context_length != 0 and left_context_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "left_context_length must be 0 or a mutiple of subsampling_factor."  # noqa
+            )
+        if right_context_length != 0 and right_context_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "right_context_length must be 0 or a mutiple of subsampling_factor."  # noqa
+            )
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder = EmformerEncoder(
+            chunk_length=chunk_length // subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+            cnn_module_kernel=cnn_module_kernel,
+            left_context_length=left_context_length // subsampling_factor,
+            right_context_length=right_context_length // subsampling_factor,
+            memory_size=memory_size,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass for training and non-streaming inference.
+
+        B: batch size;
+        D: feature dimension;
+        T: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (B, T, D).
+          x_lens (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, containing the
+            right_context at the end.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          (Tensor, Tensor):
+            - output embedding, with shape (B, T', D), where
+              T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4.
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+        """
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+        assert x.size(0) == x_lens.max().item()
+
+        output, output_lengths = self.encoder(x, x_lens, warmup=warmup)  # (T, N, C)
+
+        output = output.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+
+        return output, output_lengths
+
+    @torch.jit.export
+    def infer(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        num_processed_frames: torch.Tensor,
+        states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ]:
+        """Forward pass for streaming inference.
+
+        B: batch size;
+        D: feature dimension;
+        T: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (B, T, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, containing the
+            right_context at the end.
+          states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa
+            Cached states containing:
+            - past_lens: number of past frames for each sample in batch
+            - attn_caches: attention states from preceding chunk's computation,
+              where each element corresponds to each emformer layer
+            - conv_caches: left context for causal convolution, where each
+              element corresponds to each layer.
+        Returns:
+          (Tensor, Tensor):
+            - output embedding, with shape (B, T', D), where
+              T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4.
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+            - updated states from current chunk's computation.
+        """
+        x = self.encoder_embed(x)
+        # drop the first and last frames
+        x = x[:, 1:-1, :]
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+        x_lens -= 2
+        assert x.size(0) == x_lens.max().item()
+
+        num_processed_frames = num_processed_frames >> 2
+
+        output, output_lengths, output_states = self.encoder.infer(
+            x, x_lens, num_processed_frames, states
+        )
+
+        output = output.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+
+        return output, output_lengths, output_states
+
+    @torch.jit.export
+    def init_states(self, device: torch.device = torch.device("cpu")):
+        """Create initial states."""
+        return self.encoder.init_states(device)
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/export.py
new file mode 100755
index 000000000..67fcc35a4
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/export.py
@@ -0,0 +1,284 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./conv_emformer_transducer_stateless/export.py \
+  --exp-dir ./conv_emformer_transducer_stateless/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --jit False
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `conv_emformer_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./conv_emformer_transducer_stateless/decode.py \
+        --exp-dir ./conv_emformer_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model \
+        --use-averaged-model=False \
+        --num-encoder-layers 12 \
+        --chunk-length 32 \
+        --cnn-module-kernel 31 \
+        --left-context-length 32 \
+        --right-context-length 8 \
+        --memory-size 32
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py
new file mode 120000
index 000000000..ebb6d774d
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py
new file mode 100644
index 000000000..c211b215e
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py
@@ -0,0 +1,154 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class Stream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+        LOG_EPS: float = math.log(1e-10),
+    ) -> None:
+        """
+        Args:
+          params:
+            It's the return value of :func:`get_params`.
+          decoding_graph:
+            The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+            only when --decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+        """
+        self.LOG_EPS = LOG_EPS
+        self.cut_id = cut_id
+
+        # Containing attention caches and convolution caches
+        self.states: Optional[
+            Tuple[List[List[torch.Tensor]], List[torch.Tensor]]
+        ] = None
+
+        # It uses different attributes for different decoding methods.
+        self.context_size = params.context_size
+        self.decoding_method = params.decoding_method
+        if params.decoding_method == "greedy_search":
+            self.hyp = [params.blank_id] * params.context_size
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[params.blank_id] * params.context_size,
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # feature_len is needed to get partial results.
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+            self.hyp: Optional[List[int]] = None
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+        self.ground_truth: str = ""
+
+        self.feature: Optional[torch.Tensor] = None
+        # Make sure all feature frames can be used.
+        # Add 2 here since we will drop the first and last after subsampling.
+        self.chunk_length = params.chunk_length
+        self.pad_length = (
+            params.right_context_length + 2 * params.subsampling_factor + 3
+        )
+        self.num_frames = 0
+        self.num_processed_frames = 0
+
+        # After all feature frames are processed, we set this flag to True
+        self._done = False
+
+    def set_feature(self, feature: torch.Tensor) -> None:
+        assert feature.dim() == 2, feature.dim()
+        self.num_frames = feature.size(0)
+        # tail padding
+        self.feature = torch.nn.functional.pad(
+            feature,
+            (0, 0, 0, self.pad_length),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+
+    def set_ground_truth(self, ground_truth: str) -> None:
+        self.ground_truth = ground_truth
+
+    def set_states(
+        self, states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]]
+    ) -> None:
+        """Set states."""
+        self.states = states
+
+    def get_feature_chunk(self) -> torch.Tensor:
+        """Get a chunk of feature frames.
+
+        Returns:
+          A tensor of shape (ret_length, feature_dim).
+        """
+        update_length = min(
+            self.num_frames - self.num_processed_frames, self.chunk_length
+        )
+        ret_length = update_length + self.pad_length
+
+        ret_feature = self.feature[
+            self.num_processed_frames : self.num_processed_frames + ret_length
+        ]
+        # Cut off used frames.
+        # self.feature = self.feature[update_length:]
+
+        self.num_processed_frames += update_length
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_feature
+
+    @property
+    def done(self) -> bool:
+        """Return True if all feature frames are processed."""
+        return self._done
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.decoding_method == "greedy_search":
+            return self.hyp[self.context_size :]
+        elif self.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.context_size :]
+        else:
+            assert self.decoding_method == "fast_beam_search"
+            return self.hyp
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py
new file mode 100755
index 000000000..e5a7c7116
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py
@@ -0,0 +1,963 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./conv_emformer_transducer_stateless/streaming_decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method greedy_search \
+      --use-averaged-model True
+
+(2) modified beam search
+./conv_emformer_transducer_stateless/streaming_decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method modified_beam_search \
+      --use-averaged-model True \
+      --beam-size 4
+
+(3) fast beam search
+./conv_emformer_transducer_stateless/streaming_decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method fast_beam_search \
+      --use-averaged-model True \
+      --beam 4 \
+      --max-contexts 4 \
+      --max-states 8
+"""
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from emformer import LOG_EPSILON, stack_states, unstack_states
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from stream import Stream
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import one_best_decoding
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--sampling-rate",
+        type=float,
+        default=16000,
+        help="Sample rate of the audio",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    T = encoder_out.size(1)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (batch_size, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+    beam: int = 4,
+):
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      beam:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    streams: List[Stream],
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      streams:
+        A list of stream objects.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyps[i]
+
+
+def decode_one_chunk(
+    model: nn.Module,
+    streams: List[Stream],
+    params: AttributeDict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> List[int]:
+    """
+    Args:
+      model:
+        The Transducer model.
+      streams:
+        A list of Stream objects.
+      params:
+        It is returned by :func:`get_params`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+       A list of indexes indicating the finished streams.
+    """
+    device = next(model.parameters()).device
+
+    feature_list = []
+    feature_len_list = []
+    state_list = []
+    num_processed_frames_list = []
+
+    for stream in streams:
+        # We should first get `stream.num_processed_frames`
+        # before calling `stream.get_feature_chunk()`
+        # since `stream.num_processed_frames` would be updated
+        num_processed_frames_list.append(stream.num_processed_frames)
+        feature = stream.get_feature_chunk()
+        feature_len = feature.size(0)
+        feature_list.append(feature)
+        feature_len_list.append(feature_len)
+        state_list.append(stream.states)
+
+    features = pad_sequence(
+        feature_list, batch_first=True, padding_value=LOG_EPSILON
+    ).to(device)
+    feature_lens = torch.tensor(feature_len_list, device=device)
+    num_processed_frames = torch.tensor(num_processed_frames_list, device=device)
+
+    # Make sure it has at least 1 frame after subsampling, first-and-last-frame cutting, and right context cutting  # noqa
+    tail_length = 3 * params.subsampling_factor + params.right_context_length + 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPSILON,
+        )
+
+    # Stack states of all streams
+    states = stack_states(state_list)
+
+    encoder_out, encoder_out_lens, states = model.encoder.infer(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        num_processed_frames=num_processed_frames,
+    )
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+    elif params.decoding_method == "fast_beam_search":
+        # feature_len is needed to get partial results.
+        # The rnnt_decoding_stream for fast_beam_search.
+        fast_beam_search_one_best(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            processed_lens=(num_processed_frames >> 2) + encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    # Update cached states of each stream
+    state_list = unstack_states(states)
+    for i, s in enumerate(state_list):
+        streams[i].states = s
+
+    finished_streams = [i for i, stream in enumerate(streams) if stream.done]
+    return finished_streams
+
+
+def create_streaming_feature_extractor() -> Fbank:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return Fbank(opts)
+
+
+def decode_dataset(
+    cuts: CutSet,
+    model: nn.Module,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+):
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The Transducer model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = next(model.parameters()).device
+
+    log_interval = 300
+
+    fbank = create_streaming_feature_extractor()
+
+    decode_results = []
+    streams = []
+    for num, cut in enumerate(cuts):
+        # Each utterance has a Stream.
+        stream = Stream(
+            params=params,
+            cut_id=cut.id,
+            decoding_graph=decoding_graph,
+            device=device,
+            LOG_EPS=LOG_EPSILON,
+        )
+
+        stream.set_states(model.encoder.init_states(device))
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+        feature = fbank(samples)
+        stream.set_feature(feature)
+        stream.set_ground_truth(cut.supervisions[0].text)
+
+        streams.append(stream)
+
+        while len(streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                model=model,
+                streams=streams,
+                params=params,
+                decoding_graph=decoding_graph,
+            )
+
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        streams[i].id,
+                        streams[i].ground_truth.split(),
+                        sp.decode(streams[i].decoding_result()).split(),
+                    )
+                )
+                del streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    while len(streams) > 0:
+        finished_streams = decode_one_chunk(
+            model=model,
+            streams=streams,
+            params=params,
+            decoding_graph=decoding_graph,
+        )
+
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    streams[i].id,
+                    streams[i].ground_truth.split(),
+                    sp.decode(streams[i].decoding_result()).split(),
+                )
+            )
+            del streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    else:
+        key = f"beam_size_{params.beam_size}"
+
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=sorted(results))
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-length-{params.chunk_length}"
+    params.suffix += f"-left-context-length-{params.left_context_length}"
+    params.suffix += f"-right-context-length-{params.right_context_length}"
+    params.suffix += f"-memory-size-{params.memory_size}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-streaming-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.device = device
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            model=model,
+            params=params,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220410)
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py
new file mode 100644
index 000000000..8cde6205b
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py
@@ -0,0 +1,194 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from emformer import ConvolutionModule, Emformer, stack_states, unstack_states
+
+
+def test_convolution_module_forward():
+    B, D = 2, 256
+    chunk_length = 4
+    right_context_length = 2
+    num_chunks = 3
+    U = num_chunks * chunk_length
+    R = num_chunks * right_context_length
+    kernel_size = 31
+    conv_module = ConvolutionModule(
+        chunk_length,
+        right_context_length,
+        D,
+        kernel_size,
+    )
+
+    utterance = torch.randn(U, B, D)
+    right_context = torch.randn(R, B, D)
+
+    utterance, right_context = conv_module(utterance, right_context)
+    assert utterance.shape == (U, B, D), utterance.shape
+    assert right_context.shape == (R, B, D), right_context.shape
+
+
+def test_convolution_module_infer():
+    from emformer import ConvolutionModule
+
+    B, D = 2, 256
+    chunk_length = 4
+    right_context_length = 2
+    num_chunks = 1
+    U = num_chunks * chunk_length
+    R = num_chunks * right_context_length
+    kernel_size = 31
+    conv_module = ConvolutionModule(
+        chunk_length,
+        right_context_length,
+        D,
+        kernel_size,
+    )
+
+    utterance = torch.randn(U, B, D)
+    right_context = torch.randn(R, B, D)
+    cache = torch.randn(B, D, kernel_size - 1)
+
+    utterance, right_context, new_cache = conv_module.infer(
+        utterance, right_context, cache
+    )
+    assert utterance.shape == (U, B, D), utterance.shape
+    assert right_context.shape == (R, B, D), right_context.shape
+    assert new_cache.shape == (B, D, kernel_size - 1), new_cache.shape
+
+
+def test_state_stack_unstack():
+    num_features = 80
+    chunk_length = 32
+    encoder_dim = 512
+    num_encoder_layers = 2
+    kernel_size = 31
+    left_context_length = 32
+    right_context_length = 8
+    memory_size = 32
+
+    model = Emformer(
+        num_features=num_features,
+        chunk_length=chunk_length,
+        subsampling_factor=4,
+        d_model=encoder_dim,
+        num_encoder_layers=num_encoder_layers,
+        cnn_module_kernel=kernel_size,
+        left_context_length=left_context_length,
+        right_context_length=right_context_length,
+        memory_size=memory_size,
+    )
+
+    for batch_size in [1, 2]:
+        attn_caches = [
+            [
+                torch.zeros(memory_size, batch_size, encoder_dim),
+                torch.zeros(left_context_length // 4, batch_size, encoder_dim),
+                torch.zeros(
+                    left_context_length // 4,
+                    batch_size,
+                    encoder_dim,
+                ),
+            ]
+            for _ in range(num_encoder_layers)
+        ]
+        conv_caches = [
+            torch.zeros(batch_size, encoder_dim, kernel_size - 1)
+            for _ in range(num_encoder_layers)
+        ]
+        states = [attn_caches, conv_caches]
+        x = torch.randn(batch_size, 23, num_features)
+        x_lens = torch.full((batch_size,), 23)
+        num_processed_frames = torch.full((batch_size,), 0)
+        y, y_lens, states = model.infer(
+            x, x_lens, num_processed_frames=num_processed_frames, states=states
+        )
+
+        state_list = unstack_states(states)
+        states2 = stack_states(state_list)
+
+        for ss, ss2 in zip(states[0], states2[0]):
+            for s, s2 in zip(ss, ss2):
+                assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+        for s, s2 in zip(states[1], states2[1]):
+            assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+
+def test_torchscript_consistency_infer():
+    r"""Verify that scripting Emformer does not change the behavior of method `infer`."""  # noqa
+    num_features = 80
+    chunk_length = 32
+    encoder_dim = 512
+    num_encoder_layers = 2
+    kernel_size = 31
+    left_context_length = 32
+    right_context_length = 8
+    memory_size = 32
+    batch_size = 2
+
+    model = Emformer(
+        num_features=num_features,
+        chunk_length=chunk_length,
+        subsampling_factor=4,
+        d_model=encoder_dim,
+        num_encoder_layers=num_encoder_layers,
+        cnn_module_kernel=kernel_size,
+        left_context_length=left_context_length,
+        right_context_length=right_context_length,
+        memory_size=memory_size,
+    ).eval()
+    attn_caches = [
+        [
+            torch.zeros(memory_size, batch_size, encoder_dim),
+            torch.zeros(left_context_length // 4, batch_size, encoder_dim),
+            torch.zeros(
+                left_context_length // 4,
+                batch_size,
+                encoder_dim,
+            ),
+        ]
+        for _ in range(num_encoder_layers)
+    ]
+    conv_caches = [
+        torch.zeros(batch_size, encoder_dim, kernel_size - 1)
+        for _ in range(num_encoder_layers)
+    ]
+    states = [attn_caches, conv_caches]
+    x = torch.randn(batch_size, 23, num_features)
+    x_lens = torch.full((batch_size,), 23)
+    num_processed_frames = torch.full((batch_size,), 0)
+    y, y_lens, out_states = model.infer(
+        x, x_lens, num_processed_frames=num_processed_frames, states=states
+    )
+
+    sc_model = torch.jit.script(model).eval()
+    sc_y, sc_y_lens, sc_out_states = sc_model.infer(
+        x, x_lens, num_processed_frames=num_processed_frames, states=states
+    )
+
+    assert torch.allclose(y, sc_y)
+
+
+if __name__ == "__main__":
+    test_convolution_module_forward()
+    test_convolution_module_infer()
+    test_state_stack_unstack()
+    test_torchscript_consistency_infer()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
new file mode 100755
index 000000000..ca21bd6bf
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
@@ -0,0 +1,1116 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conv_emformer_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 280 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+# For mix precision training:
+./conv_emformer_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir conv_emformer_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from emformer import Emformer
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dim for the Emformer",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads for the Emformer",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feed-forward dimension for the Emformer",
+    )
+
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of encoder layers for the Emformer",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=int,
+        default=31,
+        help="Kernel size for the convolution module.",
+    )
+
+    parser.add_argument(
+        "--left-context-length",
+        type=int,
+        default=32,
+        help="""Number of frames before subsampling for left context
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--chunk-length",
+        type=int,
+        default=32,
+        help="""Number of frames before subsampling for each chunk
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--right-context-length",
+        type=int,
+        default=8,
+        help="""Number of frames before subsampling for right context
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--memory-size",
+        type=int,
+        default=0,
+        help="Number of entries in the memory for the Emformer",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for Emformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Emformer(
+        num_features=params.feature_dim,
+        chunk_length=params.chunk_length,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        cnn_module_kernel=params.cnn_module_kernel,
+        left_context_length=params.left_context_length,
+        right_context_length=params.right_context_length,
+        memory_size=params.memory_size,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py
new file mode 120000
index 000000000..104eeea5d
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py
new file mode 100755
index 000000000..d022d463e
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py
@@ -0,0 +1,644 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./conv_emformer_transducer_stateless2/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless2/exp \
+      --max-duration 300 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method greedy_search \
+      --use-averaged-model True
+
+(2) modified beam search
+./conv_emformer_transducer_stateless2/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless2/exp \
+      --max-duration 300 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method modified_beam_search \
+      --use-averaged-model True \
+      --beam-size 4
+
+(3) fast beam search
+./conv_emformer_transducer_stateless2/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless2/exp \
+      --max-duration 300 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method fast_beam_search \
+      --use-averaged-model True \
+      --beam 4 \
+      --max-contexts 4 \
+      --max-states 8
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += params.chunk_length
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.chunk_length),
+        value=LOG_EPS,
+    )
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0
+            start = params.epoch - params.avg
+            assert start >= 1
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..1db262df7
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/do_not_use_it_directly.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/do_not_use_it_directly.py
new file mode 100755
index 000000000..d614f0914
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/do_not_use_it_directly.py
@@ -0,0 +1,1120 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conv_emformer_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 280 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+# For mix precision training:
+./conv_emformer_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from emformer2 import Emformer
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dim for the Emformer",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads for the Emformer",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feed-forward dimension for the Emformer",
+    )
+
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of encoder layers for the Emformer",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=int,
+        default=31,
+        help="Kernel size for the convolution module.",
+    )
+
+    parser.add_argument(
+        "--left-context-length",
+        type=int,
+        default=32,
+        help="""Number of frames before subsampling for left context
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--chunk-length",
+        type=int,
+        default=32,
+        help="""Number of frames before subsampling for each chunk
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--right-context-length",
+        type=int,
+        default=8,
+        help="""Number of frames before subsampling for right context
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--memory-size",
+        type=int,
+        default=0,
+        help="Number of entries in the memory for the Emformer",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for Emformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "is_pnnx": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Emformer(
+        num_features=params.feature_dim,
+        chunk_length=params.chunk_length,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        cnn_module_kernel=params.cnn_module_kernel,
+        left_context_length=params.left_context_length,
+        right_context_length=params.right_context_length,
+        memory_size=params.memory_size,
+        is_pnnx=params.is_pnnx,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py
new file mode 100644
index 000000000..d9ef5a2da
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py
@@ -0,0 +1,1786 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# It is modified based on
+# 1) https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py  # noqa
+# 2) https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py  # noqa
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+from icefall.utils import make_pad_mask
+
+LOG_EPSILON = math.log(1e-10)
+
+
+def unstack_states(
+    states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]]
+) -> List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]:
+    """Unstack the emformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Args:
+      states:
+        A tuple of 2 elements.
+        ``states[0]`` is the attention caches of a batch of utterance.
+        ``states[1]`` is the convolution caches of a batch of utterance.
+        ``len(states[0])`` and ``len(states[1])`` both eqaul to number of layers.  # noqa
+
+    Returns:
+      A list of states.
+      ``states[i]`` is a tuple of 2 elements of i-th utterance.
+      ``states[i][0]`` is the attention caches of i-th utterance.
+      ``states[i][1]`` is the convolution caches of i-th utterance.
+      ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers.  # noqa
+    """
+
+    attn_caches, conv_caches = states
+    batch_size = conv_caches[0].size(0)
+    num_layers = len(attn_caches)
+
+    list_attn_caches = [None] * batch_size
+    for i in range(batch_size):
+        list_attn_caches[i] = [[] for _ in range(num_layers)]
+    for li, layer in enumerate(attn_caches):
+        for s in layer:
+            s_list = s.unbind(dim=1)
+            for bi, b in enumerate(list_attn_caches):
+                b[li].append(s_list[bi])
+
+    list_conv_caches = [None] * batch_size
+    for i in range(batch_size):
+        list_conv_caches[i] = [None] * num_layers
+    for li, layer in enumerate(conv_caches):
+        c_list = layer.unbind(dim=0)
+        for bi, b in enumerate(list_conv_caches):
+            b[li] = c_list[bi]
+
+    ans = [None] * batch_size
+    for i in range(batch_size):
+        ans[i] = [list_attn_caches[i], list_conv_caches[i]]
+
+    return ans
+
+
+def stack_states(
+    state_list: List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]
+) -> Tuple[List[List[torch.Tensor]], List[torch.Tensor]]:
+    """Stack list of emformer states that correspond to separate utterances
+    into a single emformer state so that it can be used as an input for
+    emformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the emformer model for a single utterance.
+        ``states[i]`` is a tuple of 2 elements of i-th utterance.
+        ``states[i][0]`` is the attention caches of i-th utterance.
+        ``states[i][1]`` is the convolution caches of i-th utterance.
+        ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers.  # noqa
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+
+    attn_caches = []
+    for layer in state_list[0][0]:
+        if batch_size > 1:
+            # Note: We will stack attn_caches[layer][s][] later to get attn_caches[layer][s]  # noqa
+            attn_caches.append([[s] for s in layer])
+        else:
+            attn_caches.append([s.unsqueeze(1) for s in layer])
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states[0]):
+            for si, s in enumerate(layer):
+                attn_caches[li][si].append(s)
+                if b == batch_size - 1:
+                    attn_caches[li][si] = torch.stack(attn_caches[li][si], dim=1)
+
+    conv_caches = []
+    for layer in state_list[0][1]:
+        if batch_size > 1:
+            # Note: We will stack conv_caches[layer][] later to get conv_caches[layer]  # noqa
+            conv_caches.append([layer])
+        else:
+            conv_caches.append(layer.unsqueeze(0))
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states[1]):
+            conv_caches[li].append(layer)
+            if b == batch_size - 1:
+                conv_caches[li] = torch.stack(conv_caches[li], dim=0)
+
+    return [attn_caches, conv_caches]
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule.
+
+    Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa
+
+    Args:
+      chunk_length (int):
+        Length of each chunk.
+      right_context_length (int):
+        Length of right context.
+      channels (int):
+        The number of input channels and output channels of conv layers.
+      kernel_size (int):
+        Kernerl size of conv layers.
+      bias (bool):
+        Whether to use bias in conv layers (default=True).
+    """
+
+    def __init__(
+        self,
+        chunk_length: int,
+        right_context_length: int,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super().__init__()
+        # kernerl_size should be an odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0, kernel_size
+
+        self.chunk_length = chunk_length
+        self.right_context_length = right_context_length
+        self.channels = channels
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        # After pointwise_conv1 we put x through a gated linear unit
+        # (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in
+        # the range 1 to 4, but sometimes, for some reason, for layer 0 the rms
+        # ends up being very large, between 50 and 100 for different channels.
+        # This will cause very peaky and sparse derivatives for the sigmoid
+        # gating function, which will tend to make the loss function not learn
+        # effectively.  (for most layers the average absolute values are in the
+        # range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for
+        # different layers, which likely breaks down as 0.5 for the "linear"
+        # half and 0.2 to 0.3 for the part that goes into the sigmoid.
+        # The idea is that if we constrain the rms values to a reasonable range
+        # via a constraint of max_abs=10.0, it will be in a better position to
+        # start learning something, i.e. to latch onto the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        # make it causal by padding cached (kernel_size - 1) frames on the left
+        self.cache_size = kernel_size - 1
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def _split_right_context(
+        self,
+        pad_utterance: torch.Tensor,
+        right_context: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          pad_utterance:
+            Its shape is (cache_size + U, B, D).
+          right_context:
+            Its shape is (R, B, D).
+
+        Returns:
+          Right context segments padding with corresponding context.
+          Its shape is (num_segs * B, D, cache_size + right_context_length).
+        """
+        U_, B, D = pad_utterance.size()
+        R = right_context.size(0)
+        assert self.right_context_length != 0
+        assert R % self.right_context_length == 0
+        num_chunks = R // self.right_context_length
+        right_context = right_context.reshape(
+            num_chunks, self.right_context_length, B, D
+        )
+        right_context = right_context.permute(0, 2, 1, 3).reshape(
+            num_chunks * B, self.right_context_length, D
+        )
+
+        intervals = torch.arange(
+            0, self.chunk_length * (num_chunks - 1), self.chunk_length
+        )
+        first = torch.arange(self.chunk_length, self.chunk_length + self.cache_size)
+        indexes = intervals.unsqueeze(1) + first.unsqueeze(0)
+        indexes = torch.cat(
+            [indexes, torch.arange(U_ - self.cache_size, U_).unsqueeze(0)]
+        )
+        padding = pad_utterance[indexes]  # (num_chunks, cache_size, B, D)
+        padding = padding.permute(0, 2, 1, 3).reshape(
+            num_chunks * B, self.cache_size, D
+        )
+
+        pad_right_context = torch.cat([padding, right_context], dim=1)
+        # (num_chunks * B, cache_size + right_context_length, D)
+        return pad_right_context.permute(0, 2, 1)
+
+    def _merge_right_context(self, right_context: torch.Tensor, B: int) -> torch.Tensor:
+        """
+        Args:
+          right_context:
+            Right context segments.
+            It shape is (num_segs * B, D, right_context_length).
+          B:
+            Batch size.
+
+        Returns:
+          A tensor of shape (B, D, R), where
+          R = num_segs * right_context_length.
+        """
+        right_context = right_context.reshape(
+            -1, B, self.channels, self.right_context_length
+        )
+        right_context = right_context.permute(1, 2, 0, 3)
+        right_context = right_context.reshape(B, self.channels, -1)
+        return right_context
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Causal convolution module.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance tensor of shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context tensor of shape (R, B, D).
+
+        Returns:
+          A tuple of 2 tensors:
+          - output utterance of shape (U, B, D).
+          - output right_context of shape (R, B, D).
+        """
+        U, B, D = utterance.size()
+        R, _, _ = right_context.size()
+
+        # point-wise conv and GLU mechanism
+        x = torch.cat([right_context, utterance], dim=0)  # (R + U, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, R + U)
+        x = self.pointwise_conv1(x)  # (B, 2 * D, R + U)
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (B, D, R + U)
+        utterance = x[:, :, R:]  # (B, D, U)
+        right_context = x[:, :, :R]  # (B, D, R)
+
+        # make causal convolution
+        cache = torch.zeros(B, D, self.cache_size, device=x.device, dtype=x.dtype)
+        pad_utterance = torch.cat([cache, utterance], dim=2)  # (B, D, cache + U)
+
+        # depth-wise conv on utterance
+        utterance = self.depthwise_conv(pad_utterance)  # (B, D, U)
+
+        if self.right_context_length > 0:
+            # depth-wise conv on right_context
+            pad_right_context = self._split_right_context(
+                pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1)
+            )  # (num_segs * B, D, cache_size + right_context_length)
+            right_context = self.depthwise_conv(
+                pad_right_context
+            )  # (num_segs * B, D, right_context_length)
+            right_context = self._merge_right_context(right_context, B)  # (B, D, R)
+
+        x = torch.cat([right_context, utterance], dim=2)  # (B, D, R + U)
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        # point-wise conv
+        x = self.pointwise_conv2(x)  # (B, D, R + U)
+
+        right_context = x[:, :, :R]  # (B, D, R)
+        utterance = x[:, :, R:]  # (B, D, U)
+        return (
+            utterance.permute(2, 0, 1),
+            right_context.permute(2, 0, 1),
+        )
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        cache: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Causal convolution module applied on both utterance and right_context.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance tensor of shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context tensor of shape (R, B, D).
+          cache (torch.Tensor, optional):
+            Cached tensor for left padding of shape (B, D, cache_size).
+
+        Returns:
+          A tuple of 3 tensors:
+            - output utterance of shape (U, B, D).
+            - output right_context of shape (R, B, D).
+            - updated cache tensor of shape (B, D, cache_size).
+        """
+        U, B, D = utterance.size()
+        R, _, _ = right_context.size()
+
+        # point-wise conv
+        x = torch.cat([utterance, right_context], dim=0)  # (U + R, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, U + R)
+        x = self.pointwise_conv1(x)  # (B, 2 * D, U + R)
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (B, D, U + R)
+
+        # make causal convolution
+        assert cache.shape == (B, D, self.cache_size), cache.shape
+        x = torch.cat([cache, x], dim=2)  # (B, D, cache_size + U + R)
+        # update cache
+        new_cache = x[:, :, -R - self.cache_size : -R]
+
+        # 1-D depth-wise conv
+        x = self.depthwise_conv(x)  # (B, D, U + R)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        # point-wise conv
+        x = self.pointwise_conv2(x)  # (B, D, U + R)
+
+        utterance = x[:, :, :U]  # (B, D, U)
+        right_context = x[:, :, U:]  # (B, D, R)
+        return (
+            utterance.permute(2, 0, 1),
+            right_context.permute(2, 0, 1),
+            new_cache,
+        )
+
+
+class EmformerAttention(nn.Module):
+    r"""Emformer layer attention module.
+
+    Args:
+      embed_dim (int):
+        Embedding dimension.
+      nhead (int):
+        Number of attention heads in each Emformer layer.
+      dropout (float, optional):
+        Dropout probability. (Default: 0.0)
+      tanh_on_mem (bool, optional):
+        If ``True``, applies tanh to memory elements. (Default: ``False``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (Default: -1e8)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        nhead: int,
+        dropout: float = 0.0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        if embed_dim % nhead != 0:
+            raise ValueError(
+                f"embed_dim ({embed_dim}) is not a multiple of nhead ({nhead})."
+            )
+
+        self.embed_dim = embed_dim
+        self.nhead = nhead
+        self.tanh_on_mem = tanh_on_mem
+        self.negative_inf = negative_inf
+        self.head_dim = embed_dim // nhead
+        self.dropout = dropout
+
+        self.emb_to_key_value = ScaledLinear(embed_dim, 2 * embed_dim, bias=True)
+        self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+    def _gen_attention_probs(
+        self,
+        attention_weights: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Given the entire attention weights, mask out unecessary connections
+        and optionally with padding positions, to obtain underlying chunk-wise
+        attention probabilities.
+
+        B: batch size;
+        Q: length of query;
+        KV: length of key and value.
+
+        Args:
+          attention_weights (torch.Tensor):
+            Attention weights computed on the entire concatenated tensor
+            with shape (B * nhead, Q, KV).
+          attention_mask (torch.Tensor):
+            Mask tensor where chunk-wise connections are filled with `False`,
+            and other unnecessary connections are filled with `True`,
+            with shape (Q, KV).
+          padding_mask (torch.Tensor, optional):
+            Mask tensor where the padding positions are fill with `True`,
+            and other positions are filled with `False`, with shapa `(B, KV)`.
+
+        Returns:
+          A tensor of shape (B * nhead, Q, KV).
+        """
+        attention_weights_float = attention_weights.float()
+        attention_weights_float = attention_weights_float.masked_fill(
+            attention_mask.unsqueeze(0), self.negative_inf
+        )
+        if padding_mask is not None:
+            Q = attention_weights.size(1)
+            B = attention_weights.size(0) // self.nhead
+            attention_weights_float = attention_weights_float.view(B, self.nhead, Q, -1)
+            attention_weights_float = attention_weights_float.masked_fill(
+                padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+                self.negative_inf,
+            )
+            attention_weights_float = attention_weights_float.view(
+                B * self.nhead, Q, -1
+            )
+
+        attention_probs = nn.functional.softmax(
+            attention_weights_float, dim=-1
+        ).type_as(attention_weights)
+
+        attention_probs = nn.functional.dropout(
+            attention_probs, p=self.dropout, training=self.training
+        )
+        return attention_probs
+
+    def _forward_impl(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        left_context_key: Optional[torch.Tensor] = None,
+        left_context_val: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Underlying chunk-wise attention implementation."""
+        U, B, _ = utterance.size()
+        R = right_context.size(0)
+        M = memory.size(0)
+        scaling = float(self.head_dim) ** -0.5
+
+        # compute query with [right_context, utterance].
+        query = self.emb_to_query(torch.cat([right_context, utterance]))
+        # compute key and value with [memory, right_context, utterance].
+        key, value = self.emb_to_key_value(
+            torch.cat([memory, right_context, utterance])
+        ).chunk(chunks=2, dim=2)
+
+        if left_context_key is not None and left_context_val is not None:
+            # now compute key and value with
+            #   [memory, right context, left context, uttrance]
+            # this is used in inference mode
+            key = torch.cat([key[: M + R], left_context_key, key[M + R :]])
+            value = torch.cat([value[: M + R], left_context_val, value[M + R :]])
+        Q = query.size(0)
+        # KV = key.size(0)
+
+        reshaped_query, reshaped_key, reshaped_value = [
+            tensor.contiguous().view(-1, B * self.nhead, self.head_dim).transpose(0, 1)
+            for tensor in [query, key, value]
+        ]  # (B * nhead, Q or KV, head_dim)
+        attention_weights = torch.bmm(
+            reshaped_query * scaling, reshaped_key.transpose(1, 2)
+        )  # (B * nhead, Q, KV)
+
+        # compute attention probabilities
+        attention_probs = self._gen_attention_probs(
+            attention_weights, attention_mask, padding_mask
+        )
+
+        # compute attention outputs
+        attention = torch.bmm(attention_probs, reshaped_value)
+        assert attention.shape == (B * self.nhead, Q, self.head_dim)
+        attention = attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim)
+
+        # apply output projection
+        output_right_context_utterance = self.out_proj(attention)
+
+        return output_right_context_utterance, key, value
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # TODO: Modify docs.
+        """Forward pass for training and validation mode.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of the hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors.
+
+        It computes a `big` attention matrix on full utterance and
+        then utilizes a pre-computed mask to simulate chunk-wise attention.
+
+        It concatenates three blocks: hard-copied right contexts,
+        and full utterance, as a `big` block,
+        to compute the query tensor:
+        query = [right_context, utterance],
+        with length Q = R + U.
+        It concatenates the three blocks: memory vectors,
+        hard-copied right contexts, and full utterance as another `big` block,
+        to compute the key and value tensors:
+        key & value = [memory, right_context, utterance],
+        with length KV = M + R + U.
+        Attention scores is computed with above `big` query and key.
+
+        Then the underlying chunk-wise attention is obtained by applying
+        the attention mask. Suppose
+        c_i: chunk at index i;
+        r_i: right context that c_i can use;
+        l_i: left context that c_i can use;
+        m_i: past memory vectors from previous layer that c_i can use;
+        The target chunk-wise attention is:
+        c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key)
+
+        Args:
+          utterance (torch.Tensor):
+            Full utterance frames, with shape (U, B, D).
+          right_context (torch.Tensor):
+            Hard-copied right context frames, with shape (R, B, D),
+            where R = num_chunks * right_context_length
+          memory (torch.Tensor):
+            Memory elements, with shape (M, B, D), where M = num_chunks - 1.
+            It is an empty tensor without using memory.
+          attention_mask (torch.Tensor):
+            Pre-computed attention mask to simulate underlying chunk-wise
+            attention, with shape (Q, KV).
+          padding_mask (torch.Tensor):
+            Padding mask of key tensor, with shape (B, KV).
+
+        Returns:
+          Output of right context and utterance, with shape (R + U, B, D).
+        """
+        output_right_context_utterance, _, _ = self._forward_impl(
+            utterance,
+            right_context,
+            memory,
+            attention_mask,
+            padding_mask=padding_mask,
+        )
+        return output_right_context_utterance
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        left_context_key: torch.Tensor,
+        left_context_val: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward pass for inference.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of right context;
+        U: length of utterance, i.e., current chunk;
+        L: length of cached left context;
+        M: length of cached memory vectors.
+
+        It concatenates the right context and utterance (i.e., current chunk)
+        of current chunk, to compute the query tensor:
+        query = [right_context, utterance],
+        with length Q = R + U.
+        It concatenates the memory vectors, right context, left context, and
+        current chunk, to compute the key and value tensors:
+        key & value = [memory, right_context, left_context, utterance],
+        with length KV = M + R + L + U.
+
+        The chunk-wise attention is:
+        chunk, right context (in query) ->
+          left context, chunk, right context, memory vectors (in key).
+
+        Args:
+          utterance (torch.Tensor):
+            Current chunk frames, with shape (U, B, D), where U = chunk_length.
+          right_context (torch.Tensor):
+            Right context frames, with shape (R, B, D),
+            where R = right_context_length.
+          memory (torch.Tensor):
+            Memory vectors, with shape (M, B, D), or empty tensor.
+          left_context_key (torch,Tensor):
+            Cached attention key of left context from preceding computation,
+            with shape (L, B, D).
+          left_context_val (torch.Tensor):
+            Cached attention value of left context from preceding computation,
+            with shape (L, B, D).
+          padding_mask (torch.Tensor):
+            Padding mask of key tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 4 tensors:
+            - output of right context and utterance, with shape (R + U, B, D).
+            - attention key of left context and utterance, which would be cached
+              for next computation, with shape (L + U, B, D).
+            - attention value of left context and utterance, which would be
+              cached for next computation, with shape (L + U, B, D).
+        """
+        U = utterance.size(0)
+        R = right_context.size(0)
+        L = left_context_key.size(0)
+        M = memory.size(0)
+
+        # query = [right context, utterance]
+        Q = R + U
+        # key, value = [memory, right context, left context, uttrance]
+        KV = M + R + L + U
+        attention_mask = torch.zeros(Q, KV).to(
+            dtype=torch.bool, device=utterance.device
+        )
+
+        output_right_context_utterance, key, value = self._forward_impl(
+            utterance,
+            right_context,
+            memory,
+            attention_mask,
+            padding_mask=padding_mask,
+            left_context_key=left_context_key,
+            left_context_val=left_context_val,
+        )
+        return (
+            output_right_context_utterance,
+            key[M + R :],
+            value[M + R :],
+        )
+
+
+class EmformerEncoderLayer(nn.Module):
+    """Emformer layer that constitutes Emformer.
+
+    Args:
+      d_model (int):
+        Input dimension.
+      nhead (int):
+        Number of attention heads.
+      dim_feedforward (int):
+        Hidden layer dimension of feedforward network.
+      chunk_length (int):
+        Length of each input segment.
+      dropout (float, optional):
+        Dropout probability. (Default: 0.0)
+      layer_dropout (float, optional):
+        Layer dropout probability. (Default: 0.0)
+      cnn_module_kernel (int):
+        Kernel size of convolution module.
+      left_context_length (int, optional):
+        Length of left context. (Default: 0)
+      right_context_length (int, optional):
+        Length of right context. (Default: 0)
+      memory_size (int, optional):
+        Number of memory elements to use. (Default: 0)
+      tanh_on_mem (bool, optional):
+        If ``True``, applies tanh to memory elements. (Default: ``False``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (Default: -1e8)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        chunk_length: int,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        self.attention = EmformerAttention(
+            embed_dim=d_model,
+            nhead=nhead,
+            dropout=dropout,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+        )
+        self.summary_op = nn.AvgPool1d(
+            kernel_size=chunk_length, stride=chunk_length, ceil_mode=True
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(
+            chunk_length,
+            right_context_length,
+            d_model,
+            cnn_module_kernel,
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean
+        # (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.layer_dropout = layer_dropout
+        self.left_context_length = left_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+        self.d_model = d_model
+        self.use_memory = memory_size > 0
+
+    def _update_attn_cache(
+        self,
+        next_key: torch.Tensor,
+        next_val: torch.Tensor,
+        memory: torch.Tensor,
+        attn_cache: List[torch.Tensor],
+    ) -> List[torch.Tensor]:
+        """Update cached attention state:
+        1) output memory of current chunk in the lower layer;
+        2) attention key and value in current chunk's computation, which would
+        be resued in next chunk's computation.
+        """
+        new_memory = torch.cat([attn_cache[0], memory])
+        new_key = torch.cat([attn_cache[1], next_key])
+        new_val = torch.cat([attn_cache[2], next_val])
+        attn_cache[0] = new_memory[new_memory.size(0) - self.memory_size :]
+        attn_cache[1] = new_key[new_key.size(0) - self.left_context_length :]
+        attn_cache[2] = new_val[new_val.size(0) - self.left_context_length :]
+        return attn_cache
+
+    def _apply_conv_module_forward(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+    ) -> torch.Tensor:
+        """Apply convolution module in training and validation mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+        utterance, right_context = self.conv_module(utterance, right_context)
+        right_context_utterance = torch.cat([right_context, utterance])
+        return right_context_utterance
+
+    def _apply_conv_module_infer(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        conv_cache: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply convolution module on utterance in inference mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+        utterance, right_context, conv_cache = self.conv_module.infer(
+            utterance, right_context, conv_cache
+        )
+        right_context_utterance = torch.cat([right_context, utterance])
+        return right_context_utterance, conv_cache
+
+    def _apply_attention_module_forward(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Apply attention module in training and validation mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+
+        if self.use_memory:
+            memory = self.summary_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[
+                :-1, :, :
+            ]
+        else:
+            memory = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+        output_right_context_utterance = self.attention(
+            utterance=utterance,
+            right_context=right_context,
+            memory=memory,
+            attention_mask=attention_mask,
+            padding_mask=padding_mask,
+        )
+
+        return output_right_context_utterance
+
+    def _apply_attention_module_infer(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        attn_cache: List[torch.Tensor],
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Apply attention module in inference mode.
+        1) Unpack cached states including:
+           - memory from previous chunks;
+           - attention key and value of left context from preceding
+             chunk's compuation;
+        2) Apply attention computation;
+        3) Update cached attention states including:
+           - memory of current chunk;
+           - attention key and value in current chunk's computation, which would
+             be resued in next chunk's computation.
+        """
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+
+        pre_memory = attn_cache[0]
+        left_context_key = attn_cache[1]
+        left_context_val = attn_cache[2]
+
+        if self.use_memory:
+            memory = self.summary_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[
+                :1, :, :
+            ]
+        else:
+            memory = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+        (output_right_context_utterance, next_key, next_val,) = self.attention.infer(
+            utterance=utterance,
+            right_context=right_context,
+            memory=pre_memory,
+            left_context_key=left_context_key,
+            left_context_val=left_context_val,
+            padding_mask=padding_mask,
+        )
+        attn_cache = self._update_attn_cache(next_key, next_val, memory, attn_cache)
+        return output_right_context_utterance, attn_cache
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        r"""Forward pass for training and validation mode.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance frames, with shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context frames, with shape (R, B, D).
+          attention_mask (torch.Tensor):
+            Attention mask for underlying attention module,
+            with shape (Q, KV), where Q = R + U, KV = M + R + U.
+          padding_mask (torch.Tensor):
+            Padding mask of ker tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 2 tensors:
+            - output utterance, with shape (U, B, D).
+            - output right context, with shape (R, B, D).
+        """
+        R = right_context.size(0)
+        src = torch.cat([right_context, utterance])
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # emformer attention module
+        src_att = self._apply_attention_module_forward(
+            src, R, attention_mask, padding_mask=padding_mask
+        )
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src_conv = self._apply_conv_module_forward(src, R)
+        src = src + self.dropout(src_conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        output_utterance = src[R:]
+        output_right_context = src[:R]
+        return output_utterance, output_right_context
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        attn_cache: List[torch.Tensor],
+        conv_cache: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor], torch.Tensor]:
+        """Forward pass for inference.
+
+         B: batch size;
+         D: embedding dimension;
+         R: length of right_context;
+         U: length of utterance;
+         M: length of memory.
+
+        Args:
+           utterance (torch.Tensor):
+             Utterance frames, with shape (U, B, D).
+           right_context (torch.Tensor):
+             Right context frames, with shape (R, B, D).
+           attn_cache (List[torch.Tensor]):
+             Cached attention tensors generated in preceding computation,
+             including memory, key and value of left context.
+           conv_cache (torch.Tensor, optional):
+             Cache tensor of left context for causal convolution.
+           padding_mask (torch.Tensor):
+             Padding mask of ker tensor.
+
+         Returns:
+           (Tensor, Tensor, List[torch.Tensor], Tensor):
+             - output utterance, with shape (U, B, D);
+             - output right_context, with shape (R, B, D);
+             - output attention cache;
+             - output convolution cache.
+        """
+        R = right_context.size(0)
+        src = torch.cat([right_context, utterance])
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # emformer attention module
+        src_att, attn_cache = self._apply_attention_module_infer(
+            src, R, attn_cache, padding_mask=padding_mask
+        )
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache)
+        src = src + self.dropout(src_conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        output_utterance = src[R:]
+        output_right_context = src[:R]
+        return (
+            output_utterance,
+            output_right_context,
+            attn_cache,
+            conv_cache,
+        )
+
+
+def _gen_attention_mask_block(
+    col_widths: List[int],
+    col_mask: List[bool],
+    num_rows: int,
+    device: torch.device,
+) -> torch.Tensor:
+    assert len(col_widths) == len(
+        col_mask
+    ), "Length of col_widths must match that of col_mask"
+
+    mask_block = [
+        torch.ones(num_rows, col_width, device=device)
+        if is_ones_col
+        else torch.zeros(num_rows, col_width, device=device)
+        for col_width, is_ones_col in zip(col_widths, col_mask)
+    ]
+    return torch.cat(mask_block, dim=1)
+
+
+class EmformerEncoder(nn.Module):
+    """Implements the Emformer architecture introduced in
+    *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency
+    Streaming Speech Recognition*
+    [:footcite:`shi2021emformer`].
+
+    In this model, the memory bank computation is simplifed, using the averaged
+    value of each chunk as its memory vector.
+
+    Args:
+      d_model (int):
+        Input dimension.
+      nhead (int):
+        Number of attention heads in each emformer layer.
+      dim_feedforward (int):
+        Hidden layer dimension of each emformer layer's feedforward network.
+      num_encoder_layers (int):
+        Number of emformer layers to instantiate.
+      chunk_length (int):
+        Length of each input segment.
+      dropout (float, optional):
+        Dropout probability. (default: 0.0)
+      layer_dropout (float, optional):
+        Layer dropout probability. (default: 0.0)
+      cnn_module_kernel (int):
+        Kernel size of convolution module.
+      left_context_length (int, optional):
+        Length of left context. (default: 0)
+      right_context_length (int, optional):
+        Length of right context. (default: 0)
+      memory_size (int, optional):
+        Number of memory elements to use. (default: 0)
+      tanh_on_mem (bool, optional):
+        If ``true``, applies tanh to memory elements. (default: ``false``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (default: -1e8)
+    """
+
+    def __init__(
+        self,
+        chunk_length: int,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        assert (
+            chunk_length - 1
+        ) & chunk_length == 0, "chunk_length should be a power of 2."
+        self.shift = int(math.log(chunk_length, 2))
+
+        self.use_memory = memory_size > 0
+
+        self.emformer_layers = nn.ModuleList(
+            [
+                EmformerEncoderLayer(
+                    d_model=d_model,
+                    nhead=nhead,
+                    dim_feedforward=dim_feedforward,
+                    chunk_length=chunk_length,
+                    dropout=dropout,
+                    layer_dropout=layer_dropout,
+                    cnn_module_kernel=cnn_module_kernel,
+                    left_context_length=left_context_length,
+                    right_context_length=right_context_length,
+                    memory_size=memory_size,
+                    tanh_on_mem=tanh_on_mem,
+                    negative_inf=negative_inf,
+                )
+                for layer_idx in range(num_encoder_layers)
+            ]
+        )
+
+        self.num_encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.left_context_length = left_context_length
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+        self.cnn_module_kernel = cnn_module_kernel
+
+    def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor:
+        """Hard copy each chunk's right context and concat them."""
+        T = x.shape[0]
+        num_chunks = math.ceil((T - self.right_context_length) / self.chunk_length)
+        # first (num_chunks - 1) right context block
+        intervals = torch.arange(
+            0, self.chunk_length * (num_chunks - 1), self.chunk_length
+        )
+        first = torch.arange(
+            self.chunk_length, self.chunk_length + self.right_context_length
+        )
+        indexes = intervals.unsqueeze(1) + first.unsqueeze(0)
+        # cat last right context block
+        indexes = torch.cat(
+            [
+                indexes,
+                torch.arange(T - self.right_context_length, T).unsqueeze(0),
+            ]
+        )
+        right_context_blocks = x[indexes.reshape(-1)]
+        return right_context_blocks
+
+    def _gen_attention_mask_col_widths(self, chunk_idx: int, U: int) -> List[int]:
+        """Calculate column widths (key, value) in attention mask for the
+        chunk_idx chunk."""
+        num_chunks = math.ceil(U / self.chunk_length)
+        rc = self.right_context_length
+        lc = self.left_context_length
+        rc_start = chunk_idx * rc
+        rc_end = rc_start + rc
+        chunk_start = max(chunk_idx * self.chunk_length - lc, 0)
+        chunk_end = min((chunk_idx + 1) * self.chunk_length, U)
+        R = rc * num_chunks
+
+        if self.use_memory:
+            m_start = max(chunk_idx - self.memory_size, 0)
+            M = num_chunks - 1
+            col_widths = [
+                m_start,  # before memory
+                chunk_idx - m_start,  # memory
+                M - chunk_idx,  # after memory
+                rc_start,  # before right context
+                rc,  # right context
+                R - rc_end,  # after right context
+                chunk_start,  # before chunk
+                chunk_end - chunk_start,  # chunk
+                U - chunk_end,  # after chunk
+            ]
+        else:
+            col_widths = [
+                rc_start,  # before right context
+                rc,  # right context
+                R - rc_end,  # after right context
+                chunk_start,  # before chunk
+                chunk_end - chunk_start,  # chunk
+                U - chunk_end,  # after chunk
+            ]
+
+        return col_widths
+
+    def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor:
+        """Generate attention mask to simulate underlying chunk-wise attention
+        computation, where chunk-wise connections are filled with `False`,
+        and other unnecessary connections beyond chunk are filled with `True`.
+
+        R: length of hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors;
+        Q: length of attention query;
+        KV: length of attention key and value.
+
+        The shape of attention mask is (Q, KV).
+        If self.use_memory is `True`:
+          query = [right_context, utterance];
+          key, value = [memory, right_context, utterance];
+          Q = R + U, KV = M + R + U.
+        Otherwise:
+          query = [right_context, utterance]
+          key, value = [right_context, utterance]
+          Q = R + U, KV = R + U.
+
+        Suppose:
+          c_i: chunk at index i;
+          r_i: right context that c_i can use;
+          l_i: left context that c_i can use;
+          m_i: past memory vectors from previous layer that c_i can use;
+        The target chunk-wise attention is:
+          c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key).
+        """
+        U = utterance.size(0)
+        num_chunks = math.ceil(U / self.chunk_length)
+
+        right_context_mask = []
+        utterance_mask = []
+
+        if self.use_memory:
+            num_cols = 9
+            # right context and utterance both attend to memory, right context,
+            # utterance
+            right_context_utterance_cols_mask = [
+                idx in [1, 4, 7] for idx in range(num_cols)
+            ]
+        else:
+            num_cols = 6
+            # right context and utterance both attend to right context and
+            # utterance
+            right_context_utterance_cols_mask = [
+                idx in [1, 4] for idx in range(num_cols)
+            ]
+        masks_to_concat = [right_context_mask, utterance_mask]
+
+        for chunk_idx in range(num_chunks):
+            col_widths = self._gen_attention_mask_col_widths(chunk_idx, U)
+
+            right_context_mask_block = _gen_attention_mask_block(
+                col_widths,
+                right_context_utterance_cols_mask,
+                self.right_context_length,
+                utterance.device,
+            )
+            right_context_mask.append(right_context_mask_block)
+
+            utterance_mask_block = _gen_attention_mask_block(
+                col_widths,
+                right_context_utterance_cols_mask,
+                min(
+                    self.chunk_length,
+                    U - chunk_idx * self.chunk_length,
+                ),
+                utterance.device,
+            )
+            utterance_mask.append(utterance_mask_block)
+
+        attention_mask = (
+            1 - torch.cat([torch.cat(mask) for mask in masks_to_concat])
+        ).to(torch.bool)
+        return attention_mask
+
+    def forward(
+        self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass for training and validation mode.
+
+        B: batch size;
+        D: input dimension;
+        U: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (U + right_context_length, B, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, which contains the
+            right_context at the end.
+
+        Returns:
+          A tuple of 2 tensors:
+            - output utterance frames, with shape (U, B, D).
+            - output_lengths, with shape (B,), without containing the
+              right_context at the end.
+        """
+        U = x.size(0) - self.right_context_length
+
+        right_context = self._gen_right_context(x)
+        utterance = x[:U]
+        output_lengths = torch.clamp(lengths - self.right_context_length, min=0)
+        attention_mask = self._gen_attention_mask(utterance)
+
+        padding_mask = make_pad_mask(attention_mask.shape[1] - U + output_lengths)
+
+        output = utterance
+        for layer in self.emformer_layers:
+            output, right_context = layer(
+                output,
+                right_context,
+                attention_mask,
+                padding_mask=padding_mask,
+                warmup=warmup,
+            )
+
+        return output, output_lengths
+
+    @torch.jit.export
+    def infer(
+        self,
+        x: torch.Tensor,
+        lengths: torch.Tensor,
+        num_processed_frames: torch.Tensor,
+        states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ]:
+        """Forward pass for streaming inference.
+
+        B: batch size;
+        D: input dimension;
+        U: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (U + right_context_length, B, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, which contains the
+            right_context at the end.
+          states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa
+            Cached states containing:
+            - attn_caches: attention states from preceding chunk's computation,
+              where each element corresponds to each emformer layer
+            - conv_caches: left context for causal convolution, where each
+              element corresponds to each layer.
+
+        Returns:
+          (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]):
+            - output utterance frames, with shape (U, B, D).
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+            - updated states from current chunk's computation.
+        """
+        assert num_processed_frames.shape == (x.size(1),)
+
+        attn_caches = states[0]
+        assert len(attn_caches) == self.num_encoder_layers, len(attn_caches)
+        for i in range(len(attn_caches)):
+            assert attn_caches[i][0].shape == (
+                self.memory_size,
+                x.size(1),
+                self.d_model,
+            ), attn_caches[i][0].shape
+            assert attn_caches[i][1].shape == (
+                self.left_context_length,
+                x.size(1),
+                self.d_model,
+            ), attn_caches[i][1].shape
+            assert attn_caches[i][2].shape == (
+                self.left_context_length,
+                x.size(1),
+                self.d_model,
+            ), attn_caches[i][2].shape
+
+        conv_caches = states[1]
+        assert len(conv_caches) == self.num_encoder_layers, len(conv_caches)
+        for i in range(len(conv_caches)):
+            assert conv_caches[i].shape == (
+                x.size(1),
+                self.d_model,
+                self.cnn_module_kernel - 1,
+            ), conv_caches[i].shape
+
+        right_context = x[-self.right_context_length :]
+        utterance = x[: -self.right_context_length]
+        output_lengths = torch.clamp(lengths - self.right_context_length, min=0)
+
+        # calcualte padding mask to mask out initial zero caches
+        chunk_mask = make_pad_mask(output_lengths).to(x.device)
+        memory_mask = (
+            (
+                (num_processed_frames >> self.shift).view(x.size(1), 1)
+                <= torch.arange(self.memory_size, device=x.device).expand(
+                    x.size(1), self.memory_size
+                )
+            ).flip(1)
+            if self.use_memory
+            else torch.empty(0).to(dtype=torch.bool, device=x.device)
+        )
+        left_context_mask = (
+            num_processed_frames.view(x.size(1), 1)
+            <= torch.arange(self.left_context_length, device=x.device).expand(
+                x.size(1), self.left_context_length
+            )
+        ).flip(1)
+        right_context_mask = torch.zeros(
+            x.size(1),
+            self.right_context_length,
+            dtype=torch.bool,
+            device=x.device,
+        )
+        padding_mask = torch.cat(
+            [memory_mask, right_context_mask, left_context_mask, chunk_mask],
+            dim=1,
+        )
+
+        output = utterance
+        output_attn_caches: List[List[torch.Tensor]] = []
+        output_conv_caches: List[torch.Tensor] = []
+        for layer_idx, layer in enumerate(self.emformer_layers):
+            (
+                output,
+                right_context,
+                output_attn_cache,
+                output_conv_cache,
+            ) = layer.infer(
+                output,
+                right_context,
+                padding_mask=padding_mask,
+                attn_cache=attn_caches[layer_idx],
+                conv_cache=conv_caches[layer_idx],
+            )
+            output_attn_caches.append(output_attn_cache)
+            output_conv_caches.append(output_conv_cache)
+
+        output_states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = (
+            output_attn_caches,
+            output_conv_caches,
+        )
+        return output, output_lengths, output_states
+
+    @torch.jit.export
+    def init_states(self, device: torch.device = torch.device("cpu")):
+        """Create initial states."""
+        attn_caches = [
+            [
+                torch.zeros(self.memory_size, self.d_model, device=device),
+                torch.zeros(self.left_context_length, self.d_model, device=device),
+                torch.zeros(self.left_context_length, self.d_model, device=device),
+            ]
+            for _ in range(self.num_encoder_layers)
+        ]
+        conv_caches = [
+            torch.zeros(self.d_model, self.cnn_module_kernel - 1, device=device)
+            for _ in range(self.num_encoder_layers)
+        ]
+        states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = (
+            attn_caches,
+            conv_caches,
+        )
+        return states
+
+
+class Emformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        chunk_length: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 3,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+    ):
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+        if chunk_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "chunk_length must be a mutiple of subsampling_factor."
+            )
+        if left_context_length != 0 and left_context_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "left_context_length must be 0 or a mutiple of subsampling_factor."  # noqa
+            )
+        if right_context_length != 0 and right_context_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "right_context_length must be 0 or a mutiple of subsampling_factor."  # noqa
+            )
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder = EmformerEncoder(
+            chunk_length=chunk_length // subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+            cnn_module_kernel=cnn_module_kernel,
+            left_context_length=left_context_length // subsampling_factor,
+            right_context_length=right_context_length // subsampling_factor,
+            memory_size=memory_size,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass for training and non-streaming inference.
+
+        B: batch size;
+        D: feature dimension;
+        T: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (B, T, D).
+          x_lens (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, containing the
+            right_context at the end.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          (Tensor, Tensor):
+            - output embedding, with shape (B, T', D), where
+              T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4.
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+        """
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+        assert x.size(0) == x_lens.max().item()
+
+        output, output_lengths = self.encoder(x, x_lens, warmup=warmup)  # (T, N, C)
+
+        output = output.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+
+        return output, output_lengths
+
+    @torch.jit.export
+    def infer(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        num_processed_frames: torch.Tensor,
+        states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        Tuple[List[List[torch.Tensor]], List[torch.Tensor]],
+    ]:
+        """Forward pass for streaming inference.
+
+        B: batch size;
+        D: feature dimension;
+        T: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (B, T, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, containing the
+            right_context at the end.
+          states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa
+            Cached states containing:
+            - past_lens: number of past frames for each sample in batch
+            - attn_caches: attention states from preceding chunk's computation,
+              where each element corresponds to each emformer layer
+            - conv_caches: left context for causal convolution, where each
+              element corresponds to each layer.
+        Returns:
+          (Tensor, Tensor):
+            - output embedding, with shape (B, T', D), where
+              T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4.
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+            - updated states from current chunk's computation.
+        """
+        x = self.encoder_embed(x)
+        # drop the first and last frames
+        x = x[:, 1:-1, :]
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+        x_lens -= 2
+        assert x.size(0) == x_lens.max().item()
+
+        num_processed_frames = num_processed_frames >> 2
+
+        output, output_lengths, output_states = self.encoder.infer(
+            x, x_lens, num_processed_frames, states
+        )
+
+        output = output.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+
+        return output, output_lengths, output_states
+
+    @torch.jit.export
+    def init_states(self, device: torch.device = torch.device("cpu")):
+        """Create initial states."""
+        return self.encoder.init_states(device)
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer2.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer2.py
new file mode 100644
index 000000000..4a844b79f
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer2.py
@@ -0,0 +1,1821 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# It is modified based on
+# 1) https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py  # noqa
+# 2) https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py  # noqa
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+from icefall.utils import make_pad_mask
+
+LOG_EPSILON = math.log(1e-10)
+
+
+def unstack_states(
+    states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]]
+) -> List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]:
+    """Unstack the emformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Args:
+      states:
+        A tuple of 2 elements.
+        ``states[0]`` is the attention caches of a batch of utterance.
+        ``states[1]`` is the convolution caches of a batch of utterance.
+        ``len(states[0])`` and ``len(states[1])`` both eqaul to number of layers.  # noqa
+
+    Returns:
+      A list of states.
+      ``states[i]`` is a tuple of 2 elements of i-th utterance.
+      ``states[i][0]`` is the attention caches of i-th utterance.
+      ``states[i][1]`` is the convolution caches of i-th utterance.
+      ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers.  # noqa
+    """
+
+    attn_caches, conv_caches = states
+    batch_size = conv_caches[0].size(0)
+    num_layers = len(attn_caches)
+
+    list_attn_caches = [None] * batch_size
+    for i in range(batch_size):
+        list_attn_caches[i] = [[] for _ in range(num_layers)]
+    for li, layer in enumerate(attn_caches):
+        for s in layer:
+            s_list = s.unbind(dim=1)
+            for bi, b in enumerate(list_attn_caches):
+                b[li].append(s_list[bi])
+
+    list_conv_caches = [None] * batch_size
+    for i in range(batch_size):
+        list_conv_caches[i] = [None] * num_layers
+    for li, layer in enumerate(conv_caches):
+        c_list = layer.unbind(dim=0)
+        for bi, b in enumerate(list_conv_caches):
+            b[li] = c_list[bi]
+
+    ans = [None] * batch_size
+    for i in range(batch_size):
+        ans[i] = [list_attn_caches[i], list_conv_caches[i]]
+
+    return ans
+
+
+def stack_states(
+    state_list: List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]
+) -> Tuple[List[List[torch.Tensor]], List[torch.Tensor]]:
+    """Stack list of emformer states that correspond to separate utterances
+    into a single emformer state so that it can be used as an input for
+    emformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the emformer model for a single utterance.
+        ``states[i]`` is a tuple of 2 elements of i-th utterance.
+        ``states[i][0]`` is the attention caches of i-th utterance.
+        ``states[i][1]`` is the convolution caches of i-th utterance.
+        ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers.  # noqa
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+
+    attn_caches = []
+    for layer in state_list[0][0]:
+        if batch_size > 1:
+            # Note: We will stack attn_caches[layer][s][] later to get attn_caches[layer][s]  # noqa
+            attn_caches.append([[s] for s in layer])
+        else:
+            attn_caches.append([s.unsqueeze(1) for s in layer])
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states[0]):
+            for si, s in enumerate(layer):
+                attn_caches[li][si].append(s)
+                if b == batch_size - 1:
+                    attn_caches[li][si] = torch.stack(attn_caches[li][si], dim=1)
+
+    conv_caches = []
+    for layer in state_list[0][1]:
+        if batch_size > 1:
+            # Note: We will stack conv_caches[layer][] later to get conv_caches[layer]  # noqa
+            conv_caches.append([layer])
+        else:
+            conv_caches.append(layer.unsqueeze(0))
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states[1]):
+            conv_caches[li].append(layer)
+            if b == batch_size - 1:
+                conv_caches[li] = torch.stack(conv_caches[li], dim=0)
+
+    return [attn_caches, conv_caches]
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule.
+
+    Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa
+
+    Args:
+      chunk_length (int):
+        Length of each chunk.
+      right_context_length (int):
+        Length of right context.
+      channels (int):
+        The number of input channels and output channels of conv layers.
+      kernel_size (int):
+        Kernerl size of conv layers.
+      bias (bool):
+        Whether to use bias in conv layers (default=True).
+    """
+
+    def __init__(
+        self,
+        chunk_length: int,
+        right_context_length: int,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        is_pnnx: bool = True,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super().__init__()
+        self.is_pnnx = is_pnnx
+        # kernerl_size should be an odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0, kernel_size
+
+        self.chunk_length = chunk_length
+        self.right_context_length = right_context_length
+        self.channels = channels
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        # After pointwise_conv1 we put x through a gated linear unit
+        # (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in
+        # the range 1 to 4, but sometimes, for some reason, for layer 0 the rms
+        # ends up being very large, between 50 and 100 for different channels.
+        # This will cause very peaky and sparse derivatives for the sigmoid
+        # gating function, which will tend to make the loss function not learn
+        # effectively.  (for most layers the average absolute values are in the
+        # range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for
+        # different layers, which likely breaks down as 0.5 for the "linear"
+        # half and 0.2 to 0.3 for the part that goes into the sigmoid.
+        # The idea is that if we constrain the rms values to a reasonable range
+        # via a constraint of max_abs=10.0, it will be in a better position to
+        # start learning something, i.e. to latch onto the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        # make it causal by padding cached (kernel_size - 1) frames on the left
+        self.cache_size = kernel_size - 1
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def _split_right_context(
+        self,
+        pad_utterance: torch.Tensor,
+        right_context: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          pad_utterance:
+            Its shape is (cache_size + U, B, D).
+          right_context:
+            Its shape is (R, B, D).
+
+        Returns:
+          Right context segments padding with corresponding context.
+          Its shape is (num_segs * B, D, cache_size + right_context_length).
+        """
+        U_, B, D = pad_utterance.size()
+        R = right_context.size(0)
+        assert self.right_context_length != 0
+        assert R % self.right_context_length == 0
+        num_chunks = R // self.right_context_length
+        right_context = right_context.reshape(
+            num_chunks, self.right_context_length, B, D
+        )
+        right_context = right_context.permute(0, 2, 1, 3).reshape(
+            num_chunks * B, self.right_context_length, D
+        )
+
+        intervals = torch.arange(
+            0, self.chunk_length * (num_chunks - 1), self.chunk_length
+        )
+        first = torch.arange(self.chunk_length, self.chunk_length + self.cache_size)
+        indexes = intervals.unsqueeze(1) + first.unsqueeze(0)
+        indexes = torch.cat(
+            [indexes, torch.arange(U_ - self.cache_size, U_).unsqueeze(0)]
+        )
+        padding = pad_utterance[indexes]  # (num_chunks, cache_size, B, D)
+        padding = padding.permute(0, 2, 1, 3).reshape(
+            num_chunks * B, self.cache_size, D
+        )
+
+        pad_right_context = torch.cat([padding, right_context], dim=1)
+        # (num_chunks * B, cache_size + right_context_length, D)
+        return pad_right_context.permute(0, 2, 1)
+
+    def _merge_right_context(self, right_context: torch.Tensor, B: int) -> torch.Tensor:
+        """
+        Args:
+          right_context:
+            Right context segments.
+            It shape is (num_segs * B, D, right_context_length).
+          B:
+            Batch size.
+
+        Returns:
+          A tensor of shape (B, D, R), where
+          R = num_segs * right_context_length.
+        """
+        right_context = right_context.reshape(
+            -1, B, self.channels, self.right_context_length
+        )
+        right_context = right_context.permute(1, 2, 0, 3)
+        right_context = right_context.reshape(B, self.channels, -1)
+        return right_context
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Causal convolution module.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance tensor of shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context tensor of shape (R, B, D).
+
+        Returns:
+          A tuple of 2 tensors:
+          - output utterance of shape (U, B, D).
+          - output right_context of shape (R, B, D).
+        """
+        U, B, D = utterance.size()
+        R, _, _ = right_context.size()
+
+        # point-wise conv and GLU mechanism
+        x = torch.cat([right_context, utterance], dim=0)  # (R + U, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, R + U)
+        x = self.pointwise_conv1(x)  # (B, 2 * D, R + U)
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (B, D, R + U)
+        utterance = x[:, :, R:]  # (B, D, U)
+        right_context = x[:, :, :R]  # (B, D, R)
+
+        # make causal convolution
+        cache = torch.zeros(B, D, self.cache_size, device=x.device, dtype=x.dtype)
+        pad_utterance = torch.cat([cache, utterance], dim=2)  # (B, D, cache + U)
+
+        # depth-wise conv on utterance
+        utterance = self.depthwise_conv(pad_utterance)  # (B, D, U)
+
+        if self.right_context_length > 0:
+            # depth-wise conv on right_context
+            pad_right_context = self._split_right_context(
+                pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1)
+            )  # (num_segs * B, D, cache_size + right_context_length)
+            right_context = self.depthwise_conv(
+                pad_right_context
+            )  # (num_segs * B, D, right_context_length)
+            right_context = self._merge_right_context(right_context, B)  # (B, D, R)
+
+        x = torch.cat([right_context, utterance], dim=2)  # (B, D, R + U)
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        # point-wise conv
+        x = self.pointwise_conv2(x)  # (B, D, R + U)
+
+        right_context = x[:, :, :R]  # (B, D, R)
+        utterance = x[:, :, R:]  # (B, D, U)
+        return (
+            utterance.permute(2, 0, 1),
+            right_context.permute(2, 0, 1),
+        )
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        cache: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Causal convolution module applied on both utterance and right_context.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance tensor of shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context tensor of shape (R, B, D).
+          cache (torch.Tensor, optional):
+            Cached tensor for left padding of shape (B, D, cache_size).
+
+        Returns:
+          A tuple of 3 tensors:
+            - output utterance of shape (U, B, D).
+            - output right_context of shape (R, B, D).
+            - updated cache tensor of shape (B, D, cache_size).
+        """
+        if self.is_pnnx is False:
+            U, B, D = utterance.size()
+            R, _, _ = right_context.size()
+        else:
+            U = self.chunk_length
+            B = 1
+            D = self.channels
+            R = self.right_context_length
+
+        # point-wise conv
+        x = torch.cat([utterance, right_context], dim=0)  # (U + R, B, D)
+        x = x.permute(1, 2, 0)  # (B, D, U + R)
+        x = self.pointwise_conv1(x)  # (B, 2 * D, U + R)
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (B, D, U + R)
+
+        # make causal convolution
+        assert cache.shape == (B, D, self.cache_size), cache.shape
+        x = torch.cat([cache, x], dim=2)  # (B, D, cache_size + U + R)
+        # update cache
+        new_cache = x[:, :, -R - self.cache_size : -R]
+
+        # 1-D depth-wise conv
+        x = self.depthwise_conv(x)  # (B, D, U + R)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        # point-wise conv
+        x = self.pointwise_conv2(x)  # (B, D, U + R)
+
+        utterance = x[:, :, :U]  # (B, D, U)
+        right_context = x[:, :, U:]  # (B, D, R)
+        return (
+            utterance.permute(2, 0, 1),
+            right_context.permute(2, 0, 1),
+            new_cache,
+        )
+
+
+class EmformerAttention(nn.Module):
+    r"""Emformer layer attention module.
+
+    Args:
+      embed_dim (int):
+        Embedding dimension.
+      nhead (int):
+        Number of attention heads in each Emformer layer.
+      dropout (float, optional):
+        Dropout probability. (Default: 0.0)
+      tanh_on_mem (bool, optional):
+        If ``True``, applies tanh to memory elements. (Default: ``False``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (Default: -1e8)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        nhead: int,
+        left_context_length: int,
+        chunk_length: int,
+        right_context_length: int,
+        memory_size: int,
+        dropout: float = 0.0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+        is_pnnx: bool = True,
+    ):
+        super().__init__()
+        self.is_pnnx = is_pnnx
+
+        if embed_dim % nhead != 0:
+            raise ValueError(
+                f"embed_dim ({embed_dim}) is not a multiple of nhead ({nhead})."
+            )
+
+        self.embed_dim = embed_dim
+        self.nhead = nhead
+        self.tanh_on_mem = tanh_on_mem
+        self.negative_inf = negative_inf
+        self.head_dim = embed_dim // nhead
+        self.dropout = dropout
+
+        self.left_context_length = left_context_length
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+
+        self.emb_to_key_value = ScaledLinear(embed_dim, 2 * embed_dim, bias=True)
+        self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+    def _gen_attention_probs(
+        self,
+        attention_weights: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Given the entire attention weights, mask out unecessary connections
+        and optionally with padding positions, to obtain underlying chunk-wise
+        attention probabilities.
+
+        B: batch size;
+        Q: length of query;
+        KV: length of key and value.
+
+        Args:
+          attention_weights (torch.Tensor):
+            Attention weights computed on the entire concatenated tensor
+            with shape (B * nhead, Q, KV).
+          attention_mask (torch.Tensor):
+            Mask tensor where chunk-wise connections are filled with `False`,
+            and other unnecessary connections are filled with `True`,
+            with shape (Q, KV).
+          padding_mask (torch.Tensor, optional):
+            Mask tensor where the padding positions are fill with `True`,
+            and other positions are filled with `False`, with shapa `(B, KV)`.
+
+        Returns:
+          A tensor of shape (B * nhead, Q, KV).
+        """
+        attention_weights_float = attention_weights.float()
+        attention_weights_float = attention_weights_float.masked_fill(
+            attention_mask.unsqueeze(0), self.negative_inf
+        )
+        if padding_mask is not None:
+            Q = attention_weights.size(1)
+            B = attention_weights.size(0) // self.nhead
+            attention_weights_float = attention_weights_float.view(B, self.nhead, Q, -1)
+            attention_weights_float = attention_weights_float.masked_fill(
+                padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+                self.negative_inf,
+            )
+            attention_weights_float = attention_weights_float.view(
+                B * self.nhead, Q, -1
+            )
+
+        attention_probs = nn.functional.softmax(
+            attention_weights_float, dim=-1
+        ).type_as(attention_weights)
+
+        attention_probs = nn.functional.dropout(
+            attention_probs, p=self.dropout, training=self.training
+        )
+        return attention_probs
+
+    def _forward_impl(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        left_context_key: Optional[torch.Tensor] = None,
+        left_context_val: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Underlying chunk-wise attention implementation."""
+        if self.is_pnnx is False:
+            U, B, _ = utterance.size()
+            R = right_context.size(0)
+            M = memory.size(0)
+        else:
+            U = self.chunk_length
+            B = 1
+            R = self.right_context_length
+            M = self.memory_size
+        L = self.left_context_length
+
+        scaling = float(self.head_dim) ** -0.5
+
+        # compute query with [right_context, utterance].
+        query = self.emb_to_query(torch.cat([right_context, utterance]))
+        # compute key and value with [memory, right_context, utterance].
+        key, value = self.emb_to_key_value(
+            torch.cat([memory, right_context, utterance])
+        ).chunk(chunks=2, dim=2)
+
+        if left_context_key is not None and left_context_val is not None:
+            # now compute key and value with
+            #   [memory, right context, left context, uttrance]
+            # this is used in inference mode
+            key = torch.cat([key[: M + R], left_context_key, key[M + R :]])
+            value = torch.cat([value[: M + R], left_context_val, value[M + R :]])
+
+        #  Q = query.size(0)
+        Q = U + R
+
+        # KV = key.size(0)
+
+        if self.is_pnnx is True:
+            reshaped_query = query.view(Q, self.nhead, self.head_dim).permute(1, 0, 2)
+            reshaped_key = key.view(M + R + U + L, self.nhead, self.head_dim).permute(
+                1, 0, 2
+            )
+            reshaped_value = value.view(
+                M + R + U + L, self.nhead, self.head_dim
+            ).permute(1, 0, 2)
+        else:
+            reshaped_query, reshaped_key, reshaped_value = [
+                tensor.contiguous()
+                .view(-1, B * self.nhead, self.head_dim)
+                .transpose(0, 1)
+                for tensor in [query, key, value]
+            ]  # (B * nhead, Q or KV, head_dim)
+        if self.is_pnnx is True:
+            attention_weights = torch.bmm(
+                reshaped_query * scaling, reshaped_key.permute(0, 2, 1)
+            )  # (B * nhead, Q, KV)
+        else:
+            attention_weights = torch.bmm(
+                reshaped_query * scaling, reshaped_key.transpose(1, 2)
+            )  # (B * nhead, Q, KV)
+
+        # compute attention probabilities
+        if False:
+            attention_probs = self._gen_attention_probs(
+                attention_weights, attention_mask, padding_mask
+            )
+        else:
+            attention_probs = nn.functional.softmax(attention_weights, dim=-1)
+
+        # compute attention outputs
+        attention = torch.bmm(attention_probs, reshaped_value)
+        assert attention.shape == (B * self.nhead, Q, self.head_dim)
+        if self.is_pnnx is True:
+            attention = attention.permute(1, 0, 2).reshape(-1, self.embed_dim)
+            # TODO(fangjun): ncnn does not support reshape(-1, 1, self.embed_dim)
+            # We have to change InnerProduct in ncnn to ignore the extra dim below
+            attention = attention.unsqueeze(1)
+        else:
+            attention = (
+                attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim)
+            )
+
+        # apply output projection
+        output_right_context_utterance = self.out_proj(attention)
+        # The return shape of output_right_context_utterance is (10, 1, 512)
+
+        return output_right_context_utterance, key, value
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # TODO: Modify docs.
+        """Forward pass for training and validation mode.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of the hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors.
+
+        It computes a `big` attention matrix on full utterance and
+        then utilizes a pre-computed mask to simulate chunk-wise attention.
+
+        It concatenates three blocks: hard-copied right contexts,
+        and full utterance, as a `big` block,
+        to compute the query tensor:
+        query = [right_context, utterance],
+        with length Q = R + U.
+        It concatenates the three blocks: memory vectors,
+        hard-copied right contexts, and full utterance as another `big` block,
+        to compute the key and value tensors:
+        key & value = [memory, right_context, utterance],
+        with length KV = M + R + U.
+        Attention scores is computed with above `big` query and key.
+
+        Then the underlying chunk-wise attention is obtained by applying
+        the attention mask. Suppose
+        c_i: chunk at index i;
+        r_i: right context that c_i can use;
+        l_i: left context that c_i can use;
+        m_i: past memory vectors from previous layer that c_i can use;
+        The target chunk-wise attention is:
+        c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key)
+
+        Args:
+          utterance (torch.Tensor):
+            Full utterance frames, with shape (U, B, D).
+          right_context (torch.Tensor):
+            Hard-copied right context frames, with shape (R, B, D),
+            where R = num_chunks * right_context_length
+          memory (torch.Tensor):
+            Memory elements, with shape (M, B, D), where M = num_chunks - 1.
+            It is an empty tensor without using memory.
+          attention_mask (torch.Tensor):
+            Pre-computed attention mask to simulate underlying chunk-wise
+            attention, with shape (Q, KV).
+          padding_mask (torch.Tensor):
+            Padding mask of key tensor, with shape (B, KV).
+
+        Returns:
+          Output of right context and utterance, with shape (R + U, B, D).
+        """
+        output_right_context_utterance, _, _ = self._forward_impl(
+            utterance,
+            right_context,
+            memory,
+            attention_mask,
+            padding_mask=padding_mask,
+        )
+        return output_right_context_utterance
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        memory: torch.Tensor,
+        left_context_key: torch.Tensor,
+        left_context_val: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward pass for inference.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of right context;
+        U: length of utterance, i.e., current chunk;
+        L: length of cached left context;
+        M: length of cached memory vectors.
+
+        It concatenates the right context and utterance (i.e., current chunk)
+        of current chunk, to compute the query tensor:
+        query = [right_context, utterance],
+        with length Q = R + U.
+        It concatenates the memory vectors, right context, left context, and
+        current chunk, to compute the key and value tensors:
+        key & value = [memory, right_context, left_context, utterance],
+        with length KV = M + R + L + U.
+
+        The chunk-wise attention is:
+        chunk, right context (in query) ->
+          left context, chunk, right context, memory vectors (in key).
+
+        Args:
+          utterance (torch.Tensor):
+            Current chunk frames, with shape (U, B, D), where U = chunk_length.
+          right_context (torch.Tensor):
+            Right context frames, with shape (R, B, D),
+            where R = right_context_length.
+          memory (torch.Tensor):
+            Memory vectors, with shape (M, B, D), or empty tensor.
+          left_context_key (torch,Tensor):
+            Cached attention key of left context from preceding computation,
+            with shape (L, B, D).
+          left_context_val (torch.Tensor):
+            Cached attention value of left context from preceding computation,
+            with shape (L, B, D).
+          padding_mask (torch.Tensor):
+            Padding mask of key tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 4 tensors:
+            - output of right context and utterance, with shape (R + U, B, D).
+            - attention key of left context and utterance, which would be cached
+              for next computation, with shape (L + U, B, D).
+            - attention value of left context and utterance, which would be
+              cached for next computation, with shape (L + U, B, D).
+        """
+        if self.is_pnnx is False:
+            U = utterance.size(0)
+            R = right_context.size(0)
+            L = left_context_key.size(0)
+            M = memory.size(0)
+        else:
+            U = self.chunk_length
+            R = self.right_context_length
+            L = self.left_context_length
+            M = self.memory_size
+
+        # query = [right context, utterance]
+        Q = R + U
+        # key, value = [memory, right context, left context, utterance]
+        KV = M + R + L + U
+        attention_mask = torch.zeros(Q, KV).to(
+            dtype=torch.bool, device=utterance.device
+        )
+
+        output_right_context_utterance, key, value = self._forward_impl(
+            utterance,
+            right_context,
+            memory,
+            attention_mask,
+            padding_mask=padding_mask,
+            left_context_key=left_context_key,
+            left_context_val=left_context_val,
+        )
+        return (
+            output_right_context_utterance,
+            key[M + R :],
+            value[M + R :],
+        )
+
+
+class EmformerEncoderLayer(nn.Module):
+    """Emformer layer that constitutes Emformer.
+
+    Args:
+      d_model (int):
+        Input dimension.
+      nhead (int):
+        Number of attention heads.
+      dim_feedforward (int):
+        Hidden layer dimension of feedforward network.
+      chunk_length (int):
+        Length of each input segment.
+      dropout (float, optional):
+        Dropout probability. (Default: 0.0)
+      layer_dropout (float, optional):
+        Layer dropout probability. (Default: 0.0)
+      cnn_module_kernel (int):
+        Kernel size of convolution module.
+      left_context_length (int, optional):
+        Length of left context. (Default: 0)
+      right_context_length (int, optional):
+        Length of right context. (Default: 0)
+      memory_size (int, optional):
+        Number of memory elements to use. (Default: 0)
+      tanh_on_mem (bool, optional):
+        If ``True``, applies tanh to memory elements. (Default: ``False``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (Default: -1e8)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        chunk_length: int,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+        is_pnnx: bool = True,
+    ):
+        super().__init__()
+
+        self.attention = EmformerAttention(
+            embed_dim=d_model,
+            nhead=nhead,
+            left_context_length=left_context_length,
+            chunk_length=chunk_length,
+            memory_size=memory_size,
+            right_context_length=right_context_length,
+            dropout=dropout,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+            is_pnnx=is_pnnx,
+        )
+        self.summary_op = nn.AvgPool1d(
+            kernel_size=chunk_length, stride=chunk_length, ceil_mode=True
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(
+            chunk_length,
+            right_context_length,
+            d_model,
+            cnn_module_kernel,
+            is_pnnx=is_pnnx,
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean
+        # (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.layer_dropout = layer_dropout
+        self.left_context_length = left_context_length
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+        self.d_model = d_model
+        self.use_memory = memory_size > 0
+
+    def _update_attn_cache(
+        self,
+        next_key: torch.Tensor,
+        next_val: torch.Tensor,
+        memory: torch.Tensor,
+        attn_cache: List[torch.Tensor],
+    ) -> List[torch.Tensor]:
+        """Update cached attention state:
+        1) output memory of current chunk in the lower layer;
+        2) attention key and value in current chunk's computation, which would
+        be reused in next chunk's computation.
+        """
+        # attn_cache[0].shape (self.memory_size, 1, 512)
+        # memory.shape (1, 1, 512)
+        # attn_cache[1].shape (self.left_context_length, 1, 512)
+        # attn_cache[2].shape (self.left_context_length, 1, 512)
+        # next_key.shape (self.left_context_length + self.right_context_utterance, 1, 512)
+        # next_value.shape (self.left_context_length + self.right_context_utterance, 1, 512)
+        new_memory = torch.cat([attn_cache[0], memory])
+        # TODO(fangjun): Remove torch.cat
+        #  new_key = torch.cat([attn_cache[1], next_key])
+        #  new_val = torch.cat([attn_cache[2], next_val])
+        attn_cache[0] = new_memory[1:]
+        attn_cache[1] = next_key[-self.left_context_length :]
+        attn_cache[2] = next_val[-self.left_context_length :]
+        return attn_cache
+
+    def _apply_conv_module_forward(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+    ) -> torch.Tensor:
+        """Apply convolution module in training and validation mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+        utterance, right_context = self.conv_module(utterance, right_context)
+        right_context_utterance = torch.cat([right_context, utterance])
+        return right_context_utterance
+
+    def _apply_conv_module_infer(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        conv_cache: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Apply convolution module on utterance in inference mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+        utterance, right_context, conv_cache = self.conv_module.infer(
+            utterance, right_context, conv_cache
+        )
+        right_context_utterance = torch.cat([right_context, utterance])
+        return right_context_utterance, conv_cache
+
+    def _apply_attention_module_forward(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Apply attention module in training and validation mode."""
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+
+        if self.use_memory:
+            memory = self.summary_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[
+                :-1, :, :
+            ]
+        else:
+            memory = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+        output_right_context_utterance = self.attention(
+            utterance=utterance,
+            right_context=right_context,
+            memory=memory,
+            attention_mask=attention_mask,
+            padding_mask=padding_mask,
+        )
+
+        return output_right_context_utterance
+
+    def _apply_attention_module_infer(
+        self,
+        right_context_utterance: torch.Tensor,
+        R: int,
+        attn_cache: List[torch.Tensor],
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Apply attention module in inference mode.
+        1) Unpack cached states including:
+           - memory from previous chunks;
+           - attention key and value of left context from preceding
+             chunk's compuation;
+        2) Apply attention computation;
+        3) Update cached attention states including:
+           - memory of current chunk;
+           - attention key and value in current chunk's computation, which would
+             be resued in next chunk's computation.
+        """
+        utterance = right_context_utterance[R:]
+        right_context = right_context_utterance[:R]
+
+        pre_memory = attn_cache[0]
+        left_context_key = attn_cache[1]
+        left_context_val = attn_cache[2]
+
+        if self.use_memory:
+            memory = torch.mean(utterance, dim=0, keepdim=True)
+
+            #  memory = self.summary_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[
+            #          :1, :, :
+            #  ]
+        else:
+            memory = torch.empty(0).to(dtype=utterance.dtype, device=utterance.device)
+        (output_right_context_utterance, next_key, next_val) = self.attention.infer(
+            utterance=utterance,
+            right_context=right_context,
+            memory=pre_memory,
+            left_context_key=left_context_key,
+            left_context_val=left_context_val,
+            padding_mask=padding_mask,
+        )
+        attn_cache = self._update_attn_cache(next_key, next_val, memory, attn_cache)
+        return output_right_context_utterance, attn_cache
+
+    def forward(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        r"""Forward pass for training and validation mode.
+
+        B: batch size;
+        D: embedding dimension;
+        R: length of hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors.
+
+        Args:
+          utterance (torch.Tensor):
+            Utterance frames, with shape (U, B, D).
+          right_context (torch.Tensor):
+            Right context frames, with shape (R, B, D).
+          attention_mask (torch.Tensor):
+            Attention mask for underlying attention module,
+            with shape (Q, KV), where Q = R + U, KV = M + R + U.
+          padding_mask (torch.Tensor):
+            Padding mask of ker tensor, with shape (B, KV).
+
+        Returns:
+          A tuple containing 2 tensors:
+            - output utterance, with shape (U, B, D).
+            - output right context, with shape (R, B, D).
+        """
+        R = right_context.size(0)
+        src = torch.cat([right_context, utterance])
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # emformer attention module
+        src_att = self._apply_attention_module_forward(
+            src, R, attention_mask, padding_mask=padding_mask
+        )
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src_conv = self._apply_conv_module_forward(src, R)
+        src = src + self.dropout(src_conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        output_utterance = src[R:]
+        output_right_context = src[:R]
+        return output_utterance, output_right_context
+
+    @torch.jit.export
+    def infer(
+        self,
+        utterance: torch.Tensor,
+        right_context: torch.Tensor,
+        cache: List[torch.Tensor],
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """Forward pass for inference.
+
+         B: batch size;
+         D: embedding dimension;
+         R: length of right_context;
+         U: length of utterance;
+         M: length of memory.
+
+        Args:
+           utterance (torch.Tensor):
+             Utterance frames, with shape (U, B, D).
+           right_context (torch.Tensor):
+             Right context frames, with shape (R, B, D).
+           attn_cache (List[torch.Tensor]):
+             Cached attention tensors generated in preceding computation,
+             including memory, key and value of left context.
+           conv_cache (torch.Tensor, optional):
+             Cache tensor of left context for causal convolution.
+           padding_mask (torch.Tensor):
+             Padding mask of ker tensor.
+
+         Returns:
+           (Tensor, Tensor, List[torch.Tensor], Tensor):
+             - output utterance, with shape (U, B, D);
+             - output right_context, with shape (R, B, D);
+             - output attention cache;
+             - output convolution cache.
+        """
+        R = self.right_context_length
+        src = torch.cat([right_context, utterance])
+        attn_cache = cache[:3]
+        conv_cache = cache[3]
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # emformer attention module
+        src_att, attn_cache = self._apply_attention_module_infer(
+            src, R, attn_cache, padding_mask=padding_mask
+        )
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache)
+        src = src + self.dropout(src_conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        output_utterance = src[R:]
+        output_right_context = src[:R]
+        return (output_utterance, output_right_context, attn_cache + [conv_cache])
+
+
+def _gen_attention_mask_block(
+    col_widths: List[int],
+    col_mask: List[bool],
+    num_rows: int,
+    device: torch.device,
+) -> torch.Tensor:
+    assert len(col_widths) == len(
+        col_mask
+    ), "Length of col_widths must match that of col_mask"
+
+    mask_block = [
+        torch.ones(num_rows, col_width, device=device)
+        if is_ones_col
+        else torch.zeros(num_rows, col_width, device=device)
+        for col_width, is_ones_col in zip(col_widths, col_mask)
+    ]
+    return torch.cat(mask_block, dim=1)
+
+
+class EmformerEncoder(nn.Module):
+    """Implements the Emformer architecture introduced in
+    *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency
+    Streaming Speech Recognition*
+    [:footcite:`shi2021emformer`].
+
+    In this model, the memory bank computation is simplifed, using the averaged
+    value of each chunk as its memory vector.
+
+    Args:
+      d_model (int):
+        Input dimension.
+      nhead (int):
+        Number of attention heads in each emformer layer.
+      dim_feedforward (int):
+        Hidden layer dimension of each emformer layer's feedforward network.
+      num_encoder_layers (int):
+        Number of emformer layers to instantiate.
+      chunk_length (int):
+        Length of each input segment.
+      dropout (float, optional):
+        Dropout probability. (default: 0.0)
+      layer_dropout (float, optional):
+        Layer dropout probability. (default: 0.0)
+      cnn_module_kernel (int):
+        Kernel size of convolution module.
+      left_context_length (int, optional):
+        Length of left context. (default: 0)
+      right_context_length (int, optional):
+        Length of right context. (default: 0)
+      memory_size (int, optional):
+        Number of memory elements to use. (default: 0)
+      tanh_on_mem (bool, optional):
+        If ``true``, applies tanh to memory elements. (default: ``false``)
+      negative_inf (float, optional):
+        Value to use for negative infinity in attention weights. (default: -1e8)
+    """
+
+    def __init__(
+        self,
+        chunk_length: int,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+        is_pnnx: bool = True,
+    ):
+        super().__init__()
+
+        assert (
+            chunk_length - 1
+        ) & chunk_length == 0, "chunk_length should be a power of 2."
+        self.shift = int(math.log(chunk_length, 2))
+
+        self.use_memory = memory_size > 0
+
+        self.emformer_layers = nn.ModuleList(
+            [
+                EmformerEncoderLayer(
+                    d_model=d_model,
+                    nhead=nhead,
+                    dim_feedforward=dim_feedforward,
+                    chunk_length=chunk_length,
+                    dropout=dropout,
+                    layer_dropout=layer_dropout,
+                    cnn_module_kernel=cnn_module_kernel,
+                    left_context_length=left_context_length,
+                    right_context_length=right_context_length,
+                    memory_size=memory_size,
+                    tanh_on_mem=tanh_on_mem,
+                    negative_inf=negative_inf,
+                    is_pnnx=is_pnnx,
+                )
+                for layer_idx in range(num_encoder_layers)
+            ]
+        )
+
+        self.num_encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.left_context_length = left_context_length
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        self.memory_size = memory_size
+        self.cnn_module_kernel = cnn_module_kernel
+
+    def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor:
+        """Hard copy each chunk's right context and concat them."""
+        T = x.shape[0]
+        num_chunks = math.ceil((T - self.right_context_length) / self.chunk_length)
+        # first (num_chunks - 1) right context block
+        intervals = torch.arange(
+            0, self.chunk_length * (num_chunks - 1), self.chunk_length
+        )
+        first = torch.arange(
+            self.chunk_length, self.chunk_length + self.right_context_length
+        )
+        indexes = intervals.unsqueeze(1) + first.unsqueeze(0)
+        # cat last right context block
+        indexes = torch.cat(
+            [
+                indexes,
+                torch.arange(T - self.right_context_length, T).unsqueeze(0),
+            ]
+        )
+        right_context_blocks = x[indexes.reshape(-1)]
+        return right_context_blocks
+
+    def _gen_attention_mask_col_widths(self, chunk_idx: int, U: int) -> List[int]:
+        """Calculate column widths (key, value) in attention mask for the
+        chunk_idx chunk."""
+        num_chunks = math.ceil(U / self.chunk_length)
+        rc = self.right_context_length
+        lc = self.left_context_length
+        rc_start = chunk_idx * rc
+        rc_end = rc_start + rc
+        chunk_start = max(chunk_idx * self.chunk_length - lc, 0)
+        chunk_end = min((chunk_idx + 1) * self.chunk_length, U)
+        R = rc * num_chunks
+
+        if self.use_memory:
+            m_start = max(chunk_idx - self.memory_size, 0)
+            M = num_chunks - 1
+            col_widths = [
+                m_start,  # before memory
+                chunk_idx - m_start,  # memory
+                M - chunk_idx,  # after memory
+                rc_start,  # before right context
+                rc,  # right context
+                R - rc_end,  # after right context
+                chunk_start,  # before chunk
+                chunk_end - chunk_start,  # chunk
+                U - chunk_end,  # after chunk
+            ]
+        else:
+            col_widths = [
+                rc_start,  # before right context
+                rc,  # right context
+                R - rc_end,  # after right context
+                chunk_start,  # before chunk
+                chunk_end - chunk_start,  # chunk
+                U - chunk_end,  # after chunk
+            ]
+
+        return col_widths
+
+    def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor:
+        """Generate attention mask to simulate underlying chunk-wise attention
+        computation, where chunk-wise connections are filled with `False`,
+        and other unnecessary connections beyond chunk are filled with `True`.
+
+        R: length of hard-copied right contexts;
+        U: length of full utterance;
+        M: length of memory vectors;
+        Q: length of attention query;
+        KV: length of attention key and value.
+
+        The shape of attention mask is (Q, KV).
+        If self.use_memory is `True`:
+          query = [right_context, utterance];
+          key, value = [memory, right_context, utterance];
+          Q = R + U, KV = M + R + U.
+        Otherwise:
+          query = [right_context, utterance]
+          key, value = [right_context, utterance]
+          Q = R + U, KV = R + U.
+
+        Suppose:
+          c_i: chunk at index i;
+          r_i: right context that c_i can use;
+          l_i: left context that c_i can use;
+          m_i: past memory vectors from previous layer that c_i can use;
+        The target chunk-wise attention is:
+          c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key).
+        """
+        U = utterance.size(0)
+        num_chunks = math.ceil(U / self.chunk_length)
+
+        right_context_mask = []
+        utterance_mask = []
+
+        if self.use_memory:
+            num_cols = 9
+            # right context and utterance both attend to memory, right context,
+            # utterance
+            right_context_utterance_cols_mask = [
+                idx in [1, 4, 7] for idx in range(num_cols)
+            ]
+        else:
+            num_cols = 6
+            # right context and utterance both attend to right context and
+            # utterance
+            right_context_utterance_cols_mask = [
+                idx in [1, 4] for idx in range(num_cols)
+            ]
+        masks_to_concat = [right_context_mask, utterance_mask]
+
+        for chunk_idx in range(num_chunks):
+            col_widths = self._gen_attention_mask_col_widths(chunk_idx, U)
+
+            right_context_mask_block = _gen_attention_mask_block(
+                col_widths,
+                right_context_utterance_cols_mask,
+                self.right_context_length,
+                utterance.device,
+            )
+            right_context_mask.append(right_context_mask_block)
+
+            utterance_mask_block = _gen_attention_mask_block(
+                col_widths,
+                right_context_utterance_cols_mask,
+                min(
+                    self.chunk_length,
+                    U - chunk_idx * self.chunk_length,
+                ),
+                utterance.device,
+            )
+            utterance_mask.append(utterance_mask_block)
+
+        attention_mask = (
+            1 - torch.cat([torch.cat(mask) for mask in masks_to_concat])
+        ).to(torch.bool)
+        return attention_mask
+
+    def _forward(
+        self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass for training and validation mode.
+
+        B: batch size;
+        D: input dimension;
+        U: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (U + right_context_length, B, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, which contains the
+            right_context at the end.
+
+        Returns:
+          A tuple of 2 tensors:
+            - output utterance frames, with shape (U, B, D).
+            - output_lengths, with shape (B,), without containing the
+              right_context at the end.
+        """
+        U = x.size(0) - self.right_context_length
+
+        right_context = self._gen_right_context(x)
+        utterance = x[:U]
+        output_lengths = torch.clamp(lengths - self.right_context_length, min=0)
+        attention_mask = self._gen_attention_mask(utterance)
+
+        M = (
+            right_context.size(0) // self.right_context_length - 1
+            if self.use_memory
+            else 0
+        )
+        padding_mask = make_pad_mask(M + right_context.size(0) + output_lengths)
+
+        output = utterance
+        for layer in self.emformer_layers:
+            output, right_context = layer(
+                output,
+                right_context,
+                attention_mask,
+                padding_mask=padding_mask,
+                warmup=warmup,
+            )
+
+        return output, output_lengths
+
+    @torch.jit.export
+    def infer(
+        self,
+        x: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Forward pass for streaming inference.
+
+        B: batch size;
+        D: input dimension;
+        U: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (U + right_context_length, B, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, which contains the
+            right_context at the end.
+          states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa
+            Cached states containing:
+            - attn_caches: attention states from preceding chunk's computation,
+              where each element corresponds to each emformer layer
+            - conv_caches: left context for causal convolution, where each
+              element corresponds to each layer.
+
+        Returns:
+          (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]):
+            - output utterance frames, with shape (U, B, D).
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+            - updated states from current chunk's computation.
+        """
+        # lengths = chunk_length + right_context_length
+        utterance = x[: self.chunk_length]
+        right_context = x[self.chunk_length :]
+        #  right_context_utterance = torch.cat([right_context, utterance])
+
+        output = utterance
+        output_states: List[torch.Tensor] = []
+        for layer_idx, layer in enumerate(self.emformer_layers):
+            start = layer_idx * 4
+            end = start + 4
+            cache = states[start:end]
+
+            (output, right_context, output_cache,) = layer.infer(
+                output,
+                right_context,
+                padding_mask=None,
+                cache=cache,
+            )
+            output_states.extend(output_cache)
+
+        return output, output_states
+
+    @torch.jit.export
+    def init_states(
+        self, device: torch.device = torch.device("cpu")
+    ) -> List[torch.Tensor]:
+        """Create initial states."""
+        #
+        states = []
+        # layer0: attn cache, conv cache, 3 tensors + 1 tensor
+        # layer1: attn cache, conv cache, 3 tensors +  1 tensor
+        # layer2: attn cache, conv cache, 3 tensors + 1 tensor
+        # ...
+        # last layer: attn cache, conv cache, 3 tensors + 1 tensor
+        for i in range(self.num_encoder_layers):
+            states.append(torch.zeros(self.memory_size, 1, self.d_model, device=device))
+            states.append(
+                torch.zeros(self.left_context_length, 1, self.d_model, device=device)
+            )
+            states.append(
+                torch.zeros(self.left_context_length, 1, self.d_model, device=device)
+            )
+
+            states.append(
+                torch.zeros(1, self.d_model, self.cnn_module_kernel - 1, device=device)
+            )
+        return states
+
+
+class Emformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        chunk_length: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 3,
+        left_context_length: int = 0,
+        right_context_length: int = 0,
+        memory_size: int = 0,
+        tanh_on_mem: bool = False,
+        negative_inf: float = -1e8,
+        is_pnnx: bool = True,
+    ):
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        self.right_context_length = right_context_length
+        self.chunk_length = chunk_length
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+        if chunk_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "chunk_length must be a mutiple of subsampling_factor."
+            )
+        if left_context_length != 0 and left_context_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "left_context_length must be 0 or a mutiple of subsampling_factor."  # noqa
+            )
+        if right_context_length != 0 and right_context_length % subsampling_factor != 0:
+            raise NotImplementedError(
+                "right_context_length must be 0 or a mutiple of subsampling_factor."  # noqa
+            )
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model, is_pnnx=is_pnnx)
+        self.is_pnnx = is_pnnx
+
+        self.num_encoder_layers = num_encoder_layers
+        self.memory_size = memory_size
+        self.d_model = d_model
+        self.cnn_module_kernel = cnn_module_kernel
+        self.left_context_length = left_context_length // subsampling_factor
+        self.right_context_length = right_context_length
+        self.subsampling_factor = subsampling_factor
+
+        assert subsampling_factor == 4, subsampling_factor
+        pad_length = right_context_length + 2 * 4 + 3
+
+        # before subsampling
+        self.T = self.chunk_length + pad_length
+
+        self.encoder = EmformerEncoder(
+            chunk_length=chunk_length // subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+            cnn_module_kernel=cnn_module_kernel,
+            left_context_length=left_context_length // subsampling_factor,
+            right_context_length=right_context_length // subsampling_factor,
+            memory_size=memory_size,
+            tanh_on_mem=tanh_on_mem,
+            negative_inf=negative_inf,
+            is_pnnx=is_pnnx,
+        )
+
+    def _forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass for training and non-streaming inference.
+
+        B: batch size;
+        D: feature dimension;
+        T: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (B, T, D).
+          x_lens (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, containing the
+            right_context at the end.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          (Tensor, Tensor):
+            - output embedding, with shape (B, T', D), where
+              T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4.
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+        """
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+        assert x.size(0) == x_lens.max().item()
+
+        output, output_lengths = self.encoder(x, x_lens, warmup=warmup)  # (T, N, C)
+
+        output = output.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+
+        return output, output_lengths
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Forward pass for streaming inference.
+
+        B: batch size;
+        D: feature dimension;
+        T: length of utterance.
+
+        Args:
+          x (torch.Tensor):
+            Utterance frames right-padded with right context frames,
+            with shape (B, T, D).
+          lengths (torch.Tensor):
+            With shape (B,) and i-th element representing number of valid
+            utterance frames for i-th batch element in x, containing the
+            right_context at the end.
+          states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa
+            Cached states containing:
+            - past_lens: number of past frames for each sample in batch
+            - attn_caches: attention states from preceding chunk's computation,
+              where each element corresponds to each emformer layer
+            - conv_caches: left context for causal convolution, where each
+              element corresponds to each layer.
+        Returns:
+          (Tensor, Tensor):
+            - output embedding, with shape (B, T', D), where
+              T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4.
+            - output lengths, with shape (B,), without containing the
+              right_context at the end.
+            - updated states from current chunk's computation.
+        """
+        x = self.encoder_embed(x)
+        # drop the first and last frames
+        x = x[:, 1:-1, :]
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        output, output_states = self.encoder.infer(x, states)
+
+        output = output.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+
+        return output, output_states
+
+    @torch.jit.export
+    def init_states(
+        self, device: torch.device = torch.device("cpu")
+    ) -> List[torch.Tensor]:
+        """Create initial states."""
+        return self.encoder.init_states(device)
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        is_pnnx: bool = True,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+          is_pnnx:
+            True if we are converting the model to PNNX format.
+            False otherwise.
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+        # ncnn supports only batch size == 1
+        self.is_pnnx = is_pnnx
+        self.conv_out_dim = self.out.weight.shape[1]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+
+        if torch.jit.is_tracing() and self.is_pnnx is True:
+            x = x.permute(0, 2, 1, 3).reshape(1, -1, self.conv_out_dim)
+            x = self.out(x)
+        else:
+            # Now x is of shape (N, odim, ((T-1)//2-1)//2, ((idim-1)//2-1)//2)
+            b, c, t, f = x.size()
+            x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..ee2f09151
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py
new file mode 100755
index 000000000..953f95c45
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py
@@ -0,0 +1,341 @@
+#!/usr/bin/env python3
+
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for more details about how to use this file.
+
+Usage:
+./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir ./conv_emformer_transducer_stateless2/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+
+cd ./conv_emformer_transducer_stateless2/exp
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+You can find converted models at
+https://huggingface.co/csukuangfj/sherpa-ncnn-conv-emformer-transducer-2022-12-04
+
+See ./streaming-ncnn-decode.py
+and
+https://github.com/k2-fsa/sherpa-ncnn
+for usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+from scaling_converter import convert_scaled_to_non_scaled
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    chunk_length = encoder_model.chunk_length  # before subsampling
+    right_context_length = encoder_model.right_context_length  # before subsampling
+    pad_length = right_context_length + 2 * 4 + 3
+    s = f"chunk_length: {chunk_length}, "
+    s += f"right_context_length: {right_context_length}\n"
+    logging.info(s)
+
+    T = chunk_length + pad_length
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.init_states()
+
+    traced_model = torch.jit.trace(encoder_model, (x, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py
new file mode 100755
index 000000000..1e59e0858
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py
@@ -0,0 +1,644 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./conv_emformer_transducer_stateless2/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from emformer import Emformer
+from scaling_converter import convert_scaled_to_non_scaled
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Emformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Emformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Emformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Emformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has the following inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - a list of states (each layers has 4 states)
+
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - a list of states (each layers has 4 states)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    num_encoder_layers = encoder_model.encoder.num_encoder_layers
+    memory_size = encoder_model.encoder.memory_size
+    cnn_module_kernel = encoder_model.encoder.cnn_module_kernel
+    chunk_length = encoder_model.encoder.chunk_length
+    right_context_length = encoder_model.encoder.right_context_length
+    encoder_dim = encoder_model.encoder.d_model
+    left_context_length = encoder_model.encoder.left_context_length
+
+    T = encoder_model.encoder.T
+
+    logging.info(f"num_encoder_layers={num_encoder_layers}")
+    logging.info(f"memory_size={memory_size}")
+    logging.info(f"cnn_module_kernel={cnn_module_kernel}")
+    logging.info(f"chunk_length={chunk_length}")
+    logging.info(f"right_context_length={right_context_length}")
+    logging.info(f"encoder_dim={encoder_dim}")
+    logging.info(f"left_context_length={left_context_length} (after subsampling)")
+    logging.info(f"T={T}")
+
+    meta_data = {
+        "model_type": "conv-emformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "decode_chunk_len": str(chunk_length),  # 32
+        "T": str(T),  # 32
+        "num_encoder_layers": str(num_encoder_layers),
+        "memory_size": str(memory_size),
+        "cnn_module_kernel": str(cnn_module_kernel),
+        "right_context_length": str(right_context_length),
+        "left_context_length": str(left_context_length),
+        "encoder_dim": str(encoder_dim),
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.encoder.init_states()
+
+    # Each layer has 4 states
+    assert len(states) == num_encoder_layers * 4, (len(states), num_encoder_layers)
+    # layer 0:
+    #  state0: (memory_size, 1, encoder_dim)
+    #  state1: (left_context_length, 1, encoder_dim)
+    #  state2: (left_context_length, 1, encoder_dim)
+    #  state3: (1, encoder_dim, cnn_module_kernel-1)
+
+    inputs = {}
+    input_names = ["x"]
+
+    outputs = {}
+    output_names = ["encoder_out"]
+
+    def build_inputs_outputs(s, name):
+        assert len(s) == 4, len(s)
+        logging.info(f"{name}_0.shape: {s[0].shape}")
+        input_names.append(f"{name}_0")
+        inputs[f"{name}_0"] = {1: "N"}
+        output_names.append(f"new_{name}_0")
+
+        logging.info(f"{name}_1.shape: {s[1].shape}")
+        input_names.append(f"{name}_1")
+        inputs[f"{name}_1"] = {1: "N"}
+        output_names.append(f"new_{name}_1")
+
+        logging.info(f"{name}_2.shape: {s[2].shape}")
+        input_names.append(f"{name}_2")
+        inputs[f"{name}_2"] = {1: "N"}
+        output_names.append(f"new_{name}_2")
+
+        logging.info(f"{name}_3.shape: {s[3].shape}")
+        input_names.append(f"{name}_3")
+        inputs[f"{name}_3"] = {0: "N"}
+        output_names.append(f"new_{name}_3")
+
+    for i in range(num_encoder_layers):
+        base_name = f"layer{i}"
+        s = states[i * 4 : (i + 1) * 4]
+        build_inputs_outputs(s, base_name)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, states),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes={
+            "x": {0: "N"},
+            "encoder_out": {0: "N"},
+            **inputs,
+            **outputs,
+        },
+    )
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+    params.is_pnnx = False
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py
new file mode 100755
index 000000000..8e5b14903
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py
@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./conv_emformer_transducer_stateless2/export.py \
+  --exp-dir ./conv_emformer_transducer_stateless2/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32 \
+  --jit False
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `conv_emformer_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./conv_emformer_transducer_stateless2/decode.py \
+        --exp-dir ./conv_emformer_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model \
+        --use-averaged-model=False \
+        --num-encoder-layers 12 \
+        --chunk-length 32 \
+        --cnn-module-kernel 31 \
+        --left-context-length 32 \
+        --right-context-length 8 \
+        --memory-size 32
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py
new file mode 100755
index 000000000..1fe358c79
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py
@@ -0,0 +1,292 @@
+#!/usr/bin/env python3
+# flake8: noqa
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models exported by `torch.jit.trace()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir ./conv_emformer_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+./conv_emformer_transducer_stateless2/jit_pretrained.py \
+  --encoder-model-filename ./conv_emformer_transducer_stateless2/exp/encoder_jit_trace-pnnx.pt \
+  --decoder-model-filename ./conv_emformer_transducer_stateless2/exp/decoder_jit_trace-pnnx.pt \
+  --joiner-model-filename ./conv_emformer_transducer_stateless2/exp/joiner_jit_trace-pnnx.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+from torch.nn.utils.rnn import pad_sequence
+from typing import Optional, List
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    assert encoder_out.ndim == 2
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32).unsqueeze(0)
+        # decoder_input.shape (1,, 1 context_size)
+        decoder_out = decoder(decoder_input, torch.tensor([0])).squeeze(1)
+    else:
+        assert decoder_out.ndim == 2
+        assert hyp is not None, hyp
+
+    T = encoder_out.size(0)
+    for i in range(T):
+        cur_encoder_out = encoder_out[i : i + 1]
+        joiner_out = joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32).unsqueeze(0)
+            decoder_out = decoder(decoder_input, torch.tensor([0])).squeeze(1)
+
+    return hyp, decoder_out
+
+
+def create_streaming_feature_extractor(sample_rate) -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor(args.sample_rate)
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=args.sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    logging.info("Decoding started")
+    chunk_length = encoder.chunk_length
+    right_context_length = encoder.right_context_length
+
+    # Assume the subsampling factor is 4
+    pad_length = right_context_length + 2 * 4 + 3
+    T = chunk_length + pad_length
+
+    logging.info(f"chunk_length: {chunk_length}")
+    logging.info(f"right_context_length: {right_context_length}")
+
+    states = encoder.init_states(device)
+    logging.info(f"num layers: {len(states)//4}")
+
+    tail_padding = torch.zeros(int(0.3 * args.sample_rate), dtype=torch.float32)
+
+    wave_samples = torch.cat([wave_samples, tail_padding])
+
+    chunk = int(0.25 * args.sample_rate)  # 0.2 second
+    num_processed_frames = 0
+
+    hyp = None
+    decoder_out = None
+
+    start = 0
+    while start < wave_samples.numel():
+        logging.info(f"{start}/{wave_samples.numel()}")
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+        online_fbank.accept_waveform(
+            sampling_rate=args.sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= T:
+            frames = []
+            for i in range(T):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += chunk_length
+            frames = torch.cat(frames, dim=0).unsqueeze(0)
+            # TODO(fangjun): remove x_lens
+            x_lens = torch.tensor([T])
+            encoder_out, _, states = encoder(frames, x_lens, states)
+
+            hyp, decoder_out = greedy_search(
+                decoder, joiner, encoder_out.squeeze(0), decoder_out, hyp
+            )
+
+    context_size = 2
+
+    logging.info(args.sound_file)
+    logging.info(sp.decode(hyp[context_size:]))
+
+    logging.info("Decoding Done")
+
+
+torch.set_num_threads(4)
+torch.set_num_interop_threads(1)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..1eb4dcc83
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/lstmp.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py
new file mode 120000
index 000000000..322b694e0
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/onnx_pretrained.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/onnx_pretrained.py
new file mode 100755
index 000000000..a6c69d54f
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/onnx_pretrained.py
@@ -0,0 +1,459 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This script loads ONNX models exported by ./export-onnx.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./conv_emformer_transducer_stateless2/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file with the exported ONNX models
+
+./conv_emformer_transducer_stateless2/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_encoder_states()
+
+    def init_encoder_states(self, batch_size: int = 1):
+        encoder_meta = self.encoder.get_modelmeta().custom_metadata_map
+
+        model_type = encoder_meta["model_type"]
+        assert model_type == "conv-emformer", model_type
+
+        decode_chunk_len = int(encoder_meta["decode_chunk_len"])
+        T = int(encoder_meta["T"])
+
+        num_encoder_layers = int(encoder_meta["num_encoder_layers"])
+        memory_size = int(encoder_meta["memory_size"])
+        cnn_module_kernel = int(encoder_meta["cnn_module_kernel"])
+        right_context_length = int(encoder_meta["right_context_length"])
+        left_context_length = int(encoder_meta["left_context_length"])
+        encoder_dim = int(encoder_meta["encoder_dim"])
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+        logging.info(f"memory_size: {memory_size}")
+        logging.info(f"cnn_module_kernel: {cnn_module_kernel}")
+        logging.info(f"left_context_length: {left_context_length} (after subsampling)")
+        logging.info(f"right_context_length: {right_context_length}")
+        logging.info(f"encoder_dim: {encoder_dim}")
+
+        N = batch_size
+
+        states = []
+        for i in range(num_encoder_layers):
+            s0 = torch.zeros(memory_size, N, encoder_dim)
+            s1 = torch.zeros(left_context_length, N, encoder_dim)
+            s2 = torch.zeros(left_context_length, N, encoder_dim)
+            s3 = torch.zeros(N, encoder_dim, cnn_module_kernel - 1)
+            states.extend([s0, s1, s2, s3])
+
+        self.states = states
+
+        self.segment = T
+        self.offset = decode_chunk_len
+        self.num_encoder_layers = num_encoder_layers
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def _build_encoder_input_output(
+        self,
+        x: torch.Tensor,
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        encoder_input = {"x": x.numpy()}
+        encoder_output = ["encoder_out"]
+
+        def build_inputs_outputs(states: List[torch.Tensor], name: str):
+            for i in range(4):
+                if isinstance(states[i], torch.Tensor):
+                    encoder_input[f"{name}_{i}"] = states[i].numpy()
+                else:
+                    encoder_input[f"{name}_{i}"] = states[i]
+
+                encoder_output.append(f"new_{name}_{i}")
+
+        for i in range(self.num_encoder_layers):
+            base_name = f"layer{i}"
+            s = self.states[i * 4 : (i + 1) * 4]
+            build_inputs_outputs(s, base_name)
+
+        return encoder_input, encoder_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        self.states = states
+
+    def run_encoder(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+        Returns:
+          Return a 3-D tensor of shape (N, T', joiner_dim) where
+          T' is usually equal to ((T-7)//2+1)//2
+        """
+        encoder_input, encoder_output_names = self._build_encoder_input_output(x)
+        out = self.encoder.run(encoder_output_names, encoder_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    context_size: int,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+) -> List[int]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (1, T, joiner_dim)
+      context_size:
+        The context size of the decoder model.
+      decoder_out:
+        Optional. Decoder output of the previous chunk.
+      hyp:
+        Decoding results for previous chunks.
+    Returns:
+      Return the decoded results so far.
+    """
+
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert hyp is not None, hyp
+
+    encoder_out = encoder_out.squeeze(0)
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t : t + 1]
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+            decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    context_size = model.context_size
+    hyp = None
+    decoder_out = None
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            encoder_out = model.run_encoder(frames)
+            hyp, decoder_out = greedy_search(
+                model,
+                encoder_out,
+                context_size,
+                decoder_out,
+                hyp,
+            )
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py
new file mode 120000
index 000000000..8f19a99da
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..12f22cf9c
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling_converter.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py
new file mode 120000
index 000000000..bf9cbbe2e
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py
@@ -0,0 +1 @@
+../conv_emformer_transducer_stateless/stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming-ncnn-decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming-ncnn-decode.py
new file mode 100755
index 000000000..74da9e6c8
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming-ncnn-decode.py
@@ -0,0 +1,386 @@
+#!/usr/bin/env python3
+#
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./conv_emformer_transducer_stateless2/streaming-ncnn-decode.py \
+  --tokens ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/tokens.txt \
+  --encoder-param-filename ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/encoder_jit_trace-epoch-30-avg-10-pnnx.ncnn.param \
+  --encoder-bin-filename ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/encoder_jit_trace-epoch-30-avg-10-pnnx.ncnn.bin \
+  --decoder-param-filename ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/decoder_jit_trace-epoch-30-avg-10-pnnx.ncnn.param \
+  --decoder-bin-filename ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/decoder_jit_trace-epoch-30-avg-10-pnnx.ncnn.bin \
+  --joiner-param-filename ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/joiner_jit_trace-epoch-30-avg-10-pnnx.ncnn.param \
+  --joiner-bin-filename ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/joiner_jit_trace-epoch-30-avg-10-pnnx.ncnn.bin \
+  ./sherpa-ncnn-conv-emformer-transducer-2022-12-04/test_wavs/1089-134686-0001.wav
+
+You can find pretrained models at
+https://huggingface.co/csukuangfj/sherpa-ncnn-conv-emformer-transducer-2022-12-04
+"""
+
+import argparse
+import logging
+from typing import List, Optional, Tuple
+
+import k2
+import ncnn
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument(
+        "--encoder-param-filename",
+        type=str,
+        help="Path to encoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--encoder-bin-filename",
+        type=str,
+        help="Path to encoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--decoder-param-filename",
+        type=str,
+        help="Path to decoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--decoder-bin-filename",
+        type=str,
+        help="Path to decoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--joiner-param-filename",
+        type=str,
+        help="Path to joiner.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--joiner-bin-filename",
+        type=str,
+        help="Path to joiner.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="Path to foo.wav",
+    )
+
+    return parser.parse_args()
+
+
+class Model:
+    def __init__(self, args):
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+
+        self.num_layers = 12
+        self.memory_size = 32
+        self.d_model = 512
+        self.cnn_module_kernel = 31
+
+        self.left_context_length = 32 // 4  # after subsampling
+        self.chunk_length = 32  # before subsampling
+        right_context_length = 8  # before subsampling
+        pad_length = right_context_length + 2 * 4 + 3
+        self.T = self.chunk_length + pad_length
+        print("T", self.T, self.chunk_length)
+
+    def get_init_states(self) -> List[torch.Tensor]:
+        states = []
+
+        for i in range(self.num_layers):
+            s0 = torch.zeros(self.memory_size, self.d_model)
+            s1 = torch.zeros(self.left_context_length, self.d_model)
+            s2 = torch.zeros(self.left_context_length, self.d_model)
+            s3 = torch.zeros(self.d_model, self.cnn_module_kernel - 1)
+            states.extend([s0, s1, s2, s3])
+
+        return states
+
+    def init_encoder(self, args):
+        encoder_net = ncnn.Net()
+        encoder_net.opt.use_packing_layout = False
+        encoder_net.opt.use_fp16_storage = False
+        encoder_net.opt.num_threads = 4
+
+        encoder_param = args.encoder_param_filename
+        encoder_model = args.encoder_bin_filename
+
+        encoder_net.load_param(encoder_param)
+        encoder_net.load_model(encoder_model)
+
+        self.encoder_net = encoder_net
+
+    def init_decoder(self, args):
+        decoder_param = args.decoder_param_filename
+        decoder_model = args.decoder_bin_filename
+
+        decoder_net = ncnn.Net()
+        decoder_net.opt.num_threads = 4
+
+        decoder_net.load_param(decoder_param)
+        decoder_net.load_model(decoder_model)
+
+        self.decoder_net = decoder_net
+
+    def init_joiner(self, args):
+        joiner_param = args.joiner_param_filename
+        joiner_model = args.joiner_bin_filename
+        joiner_net = ncnn.Net()
+        joiner_net.opt.num_threads = 4
+
+        joiner_net.load_param(joiner_param)
+        joiner_net.load_model(joiner_model)
+
+        self.joiner_net = joiner_net
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          x:
+            A tensor of shape (T, C)
+          states:
+            A list of tensors. len(states) == self.num_layers * 4
+        Returns:
+          Return a tuple containing:
+           - encoder_out, a tensor of shape (T, encoder_dim).
+           - next_states, a list of tensors containing the next states
+        """
+        with self.encoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(x.numpy()).clone())
+
+            # layer0 in2-in5
+            # layer1 in6-in9
+            for i in range(self.num_layers):
+                offset = 1 + i * 4
+                name = f"in{offset}"
+                # (32, 1, 512) -> (32, 512)
+                ex.input(name, ncnn.Mat(states[i * 4 + 0].numpy()).clone())
+
+                name = f"in{offset+1}"
+                #  (8, 1, 512) -> (8, 512)
+                ex.input(name, ncnn.Mat(states[i * 4 + 1].numpy()).clone())
+
+                name = f"in{offset+2}"
+                #  (8, 1, 512) -> (8, 512)
+                ex.input(name, ncnn.Mat(states[i * 4 + 2].numpy()).clone())
+
+                name = f"in{offset+3}"
+                #  (1, 512, 2) -> (512, 2)
+                ex.input(name, ncnn.Mat(states[i * 4 + 3].numpy()).clone())
+
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            encoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+
+            out_states: List[torch.Tensor] = []
+            for i in range(4 * self.num_layers):
+                name = f"out{i+1}"
+                ret, ncnn_out_state = ex.extract(name)
+                assert ret == 0, ret
+                ncnn_out_state = torch.from_numpy(ncnn_out_state.numpy())
+                out_states.append(ncnn_out_state)
+
+            return encoder_out, out_states
+
+    def run_decoder(self, decoder_input):
+        assert decoder_input.dtype == torch.int32
+
+        with self.decoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(decoder_input.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            decoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return decoder_out
+
+    def run_joiner(self, encoder_out, decoder_out):
+        with self.joiner_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(encoder_out.numpy()).clone())
+            ex.input("in1", ncnn.Mat(decoder_out.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            joiner_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return joiner_out
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: Model,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32)  # (1, context_size)
+        decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    else:
+        assert decoder_out.ndim == 1
+        assert hyp is not None, hyp
+
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t]
+
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32)
+            decoder_out = model.run_decoder(decoder_input).squeeze(0)
+
+    return hyp, decoder_out
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+
+    wave_samples = torch.cat([wave_samples, tail_padding])
+
+    states = model.get_init_states()
+
+    hyp = None
+    decoder_out = None
+
+    num_processed_frames = 0
+    segment = model.T
+    offset = model.chunk_length
+
+    chunk = int(1 * sample_rate)  # 0.2 second
+
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            encoder_out, states = model.run_encoder(frames, states)
+            hyp, decoder_out = greedy_search(model, encoder_out, decoder_out, hyp)
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    context_size = 2
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(sound_file)
+    logging.info(text)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py
new file mode 100755
index 000000000..f5d894a7b
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py
@@ -0,0 +1,963 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless2/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method greedy_search \
+      --use-averaged-model True
+
+(2) modified beam search
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless2/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method modified_beam_search \
+      --use-averaged-model True \
+      --beam-size 4
+
+(3) fast beam search
+./conv_emformer_transducer_stateless2/streaming_decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir conv_emformer_transducer_stateless2/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --chunk-length 32 \
+      --cnn-module-kernel 31 \
+      --left-context-length 32 \
+      --right-context-length 8 \
+      --memory-size 32 \
+      --decoding-method fast_beam_search \
+      --use-averaged-model True \
+      --beam 4 \
+      --max-contexts 4 \
+      --max-states 8
+"""
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from emformer import LOG_EPSILON, stack_states, unstack_states
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from stream import Stream
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import one_best_decoding
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--sampling-rate",
+        type=float,
+        default=16000,
+        help="Sample rate of the audio",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    T = encoder_out.size(1)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (batch_size, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+    beam: int = 4,
+):
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      beam:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    streams: List[Stream],
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      streams:
+        A list of stream objects.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyps[i]
+
+
+def decode_one_chunk(
+    model: nn.Module,
+    streams: List[Stream],
+    params: AttributeDict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> List[int]:
+    """
+    Args:
+      model:
+        The Transducer model.
+      streams:
+        A list of Stream objects.
+      params:
+        It is returned by :func:`get_params`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+       A list of indexes indicating the finished streams.
+    """
+    device = next(model.parameters()).device
+
+    feature_list = []
+    feature_len_list = []
+    state_list = []
+    num_processed_frames_list = []
+
+    for stream in streams:
+        # We should first get `stream.num_processed_frames`
+        # before calling `stream.get_feature_chunk()`
+        # since `stream.num_processed_frames` would be updated
+        num_processed_frames_list.append(stream.num_processed_frames)
+        feature = stream.get_feature_chunk()
+        feature_len = feature.size(0)
+        feature_list.append(feature)
+        feature_len_list.append(feature_len)
+        state_list.append(stream.states)
+
+    features = pad_sequence(
+        feature_list, batch_first=True, padding_value=LOG_EPSILON
+    ).to(device)
+    feature_lens = torch.tensor(feature_len_list, device=device)
+    num_processed_frames = torch.tensor(num_processed_frames_list, device=device)
+
+    # Make sure it has at least 1 frame after subsampling, first-and-last-frame cutting, and right context cutting  # noqa
+    tail_length = 3 * params.subsampling_factor + params.right_context_length + 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPSILON,
+        )
+
+    # Stack states of all streams
+    states = stack_states(state_list)
+
+    encoder_out, encoder_out_lens, states = model.encoder.infer(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        num_processed_frames=num_processed_frames,
+    )
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+    elif params.decoding_method == "fast_beam_search":
+        # feature_len is needed to get partial results.
+        # The rnnt_decoding_stream for fast_beam_search.
+        fast_beam_search_one_best(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            processed_lens=(num_processed_frames >> 2) + encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    # Update cached states of each stream
+    state_list = unstack_states(states)
+    for i, s in enumerate(state_list):
+        streams[i].states = s
+
+    finished_streams = [i for i, stream in enumerate(streams) if stream.done]
+    return finished_streams
+
+
+def create_streaming_feature_extractor() -> Fbank:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return Fbank(opts)
+
+
+def decode_dataset(
+    cuts: CutSet,
+    model: nn.Module,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+):
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The Transducer model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = next(model.parameters()).device
+
+    log_interval = 300
+
+    fbank = create_streaming_feature_extractor()
+
+    decode_results = []
+    streams = []
+    for num, cut in enumerate(cuts):
+        # Each utterance has a Stream.
+        stream = Stream(
+            params=params,
+            cut_id=cut.id,
+            decoding_graph=decoding_graph,
+            device=device,
+            LOG_EPS=LOG_EPSILON,
+        )
+
+        stream.set_states(model.encoder.init_states(device))
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+        feature = fbank(samples)
+        stream.set_feature(feature)
+        stream.set_ground_truth(cut.supervisions[0].text)
+
+        streams.append(stream)
+
+        while len(streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                model=model,
+                streams=streams,
+                params=params,
+                decoding_graph=decoding_graph,
+            )
+
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        streams[i].id,
+                        streams[i].ground_truth.split(),
+                        sp.decode(streams[i].decoding_result()).split(),
+                    )
+                )
+                del streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    while len(streams) > 0:
+        finished_streams = decode_one_chunk(
+            model=model,
+            streams=streams,
+            params=params,
+            decoding_graph=decoding_graph,
+        )
+
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    streams[i].id,
+                    streams[i].ground_truth.split(),
+                    sp.decode(streams[i].decoding_result()).split(),
+                )
+            )
+            del streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    else:
+        key = f"beam_size_{params.beam_size}"
+
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=sorted(results))
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-length-{params.chunk_length}"
+    params.suffix += f"-left-context-length-{params.left_context_length}"
+    params.suffix += f"-right-context-length-{params.right_context_length}"
+    params.suffix += f"-memory-size-{params.memory_size}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-streaming-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.device = device
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            model=model,
+            params=params,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220410)
+    main()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py
new file mode 100644
index 000000000..8cde6205b
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py
@@ -0,0 +1,194 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from emformer import ConvolutionModule, Emformer, stack_states, unstack_states
+
+
+def test_convolution_module_forward():
+    B, D = 2, 256
+    chunk_length = 4
+    right_context_length = 2
+    num_chunks = 3
+    U = num_chunks * chunk_length
+    R = num_chunks * right_context_length
+    kernel_size = 31
+    conv_module = ConvolutionModule(
+        chunk_length,
+        right_context_length,
+        D,
+        kernel_size,
+    )
+
+    utterance = torch.randn(U, B, D)
+    right_context = torch.randn(R, B, D)
+
+    utterance, right_context = conv_module(utterance, right_context)
+    assert utterance.shape == (U, B, D), utterance.shape
+    assert right_context.shape == (R, B, D), right_context.shape
+
+
+def test_convolution_module_infer():
+    from emformer import ConvolutionModule
+
+    B, D = 2, 256
+    chunk_length = 4
+    right_context_length = 2
+    num_chunks = 1
+    U = num_chunks * chunk_length
+    R = num_chunks * right_context_length
+    kernel_size = 31
+    conv_module = ConvolutionModule(
+        chunk_length,
+        right_context_length,
+        D,
+        kernel_size,
+    )
+
+    utterance = torch.randn(U, B, D)
+    right_context = torch.randn(R, B, D)
+    cache = torch.randn(B, D, kernel_size - 1)
+
+    utterance, right_context, new_cache = conv_module.infer(
+        utterance, right_context, cache
+    )
+    assert utterance.shape == (U, B, D), utterance.shape
+    assert right_context.shape == (R, B, D), right_context.shape
+    assert new_cache.shape == (B, D, kernel_size - 1), new_cache.shape
+
+
+def test_state_stack_unstack():
+    num_features = 80
+    chunk_length = 32
+    encoder_dim = 512
+    num_encoder_layers = 2
+    kernel_size = 31
+    left_context_length = 32
+    right_context_length = 8
+    memory_size = 32
+
+    model = Emformer(
+        num_features=num_features,
+        chunk_length=chunk_length,
+        subsampling_factor=4,
+        d_model=encoder_dim,
+        num_encoder_layers=num_encoder_layers,
+        cnn_module_kernel=kernel_size,
+        left_context_length=left_context_length,
+        right_context_length=right_context_length,
+        memory_size=memory_size,
+    )
+
+    for batch_size in [1, 2]:
+        attn_caches = [
+            [
+                torch.zeros(memory_size, batch_size, encoder_dim),
+                torch.zeros(left_context_length // 4, batch_size, encoder_dim),
+                torch.zeros(
+                    left_context_length // 4,
+                    batch_size,
+                    encoder_dim,
+                ),
+            ]
+            for _ in range(num_encoder_layers)
+        ]
+        conv_caches = [
+            torch.zeros(batch_size, encoder_dim, kernel_size - 1)
+            for _ in range(num_encoder_layers)
+        ]
+        states = [attn_caches, conv_caches]
+        x = torch.randn(batch_size, 23, num_features)
+        x_lens = torch.full((batch_size,), 23)
+        num_processed_frames = torch.full((batch_size,), 0)
+        y, y_lens, states = model.infer(
+            x, x_lens, num_processed_frames=num_processed_frames, states=states
+        )
+
+        state_list = unstack_states(states)
+        states2 = stack_states(state_list)
+
+        for ss, ss2 in zip(states[0], states2[0]):
+            for s, s2 in zip(ss, ss2):
+                assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+        for s, s2 in zip(states[1], states2[1]):
+            assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+
+def test_torchscript_consistency_infer():
+    r"""Verify that scripting Emformer does not change the behavior of method `infer`."""  # noqa
+    num_features = 80
+    chunk_length = 32
+    encoder_dim = 512
+    num_encoder_layers = 2
+    kernel_size = 31
+    left_context_length = 32
+    right_context_length = 8
+    memory_size = 32
+    batch_size = 2
+
+    model = Emformer(
+        num_features=num_features,
+        chunk_length=chunk_length,
+        subsampling_factor=4,
+        d_model=encoder_dim,
+        num_encoder_layers=num_encoder_layers,
+        cnn_module_kernel=kernel_size,
+        left_context_length=left_context_length,
+        right_context_length=right_context_length,
+        memory_size=memory_size,
+    ).eval()
+    attn_caches = [
+        [
+            torch.zeros(memory_size, batch_size, encoder_dim),
+            torch.zeros(left_context_length // 4, batch_size, encoder_dim),
+            torch.zeros(
+                left_context_length // 4,
+                batch_size,
+                encoder_dim,
+            ),
+        ]
+        for _ in range(num_encoder_layers)
+    ]
+    conv_caches = [
+        torch.zeros(batch_size, encoder_dim, kernel_size - 1)
+        for _ in range(num_encoder_layers)
+    ]
+    states = [attn_caches, conv_caches]
+    x = torch.randn(batch_size, 23, num_features)
+    x_lens = torch.full((batch_size,), 23)
+    num_processed_frames = torch.full((batch_size,), 0)
+    y, y_lens, out_states = model.infer(
+        x, x_lens, num_processed_frames=num_processed_frames, states=states
+    )
+
+    sc_model = torch.jit.script(model).eval()
+    sc_y, sc_y_lens, sc_out_states = sc_model.infer(
+        x, x_lens, num_processed_frames=num_processed_frames, states=states
+    )
+
+    assert torch.allclose(y, sc_y)
+
+
+if __name__ == "__main__":
+    test_convolution_module_forward()
+    test_convolution_module_infer()
+    test_state_stack_unstack()
+    test_torchscript_consistency_infer()
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py
new file mode 100755
index 000000000..23ddb6bec
--- /dev/null
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py
@@ -0,0 +1,1117 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conv_emformer_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 280 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+# For mix precision training:
+./conv_emformer_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir conv_emformer_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --master-port 12321 \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from emformer import Emformer
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dim for the Emformer",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads for the Emformer",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feed-forward dimension for the Emformer",
+    )
+
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of encoder layers for the Emformer",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=int,
+        default=31,
+        help="Kernel size for the convolution module.",
+    )
+
+    parser.add_argument(
+        "--left-context-length",
+        type=int,
+        default=32,
+        help="""Number of frames before subsampling for left context
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--chunk-length",
+        type=int,
+        default=32,
+        help="""Number of frames before subsampling for each chunk
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--right-context-length",
+        type=int,
+        default=8,
+        help="""Number of frames before subsampling for right context
+        in the Emformer.""",
+    )
+
+    parser.add_argument(
+        "--memory-size",
+        type=int,
+        default=0,
+        help="Number of entries in the memory for the Emformer",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for Emformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Emformer(
+        num_features=params.feature_dim,
+        chunk_length=params.chunk_length,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        cnn_module_kernel=params.cnn_module_kernel,
+        left_context_length=params.left_context_length,
+        right_context_length=params.right_context_length,
+        memory_size=params.memory_size,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/distillation_with_hubert.sh b/egs/librispeech/ASR/distillation_with_hubert.sh
new file mode 100755
index 000000000..a5b0b85af
--- /dev/null
+++ b/egs/librispeech/ASR/distillation_with_hubert.sh
@@ -0,0 +1,229 @@
+#!/usr/bin/env bash
+#
+# A short introduction about distillation framework.
+#
+# A typical traditional distillation method is
+# Loss(teacher embedding, student embedding).
+#
+# Comparing to these, the proposed distillation framework contains two mainly steps:
+# codebook indexes = quantizer.encode(teacher embedding)
+# Loss(codebook indexes, student embedding)
+#
+# Things worth to meantion:
+# 1. The float type teacher embedding is quantized into a sequence of
+#    8-bit integer codebook indexes.
+# 2. a middle layer 36(1-based) out of total 48 layers is used to extract
+#    teacher embeddings.
+# 3. a middle layer 6(1-based) out of total 6 layers is used to extract
+#    student embeddings.
+#
+# To directly download the extracted codebook indexes for model distillation, you can
+# set stage=2, stop_stage=4, use_extracted_codebook=True
+#
+# To start from scratch, you can
+# set stage=0, stop_stage=4, use_extracted_codebook=False
+
+stage=0
+stop_stage=4
+
+# Set the GPUs available.
+# This script requires at least one GPU.
+# You MUST set environment variable "CUDA_VISIBLE_DEVICES",
+# even you only have ONE GPU. It needed by CodebookIndexExtractor to determine numbert of jobs to extract codebook indexes parallelly.
+
+# Suppose only one GPU exists:
+# export CUDA_VISIBLE_DEVICES="0"
+#
+# Suppose GPU 2,3,4,5 are available.
+# export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+exp_dir=./pruned_transducer_stateless6/exp
+mkdir -p $exp_dir
+
+# full_libri can be "True" or "False"
+#   "True" -> use full librispeech dataset for distillation
+#   "False" -> use train-clean-100 subset for distillation
+full_libri=True
+
+# use_extracted_codebook can be "True" or "False"
+#   "True" -> stage 0 and stage 1 would be skipped,
+#     and directly download the extracted codebook indexes for distillation
+#   "False" -> start from scratch
+use_extracted_codebook=True
+
+# teacher_model_id can be one of
+#   "hubert_xtralarge_ll60k_finetune_ls960" -> fine-tuned model, it is the one we currently use.
+#   "hubert_xtralarge_ll60k" -> pretrained model without fintuing
+teacher_model_id=hubert_xtralarge_ll60k_finetune_ls960
+
+. shared/parse_options.sh || exit 1
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ] && [ ! "$use_extracted_codebook" == "True" ]; then
+  log "Stage 0: Download HuBERT model"
+  # Preparation stage.
+
+  # Install fairseq according to:
+  # https://github.com/pytorch/fairseq
+  # when testing this code:
+  # commit 806855bf660ea748ed7ffb42fe8dcc881ca3aca0 is used.
+  has_fairseq=$(python3 -c "import importlib; print(importlib.util.find_spec('fairseq') is not None)")
+  if [ $has_fairseq == 'False' ]; then
+    log "Please install fairseq before running following stages"
+    exit 1
+  fi
+
+  # Install quantization toolkit:
+  # pip install git+https://github.com/k2-fsa/multi_quantization.git
+  # or
+  # pip install multi_quantization
+
+  has_quantization=$(python3 -c "import importlib; print(importlib.util.find_spec('multi_quantization') is not None)")
+  if [ $has_quantization == 'False' ]; then
+    log "Please install multi_quantization before running following stages"
+    exit 1
+  fi
+
+  log "Download HuBERT model."
+  # Parameters about model.
+  hubert_model_dir=${exp_dir}/hubert_models
+  hubert_model=${hubert_model_dir}/${teacher_model_id}.pt
+  mkdir -p ${hubert_model_dir}
+  # For more models refer to: https://github.com/pytorch/fairseq/tree/main/examples/hubert
+  if [ -f ${hubert_model} ]; then
+    log "HuBERT model alread exists."
+  else
+    wget -c https://dl.fbaipublicfiles.com/hubert/${teacher_model_id}.pt -P ${hubert_model_dir}
+    wget -c wget https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt -P ${hubert_model_dir}
+  fi
+fi
+
+if [ ! -d ./data/fbank ]; then
+  log "This script assumes ./data/fbank is already generated by prepare.sh"
+  exit 1
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ] && [ ! "$use_extracted_codebook" == "True" ]; then
+  log "Stage 1: Verify that the downloaded HuBERT model is correct."
+  # This stage is not directly used by codebook indexes extraction.
+  # It is a method to "prove" that the downloaed hubert model
+  # is inferenced in an correct way if WERs look like normal.
+  # Expect WERs:
+  # [test-clean-ctc_greedy_search] %WER 2.04% [1075 / 52576, 92 ins, 104 del, 879 sub ]
+  # [test-other-ctc_greedy_search] %WER 3.71% [1942 / 52343, 152 ins, 126 del, 1664 sub ]
+  ./pruned_transducer_stateless6/hubert_decode.py --exp-dir $exp_dir
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  # Analysis of disk usage:
+  # With num_codebooks==8, each teacher embedding is quantized into
+  # a sequence of eight 8-bit integers, i.e. only eight bytes are needed.
+  # Training dataset including clean-100h with speed perturb 0.9 and 1.1 has 300 hours.
+  # The output frame rates of Hubert is 50 per second.
+  # Theoretically, 412M = 300 * 3600 * 50 * 8 / 1024 / 1024 is needed.
+  # The actual size of all "*.h5" files storaging codebook index is 450M.
+  # I think the extra "48M" usage is some meta information.
+
+  # Time consumption analysis:
+  # For quantizer training data(teacher embedding) extraction, only 1000 utts from clean-100 are used.
+  # Together with quantizer training, no more than 20 minutes will be used.
+  #
+  # For codebook indexes extraction,
+  # with two pieces of NVIDIA A100 gpus, around three hours needed to process 300 hours training data,
+  # i.e. clean-100 with speed purteb 0.9 and 1.1.
+
+  # GPU usage:
+  # During quantizer's training data(teacher embedding) and it's training,
+  # only the first ONE GPU is used.
+  # During codebook indexes extraction, ALL GPUs set by CUDA_VISIBLE_DEVICES are used.
+
+  if [ "$use_extracted_codebook" == "True" ]; then
+    if [ ! "$teacher_model_id" == "hubert_xtralarge_ll60k_finetune_ls960" ]; then
+      log "Currently we only uploaded codebook indexes from teacher model hubert_xtralarge_ll60k_finetune_ls960"
+      exit 1
+    fi
+    # The codebook indexes to be downloaded are generated using the following setup:
+    embedding_layer=36
+    num_codebooks=8
+
+    mkdir -p $exp_dir/vq
+    codebook_dir=$exp_dir/vq/${teacher_model_id}
+    mkdir -p codebook_dir
+    codebook_download_dir=$exp_dir/download_codebook
+    if [ -d $codebook_download_dir ]; then
+      log "$codebook_download_dir exists, you should remove it first."
+      exit 1
+    fi
+    log "Downloading extracted codebook indexes to $codebook_download_dir"
+    # Make sure you have git-lfs installed (https://git-lfs.github.com)
+    # The codebook indexes are generated using lhotse 1.11.0, to avoid
+    # potential issues, we recommend you to use lhotse version >= 1.11.0
+    lhotse_version=$(python3 -c "import lhotse; from packaging import version; print(version.parse(lhotse.version.__version__)>=version.parse('1.11.0'))")
+    if [ "$lhotse_version" == "False" ]; then
+      log "Expecting lhotse >= 1.11.0. This may lead to potential ID mismatch."
+    fi
+    git lfs install
+    git clone https://huggingface.co/marcoyang/pruned_transducer_stateless6_hubert_xtralarge_ll60k_finetune_ls960 $codebook_download_dir
+
+    vq_fbank=data/vq_fbank_layer${embedding_layer}_cb${num_codebooks}/
+    mkdir -p $vq_fbank
+    mv $codebook_download_dir/*.jsonl.gz $vq_fbank
+    mkdir -p $codebook_dir/splits4
+    mv $codebook_download_dir/*.h5 $codebook_dir/splits4/
+    log "Remove $codebook_download_dir"
+    rm -rf $codebook_download_dir
+  fi
+
+  ./pruned_transducer_stateless6/extract_codebook_index.py \
+    --full-libri $full_libri \
+    --exp-dir $exp_dir \
+    --embedding-layer $embedding_layer \
+    --num-utts 1000 \
+    --num-codebooks $num_codebooks \
+    --max-duration 100 \
+    --teacher-model-id $teacher_model_id \
+    --use-extracted-codebook $use_extracted_codebook
+
+  if [ "$full_libri" == "True" ]; then
+    # Merge the 3 subsets and create a full one
+    rm ${vq_fbank}/librispeech_cuts_train-all-shuf.jsonl.gz
+    cat <(gunzip -c ${vq_fbank}/librispeech_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c ${vq_fbank}/librispeech_cuts_train-clean-360.jsonl.gz) \
+      <(gunzip -c ${vq_fbank}/librispeech_cuts_train-other-500.jsonl.gz) | \
+      shuf | gzip -c > ${vq_fbank}/librispeech_cuts_train-all-shuf.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  # Example training script.
+  # Note: it's better to set spec-aug-time-warpi-factor=-1
+  WORLD_SIZE=$(echo ${CUDA_VISIBLE_DEVICES} | awk '{n=split($1, _, ","); print n}')
+  ./pruned_transducer_stateless6/train.py \
+    --manifest-dir ./data/vq_fbank \
+    --master-port 12359 \
+    --full-libri $full_libri \
+    --spec-aug-time-warp-factor -1 \
+    --max-duration 300 \
+    --world-size ${WORLD_SIZE} \
+    --num-epochs 20 \
+    --exp-dir $exp_dir \
+    --enable-distillation True
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  # Results should be similar to:
+  # errs-test-clean-beam_size_4-epoch-20-avg-10-beam-4.txt:%WER = 5.67
+  # errs-test-other-beam_size_4-epoch-20-avg-10-beam-4.txt:%WER = 15.60
+  ./pruned_transducer_stateless6/decode.py \
+    --decoding-method "modified_beam_search" \
+    --epoch 20 \
+    --avg 10 \
+    --max-duration 200 \
+    --exp-dir $exp_dir \
+    --enable-distillation True
+fi
diff --git a/egs/librispeech/ASR/finetune.sh b/egs/librispeech/ASR/finetune.sh
new file mode 100755
index 000000000..bc6357312
--- /dev/null
+++ b/egs/librispeech/ASR/finetune.sh
@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# This is an example script for fine-tuning. Here, we fine-tune a model trained
+# on Librispeech on GigaSpeech. The model used for fine-tuning is 
+# pruned_transducer_stateless7 (zipformer). If you want to fine-tune model 
+# from another recipe, you can adapt ./pruned_transducer_stateless7/finetune.py 
+# for that recipe. If you have any problem, please open up an issue in https://github.com/k2-fsa/icefall/issues.
+
+# We assume that you have already prepared the GigaSpeech manfiest&features under ./data.
+# If you haven't done that, please see https://github.com/k2-fsa/icefall/blob/master/egs/gigaspeech/ASR/prepare.sh.
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download Pre-trained model"
+  
+  # clone from huggingface
+  git lfs install
+  git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Start fine-tuning"
+  
+  # The following configuration of lr schedule should work well
+  # You may also tune the following parameters to adjust learning rate schedule
+  base_lr=0.005
+  lr_epochs=100
+  lr_batches=100000
+
+  # We recommend to start from an averaged model
+  finetune_ckpt=icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp/pretrained.pt
+  export CUDA_VISIBLE_DEVICES="0,1"
+
+  ./pruned_transducer_stateless7/finetune.py \
+    --world-size 2 \
+    --master-port 18180 \
+    --num-epochs 20 \
+    --start-epoch 1 \
+    --exp-dir pruned_transducer_stateless7/exp_giga_finetune \
+    --subset S \
+    --use-fp16 1 \
+    --base-lr $base_lr \
+    --lr-epochs $lr_epochs \
+    --lr-batches $lr_batches \
+    --bpe-model icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/data/lang_bpe_500/bpe.model \
+    --do-finetune True \
+    --finetune-ckpt $finetune_ckpt \
+    --max-duration 500
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Decoding"
+
+  epoch=15
+  avg=10
+
+  for m in greedy_search modified_beam_search; do
+    python pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch $epoch \
+    --avg $avg \
+    --use-averaged-model True \
+    --beam-size 4 \
+    --exp-dir pruned_transducer_stateless7/exp_giga_finetune \
+    --bpe-model icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/data/lang_bpe_500/bpe.model \
+    --max-duration 400 \
+    --decoding-method $m
+  done
+fi
diff --git a/egs/librispeech/ASR/generate-lm.sh b/egs/librispeech/ASR/generate-lm.sh
new file mode 100755
index 000000000..dacd276d1
--- /dev/null
+++ b/egs/librispeech/ASR/generate-lm.sh
@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+
+lang_dir=data/lang_bpe_500
+
+for ngram in 2 3 4 5; do
+  if [ ! -f $lang_dir/${ngram}gram.arpa ]; then
+    ./shared/make_kn_lm.py \
+      -ngram-order ${ngram} \
+      -text $lang_dir/transcript_tokens.txt \
+      -lm $lang_dir/${ngram}gram.arpa
+  fi
+
+  if [ ! -f $lang_dir/${ngram}gram.fst.txt ]; then
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_dir/tokens.txt" \
+      --disambig-symbol='#0' \
+      --max-order=${ngram} \
+      $lang_dir/${ngram}gram.arpa > $lang_dir/${ngram}gram.fst.txt
+  fi
+done
diff --git a/egs/librispeech/ASR/local/__init__.py b/egs/librispeech/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/local/add_alignment_librispeech.py b/egs/librispeech/ASR/local/add_alignment_librispeech.py
new file mode 100755
index 000000000..cc34a72d8
--- /dev/null
+++ b/egs/librispeech/ASR/local/add_alignment_librispeech.py
@@ -0,0 +1,190 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file adds alignments from https://github.com/CorentinJ/librispeech-alignments  # noqa
+to the existing fbank features dir (e.g., data/fbank)
+and save cuts to a new dir (e.g., data/fbank_ali).
+"""
+
+import argparse
+import logging
+import zipfile
+from pathlib import Path
+from typing import List
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.recipes.librispeech import parse_alignments
+from lhotse.utils import is_module_available
+
+LIBRISPEECH_ALIGNMENTS_URL = (
+    "https://drive.google.com/uc?id=1WYfgr31T-PPwMcxuAq09XZfHQO5Mw8fE"
+)
+
+DATASET_PARTS = [
+    "dev-clean",
+    "dev-other",
+    "test-clean",
+    "test-other",
+    "train-clean-100",
+    "train-clean-360",
+    "train-other-500",
+]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--alignments-dir",
+        type=str,
+        default="data/alignment",
+        help="The dir to save alignments.",
+    )
+
+    parser.add_argument(
+        "--cuts-in-dir",
+        type=str,
+        default="data/fbank",
+        help="The dir of the existing cuts without alignments.",
+    )
+
+    parser.add_argument(
+        "--cuts-out-dir",
+        type=str,
+        default="data/fbank_ali",
+        help="The dir to save the new cuts with alignments",
+    )
+
+    return parser
+
+
+def download_alignments(
+    target_dir: str, alignments_url: str = LIBRISPEECH_ALIGNMENTS_URL
+):
+    """
+    Download and extract the alignments.
+
+    Note: If you can not access drive.google.com, you could download the file
+    `LibriSpeech-Alignments.zip` from huggingface:
+    https://huggingface.co/Zengwei/librispeech-alignments
+    and extract the zip file manually.
+
+    Args:
+      target_dir:
+        The dir to save alignments.
+      alignments_url:
+        The URL of alignments.
+    """
+    """Modified from https://github.com/lhotse-speech/lhotse/blob/master/lhotse/recipes/librispeech.py"""  # noqa
+    target_dir = Path(target_dir)
+    target_dir.mkdir(parents=True, exist_ok=True)
+    completed_detector = target_dir / ".ali_completed"
+    if completed_detector.is_file():
+        logging.info("The alignment files already exist.")
+        return
+
+    ali_zip_path = target_dir / "LibriSpeech-Alignments.zip"
+    if not ali_zip_path.is_file():
+        assert is_module_available(
+            "gdown"
+        ), 'To download LibriSpeech alignments, please install "pip install gdown"'  # noqa
+        import gdown
+
+        gdown.download(alignments_url, output=str(ali_zip_path))
+
+    with zipfile.ZipFile(str(ali_zip_path)) as f:
+        f.extractall(path=target_dir)
+        completed_detector.touch()
+
+
+def add_alignment(
+    alignments_dir: str,
+    cuts_in_dir: str = "data/fbank",
+    cuts_out_dir: str = "data/fbank_ali",
+    dataset_parts: List[str] = DATASET_PARTS,
+):
+    """
+    Add alignment info to existing cuts.
+
+    Args:
+      alignments_dir:
+        The dir of the alignments.
+      cuts_in_dir:
+        The dir of the existing cuts.
+      cuts_out_dir:
+        The dir to save the new cuts with alignments.
+      dataset_parts:
+        Librispeech parts to add alignments.
+    """
+    alignments_dir = Path(alignments_dir)
+    cuts_in_dir = Path(cuts_in_dir)
+    cuts_out_dir = Path(cuts_out_dir)
+    cuts_out_dir.mkdir(parents=True, exist_ok=True)
+
+    for part in dataset_parts:
+        logging.info(f"Processing {part}")
+
+        cuts_in_path = cuts_in_dir / f"librispeech_cuts_{part}.jsonl.gz"
+        if not cuts_in_path.is_file():
+            logging.info(f"{cuts_in_path} does not exist - skipping.")
+            continue
+        cuts_out_path = cuts_out_dir / f"librispeech_cuts_{part}.jsonl.gz"
+        if cuts_out_path.is_file():
+            logging.info(f"{part} already exists - skipping.")
+            continue
+
+        # parse alignments
+        alignments = {}
+        part_ali_dir = alignments_dir / "LibriSpeech" / part
+        for ali_path in part_ali_dir.rglob("*.alignment.txt"):
+            ali = parse_alignments(ali_path)
+            alignments.update(ali)
+        logging.info(f"{part} has {len(alignments.keys())} cuts with alignments.")
+
+        # add alignment attribute and write out
+        cuts_in = load_manifest_lazy(cuts_in_path)
+        with CutSet.open_writer(cuts_out_path) as writer:
+            for cut in cuts_in:
+                for idx, subcut in enumerate(cut.supervisions):
+                    origin_id = subcut.id.split("_")[0]
+                    if origin_id in alignments:
+                        ali = alignments[origin_id]
+                    else:
+                        logging.info(f"Warning: {origin_id} does not have alignment.")
+                        ali = []
+                    subcut.alignment = {"word": ali}
+                writer.write(cut, flush=True)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    download_alignments(args.alignments_dir)
+    add_alignment(args.alignments_dir, args.cuts_in_dir, args.cuts_out_dir)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/compile_hlg.py b/egs/librispeech/ASR/local/compile_hlg.py
new file mode 100755
index 000000000..d19d50ae6
--- /dev/null
+++ b/egs/librispeech/ASR/local/compile_hlg.py
@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_n_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lm",
+        type=str,
+        default="G_3_gram",
+        help="""Stem name for LM used in HLG compiling.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+      lm:
+        The language stem base name.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path(f"data/lm/{lm}.pt").is_file():
+        logging.info(f"Loading pre-compiled {lm}")
+        d = torch.load(f"data/lm/{lm}.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info(f"Loading {lm}.fst.txt")
+        with open(f"data/lm/{lm}.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), f"data/lm/{lm}.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    # LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # see https://github.com/k2-fsa/k2/pull/1140
+    labels = LG.labels
+    labels[labels >= first_token_disambig_id] = 0
+    LG.labels = labels
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lang_dir, args.lm)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/compile_hlg_using_openfst.py b/egs/librispeech/ASR/local/compile_hlg_using_openfst.py
new file mode 100755
index 000000000..15fc47ef1
--- /dev/null
+++ b/egs/librispeech/ASR/local/compile_hlg_using_openfst.py
@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.fst
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_n_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG_fst.pt
+
+So when to use this script instead of ./local/compile_hlg.py ?
+If you have a very large G, ./local/compile_hlg.py may throw OOM for
+determinization. In that case, you can use this script to compile HLG.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import kaldifst
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lm",
+        type=str,
+        default="G_3_gram",
+        help="""Stem name for LM used in HLG compiling.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str, lm: str = "G_3_gram") -> kaldifst.StdVectorFst:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+      lm:
+        The language stem base name.
+
+    Return:
+      An FST representing HLG.
+    """
+
+    L = kaldifst.StdVectorFst.read(f"{lang_dir}/L_disambig.fst")
+    logging.info("Arc sort L")
+    kaldifst.arcsort(L, sort_type="olabel")
+    logging.info(f"L: #states {L.num_states}")
+
+    G_filename_txt = f"data/lm/{lm}.fst.txt"
+    G_filename_binary = f"data/lm/{lm}.fst"
+    if Path(G_filename_binary).is_file():
+        logging.info(f"Loading {G_filename_binary}")
+        G = kaldifst.StdVectorFst.read(G_filename_binary)
+    else:
+        logging.info(f"Loading {G_filename_txt}")
+        with open(G_filename_txt) as f:
+            G = kaldifst.compile(s=f.read(), acceptor=False)
+            logging.info(f"Saving G to {G_filename_binary}")
+            G.write(G_filename_binary)
+
+    logging.info("Arc sort G")
+    kaldifst.arcsort(G, sort_type="ilabel")
+
+    logging.info(f"G: #states {G.num_states}")
+
+    logging.info("Compose L and G and connect LG")
+    LG = kaldifst.compose(L, G, connect=True)
+    logging.info(f"LG: #states {LG.num_states}")
+
+    logging.info("Determinizestar LG")
+    kaldifst.determinize_star(LG)
+    logging.info(f"LG after determinize_star: #states {LG.num_states}")
+
+    logging.info("Minimize encoded LG")
+    kaldifst.minimize_encoded(LG)
+    logging.info(f"LG after minimize_encoded: #states {LG.num_states}")
+
+    logging.info("Converting LG to k2 format")
+    LG = k2.Fsa.from_openfst(LG.to_str(is_acceptor=False), acceptor=False)
+    logging.info(f"LG in k2: #states: {LG.shape[0]}, #arcs: {LG.num_arcs}")
+
+    lexicon = Lexicon(lang_dir)
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+    logging.info(f"token id for #0: {first_token_disambig_id}")
+    logging.info(f"word id for #0: {first_word_disambig_id}")
+
+    max_token_id = max(lexicon.tokens)
+    modified = False
+    logging.info(
+        f"Building ctc_topo. modified: {modified}, max_token_id: {max_token_id}"
+    )
+
+    H = k2.ctc_topo(max_token_id, modified=modified)
+    logging.info(f"H: #states: {H.shape[0]}, #arcs: {H.num_arcs}")
+
+    logging.info("Removing disambiguation symbols on LG")
+    LG.labels[LG.labels >= first_token_disambig_id] = 0
+    LG.aux_labels[LG.aux_labels >= first_word_disambig_id] = 0
+
+    # See https://github.com/k2-fsa/k2/issues/874
+    # for why we need to set LG.properties to None
+    LG.__dict__["_properties"] = None
+
+    logging.info("Removing epsilons from LG")
+    LG = k2.remove_epsilon(LG)
+    logging.info(
+        f"LG after k2.remove_epsilon: #states: {LG.shape[0]}, #arcs: {LG.num_arcs}"
+    )
+
+    logging.info("Connecting LG after removing epsilons")
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+    logging.info(f"LG after k2.connect: #states: {LG.shape[0]}, #arcs: {LG.num_arcs}")
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+    logging.info(
+        f"HLG after k2.compose: #states: {HLG.shape[0]}, #arcs: {HLG.num_arcs}"
+    )
+
+    logging.info("Connecting HLG")
+    HLG = k2.connect(HLG)
+    logging.info(
+        f"HLG after k2.connect: #states: {HLG.shape[0]}, #arcs: {HLG.num_arcs}"
+    )
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    filename = lang_dir / "HLG_fst.pt"
+
+    if filename.is_file():
+        logging.info(f"{filename} already exists - skipping")
+        return
+
+    HLG = compile_HLG(lang_dir, args.lm)
+    logging.info(f"Saving HLG to {filename}")
+    torch.save(HLG.as_dict(), filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/compile_lg.py b/egs/librispeech/ASR/local/compile_lg.py
new file mode 100755
index 000000000..709b14070
--- /dev/null
+++ b/egs/librispeech/ASR/local/compile_lg.py
@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input lang_dir and generates LG from
+
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_3_gram.fst.txt
+
+The generated LG is saved in $lang_dir/LG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+    parser.add_argument(
+        "--lm",
+        type=str,
+        default="G_3_gram",
+        help="""Stem name for LM used in HLG compiling.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_LG(lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+
+    Return:
+      An FSA representing LG.
+    """
+    lexicon = Lexicon(lang_dir)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path(f"data/lm/{lm}.pt").is_file():
+        logging.info(f"Loading pre-compiled {lm}")
+        d = torch.load(f"data/lm/{lm}.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info(f"Loading {lm}.fst.txt")
+        with open(f"data/lm/{lm}.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), f"data/lm/{lm}.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG, k2.DeterminizeWeightPushingType.kLogWeightPushing)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    # LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # see https://github.com/k2-fsa/k2/pull/1140
+    labels = LG.labels
+    labels[labels >= first_token_disambig_id] = 0
+    LG.labels = labels
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    return LG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "LG.pt").is_file():
+        logging.info(f"{lang_dir}/LG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    LG = compile_LG(lang_dir, args.lm)
+    logging.info(f"Saving LG.pt to {lang_dir}")
+    torch.save(LG.as_dict(), f"{lang_dir}/LG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py
new file mode 100644
index 000000000..97750f3ea
--- /dev/null
+++ b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py
@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_gigaspeech_dev_test():
+    in_out_dir = Path("data/fbank")
+    # number of workers in dataloader
+    num_workers = 20
+
+    # number of seconds in a batch
+    batch_duration = 600
+
+    subsets = ("DEV", "TEST")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    prefix = "gigaspeech"
+    suffix = "jsonl.gz"
+
+    for partition in subsets:
+        cuts_path = in_out_dir / f"{prefix}_cuts_{partition}.{suffix}"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = in_out_dir / f"{prefix}_cuts_{partition}_raw.{suffix}"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{in_out_dir}/{prefix}_feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            overwrite=True,
+        )
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_gigaspeech_dev_test()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py
new file mode 100644
index 000000000..ce0ef24e7
--- /dev/null
+++ b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py
@@ -0,0 +1,170 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import os
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the XL subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+    return parser
+
+
+def compute_fbank_gigaspeech_splits(args):
+    num_splits = args.num_splits
+    output_dir = f"data/fbank/gigaspeech_XL_split_{num_splits}"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    prefix = "gigaspeech"
+
+    num_digits = 8  # num_digits is fixed by lhotse split-lazy
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"{prefix}_cuts_XL.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"{prefix}_cuts_XL_raw.{idx}.jsonl.gz"
+        if not raw_cuts_path.is_file():
+            logging.info(f"{raw_cuts_path} does not exist - skipping it")
+            continue
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+        if (output_dir / f"{prefix}_feats_XL_{idx}.lca").exists():
+            logging.info(f"Removing {output_dir}/{prefix}_feats_XL_{idx}.lca")
+            os.remove(output_dir / f"{prefix}_feats_XL_{idx}.lca")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_XL_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            overwrite=True,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+
+    log_filename = "log-compute_fbank_gigaspeech_splits"
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    log_filename = f"{log_filename}-{date_time}"
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=logging.INFO,
+        filemode="w",
+    )
+
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter(formatter))
+    logging.getLogger("").addHandler(console)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    compute_fbank_gigaspeech_splits(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/compute_fbank_librispeech.py b/egs/librispeech/ASR/local/compute_fbank_librispeech.py
new file mode 100755
index 000000000..25d6050bb
--- /dev/null
+++ b/egs/librispeech/ASR/local/compute_fbank_librispeech.py
@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=True,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_librispeech(
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = (
+            "dev-clean",
+            "dev-other",
+            "test-clean",
+            "test-other",
+            "train-clean-100",
+            "train-clean-360",
+            "train-other-500",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = "librispeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+
+            if "train" in partition:
+                if bpe_model:
+                    cut_set = filter_cuts(cut_set, sp)
+                if perturb_speed:
+                    logging.info(f"Doing speed perturb")
+                    cut_set = (
+                        cut_set
+                        + cut_set.perturb_speed(0.9)
+                        + cut_set.perturb_speed(1.1)
+                    )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_librispeech(
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/librispeech/ASR/local/compute_fbank_musan.py b/egs/librispeech/ASR/local/compute_fbank_musan.py
new file mode 100755
index 000000000..62036467e
--- /dev/null
+++ b/egs/librispeech/ASR/local/compute_fbank_musan.py
@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter, MonoCut, combine
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def is_cut_long(c: MonoCut) -> bool:
+    return c.duration > 5
+
+
+def compute_fbank_musan():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    musan_cuts_path = output_dir / "musan_cuts.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        # create chunks of Musan with duration 5 - 10 seconds
+        musan_cuts = (
+            CutSet.from_manifests(
+                recordings=combine(part["recordings"] for part in manifests.values())
+            )
+            .cut_into_windows(10.0)
+            .filter(is_cut_long)
+            .compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/musan_feats",
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        musan_cuts.to_file(musan_cuts_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()
diff --git a/egs/librispeech/ASR/local/convert_transcript_words_to_tokens.py b/egs/librispeech/ASR/local/convert_transcript_words_to_tokens.py
new file mode 100755
index 000000000..a8d5117c9
--- /dev/null
+++ b/egs/librispeech/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+"""
+Convert a transcript file containing words to a corpus file containing tokens
+for LM training with the help of a lexicon.
+
+If the lexicon contains phones, the resulting LM will be a phone LM; If the
+lexicon contains word pieces, the resulting LM will be a word piece LM.
+
+If a word has multiple pronunciations, the one that appears first in the lexicon
+is kept; others are removed.
+
+If the input transcript is:
+
+    hello zoo world hello
+    world zoo
+    foo zoo world hellO
+
+and if the lexicon is
+
+    <UNK> SPN
+    hello h e l l o 2
+    hello h e l l o
+    world w o r l d
+    zoo z o o
+
+Then the output is
+
+    h e l l o 2 z o o w o r l d h e l l o 2
+    w o r l d z o o
+    SPN z o o w o r l d SPN
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+from generate_unique_lexicon import filter_multiple_pronunications
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="The input transcript file."
+        "We assume that the transcript file consists of "
+        "lines. Each line consists of space separated words.",
+    )
+    parser.add_argument("--lexicon", type=str, help="The input lexicon file.")
+    parser.add_argument("--oov", type=str, default="<UNK>", help="The OOV word.")
+
+    return parser.parse_args()
+
+
+def process_line(lexicon: Dict[str, List[str]], line: str, oov_token: str) -> None:
+    """
+    Args:
+      lexicon:
+        A dict containing pronunciations. Its keys are words and values
+        are pronunciations (i.e., tokens).
+      line:
+        A line of transcript consisting of space(s) separated words.
+      oov_token:
+        The pronunciation of the oov word if a word in `line` is not present
+        in the lexicon.
+    Returns:
+      Return None.
+    """
+    s = ""
+    words = line.strip().split()
+    for i, w in enumerate(words):
+        tokens = lexicon.get(w, oov_token)
+        s += " ".join(tokens)
+        s += " "
+    print(s.strip())
+
+
+def main():
+    args = get_args()
+    assert Path(args.lexicon).is_file()
+    assert Path(args.transcript).is_file()
+    assert len(args.oov) > 0
+
+    # Only the first pronunciation of a word is kept
+    lexicon = filter_multiple_pronunications(read_lexicon(args.lexicon))
+
+    lexicon = dict(lexicon)
+
+    assert args.oov in lexicon
+
+    oov_token = lexicon[args.oov]
+
+    with open(args.transcript) as f:
+        for line in f:
+            process_line(lexicon=lexicon, line=line, oov_token=oov_token)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/display_manifest_statistics.py b/egs/librispeech/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..c3c684235
--- /dev/null
+++ b/egs/librispeech/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,215 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    #  path = "./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz"
+    #  path = "./data/fbank/librispeech_cuts_train-clean-360.jsonl.gz"
+    #  path = "./data/fbank/librispeech_cuts_train-other-500.jsonl.gz"
+    #  path = "./data/fbank/librispeech_cuts_dev-clean.jsonl.gz"
+    #  path = "./data/fbank/librispeech_cuts_dev-other.jsonl.gz"
+    #  path = "./data/fbank/librispeech_cuts_test-clean.jsonl.gz"
+    path = "./data/fbank/librispeech_cuts_test-other.jsonl.gz"
+
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+## train-clean-100
+Cuts count: 85617
+Total duration (hours): 303.8
+Speech duration (hours): 303.8 (100.0%)
+***
+Duration statistics (seconds):
+mean    12.8
+std     3.8
+min     1.3
+0.1%    1.9
+0.5%    2.2
+1%      2.5
+5%      4.2
+10%     6.4
+25%     11.4
+50%     13.8
+75%     15.3
+90%     16.7
+95%     17.3
+99%     18.1
+99.5%   18.4
+99.9%   18.8
+max     27.2
+
+## train-clean-360
+Cuts count: 312042
+Total duration (hours): 1098.2
+Speech duration (hours): 1098.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    12.7
+std     3.8
+min     1.0
+0.1%    1.8
+0.5%    2.2
+1%      2.5
+5%      4.2
+10%     6.2
+25%     11.2
+50%     13.7
+75%     15.3
+90%     16.6
+95%     17.3
+99%     18.1
+99.5%   18.4
+99.9%   18.8
+max     33.0
+
+## train-other 500
+Cuts count: 446064
+Total duration (hours): 1500.6
+Speech duration (hours): 1500.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    12.1
+std     4.2
+min     0.8
+0.1%    1.7
+0.5%    2.1
+1%      2.3
+5%      3.5
+10%     5.0
+25%     9.8
+50%     13.4
+75%     15.1
+90%     16.5
+95%     17.2
+99%     18.1
+99.5%   18.4
+99.9%   18.9
+max     31.0
+
+## dev-clean
+Cuts count: 2703
+Total duration (hours): 5.4
+Speech duration (hours): 5.4 (100.0%)
+***
+Duration statistics (seconds):
+mean    7.2
+std     4.7
+min     1.4
+0.1%    1.6
+0.5%    1.8
+1%      1.9
+5%      2.4
+10%     2.7
+25%     3.8
+50%     5.9
+75%     9.3
+90%     13.3
+95%     16.4
+99%     23.8
+99.5%   28.5
+99.9%   32.3
+max     32.6
+
+## dev-other
+Cuts count: 2864
+Total duration (hours): 5.1
+Speech duration (hours): 5.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.4
+std     4.3
+min     1.1
+0.1%    1.3
+0.5%    1.7
+1%      1.8
+5%      2.2
+10%     2.6
+25%     3.5
+50%     5.3
+75%     7.9
+90%     12.0
+95%     15.0
+99%     22.2
+99.5%   27.1
+99.9%   32.4
+max     35.2
+
+## test-clean
+Cuts count: 2620
+Total duration (hours): 5.4
+Speech duration (hours): 5.4 (100.0%)
+***
+Duration statistics (seconds):
+mean    7.4
+std     5.2
+min     1.3
+0.1%    1.6
+0.5%    1.8
+1%      2.0
+5%      2.3
+10%     2.7
+25%     3.7
+50%     5.8
+75%     9.6
+90%     14.6
+95%     17.8
+99%     25.5
+99.5%   28.4
+99.9%   32.8
+max     35.0
+
+## test-other
+Cuts count: 2939
+Total duration (hours): 5.3
+Speech duration (hours): 5.3 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.5
+std     4.4
+min     1.2
+0.1%    1.5
+0.5%    1.8
+1%      1.9
+5%      2.3
+10%     2.6
+25%     3.4
+50%     5.2
+75%     8.2
+90%     12.6
+95%     15.8
+99%     21.4
+99.5%   23.8
+99.9%   33.5
+max     34.5
+"""
diff --git a/egs/librispeech/ASR/local/download_lm.py b/egs/librispeech/ASR/local/download_lm.py
new file mode 100755
index 000000000..5a36ff2a9
--- /dev/null
+++ b/egs/librispeech/ASR/local/download_lm.py
@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file downloads the following LibriSpeech LM files:
+
+    - 3-gram.pruned.1e-7.arpa.gz
+    - 4-gram.arpa.gz
+    - librispeech-vocab.txt
+    - librispeech-lexicon.txt
+    - librispeech-lm-norm.txt.gz
+
+from http://www.openslr.org/resources/11
+and save them in the user provided directory.
+
+Files are not re-downloaded if they already exist.
+
+Usage:
+    ./local/download_lm.py --out-dir ./download/lm
+"""
+
+import argparse
+import gzip
+import logging
+import os
+import shutil
+from pathlib import Path
+
+from tqdm.auto import tqdm
+
+
+# This function is copied from lhotse
+def tqdm_urlretrieve_hook(t):
+    """Wraps tqdm instance.
+    Don't forget to close() or __exit__()
+    the tqdm instance once you're done with it (easiest using `with` syntax).
+    Example
+    -------
+    >>> from urllib.request import urlretrieve
+    >>> with tqdm(...) as t:
+    ...     reporthook = tqdm_urlretrieve_hook(t)
+    ...     urlretrieve(..., reporthook=reporthook)
+
+    Source: https://github.com/tqdm/tqdm/blob/master/examples/tqdm_wget.py
+    """
+    last_b = [0]
+
+    def update_to(b=1, bsize=1, tsize=None):
+        """
+        b  : int, optional
+            Number of blocks transferred so far [default: 1].
+        bsize  : int, optional
+            Size of each block (in tqdm units) [default: 1].
+        tsize  : int, optional
+            Total size (in tqdm units). If [default: None] or -1,
+            remains unchanged.
+        """
+        if tsize not in (None, -1):
+            t.total = tsize
+        displayed = t.update((b - last_b[0]) * bsize)
+        last_b[0] = b
+        return displayed
+
+    return update_to
+
+
+# This function is copied from lhotse
+def urlretrieve_progress(url, filename=None, data=None, desc=None):
+    """
+    Works exactly like urllib.request.urlretrieve, but attaches a tqdm hook to
+    display a progress bar of the download.
+    Use "desc" argument to display a user-readable string that informs what is
+    being downloaded.
+    """
+    from urllib.request import urlretrieve
+
+    with tqdm(unit="B", unit_scale=True, unit_divisor=1024, miniters=1, desc=desc) as t:
+        reporthook = tqdm_urlretrieve_hook(t)
+        return urlretrieve(url=url, filename=filename, reporthook=reporthook, data=data)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--out-dir", type=str, help="Output directory.")
+
+    args = parser.parse_args()
+    return args
+
+
+def main(out_dir: str):
+    url = "http://www.openslr.org/resources/11"
+    out_dir = Path(out_dir)
+
+    files_to_download = (
+        "3-gram.pruned.1e-7.arpa.gz",
+        "4-gram.arpa.gz",
+        "librispeech-vocab.txt",
+        "librispeech-lexicon.txt",
+        "librispeech-lm-norm.txt.gz",
+    )
+
+    for f in tqdm(files_to_download, desc="Downloading LibriSpeech LM files"):
+        filename = out_dir / f
+        if filename.is_file() is False:
+            urlretrieve_progress(
+                f"{url}/{f}",
+                filename=filename,
+                desc=f"Downloading {filename}",
+            )
+        else:
+            logging.info(f"{filename} already exists - skipping")
+
+        if ".gz" in str(filename):
+            unzipped = Path(os.path.splitext(filename)[0])
+            if unzipped.is_file() is False:
+                with gzip.open(filename, "rb") as f_in:
+                    with open(unzipped, "wb") as f_out:
+                        shutil.copyfileobj(f_in, f_out)
+            else:
+                logging.info(f"{unzipped} already exist - skipping")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    logging.info(f"out_dir: {args.out_dir}")
+
+    main(out_dir=args.out_dir)
diff --git a/egs/librispeech/ASR/local/filter_cuts.py b/egs/librispeech/ASR/local/filter_cuts.py
new file mode 100644
index 000000000..fbcc9e24a
--- /dev/null
+++ b/egs/librispeech/ASR/local/filter_cuts.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script removes short and long utterances from a cutset.
+
+Caution:
+  You may need to tune the thresholds for your own dataset.
+
+Usage example:
+
+  python3 ./local/filter_cuts.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --in-cuts data/fbank/librispeech_cuts_test-clean.jsonl.gz \
+    --out-cuts data/fbank-filtered/librispeech_cuts_test-clean.jsonl.gz
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=Path,
+        help="Path to the bpe.model",
+    )
+
+    parser.add_argument(
+        "--in-cuts",
+        type=Path,
+        help="Path to the input cutset",
+    )
+
+    parser.add_argument(
+        "--out-cuts",
+        type=Path,
+        help="Path to the output cutset",
+    )
+
+    return parser.parse_args()
+
+
+def filter_cuts(cut_set: CutSet, sp: spm.SentencePieceProcessor):
+    total = 0  # number of total utterances before removal
+    removed = 0  # number of removed utterances
+
+    def remove_short_and_long_utterances(c: Cut):
+        """Return False to exclude the input cut"""
+        nonlocal removed, total
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ./display_manifest_statistics.py
+        #
+        # You should use ./display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        total += 1
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            removed += 1
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./pruned_transducer_stateless2/conformer.py, the
+        # conv module uses the following expression
+        # for subsampling
+        if c.num_frames is None:
+            num_frames = c.duration * 100  # approximate
+        else:
+            num_frames = c.num_frames
+
+        T = ((num_frames - 1) // 2 - 1) // 2
+        # Note: for ./lstm_transducer_stateless/lstm.py, the formula is
+        #  T = ((num_frames - 3) // 2 - 1) // 2
+
+        # Note: for ./pruned_transducer_stateless7/zipformer.py, the formula is
+        # T = ((num_frames - 7) // 2 + 1) // 2
+
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            removed += 1
+            return False
+
+        return True
+
+    # We use to_eager() here so that we can print out the value of total
+    # and removed below.
+    ans = cut_set.filter(remove_short_and_long_utterances).to_eager()
+    ratio = removed / total * 100
+    logging.info(
+        f"Removed {removed} cuts from {total} cuts. {ratio:.3f}% data is removed."
+    )
+    return ans
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    if args.out_cuts.is_file():
+        logging.info(f"{args.out_cuts} already exists - skipping")
+        return
+
+    assert args.in_cuts.is_file(), f"{args.in_cuts} does not exist"
+    assert args.bpe_model.is_file(), f"{args.bpe_model} does not exist"
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    cut_set = load_manifest_lazy(args.in_cuts)
+    assert isinstance(cut_set, CutSet)
+
+    cut_set = filter_cuts(cut_set, sp)
+    logging.info(f"Saving to {args.out_cuts}")
+    args.out_cuts.parent.mkdir(parents=True, exist_ok=True)
+    cut_set.to_file(args.out_cuts)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/generate_unique_lexicon.py b/egs/librispeech/ASR/local/generate_unique_lexicon.py
new file mode 100755
index 000000000..3459c2f5a
--- /dev/null
+++ b/egs/librispeech/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input a lexicon.txt and output a new lexicon,
+in which each word has a unique pronunciation.
+
+The way to do this is to keep only the first pronunciation of a word
+in lexicon.txt.
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        This file will generate a new file uniq_lexicon.txt
+        in it.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def filter_multiple_pronunications(
+    lexicon: List[Tuple[str, List[str]]]
+) -> List[Tuple[str, List[str]]]:
+    """Remove multiple pronunciations of words from a lexicon.
+
+    If a word has more than one pronunciation in the lexicon, only
+    the first one is kept, while other pronunciations are removed
+    from the lexicon.
+
+    Args:
+      lexicon:
+        The input lexicon, containing a list of (word, [p1, p2, ..., pn]),
+        where "p1, p2, ..., pn" are the pronunciations of the "word".
+    Returns:
+      Return a new lexicon where each word has a unique pronunciation.
+    """
+    seen = set()
+    ans = []
+
+    for word, tokens in lexicon:
+        if word in seen:
+            continue
+        seen.add(word)
+        ans.append((word, tokens))
+    return ans
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    lexicon_filename = lang_dir / "lexicon.txt"
+
+    in_lexicon = read_lexicon(lexicon_filename)
+
+    out_lexicon = filter_multiple_pronunications(in_lexicon)
+
+    write_lexicon(lang_dir / "uniq_lexicon.txt", out_lexicon)
+
+    logging.info(f"Number of entries in lexicon.txt: {len(in_lexicon)}")
+    logging.info(f"Number of entries in uniq_lexicon.txt: {len(out_lexicon)}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/prepare_lang.py b/egs/librispeech/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..d913756a1
--- /dev/null
+++ b/egs/librispeech/ASR/local/prepare_lang.py
@@ -0,0 +1,413 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+from icefall.utils import str2bool
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    lexicon_filename = lang_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(lang_dir / "tokens.txt", token2id)
+    write_mapping(lang_dir / "words.txt", word2id)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/prepare_lang_bpe.py b/egs/librispeech/ASR/local/prepare_lang_bpe.py
new file mode 100755
index 000000000..2a2d9c219
--- /dev/null
+++ b/egs/librispeech/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.utils import str2bool
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_lexicon(
+    model_file: str, words: List[str], oov: str
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    words_pieces_ids: List[List[int]] = sp.encode(words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
+
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bpe.model"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", args.oov, "#0", "<s>", "</s>"]
+
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_lexicon(model_file, words, args.oov)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/prepare_lang_fst.py b/egs/librispeech/ASR/local/prepare_lang_fst.py
new file mode 100755
index 000000000..fb1e7f9c0
--- /dev/null
+++ b/egs/librispeech/ASR/local/prepare_lang_fst.py
@@ -0,0 +1,219 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+This script takes as input lang_dir containing lexicon_disambig.txt,
+tokens.txt, and words.txt and generates the following files:
+
+    - H.fst
+    - HL.fst
+    - HLG.fst
+
+Note that saved files are in OpenFst binary format.
+
+Usage:
+
+./local/prepare_lang_fst.py \
+  --lang-dir ./data/lang_phone \
+  --has-silence 1
+
+Or
+
+./local/prepare_lang_fst.py \
+  --lang-dir ./data/lang_bpe_500
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import kaldifst
+
+from icefall.ctc import (
+    Lexicon,
+    add_disambig_self_loops,
+    add_one,
+    build_standard_ctc_topo,
+    make_lexicon_fst_no_silence,
+    make_lexicon_fst_with_silence,
+)
+from icefall.utils import str2bool
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--has-silence",
+        type=str2bool,
+        default=False,
+        help="True if the lexicon has silence.",
+    )
+
+    parser.add_argument(
+        "--ngram-G",
+        type=str,
+        help="""If not empty, it is the filename of G used to build HLG.
+        For instance, --ngram-G=./data/lm/G_3_fst.txt
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def build_HL(
+    H: kaldifst.StdVectorFst,
+    L: kaldifst.StdVectorFst,
+    has_silence: bool,
+    lexicon: Lexicon,
+) -> kaldifst.StdVectorFst:
+    if has_silence:
+        # We also need to change the input labels of L
+        add_one(L, treat_ilabel_zero_specially=True, update_olabel=False)
+    else:
+        add_one(L, treat_ilabel_zero_specially=False, update_olabel=False)
+
+    # Invoke add_disambig_self_loops() so that it eats the disambig symbols
+    # from L after composition
+    add_disambig_self_loops(
+        H,
+        start=lexicon.token2id["#0"] + 1,
+        end=lexicon.max_disambig_id + 1,
+    )
+
+    kaldifst.arcsort(H, sort_type="olabel")
+    kaldifst.arcsort(L, sort_type="ilabel")
+
+    HL = kaldifst.compose(H, L)
+    kaldifst.determinize_star(HL)
+
+    disambig0 = lexicon.token2id["#0"] + 1
+    max_disambig = lexicon.max_disambig_id + 1
+    for state in kaldifst.StateIterator(HL):
+        for arc in kaldifst.ArcIterator(HL, state):
+            # If treat_ilabel_zero_specially is False, we always change it
+            # Otherwise, we only change non-zero input labels
+            if disambig0 <= arc.ilabel <= max_disambig:
+                arc.ilabel = 0
+
+    # Note: We are not composing L with G, so there is no need to add
+    # self-loops to L to handle #0
+
+    return HL
+
+
+def build_HLG(
+    H: kaldifst.StdVectorFst,
+    L: kaldifst.StdVectorFst,
+    G: kaldifst.StdVectorFst,
+    has_silence: bool,
+    lexicon: Lexicon,
+) -> kaldifst.StdVectorFst:
+    if has_silence:
+        # We also need to change the input labels of L
+        add_one(L, treat_ilabel_zero_specially=True, update_olabel=False)
+    else:
+        add_one(L, treat_ilabel_zero_specially=False, update_olabel=False)
+
+    # add-self-loops
+    token_disambig0 = lexicon.token2id["#0"] + 1
+    word_disambig0 = lexicon.word2id["#0"]
+
+    kaldifst.add_self_loops(L, isyms=[token_disambig0], osyms=[word_disambig0])
+
+    kaldifst.arcsort(L, sort_type="olabel")
+    kaldifst.arcsort(G, sort_type="ilabel")
+    LG = kaldifst.compose(L, G)
+    kaldifst.determinize_star(LG)
+    kaldifst.minimize_encoded(LG)
+
+    kaldifst.arcsort(LG, sort_type="ilabel")
+
+    # Invoke add_disambig_self_loops() so that it eats the disambig symbols
+    # from L after composition
+    add_disambig_self_loops(
+        H,
+        start=lexicon.token2id["#0"] + 1,
+        end=lexicon.max_disambig_id + 1,
+    )
+
+    kaldifst.arcsort(H, sort_type="olabel")
+
+    HLG = kaldifst.compose(H, LG)
+    kaldifst.determinize_star(HLG)
+
+    disambig0 = lexicon.token2id["#0"] + 1
+    max_disambig = lexicon.max_disambig_id + 1
+    for state in kaldifst.StateIterator(HLG):
+        for arc in kaldifst.ArcIterator(HLG, state):
+            # If treat_ilabel_zero_specially is False, we always change it
+            # Otherwise, we only change non-zero input labels
+            if disambig0 <= arc.ilabel <= max_disambig:
+                arc.ilabel = 0
+    return HLG
+
+
+def copy_fst(fst):
+    # Please don't use fst.copy()
+    return kaldifst.StdVectorFst(fst)
+
+
+def main():
+    args = get_args()
+    lang_dir = args.lang_dir
+
+    lexicon = Lexicon(lang_dir)
+
+    logging.info("Building standard CTC topology")
+    max_token_id = max(lexicon.tokens)
+    H = build_standard_ctc_topo(max_token_id=max_token_id)
+
+    # We need to add one to all tokens since we want to use ID 0
+    # for epsilon
+    add_one(H, treat_ilabel_zero_specially=False, update_olabel=True)
+    H.write(f"{lang_dir}/H.fst")
+
+    logging.info("Building L")
+    # Now for HL
+
+    if args.has_silence:
+        L = make_lexicon_fst_with_silence(lexicon, attach_symbol_table=False)
+    else:
+        L = make_lexicon_fst_no_silence(lexicon, attach_symbol_table=False)
+
+    logging.info("Building HL")
+    HL = build_HL(
+        H=copy_fst(H),
+        L=copy_fst(L),
+        has_silence=args.has_silence,
+        lexicon=lexicon,
+    )
+    HL.write(f"{lang_dir}/HL.fst")
+
+    if not args.ngram_G:
+        logging.info("Skip building HLG")
+        return
+
+    logging.info("Building HLG")
+    with open(args.ngram_G) as f:
+        G = kaldifst.compile(
+            s=f.read(),
+            acceptor=False,
+        )
+
+    HLG = build_HLG(H=H, L=L, G=G, has_silence=args.has_silence, lexicon=lexicon)
+    HLG.write(f"{lang_dir}/HLG.fst")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/local/prepare_lm_training_data.py b/egs/librispeech/ASR/local/prepare_lm_training_data.py
new file mode 100755
index 000000000..70343fef7
--- /dev/null
+++ b/egs/librispeech/ASR/local/prepare_lm_training_data.py
@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Daniel Povey
+#                                                   Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script takes a `bpe.model` and a text file such as
+./download/lm/librispeech-lm-norm.txt
+and outputs the LM training data to a supplied directory such
+as data/lm_training_bpe_500.  The format is as follows:
+
+It creates a PyTorch archive (.pt file), say data/lm_training.pt, which is a
+representation of a dict with the following format:
+
+  'words' -> a k2.RaggedTensor of two axes [word][token] with dtype torch.int32
+             containing the BPE representations of each word, indexed by
+             integer word ID. (These integer word IDS are present in
+             'lm_data').  The sentencepiece object can be used to turn the
+             words and BPE units into string form.
+  'sentences' -> a k2.RaggedTensor of two axes [sentence][word] with dtype
+            torch.int32 containing all the sentences, as word-ids (we don't
+            output the string form of this directly but it can be worked out
+            together with 'words' and the bpe.model).
+  'sentence_lengths' -> a 1-D torch.Tensor of dtype torch.int32, containing
+            number of BPE tokens of each sentence.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import sentencepiece as spm
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="Input BPE model, e.g. data/bpe_500/bpe.model",
+    )
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        help="""Input LM training data as text, e.g.
+        download/pb.train.txt""",
+    )
+    parser.add_argument(
+        "--lm-archive",
+        type=str,
+        help="""Path to output archive, e.g. data/bpe_500/lm_data.pt;
+        look at the source of this script to see the format.""",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    if Path(args.lm_archive).exists():
+        logging.warning(f"{args.lm_archive} exists - skipping")
+        return
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    # word2index is a dictionary from words to integer ids.  No need to reserve
+    # space for epsilon, etc.; the words are just used as a convenient way to
+    # compress the sequences of BPE pieces.
+    word2index = dict()
+
+    word2bpe = []  # Will be a list-of-list-of-int, representing BPE pieces.
+    sentences = []  # Will be a list-of-list-of-int, representing word-ids.
+
+    if "librispeech-lm-norm" in args.lm_data:
+        num_lines_in_total = 40418261.0
+        step = 5000000
+    elif "valid" in args.lm_data:
+        num_lines_in_total = 5567.0
+        step = 3000
+    elif "test" in args.lm_data:
+        num_lines_in_total = 5559.0
+        step = 3000
+    else:
+        num_lines_in_total = None
+        step = None
+
+    processed = 0
+
+    with open(args.lm_data) as f:
+        while True:
+            line = f.readline()
+            if line == "":
+                break
+
+            if step and processed % step == 0:
+                logging.info(
+                    f"Processed number of lines: {processed} "
+                    f"({processed/num_lines_in_total*100: .3f}%)"
+                )
+            processed += 1
+
+            line_words = line.split()
+            for w in line_words:
+                if w not in word2index:
+                    w_bpe = sp.encode(w)
+                    word2index[w] = len(word2bpe)
+                    word2bpe.append(w_bpe)
+            sentences.append([word2index[w] for w in line_words])
+
+    logging.info("Constructing ragged tensors")
+    words = k2.ragged.RaggedTensor(word2bpe)
+    sentences = k2.ragged.RaggedTensor(sentences)
+
+    output = dict(words=words, sentences=sentences)
+
+    num_sentences = sentences.dim0
+    logging.info(f"Computing sentence lengths, num_sentences: {num_sentences}")
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        if step and i % step == 0:
+            logging.info(
+                f"Processed number of lines: {i} ({i/num_sentences*100: .3f}%)"
+            )
+
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    output["sentence_lengths"] = torch.tensor(sentence_lengths, dtype=torch.int32)
+
+    torch.save(output, args.lm_archive)
+    logging.info(f"Saved to {args.lm_archive}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/preprocess_gigaspeech.py b/egs/librispeech/ASR/local/preprocess_gigaspeech.py
new file mode 100644
index 000000000..8aa5e461d
--- /dev/null
+++ b/egs/librispeech/ASR/local/preprocess_gigaspeech.py
@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import re
+from pathlib import Path
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Similar text filtering and normalization procedure as in:
+# https://github.com/SpeechColab/GigaSpeech/blob/main/toolkits/kaldi/gigaspeech_data_prep.sh
+
+
+def normalize_text(
+    utt: str,
+    punct_pattern=re.compile(r"<(COMMA|PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    whitespace_pattern=re.compile(r"\s\s+"),
+) -> str:
+    return whitespace_pattern.sub(" ", punct_pattern.sub("", utt))
+
+
+def has_no_oov(
+    sup: SupervisionSegment,
+    oov_pattern=re.compile(r"<(SIL|MUSIC|NOISE|OTHER)>"),
+) -> bool:
+    return oov_pattern.search(sup.text) is None
+
+
+def preprocess_giga_speech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    dataset_parts = (
+        "DEV",
+        "TEST",
+        "XS",
+        "S",
+        "M",
+        "L",
+        "XL",
+    )
+
+    logging.info("Loading manifest (may take 4 minutes)")
+    prefix = "gigaspeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"{prefix}_cuts_{partition}_raw.{suffix}"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        # Note this step makes the recipe different than LibriSpeech:
+        # We must filter out some utterances and remove punctuation
+        # to be consistent with Kaldi.
+        logging.info("Filtering OOV utterances from supervisions")
+        m["supervisions"] = m["supervisions"].filter(has_no_oov)
+        logging.info(f"Normalizing text in {partition}")
+        for sup in m["supervisions"]:
+            sup.text = normalize_text(sup.text)
+            sup.custom = {"origin": "giga"}
+
+        # Create long-recording cut manifests.
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        )
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        #  if partition not in ["DEV", "TEST"]:
+        #      logging.info(
+        #          f"Speed perturb for {partition} with factors 0.9 and 1.1 "
+        #          "(Perturbing may take 8 minutes and saving may"
+        #          " take 20 minutes)"
+        #      )
+        #      cut_set = (
+        #          cut_set
+        #          + cut_set.perturb_speed(0.9)
+        #          + cut_set.perturb_speed(1.1)
+        #      )
+        #
+        # Note: No need to perturb the training subset as not all of the
+        # data is going to be used in the training.
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    preprocess_giga_speech()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/sort_lm_training_data.py b/egs/librispeech/ASR/local/sort_lm_training_data.py
new file mode 120000
index 000000000..bb26b5f5c
--- /dev/null
+++ b/egs/librispeech/ASR/local/sort_lm_training_data.py
@@ -0,0 +1 @@
+../../../ptb/LM/local/sort_lm_training_data.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/local/test_load_XL_split.py b/egs/librispeech/ASR/local/test_load_XL_split.py
new file mode 100755
index 000000000..3982a7157
--- /dev/null
+++ b/egs/librispeech/ASR/local/test_load_XL_split.py
@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file can be used to check if any split is corrupted.
+"""
+
+import glob
+import re
+
+import lhotse
+
+
+def main():
+    d = "data/fbank/XL_split_2000"
+    filenames = list(glob.glob(f"{d}/cuts_XL.*.jsonl.gz"))
+
+    pattern = re.compile(r"cuts_XL.([0-9]+).jsonl.gz")
+
+    idx_filenames = [(int(pattern.search(c).group(1)), c) for c in filenames]
+
+    idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
+
+    print(f"Loading {len(idx_filenames)} splits")
+
+    s = 0
+    for i, f in idx_filenames:
+        cuts = lhotse.load_manifest_lazy(f)
+        print(i, "filename", f)
+        for i, c in enumerate(cuts):
+            s += c.features.load().shape[0]
+            if i > 5:
+                break
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/test_prepare_lang.py b/egs/librispeech/ASR/local/test_prepare_lang.py
new file mode 100755
index 000000000..74e025ad7
--- /dev/null
+++ b/egs/librispeech/ASR/local/test_prepare_lang.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/train_bpe_model.py b/egs/librispeech/ASR/local/train_bpe_model.py
new file mode 100755
index 000000000..43142aee4
--- /dev/null
+++ b/egs/librispeech/ASR/local/train_bpe_model.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/validate_bpe_lexicon.py b/egs/librispeech/ASR/local/validate_bpe_lexicon.py
new file mode 100755
index 000000000..c542f2fab
--- /dev/null
+++ b/egs/librispeech/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks that there are no OOV tokens in the BPE-based lexicon.
+
+Usage example:
+
+    python3 ./local/validate_bpe_lexicon.py \
+            --lexicon /path/to/lexicon.txt \
+            --bpe-model /path/to/bpe.model
+"""
+
+import argparse
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+
+from icefall.lexicon import read_lexicon
+
+# Map word to word pieces
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--lexicon",
+        required=True,
+        type=Path,
+        help="Path to lexicon.txt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        required=True,
+        type=Path,
+        help="Path to bpe.model",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    assert args.lexicon.is_file(), args.lexicon
+    assert args.bpe_model.is_file(), args.bpe_model
+
+    lexicon = read_lexicon(args.lexicon)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    word_pieces = set(sp.id_to_piece(list(range(sp.vocab_size()))))
+    for word, pieces in lexicon:
+        for p in pieces:
+            if p not in word_pieces:
+                raise ValueError(f"The word {word} contains an OOV token {p}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/local/validate_manifest.py b/egs/librispeech/ASR/local/validate_manifest.py
new file mode 100755
index 000000000..de49f5321
--- /dev/null
+++ b/egs/librispeech/ASR/local/validate_manifest.py
@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+- Supervision time bounds are within cut time bounds
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+from lhotse.dataset.speech_recognition import validate_for_asr
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def validate_one_supervision_per_cut(c: Cut):
+    if len(c.supervisions) != 1:
+        raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
+
+
+def validate_supervision_and_cut_time_bounds(c: Cut):
+    tol = 2e-3  # same tolerance as in 'validate_for_asr()'
+    s = c.supervisions[0]
+
+    # Supervision start time is relative to Cut ...
+    # https://lhotse.readthedocs.io/en/v0.10_e/cuts.html
+    if s.start < -tol:
+        raise ValueError(
+            f"{c.id}: Supervision start time {s.start} must not be negative."
+        )
+    if s.start > tol:
+        raise ValueError(
+            f"{c.id}: Supervision start time {s.start} is not at the beginning of the Cut. Please apply `lhotse cut trim-to-supervisions`."
+        )
+    if c.start + s.end > c.end + tol:
+        raise ValueError(
+            f"{c.id}: Supervision end time {c.start+s.end} is larger "
+            f"than cut end time {c.end}"
+        )
+
+
+def main():
+    args = get_args()
+
+    manifest = args.manifest
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    for c in cut_set:
+        validate_one_supervision_per_cut(c)
+        validate_supervision_and_cut_time_bounds(c)
+
+    # Validation from K2 training
+    # - checks supervision start is 0
+    # - checks supervision.duration is not longer than cut.duration
+    # - there is tolerance 2ms
+    validate_for_asr(cut_set)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/long_file_recog.sh b/egs/librispeech/ASR/long_file_recog.sh
new file mode 100755
index 000000000..acd1b1253
--- /dev/null
+++ b/egs/librispeech/ASR/long_file_recog.sh
@@ -0,0 +1,94 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+set -eou pipefail
+
+# This script is used to recogize long audios. The process is as follows:
+# 1) Split long audios into chunks with overlaps.
+# 2) Perform speech recognition on chunks, getting tokens and timestamps.
+# 3) Merge the overlapped chunks into utterances acording to the timestamps.
+
+# Each chunk (except the first and the last) is padded with extra left side and right side.
+# The chunk length is: left_side + chunk_size + right_side.
+chunk=30.0
+extra=2.0
+
+stage=1
+stop_stage=4
+
+# We assume that you have downloaded the LibriLight dataset
+# with audio files in $corpus_dir and texts in $text_dir
+corpus_dir=$PWD/download/libri-light
+text_dir=$PWD/download/librilight_text
+# Path to save the manifests
+output_dir=$PWD/data/librilight
+
+world_size=4
+
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  # We will get librilight_recodings_{subset}.jsonl.gz and librilight_supervisions_{subset}.jsonl.gz
+  # saved in $output_dir/manifests
+  log "Stage 1: Prepare LibriLight manifest"
+  lhotse prepare librilight $corpus_dir $text_dir $output_dir/manifests -j 10
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  # Chunk manifests are saved to $output_dir/manifests_chunk/librilight_cuts_{subset}.jsonl.gz
+  log "Stage 2: Split long audio into chunks"
+  ./long_file_recog/split_into_chunks.py \
+    --manifest-in-dir $output_dir/manifests \
+    --manifest-out-dir $output_dir/manifests_chunk \
+    --chunk $chunk \
+    --extra $extra  # Extra duration (in seconds) at both sides
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  # Recognized tokens and timestamps are saved to $output_dir/manifests_chunk_recog/librilight_cuts_{subset}.jsonl.gz
+
+  # This script loads torchscript models, exported by `torch.jit.script()`,
+  # and uses it to decode waves.
+  # You can download the jit model from https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+
+  log "Stage 3: Perform speech recognition on splitted chunks"
+  for subset in small median large; do
+    ./long_file_recog/recognize.py \
+      --world-size $world_size \
+      --num-workers 8 \
+      --subset $subset \
+      --manifest-in-dir $output_dir/manifests_chunk \
+      --manifest-out-dir $output_dir/manifests_chunk_recog \
+      --nn-model-filename long_file_recog/exp/jit_model.pt \
+      --bpe-model data/lang_bpe_500/bpe.model \
+      --max-duration 2400 \
+      --decoding-method greedy_search
+      --master 12345
+
+    if [ $world_size -gt 1 ]; then
+      # Combine manifests from different jobs
+      lhotse combine $(find $output_dir/manifests_chunk_recog -name librilight_cuts_${subset}_job_*.jsonl.gz | tr "\n" " ") $output_dir/manifests_chunk_recog/librilight_cuts_${subset}.jsonl.gz
+    fi
+  done
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  # Final results are saved in $output_dir/manifests/librilight_cuts_{subset}.jsonl.gz
+  log "Stage 4: Merge splitted chunks into utterances."
+  ./long_file_recog/merge_chunks.py \
+    --manifest-in-dir $output_dir/manifests_chunk_recog \
+    --manifest-out-dir $output_dir/manifests \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --extra $extra
+fi
+
+
diff --git a/egs/librispeech/ASR/long_file_recog/asr_datamodule.py b/egs/librispeech/ASR/long_file_recog/asr_datamodule.py
new file mode 100644
index 000000000..eddce7213
--- /dev/null
+++ b/egs/librispeech/ASR/long_file_recog/asr_datamodule.py
@@ -0,0 +1,189 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+# Copyright      2023  Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Dict, List, Union
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest_lazy
+from lhotse.cut import Cut
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    BatchIO,
+    OnTheFlyFeatures,
+)
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class SpeechRecognitionDataset(K2SpeechRecognitionDataset):
+    def __init__(
+        self,
+        return_cuts: bool = False,
+        input_strategy: BatchIO = PrecomputedFeatures(),
+    ):
+        super().__init__(return_cuts=return_cuts, input_strategy=input_strategy)
+
+    def __getitem__(self, cuts: CutSet) -> Dict[str, Union[torch.Tensor, List[Cut]]]:
+        """
+        Return a new batch, with the batch size automatically determined using the constraints
+        of max_frames and max_cuts.
+        """
+        self.hdf5_fix.update()
+
+        # Note: don't sort cuts here
+        # Sort the cuts by duration so that the first one determines the batch time dimensions.
+        # cuts = cuts.sort_by_duration(ascending=False)
+
+        # Get a tensor with batched feature matrices, shape (B, T, F)
+        # Collation performs auto-padding, if necessary.
+        input_tpl = self.input_strategy(cuts)
+        if len(input_tpl) == 3:
+            # An input strategy with fault tolerant audio reading mode.
+            # "cuts" may be a subset of the original "cuts" variable,
+            # that only has cuts for which we succesfully read the audio.
+            inputs, _, cuts = input_tpl
+        else:
+            inputs, _ = input_tpl
+
+        # Get a dict of tensors that encode the positional information about supervisions
+        # in the batch of feature matrices. The tensors are named "sequence_idx",
+        # "start_frame/sample" and "num_frames/samples".
+        supervision_intervals = self.input_strategy.supervision_intervals(cuts)
+
+        batch = {"inputs": inputs, "supervisions": supervision_intervals}
+        if self.return_cuts:
+            batch["supervisions"]["cut"] = [cut for cut in cuts]
+
+        return batch
+
+
+class AsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests_chunk"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=600.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=8,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+            return_cuts=self.args.return_cuts,
+        )
+
+        sampler = SimpleCutSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+            drop_last=False,
+        )
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return test_dl
+
+    @lru_cache()
+    def load_subset(self, cuts_filename: Path) -> CutSet:
+        return load_manifest_lazy(cuts_filename)
diff --git a/egs/librispeech/ASR/long_file_recog/beam_search.py b/egs/librispeech/ASR/long_file_recog/beam_search.py
new file mode 100644
index 000000000..b65e9d40a
--- /dev/null
+++ b/egs/librispeech/ASR/long_file_recog/beam_search.py
@@ -0,0 +1,613 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import k2
+import torch
+
+from icefall.decode import one_best_decoding
+from icefall.utils import DecodingResults, get_texts, get_texts_with_timestamp
+
+
+def fast_beam_search(
+    model: torch.nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    temperature: float = 1.0,
+) -> k2.Fsa:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      temperature:
+        Softmax temperature.
+    Returns:
+      Return an FsaVec with axes [utt][state][arc] containing the decoded
+      lattice. Note: When the input graph is a TrivialGraph, the returned
+      lattice is actually an acceptor.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(k2.RnntDecodingStream(decoding_graph))
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = (logits / temperature).log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+    decoding_streams.terminate_and_flush_to_streams()
+    lattice = decoding_streams.format_output(encoder_out_lens.tolist())
+
+    return lattice
+
+
+def fast_beam_search_one_best(
+    model: torch.nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using fast beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        temperature=temperature,
+    )
+
+    best_path = one_best_decoding(lattice)
+
+    if not return_timestamps:
+        return get_texts(best_path)
+    else:
+        return get_texts_with_timestamp(best_path)
+
+
+def greedy_search_batch(
+    model: torch.nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = next(model.parameters()).device
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[-1] * (context_size - 1) + [blank_id] for _ in range(N)]
+
+    # timestamp[n][i] is the frame index after subsampling
+    # on which hyp[n][i] is decoded
+    timestamps = [[] for _ in range(N)]
+    # scores[n][i] is the logits on which hyp[n][i] is decoded
+    scores = [[] for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+    # decoder_out: (N, 1, decoder_out_dim)
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits'shape (batch_size, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (batch_size, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+        assert log_probs.ndim == 2, log_probs.shape
+        y = log_probs.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v not in (blank_id, unk_id):
+                hyps[i].append(v)
+                timestamps[i].append(t)
+                scores[i].append(log_probs[i, v].item())
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    ans_timestamps = []
+    ans_scores = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(timestamps[unsorted_indices[i]])
+        ans_scores.append(scores[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+            scores=ans_scores,
+        )
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    # timestamp[i] is the frame index after subsampling
+    # on which ys[i] is decoded
+    timestamp: List[int] = field(default_factory=list)
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
+def modified_beam_search(
+    model: torch.nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: int = 4,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    new_timestamp.append(t)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(
+                    ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    sorted_timestamps = [h.timestamp for h in best_hyps]
+    ans = []
+    ans_timestamps = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(sorted_timestamps[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+        )
diff --git a/egs/librispeech/ASR/long_file_recog/merge_chunks.py b/egs/librispeech/ASR/long_file_recog/merge_chunks.py
new file mode 100755
index 000000000..9e31e00d5
--- /dev/null
+++ b/egs/librispeech/ASR/long_file_recog/merge_chunks.py
@@ -0,0 +1,239 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file merge overlapped chunks into utterances accroding to recording ids.
+"""
+
+import argparse
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import List
+
+import sentencepiece as spm
+from lhotse import (
+    CutSet,
+    MonoCut,
+    SupervisionSegment,
+    SupervisionSet,
+    load_manifest,
+    load_manifest_lazy,
+)
+from lhotse.cut import Cut
+from lhotse.serialization import SequentialJsonlWriter
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--manifest-in-dir",
+        type=Path,
+        default=Path("data/librilight/manifests_chunk_recog"),
+        help="Path to directory of chunk cuts with recognition results.",
+    )
+
+    parser.add_argument(
+        "--manifest-out-dir",
+        type=Path,
+        default=Path("data/manifests"),
+        help="Path to directory to save full utterance by merging overlapped chunks.",
+    )
+
+    parser.add_argument(
+        "--extra",
+        type=float,
+        default=2.0,
+        help="""Extra duration (in seconds) at both sides.""",
+    )
+
+    return parser.parse_args()
+
+
+def merge_chunks(
+    cuts_chunk: CutSet,
+    supervisions: SupervisionSet,
+    cuts_writer: SequentialJsonlWriter,
+    sp: spm.SentencePieceProcessor,
+    extra: float,
+) -> int:
+    """Merge chunk-wise cuts accroding to recording ids.
+
+    Args:
+      cuts_chunk:
+        The chunk-wise cuts opened in a lazy mode.
+      supervisions:
+        The supervision manifest containing text file path, opened in a lazy mode.
+      cuts_writer:
+        Writer to save the cuts with recognition results.
+      sp:
+        The BPE model.
+      extra:
+        Extra duration (in seconds) to drop at both sides of each chunk.
+    """
+
+    #  Background worker to add alignemnt and save cuts to disk.
+    def _save_worker(utt_cut: Cut, flush=False):
+        cuts_writer.write(utt_cut, flush=flush)
+
+    def _merge(cut_list: List[Cut], rec_id: str, utt_idx: int):
+        """Merge chunks with same recording_id."""
+        for cut in cut_list:
+            assert cut.recording.id == rec_id, (cut.recording.id, rec_id)
+
+        # For each group with a same recording, sort it accroding to the start time
+        # In fact, we don't need to do this since the cuts have been sorted
+        # according to the start time
+        cut_list = sorted(cut_list, key=(lambda cut: cut.start))
+
+        rec = cut_list[0].recording
+        alignments = []
+        cur_end = 0
+        for cut in cut_list:
+            # Get left and right borders
+            left = cut.start + extra if cut.start > 0 else 0
+            chunk_end = cut.start + cut.duration
+            right = chunk_end - extra if chunk_end < rec.duration else rec.duration
+
+            # Assert the chunks are continuous
+            assert left == cur_end, (left, cur_end)
+            cur_end = right
+
+            assert len(cut.supervisions) == 1, len(cut.supervisions)
+            for ali in cut.supervisions[0].alignment["symbol"]:
+                t = ali.start + cut.start
+                if left <= t < right:
+                    alignments.append(ali.with_offset(cut.start))
+
+        old_sup = supervisions[rec_id]
+        # Assuming the supervisions are sorted with the same recoding order as in cuts_chunk
+        # old_sup = supervisions[utt_idx]
+        assert old_sup.recording_id == rec_id, (old_sup.recording_id, rec_id)
+
+        new_sup = SupervisionSegment(
+            id=rec_id,
+            recording_id=rec_id,
+            start=0,
+            duration=rec.duration,
+            alignment={"symbol": alignments},
+            language=old_sup.language,
+            speaker=old_sup.speaker,
+        )
+
+        utt_cut = MonoCut(
+            id=rec_id,
+            start=0,
+            duration=rec.duration,
+            channel=0,
+            recording=rec,
+            supervisions=[new_sup],
+        )
+        # Set a custom attribute to the cut
+        utt_cut.text_path = old_sup.book
+
+        return utt_cut
+
+    last_rec_id = None
+    cut_list = []
+    utt_idx = 0
+
+    futures = []
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        for cut in cuts_chunk:
+            cur_rec_id = cut.recording.id
+            if len(cut_list) == 0:
+                # Case of the first cut
+                last_rec_id = cur_rec_id
+                cut_list.append(cut)
+            elif cur_rec_id == last_rec_id:
+                cut_list.append(cut)
+            else:
+                # Case of a cut belonging to a new recording
+                utt_cut = _merge(cut_list, last_rec_id, utt_idx)
+                utt_idx += 1
+
+                futures.append(executor.submit(_save_worker, utt_cut))
+
+                last_rec_id = cur_rec_id
+                cut_list = [cut]
+
+                if utt_idx % 5000 == 0:
+                    logging.info(f"Procesed {utt_idx} utterances.")
+
+        # For the cuts belonging to the last recording
+        if len(cut_list) != 0:
+            utt_cut = _merge(cut_list, last_rec_id, utt_idx)
+            utt_idx += 1
+
+            futures.append(executor.submit(_save_worker, utt_cut))
+            logging.info("Finished")
+
+        for f in futures:
+            f.result()
+
+    return utt_idx
+
+
+def main():
+    args = get_parser()
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    # It contains "librilight_recordings_*.jsonl.gz" and "librilight_supervisions_small.jsonl.gz"
+    manifest_out_dir = args.manifest_out_dir
+
+    subsets = ["small", "median", "large"]
+
+    for subset in subsets:
+        logging.info(f"Processing {subset} subset")
+
+        manifest_out = manifest_out_dir / f"librilight_cuts_{subset}.jsonl.gz"
+        if manifest_out.is_file():
+            logging.info(f"{manifest_out} already exists - skipping.")
+            continue
+
+        supervisions = load_manifest(
+            manifest_out_dir / f"librilight_supervisions_{subset}.jsonl.gz"
+        )  # We will use the text path from supervisions
+
+        cuts_chunk = load_manifest_lazy(
+            args.manifest_in_dir / f"librilight_cuts_{subset}.jsonl.gz"
+        )
+
+        cuts_writer = CutSet.open_writer(manifest_out, overwrite=True)
+        num_utt = merge_chunks(
+            cuts_chunk, supervisions, cuts_writer=cuts_writer, sp=sp, extra=args.extra
+        )
+        cuts_writer.close()
+        logging.info(f"{num_utt} cuts saved to {manifest_out}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/long_file_recog/recognize.py b/egs/librispeech/ASR/long_file_recog/recognize.py
new file mode 100755
index 000000000..466253446
--- /dev/null
+++ b/egs/librispeech/ASR/long_file_recog/recognize.py
@@ -0,0 +1,436 @@
+#!/usr/bin/env python3
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`,
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+You can also download the jit model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+"""
+
+import argparse
+import torch.multiprocessing as mp
+import torch
+import torch.nn as nn
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from typing import List, Optional, Tuple
+
+from pathlib import Path
+
+import k2
+import sentencepiece as spm
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    fast_beam_search_one_best,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from icefall.utils import AttributeDict, convert_timestamp, setup_logger
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+from lhotse.supervision import AlignmentItem
+from lhotse.serialization import SequentialJsonlWriter
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+    parser.add_argument(
+        "--subset",
+        type=str,
+        default="small",
+        help="Subset to process. Possible values are 'small', 'medium', 'large'",
+    )
+
+    parser.add_argument(
+        "--manifest-in-dir",
+        type=Path,
+        default=Path("data/librilight/manifests_chunk"),
+        help="Path to directory with chunks cuts.",
+    )
+
+    parser.add_argument(
+        "--manifest-out-dir",
+        type=Path,
+        default=Path("data/librilight/manifests_chunk_recog"),
+        help="Path to directory to save the chunk cuts with recognition results.",
+    )
+
+    parser.add_argument(
+        "--log-dir",
+        type=Path,
+        default=Path("long_file_recog/log"),
+        help="Path to directory to save logs.",
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing decoding parameters."""
+    params = AttributeDict(
+        {
+            "subsampling_factor": 4,
+            "frame_shift_ms": 10,
+            # Used only when --method is beam_search or modified_beam_search.
+            "beam_size": 4,
+            # Used only when --method is beam_search or fast_beam_search.
+            # A floating point value to calculate the cutoff score during beam
+            # search (i.e., `cutoff = max-score - beam`), which is the same as the
+            # `beam` in Kaldi.
+            "beam": 4,
+            "max_contexts": 4,
+            "max_states": 8,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Tuple[List[List[str]], List[List[float]], List[List[float]]]:
+    """Decode one batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      paramsmodel:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+      Return the decoding result, timestamps, and scores.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        res = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "greedy_search":
+        res = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        res = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            return_timestamps=True,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    hyps = []
+    timestamps = []
+    scores = []
+    for i in range(feature.shape[0]):
+        hyps.append(res.hyps[i])
+        timestamps.append(
+            convert_timestamp(
+                res.timestamps[i], params.subsampling_factor, params.frame_shift_ms
+            )
+        )
+        scores.append(res.scores[i])
+
+    return hyps, timestamps, scores
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    cuts_writer: SequentialJsonlWriter,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> None:
+    """Decode dataset and store the recognition results to manifest.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      cuts_writer:
+        Writer to save the cuts with recognition results.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains five elements:
+        - cut_id
+        - reference transcript
+        - predicted result
+        - timestamps of reference transcript
+        - timestamps of predicted result
+    """
+
+    #  Background worker to add alignemnt and save cuts to disk.
+    def _save_worker(
+        cuts: List[Cut],
+        hyps: List[List[str]],
+        timestamps: List[List[float]],
+        scores: List[List[float]],
+    ):
+        for cut, symbol_list, time_list, score_list in zip(
+            cuts, hyps, timestamps, scores
+        ):
+            symbol_list = sp.id_to_piece(symbol_list)
+            ali = [
+                AlignmentItem(symbol=symbol, start=start, duration=None, score=score)
+                for symbol, start, score in zip(symbol_list, time_list, score_list)
+            ]
+            assert len(cut.supervisions) == 1, len(cut.supervisions)
+            cut.supervisions[0].alignment = {"symbol": ali}
+            cuts_writer.write(cut, flush=True)
+
+    num_cuts = 0
+    log_interval = 10
+    futures = []
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        # We only want one background worker so that serialization is deterministic.
+
+        for batch_idx, batch in enumerate(dl):
+            cuts = batch["supervisions"]["cut"]
+
+            hyps, timestamps, scores = decode_one_batch(
+                params=params,
+                model=model,
+                decoding_graph=decoding_graph,
+                batch=batch,
+            )
+
+            futures.append(
+                executor.submit(_save_worker, cuts, hyps, timestamps, scores)
+            )
+
+            num_cuts += len(cuts)
+            if batch_idx % log_interval == 0:
+                logging.info(f"cuts processed until now is {num_cuts}")
+
+        for f in futures:
+            f.result()
+
+
+@torch.no_grad()
+def run(rank, world_size, args, in_cuts):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.log_dir}/log-decode")
+    logging.info("Decoding started")
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    ), params.decoding_method
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"device: {device}")
+
+    logging.info("Loading jit model")
+    model = torch.jit.load(params.nn_model_filename)
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    # we will store new cuts with recognition results.
+    args.return_cuts = True
+    asr_data_module = AsrDataModule(args)
+
+    if world_size > 1:
+        in_cuts = in_cuts[rank]
+        out_cuts_filename = params.manifest_out_dir / (
+            f"{params.cuts_filename}_job_{rank}" + params.suffix
+        )
+    else:
+        out_cuts_filename = params.manifest_out_dir / (
+            f"{params.cuts_filename}" + params.suffix
+        )
+
+    dl = asr_data_module.dataloaders(in_cuts)
+
+    cuts_writer = CutSet.open_writer(out_cuts_filename, overwrite=True)
+    decode_dataset(
+        dl=dl,
+        params=params,
+        model=model,
+        sp=sp,
+        decoding_graph=decoding_graph,
+        cuts_writer=cuts_writer,
+    )
+    cuts_writer.close()
+    logging.info(f"Cuts saved to {out_cuts_filename}")
+
+    logging.info("Done!")
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    subset = args.subset
+    assert subset in ["small", "medium", "large"], subset
+
+    manifest_out_dir = args.manifest_out_dir
+    manifest_out_dir.mkdir(parents=True, exist_ok=True)
+
+    args.suffix = ".jsonl.gz"
+    args.cuts_filename = f"librilight_cuts_{args.subset}"
+
+    out_cuts_filename = manifest_out_dir / (args.cuts_filename + args.suffix)
+    if out_cuts_filename.is_file():
+        logging.info(f"{out_cuts_filename} already exists - skipping.")
+        return
+
+    in_cuts_filename = args.manifest_in_dir / (args.cuts_filename + args.suffix)
+    in_cuts = load_manifest_lazy(in_cuts_filename)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        chunk_size = (len(in_cuts) + (world_size - 1)) // world_size
+        # Each manifest is saved at: ``{output_dir}/{prefix}.{split_idx}.jsonl.gz``
+        splits = in_cuts.split_lazy(
+            output_dir=args.manifest_in_dir / "split",
+            chunk_size=chunk_size,
+            prefix=args.cuts_filename,
+        )
+        assert len(splits) == world_size, (len(splits), world_size)
+        mp.spawn(run, args=(world_size, args, splits), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=world_size, args=args, in_cuts=in_cuts)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/long_file_recog/split_into_chunks.py b/egs/librispeech/ASR/long_file_recog/split_into_chunks.py
new file mode 100755
index 000000000..4a900831c
--- /dev/null
+++ b/egs/librispeech/ASR/long_file_recog/split_into_chunks.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script splits long utterances into chunks with overlaps.
+Each chunk (except the first and the last) is padded with extra left side and right side.
+The chunk length is: left_side + chunk_size + right_side.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest-in-dir",
+        type=Path,
+        default=Path("data/librilight/manifests"),
+        help="Path to directory of full utterances.",
+    )
+
+    parser.add_argument(
+        "--manifest-out-dir",
+        type=Path,
+        default=Path("data/librilight/manifests_chunk"),
+        help="Path to directory to save splitted chunks.",
+    )
+
+    parser.add_argument(
+        "--chunk",
+        type=float,
+        default=300.0,
+        help="""Duration (in seconds) of each chunk.""",
+    )
+
+    parser.add_argument(
+        "--extra",
+        type=float,
+        default=2.0,
+        help="""Extra duration (in seconds) at both sides.""",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    manifest_out_dir = args.manifest_out_dir
+    manifest_out_dir.mkdir(parents=True, exist_ok=True)
+
+    subsets = ["small", "medium", "large"]
+
+    for subset in subsets:
+        logging.info(f"Processing {subset} subset")
+
+        manifest_out = manifest_out_dir / f"librilight_cuts_{subset}.jsonl.gz"
+        if manifest_out.is_file():
+            logging.info(f"{manifest_out} already exists - skipping.")
+            continue
+
+        manifest_in = args.manifest_in_dir / f"librilight_recordings_{subset}.jsonl.gz"
+        recordings = load_manifest(manifest_in)
+
+        cuts = CutSet.from_manifests(recordings=recordings)
+        cuts = cuts.cut_into_windows(
+            duration=args.chunk, hop=args.chunk - args.extra * 2
+        )
+        cuts = cuts.fill_supervisions()
+
+        cuts.to_file(manifest_out)
+        logging.info(f"Cuts saved to {manifest_out}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/__init__.py b/egs/librispeech/ASR/lstm_transducer_stateless/__init__.py
new file mode 120000
index 000000000..b24e5e357
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/__init__.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/lstm_transducer_stateless/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/beam_search.py b/egs/librispeech/ASR/lstm_transducer_stateless/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/decode.py b/egs/librispeech/ASR/lstm_transducer_stateless/decode.py
new file mode 100755
index 000000000..856c9d945
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/decode.py
@@ -0,0 +1,799 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./lstm_transducer_stateless/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    # tail padding here to alleviate the tail deletion problem
+    num_tail_padded_frames = 35
+    feature = torch.nn.functional.pad(
+        feature,
+        (0, 0, 0, num_tail_padded_frames),
+        mode="constant",
+        value=LOG_EPS,
+    )
+    feature_lens += num_tail_padded_frames
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/decoder.py b/egs/librispeech/ASR/lstm_transducer_stateless/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/lstm_transducer_stateless/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/export-onnx.py b/egs/librispeech/ASR/lstm_transducer_stateless/export-onnx.py
new file mode 120000
index 000000000..9f5064deb
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/export-onnx.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/export-onnx.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/export.py b/egs/librispeech/ASR/lstm_transducer_stateless/export.py
new file mode 100755
index 000000000..c007220d5
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/export.py
@@ -0,0 +1,386 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.trace()
+
+./lstm_transducer_stateless/export.py \
+  --exp-dir ./lstm_transducer_stateless/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 35 \
+  --avg 10 \
+  --jit-trace 1
+
+It will generate 3 files: `encoder_jit_trace.pt`,
+`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`.
+
+(2) Export `model.state_dict()`
+
+./lstm_transducer_stateless/export.py \
+  --exp-dir ./lstm_transducer_stateless/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 35 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `lstm_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./lstm_transducer_stateless/decode.py \
+        --exp-dir ./lstm_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18
+    # You will find the pre-trained model in icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace.
+        It will generate 3 files:
+         - encoder_jit_trace.pt
+         - decoder_jit_trace.pt
+         - joiner_jit_trace.pt
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+    states = encoder_model.get_init_states()
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size, <blk> is
+    # defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit_trace is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    else:
+        logging.info("Not using torchscript")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/jit_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless/jit_pretrained.py
new file mode 100755
index 000000000..c07956243
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/jit_pretrained.py
@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, either exported by `torch.jit.trace()`
+or by `torch.jit.script()`, and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./lstm_transducer_stateless/export.py \
+  --exp-dir ./lstm_transducer_stateless/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit-trace 1
+
+Usage of this script:
+
+./lstm_transducer_stateless/jit_pretrained.py \
+  --encoder-model-filename ./lstm_transducer_stateless/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./lstm_transducer_stateless/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./lstm_transducer_stateless/exp/joiner_jit_trace.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_size: int,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      decoder:
+        The decoder model.
+      joiner:
+        The joiner model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+      context_size:
+        The context size of the decoder model.
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    states = encoder.get_init_states(batch_size=features.size(0), device=device)
+
+    encoder_out, encoder_out_lens, _ = encoder(
+        x=features,
+        x_lens=feature_lengths,
+        states=states,
+    )
+
+    hyps = greedy_search(
+        decoder=decoder,
+        joiner=joiner,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        context_size=args.context_size,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/joiner.py b/egs/librispeech/ASR/lstm_transducer_stateless/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/lstm.py b/egs/librispeech/ASR/lstm_transducer_stateless/lstm.py
new file mode 100644
index 000000000..bbab16af7
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/lstm.py
@@ -0,0 +1,867 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import List, Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv2d,
+    ScaledLinear,
+    ScaledLSTM,
+)
+from torch import nn
+
+LOG_EPSILON = math.log(1e-10)
+
+
+def unstack_states(
+    states: Tuple[torch.Tensor, torch.Tensor]
+) -> List[Tuple[torch.Tensor, torch.Tensor]]:
+    """
+    Unstack the lstm states corresponding to a batch of utterances into a list
+    of states, where the i-th entry is the state from the i-th utterance.
+
+    Args:
+      states:
+        A tuple of 2 elements.
+        ``states[0]`` is the lstm hidden states, of a batch of utterance.
+        ``states[1]`` is the lstm cell states, of a batch of utterances.
+
+    Returns:
+      A list of states.
+        ``states[i]`` is a tuple of 2 elememts of i-th utterance.
+        ``states[i][0]`` is the lstm hidden states of i-th utterance.
+        ``states[i][1]`` is the lstm cell states of i-th utterance.
+    """
+    hidden_states, cell_states = states
+
+    list_hidden_states = hidden_states.unbind(dim=1)
+    list_cell_states = cell_states.unbind(dim=1)
+
+    ans = [
+        (h.unsqueeze(1), c.unsqueeze(1))
+        for (h, c) in zip(list_hidden_states, list_cell_states)
+    ]
+    return ans
+
+
+def stack_states(
+    states_list: List[Tuple[torch.Tensor, torch.Tensor]]
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Stack list of lstm states corresponding to separate utterances into a single
+    lstm state so that it can be used as an input for lstm when those utterances
+    are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponds to the lstm state for a single
+        utterance.
+        ``states[i]`` is a tuple of 2 elememts of i-th utterance.
+        ``states[i][0]`` is the lstm hidden states of i-th utterance.
+        ``states[i][1]`` is the lstm cell states of i-th utterance.
+
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      It is a tuple of 2 elements.
+        ``states[0]`` is the lstm hidden states, of a batch of utterance.
+        ``states[1]`` is the lstm cell states, of a batch of utterances.
+    """
+    hidden_states = torch.cat([s[0] for s in states_list], dim=1)
+    cell_states = torch.cat([s[1] for s in states_list], dim=1)
+    ans = (hidden_states, cell_states)
+    return ans
+
+
+class RNN(EncoderInterface):
+    """
+    Args:
+      num_features (int):
+        Number of input features.
+      subsampling_factor (int):
+        Subsampling factor of encoder (convolution layers before lstm layers) (default=4).  # noqa
+      d_model (int):
+        Output dimension (default=512).
+      dim_feedforward (int):
+        Feedforward dimension (default=2048).
+      rnn_hidden_size (int):
+        Hidden dimension for lstm layers (default=1024).
+      num_encoder_layers (int):
+        Number of encoder layers (default=12).
+      dropout (float):
+        Dropout rate (default=0.1).
+      layer_dropout (float):
+        Dropout value for model-level warmup (default=0.075).
+      aux_layer_period (int):
+        Period of auxiliary layers used for random combiner during training.
+        If set to 0, will not use the random combiner (Default).
+        You can set a positive integer to use the random combiner, e.g., 3.
+      is_pnnx:
+        True to make this class exportable via PNNX.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 512,
+        dim_feedforward: int = 2048,
+        rnn_hidden_size: int = 1024,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        aux_layer_period: int = 0,
+        is_pnnx: bool = False,
+    ) -> None:
+        super(RNN, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(
+            num_features,
+            d_model,
+            is_pnnx=is_pnnx,
+        )
+
+        self.is_pnnx = is_pnnx
+
+        self.num_encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.rnn_hidden_size = rnn_hidden_size
+
+        encoder_layer = RNNEncoderLayer(
+            d_model=d_model,
+            dim_feedforward=dim_feedforward,
+            rnn_hidden_size=rnn_hidden_size,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+        self.encoder = RNNEncoder(
+            encoder_layer,
+            num_encoder_layers,
+            aux_layers=list(
+                range(
+                    num_encoder_layers // 3,
+                    num_encoder_layers - 1,
+                    aux_layer_period,
+                )
+            )
+            if aux_layer_period > 0
+            else None,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C), where N is the batch size,
+            T is the sequence length, C is the feature dimension.
+          x_lens:
+            A tensor of shape (N,), containing the number of frames in `x`
+            before padding.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (num_layers, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (num_layers, N, rnn_hidden_size).
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          A tuple of 3 tensors:
+            - embeddings: its shape is (N, T', d_model), where T' is the output
+              sequence lengths.
+            - lengths: a tensor of shape (batch_size,) containing the number of
+              frames in `embeddings` before padding.
+            - updated states, whose shape is the same as the input states.
+        """
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # lengths = ((x_lens - 3) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        if not self.is_pnnx:
+            lengths = (((x_lens - 3) >> 1) - 1) >> 1
+        else:
+            lengths1 = torch.floor((x_lens - 3) / 2)
+            lengths = torch.floor((lengths1 - 1) / 2)
+            lengths = lengths.to(x_lens)
+
+        if not torch.jit.is_tracing():
+            assert x.size(0) == lengths.max().item()
+
+        if states is None:
+            x = self.encoder(x, warmup=warmup)[0]
+            # torch.jit.trace requires returned types to be the same as annotated  # noqa
+            new_states = (torch.empty(0), torch.empty(0))
+        else:
+            assert not self.training
+            assert len(states) == 2
+            if not torch.jit.is_tracing():
+                # for hidden state
+                assert states[0].shape == (
+                    self.num_encoder_layers,
+                    x.size(1),
+                    self.d_model,
+                )
+                # for cell state
+                assert states[1].shape == (
+                    self.num_encoder_layers,
+                    x.size(1),
+                    self.rnn_hidden_size,
+                )
+            x, new_states = self.encoder(x, states)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        return x, lengths, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self, batch_size: int = 1, device: torch.device = torch.device("cpu")
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Get model initial states."""
+        # for rnn hidden states
+        hidden_states = torch.zeros(
+            (self.num_encoder_layers, batch_size, self.d_model), device=device
+        )
+        cell_states = torch.zeros(
+            (self.num_encoder_layers, batch_size, self.rnn_hidden_size),
+            device=device,
+        )
+        return (hidden_states, cell_states)
+
+
+class RNNEncoderLayer(nn.Module):
+    """
+    RNNEncoderLayer is made up of lstm and feedforward networks.
+
+    Args:
+      d_model:
+        The number of expected features in the input (required).
+      dim_feedforward:
+        The dimension of feedforward network model (default=2048).
+      rnn_hidden_size:
+        The hidden dimension of rnn layer.
+      dropout:
+        The dropout value (default=0.1).
+      layer_dropout:
+        The dropout value for model-level warmup (default=0.075).
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        dim_feedforward: int,
+        rnn_hidden_size: int,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        super(RNNEncoderLayer, self).__init__()
+        self.layer_dropout = layer_dropout
+        self.d_model = d_model
+        self.rnn_hidden_size = rnn_hidden_size
+
+        assert rnn_hidden_size >= d_model, (rnn_hidden_size, d_model)
+        self.lstm = ScaledLSTM(
+            input_size=d_model,
+            hidden_size=rnn_hidden_size,
+            proj_size=d_model if rnn_hidden_size > d_model else 0,
+            num_layers=1,
+            dropout=0.0,
+        )
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).  # noqa
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+          src:
+            The sequence to the encoder layer (required).
+            Its shape is (S, N, E), where S is the sequence length,
+            N is the batch size, and E is the feature number.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (1, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (1, N, rnn_hidden_size).
+          warmup:
+            It controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # lstm module
+        if states is None:
+            src_lstm = self.lstm(src)[0]
+            # torch.jit.trace requires returned types be the same as annotated
+            new_states = (torch.empty(0), torch.empty(0))
+        else:
+            assert not self.training
+            assert len(states) == 2
+            if not torch.jit.is_tracing():
+                # for hidden state
+                assert states[0].shape == (1, src.size(1), self.d_model)
+                # for cell state
+                assert states[1].shape == (1, src.size(1), self.rnn_hidden_size)
+            src_lstm, new_states = self.lstm(src, states)
+        src = self.dropout(src_lstm) + src
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src, new_states
+
+
+class RNNEncoder(nn.Module):
+    """
+    RNNEncoder is a stack of N encoder layers.
+
+    Args:
+      encoder_layer:
+        An instance of the RNNEncoderLayer() class (required).
+      num_layers:
+        The number of sub-encoder-layers in the encoder (required).
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        aux_layers: Optional[List[int]] = None,
+    ) -> None:
+        super(RNNEncoder, self).__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+        self.d_model = encoder_layer.d_model
+        self.rnn_hidden_size = encoder_layer.rnn_hidden_size
+
+        self.aux_layers: List[int] = []
+        self.combiner: Optional[nn.Module] = None
+        if aux_layers is not None:
+            assert len(set(aux_layers)) == len(aux_layers)
+            assert num_layers - 1 not in aux_layers
+            self.aux_layers = aux_layers + [num_layers - 1]
+            self.combiner = RandomCombine(
+                num_inputs=len(self.aux_layers),
+                final_weight=0.5,
+                pure_prob=0.333,
+                stddev=2.0,
+            )
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Pass the input through the encoder layer in turn.
+
+        Args:
+          src:
+            The sequence to the encoder layer (required).
+            Its shape is (S, N, E), where S is the sequence length,
+            N is the batch size, and E is the feature number.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (num_layers, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (num_layers, N, rnn_hidden_size).
+          warmup:
+            It controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+        """
+        if states is not None:
+            assert not self.training
+            assert len(states) == 2
+            if not torch.jit.is_tracing():
+                # for hidden state
+                assert states[0].shape == (
+                    self.num_layers,
+                    src.size(1),
+                    self.d_model,
+                )
+                # for cell state
+                assert states[1].shape == (
+                    self.num_layers,
+                    src.size(1),
+                    self.rnn_hidden_size,
+                )
+
+        output = src
+
+        outputs = []
+
+        new_hidden_states = []
+        new_cell_states = []
+
+        for i, mod in enumerate(self.layers):
+            if states is None:
+                output = mod(output, warmup=warmup)[0]
+            else:
+                layer_state = (
+                    states[0][i : i + 1, :, :],  # h: (1, N, d_model)
+                    states[1][i : i + 1, :, :],  # c: (1, N, rnn_hidden_size)
+                )
+                output, (h, c) = mod(output, layer_state)
+                new_hidden_states.append(h)
+                new_cell_states.append(c)
+
+            if self.combiner is not None and i in self.aux_layers:
+                outputs.append(output)
+
+        if self.combiner is not None:
+            output = self.combiner(outputs)
+
+        if states is None:
+            new_states = (torch.empty(0), torch.empty(0))
+        else:
+            new_states = (
+                torch.cat(new_hidden_states, dim=0),
+                torch.cat(new_cell_states, dim=0),
+            )
+
+        return output, new_states
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-3)//2-1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        is_pnnx: bool = False,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >= 9, in_channels >= 9.
+          out_channels
+            Output dim. The output shape is (N, ((T-3)//2-1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+          is_pnnx:
+            True if we are converting the model to PNNX format.
+            False otherwise.
+        """
+        assert in_channels >= 9
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=0,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 3) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+        # ncnn supports only batch size == 1
+        self.is_pnnx = is_pnnx
+        self.conv_out_dim = self.out.weight.shape[1]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-3)//2-1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+
+        if torch.jit.is_tracing() and self.is_pnnx:
+            x = x.permute(0, 2, 1, 3).reshape(1, -1, self.conv_out_dim)
+            x = self.out(x)
+        else:
+            # Now x is of shape (N, odim, ((T-3)//2-1)//2, ((idim-3)//2-1)//2)
+            b, c, t, f = x.size()
+            x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+
+        # Now x is of shape (N, ((T-3)//2-1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+class RandomCombine(nn.Module):
+    """
+    This module combines a list of Tensors, all with the same shape, to
+    produce a single output of that same shape which, in training time,
+    is a random combination of all the inputs; but which in test time
+    will be just the last input.
+
+    The idea is that the list of Tensors will be a list of outputs of multiple
+    conformer layers.  This has a similar effect as iterated loss. (See:
+    DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER
+    NETWORKS).
+    """
+
+    def __init__(
+        self,
+        num_inputs: int,
+        final_weight: float = 0.5,
+        pure_prob: float = 0.5,
+        stddev: float = 2.0,
+    ) -> None:
+        """
+        Args:
+          num_inputs:
+            The number of tensor inputs, which equals the number of layers'
+            outputs that are fed into this module.  E.g. in an 18-layer neural
+            net if we output layers 16, 12, 18, num_inputs would be 3.
+          final_weight:
+            The amount of weight or probability we assign to the
+            final layer when randomly choosing layers or when choosing
+            continuous layer weights.
+          pure_prob:
+            The probability, on each frame, with which we choose
+            only a single layer to output (rather than an interpolation)
+          stddev:
+            A standard deviation that we add to log-probs for computing
+            randomized weights.
+
+        The method of choosing which layers, or combinations of layers, to use,
+        is conceptually as follows::
+
+            With probability `pure_prob`::
+               With probability `final_weight`: choose final layer,
+               Else: choose random non-final layer.
+            Else::
+               Choose initial log-weights that correspond to assigning
+               weight `final_weight` to the final layer and equal
+               weights to other layers; then add Gaussian noise
+               with variance `stddev` to these log-weights, and normalize
+               to weights (note: the average weight assigned to the
+               final layer here will not be `final_weight` if stddev>0).
+        """
+        super().__init__()
+        assert 0 <= pure_prob <= 1, pure_prob
+        assert 0 < final_weight < 1, final_weight
+        assert num_inputs >= 1
+
+        self.num_inputs = num_inputs
+        self.final_weight = final_weight
+        self.pure_prob = pure_prob
+        self.stddev = stddev
+
+        self.final_log_weight = (
+            torch.tensor((final_weight / (1 - final_weight)) * (self.num_inputs - 1))
+            .log()
+            .item()
+        )
+
+    def forward(self, inputs: List[torch.Tensor]) -> torch.Tensor:
+        """Forward function.
+        Args:
+          inputs:
+            A list of Tensor, e.g. from various layers of a transformer.
+            All must be the same shape, of (*, num_channels)
+        Returns:
+          A Tensor of shape (*, num_channels).  In test mode
+          this is just the final input.
+        """
+        num_inputs = self.num_inputs
+        assert len(inputs) == num_inputs
+        if not self.training or torch.jit.is_scripting():
+            return inputs[-1]
+
+        # Shape of weights: (*, num_inputs)
+        num_channels = inputs[0].shape[-1]
+        num_frames = inputs[0].numel() // num_channels
+
+        ndim = inputs[0].ndim
+        # stacked_inputs: (num_frames, num_channels, num_inputs)
+        stacked_inputs = torch.stack(inputs, dim=ndim).reshape(
+            (num_frames, num_channels, num_inputs)
+        )
+
+        # weights: (num_frames, num_inputs)
+        weights = self._get_random_weights(
+            inputs[0].dtype, inputs[0].device, num_frames
+        )
+
+        weights = weights.reshape(num_frames, num_inputs, 1)
+        # ans: (num_frames, num_channels, 1)
+        ans = torch.matmul(stacked_inputs, weights)
+        # ans: (*, num_channels)
+
+        ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,))
+
+        # The following if causes errors for torch script in torch 1.6.0
+        #  if __name__ == "__main__":
+        #      # for testing only...
+        #      print("Weights = ", weights.reshape(num_frames, num_inputs))
+        return ans
+
+    def _get_random_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), such that
+          `ans.sum(dim=1)` is all ones.
+        """
+        pure_prob = self.pure_prob
+        if pure_prob == 0.0:
+            return self._get_random_mixed_weights(dtype, device, num_frames)
+        elif pure_prob == 1.0:
+            return self._get_random_pure_weights(dtype, device, num_frames)
+        else:
+            p = self._get_random_pure_weights(dtype, device, num_frames)
+            m = self._get_random_mixed_weights(dtype, device, num_frames)
+            return torch.where(
+                torch.rand(num_frames, 1, device=device) < self.pure_prob, p, m
+            )
+
+    def _get_random_pure_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ):
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A one-hot tensor of shape `(num_frames, self.num_inputs)`, with
+          exactly one weight equal to 1.0 on each frame.
+        """
+        final_prob = self.final_weight
+
+        # final contains self.num_inputs - 1 in all elements
+        final = torch.full((num_frames,), self.num_inputs - 1, device=device)
+        # nonfinal contains random integers in [0..num_inputs - 2], these are for non-final weights.  # noqa
+        nonfinal = torch.randint(self.num_inputs - 1, (num_frames,), device=device)
+
+        indexes = torch.where(
+            torch.rand(num_frames, device=device) < final_prob, final, nonfinal
+        )
+        ans = torch.nn.functional.one_hot(indexes, num_classes=self.num_inputs).to(
+            dtype=dtype
+        )
+        return ans
+
+    def _get_random_mixed_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ):
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), which elements
+          in [0..1] that sum to one over the second axis, i.e.
+          `ans.sum(dim=1)` is all ones.
+        """
+        logprobs = (
+            torch.randn(num_frames, self.num_inputs, dtype=dtype, device=device)
+            * self.stddev  # noqa
+        )
+        logprobs[:, -1] += self.final_log_weight
+        return logprobs.softmax(dim=1)
+
+
+def _test_random_combine(final_weight: float, pure_prob: float, stddev: float):
+    print(
+        f"_test_random_combine: final_weight={final_weight}, pure_prob={pure_prob}, stddev={stddev}"  # noqa
+    )
+    num_inputs = 3
+    num_channels = 50
+    m = RandomCombine(
+        num_inputs=num_inputs,
+        final_weight=final_weight,
+        pure_prob=pure_prob,
+        stddev=stddev,
+    )
+
+    x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)]
+
+    y = m(x)
+    assert y.shape == x[0].shape
+    assert torch.allclose(y, x[0])  # .. since actually all ones.
+
+
+def _test_random_combine_main():
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.0)
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.3)
+    _test_random_combine(0.5, 1, 0.3)
+    _test_random_combine(0.5, 0.5, 0.3)
+
+    feature_dim = 50
+    c = RNN(num_features=feature_dim, d_model=128)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+if __name__ == "__main__":
+    feature_dim = 80
+    m = RNN(
+        num_features=feature_dim,
+        d_model=512,
+        rnn_hidden_size=1024,
+        dim_feedforward=2048,
+        num_encoder_layers=12,
+    )
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = m(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
+    num_param = sum([p.numel() for p in m.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    _test_random_combine_main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/lstmp.py b/egs/librispeech/ASR/lstm_transducer_stateless/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/model.py b/egs/librispeech/ASR/lstm_transducer_stateless/model.py
new file mode 100644
index 000000000..e7bad7ed8
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/model.py
@@ -0,0 +1,206 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output
+            contains unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        reduction: str = "sum",
+        delay_penalty: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          delay_penalty:
+            A constant value used to penalize symbol delay, to encourage
+            streaming models to emit symbols earlier.
+            See https://github.com/k2-fsa/k2/issues/955 and
+            https://arxiv.org/pdf/2211.00490.pdf for more details.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens, _ = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                delay_penalty=delay_penalty,
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                delay_penalty=delay_penalty,
+                reduction=reduction,
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/onnx_check.py b/egs/librispeech/ASR/lstm_transducer_stateless/onnx_check.py
new file mode 120000
index 000000000..0b1ea0326
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/onnx_check.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/onnx_check.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/onnx_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless/onnx_pretrained.py
new file mode 120000
index 000000000..099c2882f
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/onnx_pretrained.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/optim.py b/egs/librispeech/ASR/lstm_transducer_stateless/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless/pretrained.py
new file mode 100755
index 000000000..119fcf1fd
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/pretrained.py
@@ -0,0 +1,351 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./lstm_transducer_stateless/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./lstm_transducer_stateless/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./lstm_transducer_stateless/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./lstm_transducer_stateless/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./lstm_transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./lstm_transducer_stateless/exp/pretrained.pt is generated by
+./lstm_transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/scaling.py b/egs/librispeech/ASR/lstm_transducer_stateless/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/stream.py b/egs/librispeech/ASR/lstm_transducer_stateless/stream.py
new file mode 100644
index 000000000..d8f7fd960
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/stream.py
@@ -0,0 +1,146 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class Stream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+        LOG_EPS: float = math.log(1e-10),
+    ) -> None:
+        """
+        Args:
+          params:
+            It's the return value of :func:`get_params`.
+          cut_id:
+            The cut id of the current stream.
+          decoding_graph:
+            The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+            only when --decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+          LOG_EPS:
+            A float value used for padding.
+        """
+        self.LOG_EPS = LOG_EPS
+        self.cut_id = cut_id
+
+        # Containing attention caches and convolution caches
+        self.states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
+
+        # It uses different attributes for different decoding methods.
+        self.context_size = params.context_size
+        self.decoding_method = params.decoding_method
+        if params.decoding_method == "greedy_search":
+            self.hyp = [params.blank_id] * params.context_size
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[params.blank_id] * params.context_size,
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # feature_len is needed to get partial results.
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+            self.hyp: Optional[List[int]] = None
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+        self.ground_truth: str = ""
+
+        self.feature: Optional[torch.Tensor] = None
+        # Make sure all feature frames can be used.
+        # We aim to obtain 1 frame after subsampling.
+        self.chunk_length = params.subsampling_factor
+        self.pad_length = 5
+        self.num_frames = 0
+        self.num_processed_frames = 0
+
+        # After all feature frames are processed, we set this flag to True
+        self._done = False
+
+    def set_feature(self, feature: torch.Tensor) -> None:
+        assert feature.dim() == 2, feature.dim()
+        # tail padding here to alleviate the tail deletion problem
+        num_tail_padded_frames = 35
+        self.num_frames = feature.size(0) + num_tail_padded_frames
+        self.feature = torch.nn.functional.pad(
+            feature,
+            (0, 0, 0, self.pad_length + num_tail_padded_frames),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+
+    def get_feature_chunk(self) -> torch.Tensor:
+        """Get a chunk of feature frames.
+
+        Returns:
+          A tensor of shape (ret_length, feature_dim).
+        """
+        update_length = min(
+            self.num_frames - self.num_processed_frames, self.chunk_length
+        )
+        ret_length = update_length + self.pad_length
+
+        ret_feature = self.feature[
+            self.num_processed_frames : self.num_processed_frames + ret_length
+        ]
+        # Cut off used frames.
+        # self.feature = self.feature[update_length:]
+
+        self.num_processed_frames += update_length
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_feature
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    @property
+    def done(self) -> bool:
+        """Return True if all feature frames are processed."""
+        return self._done
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.decoding_method == "greedy_search":
+            return self.hyp[self.context_size :]
+        elif self.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.context_size :]
+        else:
+            assert self.decoding_method == "fast_beam_search"
+            return self.hyp
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/streaming_decode.py b/egs/librispeech/ASR/lstm_transducer_stateless/streaming_decode.py
new file mode 100755
index 000000000..f989d9bc0
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/streaming_decode.py
@@ -0,0 +1,949 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./lstm_transducer_stateless/streaming_decode.py \
+      --epoch 35 \
+      --avg 10 \
+      --exp-dir lstm_transducer_stateless/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --rnn-hidden-size 1024 \
+      --decoding-method greedy_search \
+      --use-averaged-model True
+
+(2) modified beam search
+./lstm_transducer_stateless/streaming_decode.py \
+      --epoch 35 \
+      --avg 10 \
+      --exp-dir lstm_transducer_stateless/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --rnn-hidden-size 1024 \
+      --decoding-method modified_beam_search \
+      --use-averaged-model True \
+      --beam-size 4
+
+(3) fast beam search
+./lstm_transducer_stateless/streaming_decode.py \
+      --epoch 35 \
+      --avg 10 \
+      --exp-dir lstm_transducer_stateless/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --rnn-hidden-size 1024 \
+      --decoding-method fast_beam_search \
+      --use-averaged-model True \
+      --beam 4 \
+      --max-contexts 4 \
+      --max-states 8
+"""
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from lstm import LOG_EPSILON, stack_states, unstack_states
+from stream import Stream
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import one_best_decoding
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_emformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--sampling-rate",
+        type=float,
+        default=16000,
+        help="Sample rate of the audio",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded in parallel",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    T = encoder_out.size(1)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (batch_size, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+    beam: int = 4,
+):
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      beam:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    streams: List[Stream],
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      streams:
+        A list of stream objects.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyps[i]
+
+
+def decode_one_chunk(
+    model: nn.Module,
+    streams: List[Stream],
+    params: AttributeDict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> List[int]:
+    """
+    Args:
+      model:
+        The Transducer model.
+      streams:
+        A list of Stream objects.
+      params:
+        It is returned by :func:`get_params`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+       A list of indexes indicating the finished streams.
+    """
+    device = next(model.parameters()).device
+
+    feature_list = []
+    feature_len_list = []
+    state_list = []
+    num_processed_frames_list = []
+
+    for stream in streams:
+        # We should first get `stream.num_processed_frames`
+        # before calling `stream.get_feature_chunk()`
+        # since `stream.num_processed_frames` would be updated
+        num_processed_frames_list.append(stream.num_processed_frames)
+        feature = stream.get_feature_chunk()
+        feature_len = feature.size(0)
+        feature_list.append(feature)
+        feature_len_list.append(feature_len)
+        state_list.append(stream.states)
+
+    features = pad_sequence(
+        feature_list, batch_first=True, padding_value=LOG_EPSILON
+    ).to(device)
+    feature_lens = torch.tensor(feature_len_list, device=device)
+    num_processed_frames = torch.tensor(num_processed_frames_list, device=device)
+
+    # Make sure it has at least 1 frame after subsampling
+    tail_length = params.subsampling_factor + 5
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPSILON,
+        )
+
+    # Stack states of all streams
+    states = stack_states(state_list)
+
+    encoder_out, encoder_out_lens, states = model.encoder(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+    elif params.decoding_method == "fast_beam_search":
+        # feature_len is needed to get partial results.
+        # The rnnt_decoding_stream for fast_beam_search.
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            processed_lens = (
+                num_processed_frames // params.subsampling_factor + encoder_out_lens
+            )
+        fast_beam_search_one_best(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    # Update cached states of each stream
+    state_list = unstack_states(states)
+    for i, s in enumerate(state_list):
+        streams[i].states = s
+
+    finished_streams = [i for i, stream in enumerate(streams) if stream.done]
+    return finished_streams
+
+
+def create_streaming_feature_extractor() -> Fbank:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return Fbank(opts)
+
+
+def decode_dataset(
+    cuts: CutSet,
+    model: nn.Module,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+):
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The Transducer model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = next(model.parameters()).device
+
+    log_interval = 300
+
+    fbank = create_streaming_feature_extractor()
+
+    decode_results = []
+    streams = []
+    for num, cut in enumerate(cuts):
+        # Each utterance has a Stream.
+        stream = Stream(
+            params=params,
+            cut_id=cut.id,
+            decoding_graph=decoding_graph,
+            device=device,
+            LOG_EPS=LOG_EPSILON,
+        )
+
+        stream.states = model.encoder.get_init_states(device=device)
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+        feature = fbank(samples)
+        stream.set_feature(feature)
+        stream.ground_truth = cut.supervisions[0].text
+
+        streams.append(stream)
+
+        while len(streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                model=model,
+                streams=streams,
+                params=params,
+                decoding_graph=decoding_graph,
+            )
+
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        streams[i].id,
+                        streams[i].ground_truth.split(),
+                        sp.decode(streams[i].decoding_result()).split(),
+                    )
+                )
+                del streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    while len(streams) > 0:
+        finished_streams = decode_one_chunk(
+            model=model,
+            streams=streams,
+            params=params,
+            decoding_graph=decoding_graph,
+        )
+
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    streams[i].id,
+                    streams[i].ground_truth.split(),
+                    sp.decode(streams[i].decoding_result()).split(),
+                )
+            )
+            del streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    else:
+        key = f"beam_size_{params.beam_size}"
+
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=sorted(results))
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-streaming-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.device = device
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            model=model,
+            params=params,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220810)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py b/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py
new file mode 100755
index 000000000..91ef53e24
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py
@@ -0,0 +1,91 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./lstm_transducer_stateless/test_model.py
+"""
+
+import os
+from pathlib import Path
+
+import torch
+from export import (
+    export_decoder_model_jit_trace,
+    export_encoder_model_jit_trace,
+    export_joiner_model_jit_trace,
+)
+from lstm import stack_states, unstack_states
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.encoder_dim = 512
+    params.rnn_hidden_size = 1024
+    params.num_encoder_layers = 12
+    params.aux_layer_period = 0
+    params.exp_dir = Path("exp_test_model")
+
+    model = get_transducer_model(params)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    params.exp_dir.mkdir(exist_ok=True)
+
+    encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+    print("The model has been successfully exported using jit.trace.")
+
+
+def test_states_stack_and_unstack():
+    layer, batch, hidden, cell = 12, 100, 512, 1024
+    states = (
+        torch.randn(layer, batch, hidden),
+        torch.randn(layer, batch, cell),
+    )
+    states2 = stack_states(unstack_states(states))
+    assert torch.allclose(states[0], states2[0])
+    assert torch.allclose(states[1], states2[1])
+
+
+def main():
+    test_model()
+    test_states_stack_and_unstack()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/test_scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless/test_scaling_converter.py
new file mode 100644
index 000000000..7567dd58c
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/test_scaling_converter.py
@@ -0,0 +1,257 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./lstm_transducer_stateless/test_scaling_converter.py
+"""
+
+import copy
+
+import torch
+from scaling import (
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledEmbedding,
+    ScaledLinear,
+    ScaledLSTM,
+)
+from scaling_converter import (
+    convert_scaled_to_non_scaled,
+    scaled_conv1d_to_conv1d,
+    scaled_conv2d_to_conv2d,
+    scaled_embedding_to_embedding,
+    scaled_linear_to_linear,
+    scaled_lstm_to_lstm,
+)
+from train import get_params, get_transducer_model
+
+
+def get_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.encoder_dim = 512
+    params.rnn_hidden_size = 1024
+    params.num_encoder_layers = 12
+    params.aux_layer_period = -1
+
+    model = get_transducer_model(params)
+    return model
+
+
+def test_scaled_linear_to_linear():
+    N = 5
+    in_features = 10
+    out_features = 20
+    for bias in [True, False]:
+        scaled_linear = ScaledLinear(
+            in_features=in_features,
+            out_features=out_features,
+            bias=bias,
+        )
+        linear = scaled_linear_to_linear(scaled_linear)
+        x = torch.rand(N, in_features)
+
+        y1 = scaled_linear(x)
+        y2 = linear(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_linear = torch.jit.script(scaled_linear)
+        jit_linear = torch.jit.script(linear)
+
+        y3 = jit_scaled_linear(x)
+        y4 = jit_linear(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_conv1d_to_conv1d():
+    in_channels = 3
+    for bias in [True, False]:
+        scaled_conv1d = ScaledConv1d(
+            in_channels,
+            6,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        conv1d = scaled_conv1d_to_conv1d(scaled_conv1d)
+
+        x = torch.rand(20, in_channels, 10)
+        y1 = scaled_conv1d(x)
+        y2 = conv1d(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_conv1d = torch.jit.script(scaled_conv1d)
+        jit_conv1d = torch.jit.script(conv1d)
+
+        y3 = jit_scaled_conv1d(x)
+        y4 = jit_conv1d(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_conv2d_to_conv2d():
+    in_channels = 1
+    for bias in [True, False]:
+        scaled_conv2d = ScaledConv2d(
+            in_channels=in_channels,
+            out_channels=3,
+            kernel_size=3,
+            padding=1,
+            bias=bias,
+        )
+
+        conv2d = scaled_conv2d_to_conv2d(scaled_conv2d)
+
+        x = torch.rand(20, in_channels, 10, 20)
+        y1 = scaled_conv2d(x)
+        y2 = conv2d(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_conv2d = torch.jit.script(scaled_conv2d)
+        jit_conv2d = torch.jit.script(conv2d)
+
+        y3 = jit_scaled_conv2d(x)
+        y4 = jit_conv2d(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_embedding_to_embedding():
+    scaled_embedding = ScaledEmbedding(
+        num_embeddings=500,
+        embedding_dim=10,
+        padding_idx=0,
+    )
+    embedding = scaled_embedding_to_embedding(scaled_embedding)
+
+    for s in [10, 100, 300, 500, 800, 1000]:
+        x = torch.randint(low=0, high=500, size=(s,))
+        scaled_y = scaled_embedding(x)
+        y = embedding(x)
+        assert torch.equal(scaled_y, y)
+
+
+def test_scaled_lstm_to_lstm():
+    input_size = 512
+    batch_size = 20
+    for bias in [True, False]:
+        for hidden_size in [512, 1024]:
+            scaled_lstm = ScaledLSTM(
+                input_size=input_size,
+                hidden_size=hidden_size,
+                num_layers=1,
+                bias=bias,
+                proj_size=0 if hidden_size == input_size else input_size,
+            )
+
+            lstm = scaled_lstm_to_lstm(scaled_lstm)
+
+            x = torch.rand(200, batch_size, input_size)
+            h0 = torch.randn(1, batch_size, input_size)
+            c0 = torch.randn(1, batch_size, hidden_size)
+
+            y1, (h1, c1) = scaled_lstm(x, (h0, c0))
+            y2, (h2, c2) = lstm(x, (h0, c0))
+            assert torch.allclose(y1, y2)
+            assert torch.allclose(h1, h2)
+            assert torch.allclose(c1, c2)
+
+            jit_scaled_lstm = torch.jit.trace(lstm, (x, (h0, c0)))
+            y3, (h3, c3) = jit_scaled_lstm(x, (h0, c0))
+            assert torch.allclose(y1, y3)
+            assert torch.allclose(h1, h3)
+            assert torch.allclose(c1, c3)
+
+
+def test_convert_scaled_to_non_scaled():
+    for inplace in [False, True]:
+        model = get_model()
+        model.eval()
+
+        orig_model = copy.deepcopy(model)
+
+        converted_model = convert_scaled_to_non_scaled(model, inplace=inplace)
+
+        model = orig_model
+
+        # test encoder
+        N = 2
+        T = 100
+        vocab_size = model.decoder.vocab_size
+
+        x = torch.randn(N, T, 80, dtype=torch.float32)
+        x_lens = torch.full((N,), x.size(1))
+
+        e1, e1_lens, _ = model.encoder(x, x_lens)
+        e2, e2_lens, _ = converted_model.encoder(x, x_lens)
+
+        assert torch.all(torch.eq(e1_lens, e2_lens))
+        assert torch.allclose(e1, e2), (e1 - e2).abs().max()
+
+        # test decoder
+        U = 50
+        y = torch.randint(low=1, high=vocab_size - 1, size=(N, U))
+
+        d1 = model.decoder(y)
+        d2 = model.decoder(y)
+
+        assert torch.allclose(d1, d2)
+
+        # test simple projection
+        lm1 = model.simple_lm_proj(d1)
+        am1 = model.simple_am_proj(e1)
+
+        lm2 = converted_model.simple_lm_proj(d2)
+        am2 = converted_model.simple_am_proj(e2)
+
+        assert torch.allclose(lm1, lm2)
+        assert torch.allclose(am1, am2)
+
+        # test joiner
+        e = torch.rand(2, 3, 4, 512)
+        d = torch.rand(2, 3, 4, 512)
+
+        j1 = model.joiner(e, d)
+        j2 = converted_model.joiner(e, d)
+        assert torch.allclose(j1, j2)
+
+
+@torch.no_grad()
+def main():
+    test_scaled_linear_to_linear()
+    test_scaled_conv1d_to_conv1d()
+    test_scaled_conv2d_to_conv2d()
+    test_scaled_embedding_to_embedding()
+    test_scaled_lstm_to_lstm()
+    test_convert_scaled_to_non_scaled()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220730)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/train.py b/egs/librispeech/ASR/lstm_transducer_stateless/train.py
new file mode 100755
index 000000000..feb81d500
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/train.py
@@ -0,0 +1,1137 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./lstm_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir lstm_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./lstm_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir lstm_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from lstm import RNN
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of RNN encoder layers..",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Encoder output dimesion.",
+    )
+
+    parser.add_argument(
+        "--rnn-hidden-size",
+        type=int,
+        default=1024,
+        help="Hidden dim for LSTM layers.",
+    )
+
+    parser.add_argument(
+        "--aux-layer-period",
+        type=int,
+        default=0,
+        help="""Peroid of auxiliary layers used for randomly combined during training.
+        If set to 0, will not use the random combiner (Default).
+        You can set a positive integer to use the random combiner, e.g., 3.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=35,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=10,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "dim_feedforward": 2048,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = RNN(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        rnn_hidden_size=params.rnn_hidden_size,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        aux_layer_period=params.aux_layer_period,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    # # overwrite it
+    # scheduler.base_lrs = [params.initial_lr for _ in scheduler.base_lrs]
+    # print(scheduler.base_lrs)
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./lstm.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 3) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/__init__.py b/egs/librispeech/ASR/lstm_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/asr_datamodule.py b/egs/librispeech/ASR/lstm_transducer_stateless2/asr_datamodule.py
new file mode 120000
index 000000000..3ba9ada4f
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/lstm_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/decode.py b/egs/librispeech/ASR/lstm_transducer_stateless2/decode.py
new file mode 100755
index 000000000..1a724830b
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/decode.py
@@ -0,0 +1,990 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search (with LM shallow fusion)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+(9) modified beam search with LM shallow fusion + LODR
+./lstm_transducer_stateless2/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --max-duration 600 \
+    --exp-dir ./lstm_transducer_stateless2/exp \
+    --decoding-method modified_beam_search_LODR \
+    --beam 4 \
+    --lm-type rnn \
+    --lm-scale 0.4 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16 \
+"""
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+    modified_beam_search_ngram_rescoring,
+)
+from librispeech import LibriSpeech
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_ngram_rescoring
+          - modified_beam_search_lm_shallow_fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=3,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring""",
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    # tail padding here to alleviate the tail deletion problem
+    num_tail_padded_frames = 35
+    feature = torch.nn.functional.pad(
+        feature,
+        (0, 0, 0, num_tail_padded_frames),
+        mode="constant",
+        value=LOG_EPS,
+    )
+    feature_lens += num_tail_padded_frames
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_ngram_rescoring":
+        hyp_tokens = modified_beam_search_ngram_rescoring(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+        logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_ngram_rescoring",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+    else:
+        LM = None
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    # test_clean_cuts = test_clean_cuts.subset(first=500)
+    test_other_cuts = librispeech.test_other_cuts()
+    # test_other_cuts = test_other_cuts.subset(first=500)
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/decoder.py b/egs/librispeech/ASR/lstm_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/encoder_interface.py b/egs/librispeech/ASR/lstm_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/export-for-ncnn.py b/egs/librispeech/ASR/lstm_transducer_stateless2/export-for-ncnn.py
new file mode 100755
index 000000000..2b8c92208
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/export-for-ncnn.py
@@ -0,0 +1,348 @@
+#!/usr/bin/env python3
+
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for more details about how to use this file.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+
+cd exp
+ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+2. Export via torch.jit.trace()
+
+./lstm_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+
+cd ./lstm_transducer_stateless2/exp
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+See ./streaming-ncnn-decode.py
+and
+https://github.com/k2-fsa/sherpa-ncnn
+for usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+    states = encoder_model.get_init_states()
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size, <blk> is
+    # defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    params.is_pnnx = True
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx-zh.py b/egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx-zh.py
new file mode 100755
index 000000000..2a52e2eec
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx-zh.py
@@ -0,0 +1,633 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-wenetspeech-lstm-transducer-stateless-2022-10-14
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-wenetspeech-lstm-transducer-stateless-2022-10-14
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lexicon.txt"
+git lfs pull --include "data/L.pt"
+git lfs pull --include "exp/epoch-11.pt"
+git lfs pull --include "exp/epoch-10.pt"
+
+popd
+
+2. Export the model to ONNX
+
+./lstm_transducer_stateless2/export-onnx-zh.py \
+  --lang-dir ./icefall-asr-wenetspeech-lstm-transducer-stateless-2022-10-14/data/lang_char \
+  --use-averaged-model 1 \
+  --epoch 11 \
+  --avg 1 \
+  --exp-dir ./icefall-asr-wenetspeech-lstm-transducer-stateless-2022-10-14/exp \
+  --num-encoder-layers 12 \
+  --encoder-dim 512 \
+  --rnn-hidden-size 1024
+
+It will generate the following files inside $repo/exp:
+
+  - encoder-epoch-11-avg-1.onnx
+  - decoder-epoch-11-avg-1.onnx
+  - joiner-epoch-11-avg-1.onnx
+  - encoder-epoch-11-avg-1.int8.onnx
+  - decoder-epoch-11-avg-1.int8.onnx
+  - joiner-epoch-11-avg-1.int8.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from lstm import RNN
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for RNN and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: RNN, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            An RNN encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of RNN.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (num_layers, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (num_layers, N, rnn_hidden_size).
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - updated states, whose shape is the same as the input states.
+        """
+        N = x.size(0)
+        T = x.size(1)
+        x_lens = torch.tensor([T] * N, dtype=torch.int64, device=x.device)
+        encoder_out, _, next_states = self.encoder(x, x_lens, states)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, next_states
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has the following inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - state0, a tensor of shape (num_encoder_layers, batch_size, d_model)
+        - state1, a tensor of shape (num_encoder_layers, batch_size, rnn_hidden_size)
+
+    and it has 3 outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - new_state0, a tensor of shape (num_encoder_layers, batch_size, d_model)
+        - new_state1, a tensor of shape (num_encoder_layers, batch_size, rnn_hidden_size)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    num_encoder_layers = encoder_model.encoder.num_encoder_layers
+    d_model = encoder_model.encoder.d_model
+    rnn_hidden_size = encoder_model.encoder.rnn_hidden_size
+
+    decode_chunk_len = 4
+    T = 9
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.encoder.get_init_states()
+    # state0: (num_encoder_layers, batch_size, d_model)
+    # state1: (num_encoder_layers, batch_size, rnn_hidden_size)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, states),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "state0", "state1"],
+        output_names=["encoder_out", "new_state0", "new_state1"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "state0": {1: "N"},
+            "state1": {1: "N"},
+            "encoder_out": {0: "N"},
+            "new_state0": {1: "N"},
+            "new_state1": {1: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "lstm",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "decode_chunk_len": str(decode_chunk_len),  # 32
+        "T": str(T),  # 39
+        "num_encoder_layers": str(num_encoder_layers),
+        "d_model": str(d_model),
+        "rnn_hidden_size": str(rnn_hidden_size),
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx.py b/egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx.py
new file mode 100755
index 000000000..c543628ff
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/export-onnx.py
@@ -0,0 +1,628 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+
+cd exp
+ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./lstm_transducer_stateless2/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+It will generate the following files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+  - encoder-epoch-99-avg-1.int8.onnx
+  - decoder-epoch-99-avg-1.int8.onnx
+  - joiner-epoch-99-avg-1.int8.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from lstm import RNN
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for RNN and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: RNN, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            An RNN encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of RNN.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (num_layers, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (num_layers, N, rnn_hidden_size).
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - updated states, whose shape is the same as the input states.
+        """
+        N = x.size(0)
+        T = x.size(1)
+        x_lens = torch.tensor([T] * N, dtype=torch.int64, device=x.device)
+        encoder_out, _, next_states = self.encoder(x, x_lens, states)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, next_states
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has the following inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - state0, a tensor of shape (num_encoder_layers, batch_size, d_model)
+        - state1, a tensor of shape (num_encoder_layers, batch_size, rnn_hidden_size)
+
+    and it has 3 outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - new_state0, a tensor of shape (num_encoder_layers, batch_size, d_model)
+        - new_state1, a tensor of shape (num_encoder_layers, batch_size, rnn_hidden_size)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    num_encoder_layers = encoder_model.encoder.num_encoder_layers
+    d_model = encoder_model.encoder.d_model
+    rnn_hidden_size = encoder_model.encoder.rnn_hidden_size
+
+    decode_chunk_len = 4
+    T = 9
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.encoder.get_init_states()
+    # state0: (num_encoder_layers, batch_size, d_model)
+    # state1: (num_encoder_layers, batch_size, rnn_hidden_size)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, states),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "state0", "state1"],
+        output_names=["encoder_out", "new_state0", "new_state1"],
+        dynamic_axes={
+            "x": {0: "N"},
+            "state0": {1: "N"},
+            "state1": {1: "N"},
+            "encoder_out": {0: "N"},
+            "new_state0": {1: "N"},
+            "new_state1": {1: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "lstm",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "decode_chunk_len": str(decode_chunk_len),  # 32
+        "T": str(T),  # 39
+        "num_encoder_layers": str(num_encoder_layers),
+        "d_model": str(d_model),
+        "rnn_hidden_size": str(rnn_hidden_size),
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size, <blk> is
+    # defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/export.py b/egs/librispeech/ASR/lstm_transducer_stateless2/export.py
new file mode 100755
index 000000000..5712da25e
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/export.py
@@ -0,0 +1,398 @@
+#!/usr/bin/env python3
+# flake8: noqa
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.trace()
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir ./lstm_transducer_stateless2/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 35 \
+  --avg 10 \
+  --jit-trace 1
+
+It will generate 3 files: `encoder_jit_trace.pt`,
+`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`.
+
+(2) Export `model.state_dict()`
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir ./lstm_transducer_stateless2/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 35 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `lstm_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./lstm_transducer_stateless2/decode.py \
+        --exp-dir ./lstm_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+    # You will find the pre-trained models in icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace.
+        It will generate 3 files:
+         - encoder_jit_trace.pt
+         - decoder_jit_trace.pt
+         - joiner_jit_trace.pt
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+    states = encoder_model.get_init_states()
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size, <blk> is
+    # defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit_trace is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    else:
+        logging.info("Not using torchscript")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/gigaspeech.py b/egs/librispeech/ASR/lstm_transducer_stateless2/gigaspeech.py
new file mode 120000
index 000000000..5242c652a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/gigaspeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/gigaspeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/jit_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless2/jit_pretrained.py
new file mode 100755
index 000000000..728b09104
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/jit_pretrained.py
@@ -0,0 +1,320 @@
+#!/usr/bin/env python3
+# flake8: noqa
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, either exported by `torch.jit.trace()`
+or by `torch.jit.script()`, and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir ./lstm_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit-trace 1
+
+Usage of this script:
+
+./lstm_transducer_stateless2/jit_pretrained.py \
+  --encoder-model-filename ./lstm_transducer_stateless2/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./lstm_transducer_stateless2/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./lstm_transducer_stateless2/exp/joiner_jit_trace.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_size: int,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      decoder:
+        The decoder model.
+      joiner:
+        The joiner model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+      context_size:
+        The context size of the decoder model.
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    states = encoder.get_init_states(batch_size=features.size(0), device=device)
+
+    encoder_out, encoder_out_lens, _ = encoder(
+        x=features,
+        x_lens=feature_lengths,
+        states=states,
+    )
+
+    hyps = greedy_search(
+        decoder=decoder,
+        joiner=joiner,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        context_size=args.context_size,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/joiner.py b/egs/librispeech/ASR/lstm_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/librispeech.py b/egs/librispeech/ASR/lstm_transducer_stateless2/librispeech.py
new file mode 120000
index 000000000..b76723bf5
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/librispeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/librispeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/lstm.py b/egs/librispeech/ASR/lstm_transducer_stateless2/lstm.py
new file mode 120000
index 000000000..a94e35f63
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/lstm.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless/lstm.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/lstmp.py b/egs/librispeech/ASR/lstm_transducer_stateless2/lstmp.py
new file mode 100644
index 000000000..dba6eb520
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/lstmp.py
@@ -0,0 +1,102 @@
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LSTMP(nn.Module):
+    """LSTM with projection.
+
+    PyTorch does not support exporting LSTM with projection to ONNX.
+    This class reimplements LSTM with projection using basic matrix-matrix
+    and matrix-vector operations. It is not intended for training.
+    """
+
+    def __init__(self, lstm: nn.LSTM):
+        """
+        Args:
+          lstm:
+            LSTM with proj_size. We support only uni-directional,
+            1-layer LSTM with projection at present.
+        """
+        super().__init__()
+        assert lstm.bidirectional is False, lstm.bidirectional
+        assert lstm.num_layers == 1, lstm.num_layers
+        assert 0 < lstm.proj_size < lstm.hidden_size, (
+            lstm.proj_size,
+            lstm.hidden_size,
+        )
+
+        assert lstm.batch_first is False, lstm.batch_first
+
+        state_dict = lstm.state_dict()
+
+        w_ih = state_dict["weight_ih_l0"]
+        w_hh = state_dict["weight_hh_l0"]
+
+        b_ih = state_dict["bias_ih_l0"]
+        b_hh = state_dict["bias_hh_l0"]
+
+        w_hr = state_dict["weight_hr_l0"]
+        self.input_size = lstm.input_size
+        self.proj_size = lstm.proj_size
+        self.hidden_size = lstm.hidden_size
+
+        self.w_ih = w_ih
+        self.w_hh = w_hh
+        self.b = b_ih + b_hh
+        self.w_hr = w_hr
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        hx: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          input:
+            A tensor of shape [T, N, hidden_size]
+          hx:
+            A tuple containing:
+              - h0: a tensor of shape (1, N, proj_size)
+              - c0: a tensor of shape (1, N, hidden_size)
+        Returns:
+          Return a tuple containing:
+            - output: a tensor of shape (T, N, proj_size).
+            - A tuple containing:
+               - h: a tensor of shape (1, N, proj_size)
+               - c: a tensor of shape (1, N, hidden_size)
+
+        """
+        x_list = input.unbind(dim=0)  # We use batch_first=False
+
+        if hx is not None:
+            h0, c0 = hx
+        else:
+            h0 = torch.zeros(1, input.size(1), self.proj_size)
+            c0 = torch.zeros(1, input.size(1), self.hidden_size)
+        h0 = h0.squeeze(0)
+        c0 = c0.squeeze(0)
+        y_list = []
+        for x in x_list:
+            gates = F.linear(x, self.w_ih, self.b) + F.linear(h0, self.w_hh)
+            i, f, g, o = gates.chunk(4, dim=1)
+
+            i = i.sigmoid()
+            f = f.sigmoid()
+            g = g.tanh()
+            o = o.sigmoid()
+
+            c = f * c0 + i * g
+            h = o * c.tanh()
+
+            h = F.linear(h, self.w_hr)
+            y_list.append(h)
+
+            c0 = c
+            h0 = h
+
+        y = torch.stack(y_list, dim=0)
+
+        return y, (h0.unsqueeze(0), c0.unsqueeze(0))
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/model.py b/egs/librispeech/ASR/lstm_transducer_stateless2/model.py
new file mode 100644
index 000000000..4957d14b1
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/model.py
@@ -0,0 +1,245 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        decoder_giga: Optional[nn.Module] = None,
+        joiner_giga: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output
+            contains unnormalized probs, i.e., not processed by log-softmax.
+          encoder_dim:
+            Output dimension of the encoder network.
+          decoder_dim:
+            Output dimension of the decoder network.
+          joiner_dim:
+            Input dimension of the joiner network.
+          vocab_size:
+            Output dimension of the joiner network.
+          decoder_giga:
+            Optional. The decoder network for the GigaSpeech dataset.
+          joiner_giga:
+            Optional. The joiner network for the GigaSpeech dataset.
+        """
+        super().__init__()
+
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_giga = decoder_giga
+        self.joiner_giga = joiner_giga
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+        if decoder_giga is not None:
+            self.simple_am_proj_giga = ScaledLinear(
+                encoder_dim, vocab_size, initial_speed=0.5
+            )
+            self.simple_lm_proj_giga = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        libri: bool = True,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        reduction: str = "sum",
+        delay_penalty: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          libri:
+            True to use the decoder and joiner for the LibriSpeech dataset.
+            False to use the decoder and joiner for the GigaSpeech dataset.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          delay_penalty:
+            A constant value used to penalize symbol delay, to encourage
+            streaming models to emit symbols earlier.
+            See https://github.com/k2-fsa/k2/issues/955 and
+            https://arxiv.org/pdf/2211.00490.pdf for more details.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens, _ = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(x_lens > 0)
+
+        if libri:
+            decoder = self.decoder
+            simple_lm_proj = self.simple_lm_proj
+            simple_am_proj = self.simple_am_proj
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_giga
+            simple_lm_proj = self.simple_lm_proj_giga
+            simple_am_proj = self.simple_am_proj_giga
+            joiner = self.joiner_giga
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = simple_lm_proj(decoder_out)
+        am = simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                delay_penalty=delay_penalty,
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=joiner.encoder_proj(encoder_out),
+            lm=joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                delay_penalty=delay_penalty,
+                reduction=reduction,
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/ncnn-decode.py b/egs/librispeech/ASR/lstm_transducer_stateless2/ncnn-decode.py
new file mode 100755
index 000000000..3eeaa5397
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/ncnn-decode.py
@@ -0,0 +1,301 @@
+#!/usr/bin/env python3
+# flake8: noqa
+#
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+  ./lstm_transducer_stateless2/ncnn-decode.py \
+   --tokens ./data/lang_bpe_500/tokens.txt \
+   --encoder-param-filename ./lstm_transducer_stateless2/exp/encoder_jit_trace-iter-468000-avg-16-pnnx.ncnn.param \
+   --encoder-bin-filename ./lstm_transducer_stateless2/exp/encoder_jit_trace-iter-468000-avg-16-pnnx.ncnn.bin \
+   --decoder-param-filename ./lstm_transducer_stateless2/exp/decoder_jit_trace-iter-468000-avg-16-pnnx.ncnn.param \
+   --decoder-bin-filename ./lstm_transducer_stateless2/exp/decoder_jit_trace-iter-468000-avg-16-pnnx.ncnn.bin \
+   --joiner-param-filename ./lstm_transducer_stateless2/exp/joiner_jit_trace-iter-468000-avg-16-pnnx.ncnn.param \
+   --joiner-bin-filename ./lstm_transducer_stateless2/exp/joiner_jit_trace-iter-468000-avg-16-pnnx.ncnn.bin \
+   ./test_wavs/1089-134686-0001.wav
+
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for details.
+"""
+
+import argparse
+import logging
+from typing import List
+
+import k2
+import kaldifeat
+import ncnn
+import torch
+import torchaudio
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument(
+        "--encoder-param-filename",
+        type=str,
+        help="Path to encoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--encoder-bin-filename",
+        type=str,
+        help="Path to encoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--decoder-param-filename",
+        type=str,
+        help="Path to decoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--decoder-bin-filename",
+        type=str,
+        help="Path to decoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--joiner-param-filename",
+        type=str,
+        help="Path to joiner.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--joiner-bin-filename",
+        type=str,
+        help="Path to joiner.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="Path to foo.wav",
+    )
+
+    return parser.parse_args()
+
+
+class Model:
+    def __init__(self, args):
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+
+    def init_encoder(self, args):
+        encoder_net = ncnn.Net()
+        encoder_net.opt.use_packing_layout = False
+        encoder_net.opt.use_fp16_storage = False
+        encoder_net.opt.num_threads = 4
+
+        encoder_param = args.encoder_param_filename
+        encoder_model = args.encoder_bin_filename
+
+        encoder_net.load_param(encoder_param)
+        encoder_net.load_model(encoder_model)
+
+        self.encoder_net = encoder_net
+
+    def init_decoder(self, args):
+        decoder_param = args.decoder_param_filename
+        decoder_model = args.decoder_bin_filename
+
+        decoder_net = ncnn.Net()
+        decoder_net.opt.use_packing_layout = False
+        decoder_net.opt.num_threads = 4
+
+        decoder_net.load_param(decoder_param)
+        decoder_net.load_model(decoder_model)
+
+        self.decoder_net = decoder_net
+
+    def init_joiner(self, args):
+        joiner_param = args.joiner_param_filename
+        joiner_model = args.joiner_bin_filename
+        joiner_net = ncnn.Net()
+        joiner_net.opt.use_packing_layout = False
+        joiner_net.opt.num_threads = 4
+
+        joiner_net.load_param(joiner_param)
+        joiner_net.load_model(joiner_model)
+
+        self.joiner_net = joiner_net
+
+    def run_encoder(self, x, states):
+        with self.encoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(x.numpy()).clone())
+            x_lens = torch.tensor([x.size(0)], dtype=torch.float32)
+            ex.input("in1", ncnn.Mat(x_lens.numpy()).clone())
+            ex.input("in2", ncnn.Mat(states[0].numpy()).clone())
+            ex.input("in3", ncnn.Mat(states[1].numpy()).clone())
+
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+
+            ret, ncnn_out1 = ex.extract("out1")
+            assert ret == 0, ret
+
+            ret, ncnn_out2 = ex.extract("out2")
+            assert ret == 0, ret
+
+            ret, ncnn_out3 = ex.extract("out3")
+            assert ret == 0, ret
+
+            encoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            encoder_out_lens = torch.from_numpy(ncnn_out1.numpy()).to(torch.int32)
+            hx = torch.from_numpy(ncnn_out2.numpy()).clone()
+            cx = torch.from_numpy(ncnn_out3.numpy()).clone()
+            return encoder_out, encoder_out_lens, hx, cx
+
+    def run_decoder(self, decoder_input):
+        assert decoder_input.dtype == torch.int32
+
+        with self.decoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(decoder_input.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            decoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return decoder_out
+
+    def run_joiner(self, encoder_out, decoder_out):
+        with self.joiner_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(encoder_out.numpy()).clone())
+            ex.input("in1", ncnn.Mat(decoder_out.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            joiner_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return joiner_out
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(model: Model, encoder_out: torch.Tensor):
+    assert encoder_out.ndim == 2
+    T = encoder_out.size(0)
+
+    context_size = 2
+    blank_id = 0  # hard-code to 0
+    hyp = [blank_id] * context_size
+
+    decoder_input = torch.tensor(hyp, dtype=torch.int32)  # (1, context_size)
+
+    decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    #  print(decoder_out.shape)  # (512,)
+
+    for t in range(T):
+        encoder_out_t = encoder_out[t]
+        joiner_out = model.run_joiner(encoder_out_t, decoder_out)
+        #  print(joiner_out.shape) # [500]
+        y = joiner_out.argmax(dim=0).tolist()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32)
+            decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    return hyp[context_size:]
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    logging.info("Decoding started")
+    features = fbank(wave_samples)
+
+    num_encoder_layers = 12
+    d_model = 512
+    rnn_hidden_size = 1024
+
+    states = (
+        torch.zeros(num_encoder_layers, d_model),
+        torch.zeros(
+            num_encoder_layers,
+            rnn_hidden_size,
+        ),
+    )
+
+    encoder_out, encoder_out_lens, hx, cx = model.run_encoder(features, states)
+    hyp = greedy_search(model, encoder_out)
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(sound_file)
+    logging.info(text)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py b/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py
new file mode 100755
index 000000000..c83f38b2a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py
@@ -0,0 +1,261 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script checks that exported ONNX models produce the same output
+with the given torchscript model for the same input.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+
+cd exp
+ln -s pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+2. Export the model via torch.jit.trace()
+
+./lstm_transducer_stateless2/export.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --jit-trace 1
+
+It will generate the following 3 files inside $repo/exp
+
+  - encoder_jit_trace.pt
+  - decoder_jit_trace.pt
+  - joiner_jit_trace.pt
+
+3. Export the model to ONNX
+
+./lstm_transducer_stateless2/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+4. Run this file
+
+./lstm_transducer_stateless2/onnx_check.py \
+  --jit-encoder-filename $repo/exp/encoder_jit_trace.pt \
+  --jit-decoder-filename $repo/exp/decoder_jit_trace.pt \
+  --jit-joiner-filename $repo/exp/joiner_jit_trace.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+"""
+
+import argparse
+import logging
+
+from onnx_pretrained import OnnxModel
+
+from icefall import is_module_available
+
+import torch
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript encoder model",
+    )
+
+    parser.add_argument(
+        "--jit-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript decoder model",
+    )
+
+    parser.add_argument(
+        "--jit-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript joiner model",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the ONNX encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the ONNX decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the ONNX joiner model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    torch_encoder_model: torch.jit.ScriptModule,
+    torch_encoder_proj_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    N = torch.randint(1, 100, size=(1,)).item()
+    T = onnx_model.segment
+    C = 80
+    x_lens = torch.tensor([T] * N)
+    torch_states = torch_encoder_model.get_init_states(N)
+
+    onnx_model.init_encoder_states(N)
+
+    for i in range(5):
+        logging.info(f"test_encoder: iter {i}")
+        x = torch.rand(N, T, C)
+        torch_encoder_out, _, torch_states = torch_encoder_model(
+            x, x_lens, torch_states
+        )
+        torch_encoder_out = torch_encoder_proj_model(torch_encoder_out)
+
+        onnx_encoder_out = onnx_model.run_encoder(x)
+
+        assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-4), (
+            (torch_encoder_out - onnx_encoder_out).abs().max()
+        )
+
+
+def test_decoder(
+    torch_decoder_model: torch.jit.ScriptModule,
+    torch_decoder_proj_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    context_size = onnx_model.context_size
+    vocab_size = onnx_model.vocab_size
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_decoder: iter {i}, N={N}")
+        x = torch.randint(
+            low=1,
+            high=vocab_size,
+            size=(N, context_size),
+            dtype=torch.int64,
+        )
+        torch_decoder_out = torch_decoder_model(x, need_pad=torch.tensor([False]))
+        torch_decoder_out = torch_decoder_proj_model(torch_decoder_out)
+        torch_decoder_out = torch_decoder_out.squeeze(1)
+
+        onnx_decoder_out = onnx_model.run_decoder(x)
+        assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
+            (torch_decoder_out - onnx_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    torch_joiner_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    encoder_dim = torch_joiner_model.encoder_proj.weight.shape[1]
+    decoder_dim = torch_joiner_model.decoder_proj.weight.shape[1]
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_joiner: iter {i}, N={N}")
+        encoder_out = torch.rand(N, encoder_dim)
+        decoder_out = torch.rand(N, decoder_dim)
+
+        projected_encoder_out = torch_joiner_model.encoder_proj(encoder_out)
+        projected_decoder_out = torch_joiner_model.decoder_proj(decoder_out)
+
+        torch_joiner_out = torch_joiner_model(encoder_out, decoder_out)
+        onnx_joiner_out = onnx_model.run_joiner(
+            projected_encoder_out, projected_decoder_out
+        )
+
+        assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
+            (torch_joiner_out - onnx_joiner_out).abs().max()
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_encoder_model = torch.jit.load(args.jit_encoder_filename)
+    torch_decoder_model = torch.jit.load(args.jit_decoder_filename)
+    torch_joiner_model = torch.jit.load(args.jit_joiner_filename)
+
+    onnx_model = OnnxModel(
+        encoder_model_filename=args.onnx_encoder_filename,
+        decoder_model_filename=args.onnx_decoder_filename,
+        joiner_model_filename=args.onnx_joiner_filename,
+    )
+
+    logging.info("Test encoder")
+    # When exporting the model to onnx, we have already put the encoder_proj
+    # inside the encoder.
+    test_encoder(torch_encoder_model, torch_joiner_model.encoder_proj, onnx_model)
+
+    logging.info("Test decoder")
+    # When exporting the model to onnx, we have already put the decoder_proj
+    # inside the decoder.
+    test_decoder(torch_decoder_model, torch_joiner_model.decoder_proj, onnx_model)
+
+    logging.info("Test joiner")
+    test_joiner(torch_joiner_model, onnx_model)
+
+    logging.info("Finished checking ONNX models")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# See https://github.com/pytorch/pytorch/issues/38342
+# and https://github.com/pytorch/pytorch/issues/33354
+#
+# If we don't do this, the delay increases whenever there is
+# a new request that changes the actual batch size.
+# If you use `py-spy dump --pid <server-pid> --native`, you will
+# see a lot of time is spent in re-compiling the torch script model.
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    torch.manual_seed(20230207)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_pretrained.py
new file mode 100755
index 000000000..06159e56a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_pretrained.py
@@ -0,0 +1,431 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This script loads ONNX models exported by ./export-onnx.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-468000-avg-16.pt"
+cd exp
+ln -s exp/pretrained-iter-468000-avg-16.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./lstm_transducer_stateless2/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file with the exported ONNX models
+
+./lstm_transducer_stateless2/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1221-135766-0001.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_encoder_states()
+
+    def init_encoder_states(self, batch_size: int = 1):
+        encoder_meta = self.encoder.get_modelmeta().custom_metadata_map
+
+        model_type = encoder_meta["model_type"]
+        assert model_type == "lstm", model_type
+
+        decode_chunk_len = int(encoder_meta["decode_chunk_len"])
+        T = int(encoder_meta["T"])
+
+        num_encoder_layers = int(encoder_meta["num_encoder_layers"])
+        d_model = int(encoder_meta["d_model"])
+        rnn_hidden_size = int(encoder_meta["rnn_hidden_size"])
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+        logging.info(f"d_model: {d_model}")
+        logging.info(f"rnn_hidden_size: {rnn_hidden_size}")
+
+        N = batch_size
+
+        s0 = torch.zeros(num_encoder_layers, N, d_model)
+        s1 = torch.zeros(num_encoder_layers, N, rnn_hidden_size)
+        states = [s0.numpy(), s1.numpy()]
+
+        self.states = states
+
+        self.segment = T
+        self.offset = decode_chunk_len
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def _build_encoder_input_output(
+        self,
+        x: torch.Tensor,
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        encoder_input = {
+            "x": x.numpy(),
+            "state0": self.states[0],
+            "state1": self.states[1],
+        }
+        encoder_output = ["encoder_out", "new_state0", "new_state1"]
+
+        return encoder_input, encoder_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        self.states = states
+
+    def run_encoder(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+        Returns:
+          Return a 3-D tensor of shape (N, T', joiner_dim) where
+          T' is usually equal to ((T-3)//2-1)//2
+        """
+        encoder_input, encoder_output_names = self._build_encoder_input_output(x)
+        out = self.encoder.run(encoder_output_names, encoder_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    context_size: int,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+) -> List[int]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (1, T, joiner_dim)
+      context_size:
+        The context size of the decoder model.
+      decoder_out:
+        Optional. Decoder output of the previous chunk.
+      hyp:
+        Decoding results for previous chunks.
+    Returns:
+      Return the decoded results so far.
+    """
+
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert hyp is not None, hyp
+
+    encoder_out = encoder_out.squeeze(0)
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t : t + 1]
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+            decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    context_size = model.context_size
+    hyp = None
+    decoder_out = None
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            encoder_out = model.run_encoder(frames)
+            hyp, decoder_out = greedy_search(
+                model,
+                encoder_out,
+                context_size,
+                decoder_out,
+                hyp,
+            )
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/optim.py b/egs/librispeech/ASR/lstm_transducer_stateless2/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless2/pretrained.py
new file mode 100755
index 000000000..5d6d97320
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/pretrained.py
@@ -0,0 +1,354 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./lstm_transducer_stateless2/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./lstm_transducer_stateless2/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./lstm_transducer_stateless2/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./lstm_transducer_stateless2/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./lstm_transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./lstm_transducer_stateless2/exp/pretrained.pt is generated by
+./lstm_transducer_stateless2/export.py
+
+You can find pretrained models by visiting
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/tree/main/exp
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/scaling.py b/egs/librispeech/ASR/lstm_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless2/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/streaming-ncnn-decode.py b/egs/librispeech/ASR/lstm_transducer_stateless2/streaming-ncnn-decode.py
new file mode 100755
index 000000000..cbbc77928
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/streaming-ncnn-decode.py
@@ -0,0 +1,351 @@
+#!/usr/bin/env python3
+# flake8: noqa
+#
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for usage
+"""
+
+import argparse
+import logging
+from typing import List, Optional
+
+import k2
+import ncnn
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument(
+        "--encoder-param-filename",
+        type=str,
+        help="Path to encoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--encoder-bin-filename",
+        type=str,
+        help="Path to encoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--decoder-param-filename",
+        type=str,
+        help="Path to decoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--decoder-bin-filename",
+        type=str,
+        help="Path to decoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--joiner-param-filename",
+        type=str,
+        help="Path to joiner.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--joiner-bin-filename",
+        type=str,
+        help="Path to joiner.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="Path to foo.wav",
+    )
+
+    return parser.parse_args()
+
+
+class Model:
+    def __init__(self, args):
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+
+    def init_encoder(self, args):
+        encoder_net = ncnn.Net()
+        encoder_net.opt.use_packing_layout = False
+        encoder_net.opt.use_fp16_storage = False
+        encoder_net.opt.num_threads = 4
+
+        encoder_param = args.encoder_param_filename
+        encoder_model = args.encoder_bin_filename
+
+        encoder_net.load_param(encoder_param)
+        encoder_net.load_model(encoder_model)
+
+        self.encoder_net = encoder_net
+
+    def init_decoder(self, args):
+        decoder_param = args.decoder_param_filename
+        decoder_model = args.decoder_bin_filename
+
+        decoder_net = ncnn.Net()
+        decoder_net.opt.use_packing_layout = False
+        decoder_net.opt.num_threads = 4
+
+        decoder_net.load_param(decoder_param)
+        decoder_net.load_model(decoder_model)
+
+        self.decoder_net = decoder_net
+
+    def init_joiner(self, args):
+        joiner_param = args.joiner_param_filename
+        joiner_model = args.joiner_bin_filename
+        joiner_net = ncnn.Net()
+        joiner_net.opt.use_packing_layout = False
+        joiner_net.opt.num_threads = 4
+
+        joiner_net.load_param(joiner_param)
+        joiner_net.load_model(joiner_model)
+
+        self.joiner_net = joiner_net
+
+    def run_encoder(self, x, states):
+        with self.encoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(x.numpy()).clone())
+            x_lens = torch.tensor([x.size(0)], dtype=torch.float32)
+            ex.input("in1", ncnn.Mat(x_lens.numpy()).clone())
+            ex.input("in2", ncnn.Mat(states[0].numpy()).clone())
+            ex.input("in3", ncnn.Mat(states[1].numpy()).clone())
+
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+
+            ret, ncnn_out1 = ex.extract("out1")
+            assert ret == 0, ret
+
+            ret, ncnn_out2 = ex.extract("out2")
+            assert ret == 0, ret
+
+            ret, ncnn_out3 = ex.extract("out3")
+            assert ret == 0, ret
+
+            encoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            encoder_out_lens = torch.from_numpy(ncnn_out1.numpy()).to(torch.int32)
+            hx = torch.from_numpy(ncnn_out2.numpy()).clone()
+            cx = torch.from_numpy(ncnn_out3.numpy()).clone()
+            return encoder_out, encoder_out_lens, hx, cx
+
+    def run_decoder(self, decoder_input):
+        assert decoder_input.dtype == torch.int32
+
+        with self.decoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(decoder_input.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            decoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return decoder_out
+
+    def run_joiner(self, encoder_out, decoder_out):
+        with self.joiner_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(encoder_out.numpy()).clone())
+            ex.input("in1", ncnn.Mat(decoder_out.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            joiner_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return joiner_out
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: Model,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    assert encoder_out.ndim == 1
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32)  # (1, context_size)
+        decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    else:
+        assert decoder_out.ndim == 1
+        assert hyp is not None, hyp
+
+    joiner_out = model.run_joiner(encoder_out, decoder_out)
+    y = joiner_out.argmax(dim=0).item()
+    if y != blank_id:
+        hyp.append(y)
+        decoder_input = hyp[-context_size:]
+        decoder_input = torch.tensor(decoder_input, dtype=torch.int32)
+        decoder_out = model.run_decoder(decoder_input).squeeze(0)
+
+    return hyp, decoder_out
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    num_encoder_layers = 12
+    batch_size = 1
+    d_model = 512
+    rnn_hidden_size = 1024
+
+    states = (
+        torch.zeros(num_encoder_layers, batch_size, d_model),
+        torch.zeros(
+            num_encoder_layers,
+            batch_size,
+            rnn_hidden_size,
+        ),
+    )
+
+    hyp = None
+    decoder_out = None
+
+    num_processed_frames = 0
+    segment = 9
+    offset = 4
+
+    chunk = 3200  # 0.2 second
+
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            encoder_out, encoder_out_lens, hx, cx = model.run_encoder(frames, states)
+            states = (hx, cx)
+            hyp, decoder_out = greedy_search(
+                model, encoder_out.squeeze(0), decoder_out, hyp
+            )
+    online_fbank.accept_waveform(
+        sampling_rate=sample_rate, waveform=torch.zeros(8000, dtype=torch.int32)
+    )
+
+    online_fbank.input_finished()
+    while online_fbank.num_frames_ready - num_processed_frames >= segment:
+        frames = []
+        for i in range(segment):
+            frames.append(online_fbank.get_frame(num_processed_frames + i))
+        num_processed_frames += offset
+        frames = torch.cat(frames, dim=0)
+        encoder_out, encoder_out_lens, hx, cx = model.run_encoder(frames, states)
+        states = (hx, cx)
+        hyp, decoder_out = greedy_search(
+            model, encoder_out.squeeze(0), decoder_out, hyp
+        )
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    context_size = 2
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(sound_file)
+    logging.info(text)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/streaming-onnx-decode.py b/egs/librispeech/ASR/lstm_transducer_stateless2/streaming-onnx-decode.py
new file mode 100755
index 000000000..487fc2114
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/streaming-onnx-decode.py
@@ -0,0 +1,471 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir ./lstm_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --onnx 1
+
+Usage of this script:
+
+./lstm_transducer_stateless2/onnx-streaming-decode.py \
+  --encoder-model-filename ./lstm_transducer_stateless2/exp/encoder.onnx \
+  --decoder-model-filename ./lstm_transducer_stateless2/exp/decoder.onnx \
+  --joiner-model-filename ./lstm_transducer_stateless2/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename ./lstm_transducer_stateless2/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename ./lstm_transducer_stateless2/exp/joiner_decoder_proj.onnx \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+from typing import List, Optional, Tuple
+
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
+import onnxruntime as ort
+import sentencepiece as spm
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model-filename",
+        type=str,
+        help="Path to bpe.model",
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-encoder-proj-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner encoder_proj onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-decoder-proj-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner decoder_proj onnx model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser.parse_args()
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+class Model:
+    def __init__(self, args):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 5
+        session_opts.intra_op_num_threads = 5
+        self.session_opts = session_opts
+
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+        self.init_joiner_encoder_proj(args)
+        self.init_joiner_decoder_proj(args)
+
+    def init_encoder(self, args):
+        self.encoder = ort.InferenceSession(
+            args.encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_decoder(self, args):
+        self.decoder = ort.InferenceSession(
+            args.decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_joiner(self, args):
+        self.joiner = ort.InferenceSession(
+            args.joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_joiner_encoder_proj(self, args):
+        self.joiner_encoder_proj = ort.InferenceSession(
+            args.joiner_encoder_proj_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_joiner_decoder_proj(self, args):
+        self.joiner_decoder_proj = ort.InferenceSession(
+            args.joiner_decoder_proj_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def run_encoder(self, x, h0, c0) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (N, T, C)
+          h0:
+            A tensor of shape (num_layers, N, proj_size)
+          c0:
+            A tensor of shape (num_layers, N, hidden_size)
+        Returns:
+          Return a tuple containing:
+            - encoder_out: A tensor of shape (N, T', C')
+            - next_h0: A tensor of shape (num_layers, N, proj_size)
+            - next_c0: A tensor of shape (num_layers, N, hidden_size)
+        """
+        encoder_input_nodes = self.encoder.get_inputs()
+        encoder_out_nodes = self.encoder.get_outputs()
+        x_lens = torch.tensor([x.size(1)], dtype=torch.int64)
+
+        encoder_out, encoder_out_lens, next_h0, next_c0 = self.encoder.run(
+            [
+                encoder_out_nodes[0].name,
+                encoder_out_nodes[1].name,
+                encoder_out_nodes[2].name,
+                encoder_out_nodes[3].name,
+            ],
+            {
+                encoder_input_nodes[0].name: x.numpy(),
+                encoder_input_nodes[1].name: x_lens.numpy(),
+                encoder_input_nodes[2].name: h0.numpy(),
+                encoder_input_nodes[3].name: c0.numpy(),
+            },
+        )
+        return (
+            torch.from_numpy(encoder_out),
+            torch.from_numpy(next_h0),
+            torch.from_numpy(next_c0),
+        )
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A tensor of shape (N, context_size). Its dtype is torch.int64.
+        Returns:
+          Return a tensor of shape (N, 1, decoder_out_dim).
+        """
+        decoder_input_nodes = self.decoder.get_inputs()
+        decoder_output_nodes = self.decoder.get_outputs()
+
+        decoder_out = self.decoder.run(
+            [decoder_output_nodes[0].name],
+            {
+                decoder_input_nodes[0].name: decoder_input.numpy(),
+            },
+        )[0]
+
+        return self.run_joiner_decoder_proj(torch.from_numpy(decoder_out).squeeze(1))
+
+    def run_joiner(
+        self,
+        projected_encoder_out: torch.Tensor,
+        projected_decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          projected_encoder_out:
+            A tensor of shape (N, joiner_dim)
+          projected_decoder_out:
+            A tensor of shape (N, joiner_dim)
+        Returns:
+          Return a tensor of shape (N, vocab_size)
+        """
+        joiner_input_nodes = self.joiner.get_inputs()
+        joiner_output_nodes = self.joiner.get_outputs()
+
+        logits = self.joiner.run(
+            [joiner_output_nodes[0].name],
+            {
+                joiner_input_nodes[0].name: projected_encoder_out.numpy(),
+                joiner_input_nodes[1].name: projected_decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(logits)
+
+    def run_joiner_encoder_proj(
+        self,
+        encoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A tensor of shape (N, encoder_out_dim)
+        Returns:
+            A tensor of shape (N, joiner_dim)
+        """
+
+        projected_encoder_out = self.joiner_encoder_proj.run(
+            [self.joiner_encoder_proj.get_outputs()[0].name],
+            {self.joiner_encoder_proj.get_inputs()[0].name: encoder_out.numpy()},
+        )[0]
+
+        return torch.from_numpy(projected_encoder_out)
+
+    def run_joiner_decoder_proj(
+        self,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          decoder_out:
+            A tensor of shape (N, decoder_out_dim)
+        Returns:
+            A tensor of shape (N, joiner_dim)
+        """
+
+        projected_decoder_out = self.joiner_decoder_proj.run(
+            [self.joiner_decoder_proj.get_outputs()[0].name],
+            {self.joiner_decoder_proj.get_inputs()[0].name: decoder_out.numpy()},
+        )[0]
+
+        return torch.from_numpy(projected_decoder_out)
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: Model,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    assert encoder_out.ndim == 2
+    assert encoder_out.shape[0] == 1, "TODO: support batch_size > 1"
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)  # (1, context_size)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert decoder_out.shape[0] == 1
+        assert hyp is not None, hyp
+
+    projected_encoder_out = model.run_joiner_encoder_proj(encoder_out)
+
+    joiner_out = model.run_joiner(projected_encoder_out, decoder_out)
+    y = joiner_out.squeeze(0).argmax(dim=0).item()
+
+    if y != blank_id:
+        hyp.append(y)
+        decoder_input = hyp[-context_size:]
+        decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+    sample_rate = 16000
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model_filename)
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    num_encoder_layers = 12
+    batch_size = 1
+    d_model = 512
+    rnn_hidden_size = 1024
+
+    h0 = torch.zeros(num_encoder_layers, batch_size, d_model)
+    c0 = torch.zeros(num_encoder_layers, batch_size, rnn_hidden_size)
+
+    hyp = None
+    decoder_out = None
+
+    num_processed_frames = 0
+    segment = 9
+    offset = 4
+
+    chunk = 3200  # 0.2 second
+
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0).unsqueeze(0)
+            encoder_out, h0, c0 = model.run_encoder(frames, h0, c0)
+            hyp, decoder_out = greedy_search(
+                model, encoder_out.squeeze(0), decoder_out, hyp
+            )
+    online_fbank.accept_waveform(
+        sampling_rate=sample_rate, waveform=torch.zeros(5000, dtype=torch.float)
+    )
+
+    online_fbank.input_finished()
+    while online_fbank.num_frames_ready - num_processed_frames >= segment:
+        frames = []
+        for i in range(segment):
+            frames.append(online_fbank.get_frame(num_processed_frames + i))
+        num_processed_frames += offset
+        frames = torch.cat(frames, dim=0).unsqueeze(0)
+        encoder_out, h0, c0 = model.run_encoder(frames, h0, c0)
+        hyp, decoder_out = greedy_search(
+            model, encoder_out.squeeze(0), decoder_out, hyp
+        )
+
+    context_size = 2
+
+    logging.info(sound_file)
+    logging.info(sp.decode(hyp[context_size:]))
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/test_lstmp.py b/egs/librispeech/ASR/lstm_transducer_stateless2/test_lstmp.py
new file mode 100755
index 000000000..00ba224cd
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/test_lstmp.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+
+import torch
+import torch.nn as nn
+from lstmp import LSTMP
+
+
+def test():
+    input_size = torch.randint(low=10, high=1024, size=(1,)).item()
+    hidden_size = torch.randint(low=10, high=1024, size=(1,)).item()
+    proj_size = hidden_size - 1
+    lstm = nn.LSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=1,
+        bias=True,
+        proj_size=proj_size,
+    )
+    lstmp = LSTMP(lstm)
+
+    N = torch.randint(low=1, high=10, size=(1,)).item()
+    T = torch.randint(low=1, high=20, size=(1,)).item()
+    x = torch.rand(T, N, input_size)
+    h0 = torch.rand(1, N, proj_size)
+    c0 = torch.rand(1, N, hidden_size)
+
+    y1, (h1, c1) = lstm(x, (h0, c0))
+    y2, (h2, c2) = lstmp(x, (h0, c0))
+
+    assert torch.allclose(y1, y2, atol=1e-5), (y1 - y2).abs().max()
+    assert torch.allclose(h1, h2, atol=1e-5), (h1 - h2).abs().max()
+    assert torch.allclose(c1, c2, atol=1e-5), (c1 - c2).abs().max()
+
+    #  lstm_script = torch.jit.script(lstm) # pytorch does not support it
+    lstm_script = lstm
+    lstmp_script = torch.jit.script(lstmp)
+
+    y3, (h3, c3) = lstm_script(x, (h0, c0))
+    y4, (h4, c4) = lstmp_script(x, (h0, c0))
+
+    assert torch.allclose(y3, y4, atol=1e-5), (y3 - y4).abs().max()
+    assert torch.allclose(h3, h4, atol=1e-5), (h3 - h4).abs().max()
+    assert torch.allclose(c3, c4, atol=1e-5), (c3 - c4).abs().max()
+
+    assert torch.allclose(y3, y1, atol=1e-5), (y3 - y1).abs().max()
+    assert torch.allclose(h3, h1, atol=1e-5), (h3 - h1).abs().max()
+    assert torch.allclose(c3, c1, atol=1e-5), (c3 - c1).abs().max()
+
+    lstm_trace = torch.jit.trace(lstm, (x, (h0, c0)))
+    lstmp_trace = torch.jit.trace(lstmp, (x, (h0, c0)))
+
+    y5, (h5, c5) = lstm_trace(x, (h0, c0))
+    y6, (h6, c6) = lstmp_trace(x, (h0, c0))
+
+    assert torch.allclose(y5, y6, atol=1e-5), (y5 - y6).abs().max()
+    assert torch.allclose(h5, h6, atol=1e-5), (h5 - h6).abs().max()
+    assert torch.allclose(c5, c6, atol=1e-5), (c5 - c6).abs().max()
+
+    assert torch.allclose(y5, y1, atol=1e-5), (y5 - y1).abs().max()
+    assert torch.allclose(h5, h1, atol=1e-5), (h5 - h1).abs().max()
+    assert torch.allclose(c5, c1, atol=1e-5), (c5 - c1).abs().max()
+
+
+@torch.no_grad()
+def main():
+    test()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/train.py b/egs/librispeech/ASR/lstm_transducer_stateless2/train.py
new file mode 100755
index 000000000..4fc4fa7f8
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless2/train.py
@@ -0,0 +1,1294 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+cd egs/librispeech/ASR/
+./prepare.sh
+./prepare_giga_speech.sh
+
+./lstm_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir lstm_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./lstm_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir lstm_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decoder import Decoder
+from gigaspeech import GigaSpeech
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from librispeech import LibriSpeech
+from lstm import RNN
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of RNN encoder layers..",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Encoder output dimesion.",
+    )
+
+    parser.add_argument(
+        "--rnn-hidden-size",
+        type=int,
+        default=1024,
+        help="Hidden dim for LSTM layers.",
+    )
+
+    parser.add_argument(
+        "--aux-layer-period",
+        type=int,
+        default=0,
+        help="""Peroid of auxiliary layers used for randomly combined during training.
+        If set to 0, will not use the random combiner (Default).
+        You can set a positive integer to use the random combiner, e.g., 3.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--full-libri",
+        type=str2bool,
+        default=True,
+        help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=35,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=10,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--giga-prob",
+        type=float,
+        default=0.5,
+        help="The probability to select a batch from the GigaSpeech dataset",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "dim_feedforward": 2048,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # True to generate a model that can be exported via PNNX
+            "is_pnnx": False,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = RNN(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        rnn_hidden_size=params.rnn_hidden_size,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        aux_layer_period=params.aux_layer_period,
+        is_pnnx=params.is_pnnx,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(
+    params: AttributeDict,
+    enable_giga: bool = True,
+) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    if enable_giga:
+        logging.info("Use giga")
+        decoder_giga = get_decoder_model(params)
+        joiner_giga = get_joiner_model(params)
+    else:
+        logging.info("Disable giga")
+        decoder_giga = None
+        joiner_giga = None
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        decoder_giga=decoder_giga,
+        joiner_giga=joiner_giga,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_libri(c: Cut) -> bool:
+    """Return True if this cut is from the LibriSpeech dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of GigaSpeech to dict(origin='giga')
+      See ../local/preprocess_gigaspeech.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    libri = is_libri(supervisions["cut"][0])
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            libri=libri,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    giga_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      giga_train_dl:
+        Dataloader for the GigaSpeech training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      rng:
+        For selecting which dataset to use.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    libri_tot_loss = MetricsTracker()
+    giga_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.giga_prob, params.giga_prob]
+
+    iter_libri = iter(train_dl)
+    iter_giga = iter(giga_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_libri if idx == 0 else iter_giga
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            name = "libri" if idx == 0 else "giga"
+            logging.info(f"{name} reaches end of dataloader")
+            break
+
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        libri = is_libri(batch["supervisions"]["cut"][0])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            if libri:
+                libri_tot_loss = (
+                    libri_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "libri"  # for logging only
+            else:
+                giga_tot_loss = (
+                    giga_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "giga"
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], "
+                f"libri_tot_loss[{libri_tot_loss}], "
+                f"giga_tot_loss[{giga_tot_loss}], "
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                libri_tot_loss.write_summary(
+                    tb_writer, "train/libri_tot_", params.batch_idx_train
+                )
+                giga_tot_loss.write_summary(
+                    tb_writer, "train/giga_tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(
+    cuts: CutSet,
+    sp: spm.SentencePieceProcessor,
+) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./lstm.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 3) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    cuts = cuts.filter(remove_short_and_long_utt)
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    # # overwrite it
+    # scheduler.base_lrs = [params.initial_lr for _ in scheduler.base_lrs]
+    # print(scheduler.base_lrs)
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    train_cuts = filter_short_and_long_utterances(train_cuts, sp)
+
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+    # XL 10k hours
+    # L  2.5k hours
+    # M  1k hours
+    # S  250 hours
+    # XS 10 hours
+    # DEV 12 hours
+    # Test 40 hours
+    if params.full_libri:
+        logging.info("Using the XL subset of GigaSpeech (10k hours)")
+        train_giga_cuts = gigaspeech.train_XL_cuts()
+    else:
+        logging.info("Using the S subset of GigaSpeech (250 hours)")
+        train_giga_cuts = gigaspeech.train_S_cuts()
+
+    train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts, sp)
+    train_giga_cuts = train_giga_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    giga_train_dl = asr_datamodule.train_dataloaders(
+        train_giga_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    # It's time consuming to include `giga_train_dl` here
+    #  for dl in [train_dl, giga_train_dl]:
+    for dl in [train_dl]:
+        if (
+            params.start_batch <= 0
+            and params.start_epoch == 1
+            and not params.print_diagnostics
+            and False
+        ):
+            scan_pessimistic_batches_for_oom(
+                model=model,
+                train_dl=dl,
+                optimizer=optimizer,
+                sp=sp,
+                params=params,
+                warmup=0.0 if params.start_epoch == 0 else 1.0,
+            )
+        else:
+            logging.info("Skip scan_pessimistic_batches_for_oom")
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            giga_train_dl=giga_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert 0 <= args.giga_prob < 1, args.giga_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py b/egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py
new file mode 120000
index 000000000..b24e5e357
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py b/egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/decode.py b/egs/librispeech/ASR/lstm_transducer_stateless3/decode.py
new file mode 100755
index 000000000..a2b4f9e1a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/decode.py
@@ -0,0 +1,864 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./lstm_transducer_stateless2/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+To evaluate symbol delay, you should:
+(1) Generate cuts with word-time alignments:
+./add_alignments.sh
+(2) Set the argument "--manifest-dir data/fbank_ali" while decoding.
+For example:
+./lstm_transducer_stateless3/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search \
+    --manifest-dir data/fbank_ali
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    DecodingResults,
+    parse_hyp_and_timestamp,
+    setup_logger,
+    store_transcripts_and_timestamps,
+    str2bool,
+    write_error_stats_with_timestamps,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, Tuple[List[List[str]], List[List[float]]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It is a tuple. `len(value[0])` and `len(value[1])` are both
+                 equal to the batch size. `value[0][i]` and `value[1][i]`
+                 are the decoding result and timestamps for the i-th utterance
+                 in the given batch respectively.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result and timestamps. See above description for the
+      format of the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    # tail padding here to alleviate the tail deletion problem
+    num_tail_padded_frames = 35
+    feature = torch.nn.functional.pad(
+        feature,
+        (0, 0, 0, num_tail_padded_frames),
+        mode="constant",
+        value=LOG_EPS,
+    )
+    feature_lens += num_tail_padded_frames
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        res = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        res = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "fast_beam_search_nbest":
+        res = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        res = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        res = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        res = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            return_timestamps=True,
+        )
+    else:
+        batch_size = encoder_out.size(0)
+        tokens = []
+        timestamps = []
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                res = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                    return_timestamps=True,
+                )
+            elif params.decoding_method == "beam_search":
+                res = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                    return_timestamps=True,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            tokens.extend(res.tokens)
+            timestamps.extend(res.timestamps)
+        res = DecodingResults(hyps=tokens, timestamps=timestamps)
+
+    hyps, timestamps = parse_hyp_and_timestamp(
+        res=res,
+        sp=sp,
+        subsampling_factor=params.subsampling_factor,
+        frame_shift_ms=params.frame_shift_ms,
+        word_table=word_table,
+    )
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": (hyps, timestamps)}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: (hyps, timestamps)}
+    else:
+        return {f"beam_size_{params.beam_size}": (hyps, timestamps)}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str], List[float], List[float]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains five elements:
+      - cut_id
+      - reference transcript
+      - predicted result
+      - timestamp of reference transcript
+      - timestamp of predicted result
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        timestamps_ref = []
+        for cut in batch["supervisions"]["cut"]:
+            for s in cut.supervisions:
+                time = []
+                if s.alignment is not None and "word" in s.alignment:
+                    time = [
+                        aliword.start
+                        for aliword in s.alignment["word"]
+                        if aliword.symbol != ""
+                    ]
+                timestamps_ref.append(time)
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, (hyps, timestamps_hyp) in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts) and len(timestamps_hyp) == len(
+                timestamps_ref
+            )
+            for cut_id, hyp_words, ref_text, time_hyp, time_ref in zip(
+                cut_ids, hyps, texts, timestamps_hyp, timestamps_ref
+            ):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words, time_ref, time_hyp))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[
+        str,
+        List[Tuple[List[str], List[str], List[str], List[float], List[float]]],
+    ],
+):
+    test_set_wers = dict()
+    test_set_delays = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts_and_timestamps(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer, mean_delay, var_delay = write_error_stats_with_timestamps(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+            test_set_delays[key] = (mean_delay, var_delay)
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    test_set_delays = sorted(test_set_delays.items(), key=lambda x: x[1][0])
+    delays_info = (
+        params.res_dir / f"symbol-delay-summary-{test_set_name}-{params.suffix}.txt"
+    )
+    with open(delays_info, "w") as f:
+        print("settings\tsymbol-delay", file=f)
+        for key, val in test_set_delays:
+            print(
+                "{}\tmean: {}s, variance: {}".format(key, val[0], val[1]),
+                file=f,
+            )
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    s = "\nFor {}, symbol-delay of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_delays:
+        s += "{}\tmean: {}s, variance: {}{}\n".format(key, val[0], val[1], note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py b/egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py b/egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/export-for-ncnn.py b/egs/librispeech/ASR/lstm_transducer_stateless3/export-for-ncnn.py
new file mode 120000
index 000000000..d56cff73f
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/export-for-ncnn.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/export-for-ncnn.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/export-onnx.py b/egs/librispeech/ASR/lstm_transducer_stateless3/export-onnx.py
new file mode 120000
index 000000000..9f5064deb
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/export-onnx.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/export-onnx.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/export.py b/egs/librispeech/ASR/lstm_transducer_stateless3/export.py
new file mode 100755
index 000000000..21eaa049b
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/export.py
@@ -0,0 +1,386 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.trace()
+
+./lstm_transducer_stateless3/export.py \
+  --exp-dir ./lstm_transducer_stateless3/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 40 \
+  --avg 20 \
+  --jit-trace 1
+
+It will generate 3 files: `encoder_jit_trace.pt`,
+`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`.
+
+(2) Export `model.state_dict()`
+
+./lstm_transducer_stateless3/export.py \
+  --exp-dir ./lstm_transducer_stateless3/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 40 \
+  --avg 20
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `lstm_transducer_stateless3/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./lstm_transducer_stateless3/decode.py \
+        --exp-dir ./lstm_transducer_stateless3/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18
+    # You will find the pre-trained model in icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace.
+        It will generate 3 files:
+         - encoder_jit_trace.pt
+         - decoder_jit_trace.pt
+         - joiner_jit_trace.pt
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+    states = encoder_model.get_init_states()
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size, <blk> is
+    # defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit_trace is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    else:
+        logging.info("Not using torchscript")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py
new file mode 100755
index 000000000..237591a36
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py
@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, either exported by `torch.jit.trace()`
+or by `torch.jit.script()`, and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./lstm_transducer_stateless3/export.py \
+  --exp-dir ./lstm_transducer_stateless3/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 40 \
+  --avg 15 \
+  --jit-trace 1
+
+Usage of this script:
+
+./lstm_transducer_stateless3/jit_pretrained.py \
+  --encoder-model-filename ./lstm_transducer_stateless3/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./lstm_transducer_stateless3/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./lstm_transducer_stateless3/exp/joiner_jit_trace.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_size: int,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      decoder:
+        The decoder model.
+      joiner:
+        The joiner model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+      context_size:
+        The context size of the decoder model.
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    states = encoder.get_init_states(batch_size=features.size(0), device=device)
+
+    encoder_out, encoder_out_lens, _ = encoder(
+        x=features,
+        x_lens=feature_lengths,
+        states=states,
+    )
+
+    hyps = greedy_search(
+        decoder=decoder,
+        joiner=joiner,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        context_size=args.context_size,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py b/egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py b/egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py
new file mode 100644
index 000000000..59a835d35
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py
@@ -0,0 +1,885 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import List, Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv2d,
+    ScaledLinear,
+    ScaledLSTM,
+)
+from torch import nn
+
+LOG_EPSILON = math.log(1e-10)
+
+
+def unstack_states(
+    states: Tuple[torch.Tensor, torch.Tensor]
+) -> List[Tuple[torch.Tensor, torch.Tensor]]:
+    """
+    Unstack the lstm states corresponding to a batch of utterances into a list
+    of states, where the i-th entry is the state from the i-th utterance.
+
+    Args:
+      states:
+        A tuple of 2 elements.
+        ``states[0]`` is the lstm hidden states, of a batch of utterance.
+        ``states[1]`` is the lstm cell states, of a batch of utterances.
+
+    Returns:
+      A list of states.
+        ``states[i]`` is a tuple of 2 elememts of i-th utterance.
+        ``states[i][0]`` is the lstm hidden states of i-th utterance.
+        ``states[i][1]`` is the lstm cell states of i-th utterance.
+    """
+    hidden_states, cell_states = states
+
+    list_hidden_states = hidden_states.unbind(dim=1)
+    list_cell_states = cell_states.unbind(dim=1)
+
+    ans = [
+        (h.unsqueeze(1), c.unsqueeze(1))
+        for (h, c) in zip(list_hidden_states, list_cell_states)
+    ]
+    return ans
+
+
+def stack_states(
+    states_list: List[Tuple[torch.Tensor, torch.Tensor]]
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Stack list of lstm states corresponding to separate utterances into a single
+    lstm state so that it can be used as an input for lstm when those utterances
+    are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponds to the lstm state for a single
+        utterance.
+        ``states[i]`` is a tuple of 2 elememts of i-th utterance.
+        ``states[i][0]`` is the lstm hidden states of i-th utterance.
+        ``states[i][1]`` is the lstm cell states of i-th utterance.
+
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      It is a tuple of 2 elements.
+        ``states[0]`` is the lstm hidden states, of a batch of utterance.
+        ``states[1]`` is the lstm cell states, of a batch of utterances.
+    """
+    hidden_states = torch.cat([s[0] for s in states_list], dim=1)
+    cell_states = torch.cat([s[1] for s in states_list], dim=1)
+    ans = (hidden_states, cell_states)
+    return ans
+
+
+class RNN(EncoderInterface):
+    """
+    Args:
+      num_features (int):
+        Number of input features.
+      subsampling_factor (int):
+        Subsampling factor of encoder (convolution layers before lstm layers) (default=4).  # noqa
+      d_model (int):
+        Output dimension (default=512).
+      dim_feedforward (int):
+        Feedforward dimension (default=2048).
+      rnn_hidden_size (int):
+        Hidden dimension for lstm layers (default=1024).
+      grad_norm_threshold:
+        For each sequence element in batch, its gradient will be
+        filtered out if the gradient norm is larger than
+        `grad_norm_threshold * median`, where `median` is the median
+        value of gradient norms of all elememts in batch.
+      num_encoder_layers (int):
+        Number of encoder layers (default=12).
+      dropout (float):
+        Dropout rate (default=0.1).
+      layer_dropout (float):
+        Dropout value for model-level warmup (default=0.075).
+      aux_layer_period (int):
+        Period of auxiliary layers used for random combiner during training.
+        If set to 0, will not use the random combiner (Default).
+        You can set a positive integer to use the random combiner, e.g., 3.
+      is_pnnx:
+        True to make this class exportable via PNNX.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 512,
+        dim_feedforward: int = 2048,
+        rnn_hidden_size: int = 1024,
+        grad_norm_threshold: float = 10.0,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        aux_layer_period: int = 0,
+        is_pnnx: bool = False,
+    ) -> None:
+        super(RNN, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(
+            num_features,
+            d_model,
+            is_pnnx=is_pnnx,
+        )
+
+        self.is_pnnx = is_pnnx
+
+        self.num_encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.rnn_hidden_size = rnn_hidden_size
+
+        encoder_layer = RNNEncoderLayer(
+            d_model=d_model,
+            dim_feedforward=dim_feedforward,
+            rnn_hidden_size=rnn_hidden_size,
+            grad_norm_threshold=grad_norm_threshold,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+        self.encoder = RNNEncoder(
+            encoder_layer,
+            num_encoder_layers,
+            aux_layers=list(
+                range(
+                    num_encoder_layers // 3,
+                    num_encoder_layers - 1,
+                    aux_layer_period,
+                )
+            )
+            if aux_layer_period > 0
+            else None,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C), where N is the batch size,
+            T is the sequence length, C is the feature dimension.
+          x_lens:
+            A tensor of shape (N,), containing the number of frames in `x`
+            before padding.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (num_layers, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (num_layers, N, rnn_hidden_size).
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          A tuple of 3 tensors:
+            - embeddings: its shape is (N, T', d_model), where T' is the output
+              sequence lengths.
+            - lengths: a tensor of shape (batch_size,) containing the number of
+              frames in `embeddings` before padding.
+            - updated states, whose shape is the same as the input states.
+        """
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # lengths = ((x_lens - 3) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        if not self.is_pnnx:
+            lengths = (((x_lens - 3) >> 1) - 1) >> 1
+        else:
+            lengths1 = torch.floor((x_lens - 3) / 2)
+            lengths = torch.floor((lengths1 - 1) / 2)
+            lengths = lengths.to(x_lens)
+
+        if not torch.jit.is_tracing():
+            assert x.size(0) == lengths.max().item()
+
+        if states is None:
+            x = self.encoder(x, warmup=warmup)[0]
+            # torch.jit.trace requires returned types to be the same as annotated  # noqa
+            new_states = (torch.empty(0), torch.empty(0))
+        else:
+            assert not self.training
+            assert len(states) == 2
+            if not torch.jit.is_tracing():
+                # for hidden state
+                assert states[0].shape == (
+                    self.num_encoder_layers,
+                    x.size(1),
+                    self.d_model,
+                )
+                # for cell state
+                assert states[1].shape == (
+                    self.num_encoder_layers,
+                    x.size(1),
+                    self.rnn_hidden_size,
+                )
+            x, new_states = self.encoder(x, states)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        return x, lengths, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self, batch_size: int = 1, device: torch.device = torch.device("cpu")
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Get model initial states."""
+        # for rnn hidden states
+        hidden_states = torch.zeros(
+            (self.num_encoder_layers, batch_size, self.d_model), device=device
+        )
+        cell_states = torch.zeros(
+            (self.num_encoder_layers, batch_size, self.rnn_hidden_size),
+            device=device,
+        )
+        return (hidden_states, cell_states)
+
+
+class RNNEncoderLayer(nn.Module):
+    """
+    RNNEncoderLayer is made up of lstm and feedforward networks.
+    For stable training, in each lstm module, gradient filter
+    is applied to filter out extremely large elements in batch gradients
+    and also the module parameters with soft masks.
+
+    Args:
+      d_model:
+        The number of expected features in the input (required).
+      dim_feedforward:
+        The dimension of feedforward network model (default=2048).
+      rnn_hidden_size:
+        The hidden dimension of rnn layer.
+      grad_norm_threshold:
+        For each sequence element in batch, its gradient will be
+        filtered out if the gradient norm is larger than
+        `grad_norm_threshold * median`, where `median` is the median
+        value of gradient norms of all elememts in batch.
+      dropout:
+        The dropout value (default=0.1).
+      layer_dropout:
+        The dropout value for model-level warmup (default=0.075).
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        dim_feedforward: int,
+        rnn_hidden_size: int,
+        grad_norm_threshold: float = 10.0,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        super(RNNEncoderLayer, self).__init__()
+        self.layer_dropout = layer_dropout
+        self.d_model = d_model
+        self.rnn_hidden_size = rnn_hidden_size
+
+        assert rnn_hidden_size >= d_model, (rnn_hidden_size, d_model)
+
+        self.lstm = ScaledLSTM(
+            input_size=d_model,
+            hidden_size=rnn_hidden_size,
+            proj_size=d_model if rnn_hidden_size > d_model else 0,
+            num_layers=1,
+            dropout=0.0,
+            grad_norm_threshold=grad_norm_threshold,
+        )
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).  # noqa
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+          src:
+            The sequence to the encoder layer (required).
+            Its shape is (S, N, E), where S is the sequence length,
+            N is the batch size, and E is the feature number.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (1, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (1, N, rnn_hidden_size).
+          warmup:
+            It controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # lstm module
+        if states is None:
+            src_lstm = self.lstm(src)[0]
+            # torch.jit.trace requires returned types be the same as annotated
+            new_states = (torch.empty(0), torch.empty(0))
+        else:
+            assert not self.training
+            assert len(states) == 2
+            if not torch.jit.is_tracing():
+                # for hidden state
+                assert states[0].shape == (1, src.size(1), self.d_model)
+                # for cell state
+                assert states[1].shape == (1, src.size(1), self.rnn_hidden_size)
+            src_lstm, new_states = self.lstm(src, states)
+        src = self.dropout(src_lstm) + src
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src, new_states
+
+
+class RNNEncoder(nn.Module):
+    """
+    RNNEncoder is a stack of N encoder layers.
+
+    Args:
+      encoder_layer:
+        An instance of the RNNEncoderLayer() class (required).
+      num_layers:
+        The number of sub-encoder-layers in the encoder (required).
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        aux_layers: Optional[List[int]] = None,
+    ) -> None:
+        super(RNNEncoder, self).__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+        self.d_model = encoder_layer.d_model
+        self.rnn_hidden_size = encoder_layer.rnn_hidden_size
+
+        self.aux_layers: List[int] = []
+        self.combiner: Optional[nn.Module] = None
+        if aux_layers is not None:
+            assert len(set(aux_layers)) == len(aux_layers)
+            assert num_layers - 1 not in aux_layers
+            self.aux_layers = aux_layers + [num_layers - 1]
+            self.combiner = RandomCombine(
+                num_inputs=len(self.aux_layers),
+                final_weight=0.5,
+                pure_prob=0.333,
+                stddev=2.0,
+            )
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Pass the input through the encoder layer in turn.
+
+        Args:
+          src:
+            The sequence to the encoder layer (required).
+            Its shape is (S, N, E), where S is the sequence length,
+            N is the batch size, and E is the feature number.
+          states:
+            A tuple of 2 tensors (optional). It is for streaming inference.
+            states[0] is the hidden states of all layers,
+              with shape of (num_layers, N, d_model);
+            states[1] is the cell states of all layers,
+              with shape of (num_layers, N, rnn_hidden_size).
+          warmup:
+            It controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+        """
+        if states is not None:
+            assert not self.training
+            assert len(states) == 2
+            if not torch.jit.is_tracing():
+                # for hidden state
+                assert states[0].shape == (
+                    self.num_layers,
+                    src.size(1),
+                    self.d_model,
+                )
+                # for cell state
+                assert states[1].shape == (
+                    self.num_layers,
+                    src.size(1),
+                    self.rnn_hidden_size,
+                )
+
+        output = src
+
+        outputs = []
+
+        new_hidden_states = []
+        new_cell_states = []
+
+        for i, mod in enumerate(self.layers):
+            if states is None:
+                output = mod(output, warmup=warmup)[0]
+            else:
+                layer_state = (
+                    states[0][i : i + 1, :, :],  # h: (1, N, d_model)
+                    states[1][i : i + 1, :, :],  # c: (1, N, rnn_hidden_size)
+                )
+                output, (h, c) = mod(output, layer_state)
+                new_hidden_states.append(h)
+                new_cell_states.append(c)
+
+            if self.combiner is not None and i in self.aux_layers:
+                outputs.append(output)
+
+        if self.combiner is not None:
+            output = self.combiner(outputs)
+
+        if states is None:
+            new_states = (torch.empty(0), torch.empty(0))
+        else:
+            new_states = (
+                torch.cat(new_hidden_states, dim=0),
+                torch.cat(new_cell_states, dim=0),
+            )
+
+        return output, new_states
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-3)//2-1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        is_pnnx: bool = False,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >= 9, in_channels >= 9.
+          out_channels
+            Output dim. The output shape is (N, ((T-3)//2-1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+          is_pnnx:
+            True if we are converting the model to PNNX format.
+            False otherwise.
+        """
+        assert in_channels >= 9
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=0,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 3) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+        # ncnn supports only batch size == 1
+        self.is_pnnx = is_pnnx
+        self.conv_out_dim = self.out.weight.shape[1]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-3)//2-1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+
+        if torch.jit.is_tracing() and self.is_pnnx:
+            x = x.permute(0, 2, 1, 3).reshape(1, -1, self.conv_out_dim)
+            x = self.out(x)
+        else:
+            # Now x is of shape (N, odim, ((T-3)//2-1)//2, ((idim-3)//2-1)//2)
+            b, c, t, f = x.size()
+            x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+
+        # Now x is of shape (N, ((T-3)//2-1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+class RandomCombine(nn.Module):
+    """
+    This module combines a list of Tensors, all with the same shape, to
+    produce a single output of that same shape which, in training time,
+    is a random combination of all the inputs; but which in test time
+    will be just the last input.
+
+    The idea is that the list of Tensors will be a list of outputs of multiple
+    conformer layers.  This has a similar effect as iterated loss. (See:
+    DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER
+    NETWORKS).
+    """
+
+    def __init__(
+        self,
+        num_inputs: int,
+        final_weight: float = 0.5,
+        pure_prob: float = 0.5,
+        stddev: float = 2.0,
+    ) -> None:
+        """
+        Args:
+          num_inputs:
+            The number of tensor inputs, which equals the number of layers'
+            outputs that are fed into this module.  E.g. in an 18-layer neural
+            net if we output layers 16, 12, 18, num_inputs would be 3.
+          final_weight:
+            The amount of weight or probability we assign to the
+            final layer when randomly choosing layers or when choosing
+            continuous layer weights.
+          pure_prob:
+            The probability, on each frame, with which we choose
+            only a single layer to output (rather than an interpolation)
+          stddev:
+            A standard deviation that we add to log-probs for computing
+            randomized weights.
+
+        The method of choosing which layers, or combinations of layers, to use,
+        is conceptually as follows::
+
+            With probability `pure_prob`::
+               With probability `final_weight`: choose final layer,
+               Else: choose random non-final layer.
+            Else::
+               Choose initial log-weights that correspond to assigning
+               weight `final_weight` to the final layer and equal
+               weights to other layers; then add Gaussian noise
+               with variance `stddev` to these log-weights, and normalize
+               to weights (note: the average weight assigned to the
+               final layer here will not be `final_weight` if stddev>0).
+        """
+        super().__init__()
+        assert 0 <= pure_prob <= 1, pure_prob
+        assert 0 < final_weight < 1, final_weight
+        assert num_inputs >= 1
+
+        self.num_inputs = num_inputs
+        self.final_weight = final_weight
+        self.pure_prob = pure_prob
+        self.stddev = stddev
+
+        self.final_log_weight = (
+            torch.tensor((final_weight / (1 - final_weight)) * (self.num_inputs - 1))
+            .log()
+            .item()
+        )
+
+    def forward(self, inputs: List[torch.Tensor]) -> torch.Tensor:
+        """Forward function.
+        Args:
+          inputs:
+            A list of Tensor, e.g. from various layers of a transformer.
+            All must be the same shape, of (*, num_channels)
+        Returns:
+          A Tensor of shape (*, num_channels).  In test mode
+          this is just the final input.
+        """
+        num_inputs = self.num_inputs
+        assert len(inputs) == num_inputs
+        if not self.training or torch.jit.is_scripting():
+            return inputs[-1]
+
+        # Shape of weights: (*, num_inputs)
+        num_channels = inputs[0].shape[-1]
+        num_frames = inputs[0].numel() // num_channels
+
+        ndim = inputs[0].ndim
+        # stacked_inputs: (num_frames, num_channels, num_inputs)
+        stacked_inputs = torch.stack(inputs, dim=ndim).reshape(
+            (num_frames, num_channels, num_inputs)
+        )
+
+        # weights: (num_frames, num_inputs)
+        weights = self._get_random_weights(
+            inputs[0].dtype, inputs[0].device, num_frames
+        )
+
+        weights = weights.reshape(num_frames, num_inputs, 1)
+        # ans: (num_frames, num_channels, 1)
+        ans = torch.matmul(stacked_inputs, weights)
+        # ans: (*, num_channels)
+
+        ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,))
+
+        # The following if causes errors for torch script in torch 1.6.0
+        #  if __name__ == "__main__":
+        #      # for testing only...
+        #      print("Weights = ", weights.reshape(num_frames, num_inputs))
+        return ans
+
+    def _get_random_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), such that
+          `ans.sum(dim=1)` is all ones.
+        """
+        pure_prob = self.pure_prob
+        if pure_prob == 0.0:
+            return self._get_random_mixed_weights(dtype, device, num_frames)
+        elif pure_prob == 1.0:
+            return self._get_random_pure_weights(dtype, device, num_frames)
+        else:
+            p = self._get_random_pure_weights(dtype, device, num_frames)
+            m = self._get_random_mixed_weights(dtype, device, num_frames)
+            return torch.where(
+                torch.rand(num_frames, 1, device=device) < self.pure_prob, p, m
+            )
+
+    def _get_random_pure_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ):
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A one-hot tensor of shape `(num_frames, self.num_inputs)`, with
+          exactly one weight equal to 1.0 on each frame.
+        """
+        final_prob = self.final_weight
+
+        # final contains self.num_inputs - 1 in all elements
+        final = torch.full((num_frames,), self.num_inputs - 1, device=device)
+        # nonfinal contains random integers in [0..num_inputs - 2], these are for non-final weights.  # noqa
+        nonfinal = torch.randint(self.num_inputs - 1, (num_frames,), device=device)
+
+        indexes = torch.where(
+            torch.rand(num_frames, device=device) < final_prob, final, nonfinal
+        )
+        ans = torch.nn.functional.one_hot(indexes, num_classes=self.num_inputs).to(
+            dtype=dtype
+        )
+        return ans
+
+    def _get_random_mixed_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ):
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), which elements
+          in [0..1] that sum to one over the second axis, i.e.
+          `ans.sum(dim=1)` is all ones.
+        """
+        logprobs = (
+            torch.randn(num_frames, self.num_inputs, dtype=dtype, device=device)
+            * self.stddev
+        )
+        logprobs[:, -1] += self.final_log_weight
+        return logprobs.softmax(dim=1)
+
+
+def _test_random_combine(final_weight: float, pure_prob: float, stddev: float):
+    print(
+        f"_test_random_combine: final_weight={final_weight}, pure_prob={pure_prob}, stddev={stddev}"  # noqa
+    )
+    num_inputs = 3
+    num_channels = 50
+    m = RandomCombine(
+        num_inputs=num_inputs,
+        final_weight=final_weight,
+        pure_prob=pure_prob,
+        stddev=stddev,
+    )
+
+    x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)]
+
+    y = m(x)
+    assert y.shape == x[0].shape
+    assert torch.allclose(y, x[0])  # .. since actually all ones.
+
+
+def _test_random_combine_main():
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.0)
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.3)
+    _test_random_combine(0.5, 1, 0.3)
+    _test_random_combine(0.5, 0.5, 0.3)
+
+    feature_dim = 50
+    c = RNN(num_features=feature_dim, d_model=128)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+if __name__ == "__main__":
+    feature_dim = 80
+    m = RNN(
+        num_features=feature_dim,
+        d_model=512,
+        rnn_hidden_size=1024,
+        dim_feedforward=2048,
+        num_encoder_layers=12,
+    )
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = m(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
+    num_param = sum([p.numel() for p in m.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    _test_random_combine_main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/lstmp.py b/egs/librispeech/ASR/lstm_transducer_stateless3/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/model.py b/egs/librispeech/ASR/lstm_transducer_stateless3/model.py
new file mode 120000
index 000000000..1bf04f3a4
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/model.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/onnx_check.py b/egs/librispeech/ASR/lstm_transducer_stateless3/onnx_check.py
new file mode 120000
index 000000000..0b1ea0326
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/onnx_check.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/onnx_check.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/onnx_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless3/onnx_pretrained.py
new file mode 120000
index 000000000..099c2882f
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/onnx_pretrained.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/optim.py b/egs/librispeech/ASR/lstm_transducer_stateless3/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py
new file mode 100755
index 000000000..29a0d4d1a
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py
@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./lstm_transducer_stateless3/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./lstm_transducer_stateless3/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./lstm_transducer_stateless3/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./lstm_transducer_stateless3/pretrained.py \
+    --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./lstm_transducer_stateless3/exp/epoch-xx.pt`.
+
+Note: ./lstm_transducer_stateless3/exp/pretrained.pt is generated by
+./lstm_transducer_stateless3/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/stream.py b/egs/librispeech/ASR/lstm_transducer_stateless3/stream.py
new file mode 120000
index 000000000..71ea6dff1
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/stream.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless/stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py b/egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py
new file mode 100755
index 000000000..c737e3611
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py
@@ -0,0 +1,949 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./lstm_transducer_stateless3/streaming_decode.py \
+      --epoch 40 \
+      --avg 20 \
+      --exp-dir lstm_transducer_stateless3/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --rnn-hidden-size 1024 \
+      --decoding-method greedy_search \
+      --use-averaged-model True
+
+(2) modified beam search
+./lstm_transducer_stateless3/streaming_decode.py \
+      --epoch 40 \
+      --avg 20 \
+      --exp-dir lstm_transducer_stateless3/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --rnn-hidden-size 1024 \
+      --decoding-method modified_beam_search \
+      --use-averaged-model True \
+      --beam-size 4
+
+(3) fast beam search
+./lstm_transducer_stateless3/streaming_decode.py \
+      --epoch 40 \
+      --avg 20 \
+      --exp-dir lstm_transducer_stateless3/exp \
+      --num-decode-streams 2000 \
+      --num-encoder-layers 12 \
+      --rnn-hidden-size 1024 \
+      --decoding-method fast_beam_search \
+      --use-averaged-model True \
+      --beam 4 \
+      --max-contexts 4 \
+      --max-states 8
+"""
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from lstm import LOG_EPSILON, stack_states, unstack_states
+from stream import Stream
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import one_best_decoding
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=40,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--sampling-rate",
+        type=float,
+        default=16000,
+        help="Sample rate of the audio",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded in parallel",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    T = encoder_out.size(1)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (batch_size, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[Stream],
+    beam: int = 4,
+):
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      beam:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    streams: List[Stream],
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      streams:
+        A list of stream objects.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyps[i]
+
+
+def decode_one_chunk(
+    model: nn.Module,
+    streams: List[Stream],
+    params: AttributeDict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> List[int]:
+    """
+    Args:
+      model:
+        The Transducer model.
+      streams:
+        A list of Stream objects.
+      params:
+        It is returned by :func:`get_params`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+       A list of indexes indicating the finished streams.
+    """
+    device = next(model.parameters()).device
+
+    feature_list = []
+    feature_len_list = []
+    state_list = []
+    num_processed_frames_list = []
+
+    for stream in streams:
+        # We should first get `stream.num_processed_frames`
+        # before calling `stream.get_feature_chunk()`
+        # since `stream.num_processed_frames` would be updated
+        num_processed_frames_list.append(stream.num_processed_frames)
+        feature = stream.get_feature_chunk()
+        feature_len = feature.size(0)
+        feature_list.append(feature)
+        feature_len_list.append(feature_len)
+        state_list.append(stream.states)
+
+    features = pad_sequence(
+        feature_list, batch_first=True, padding_value=LOG_EPSILON
+    ).to(device)
+    feature_lens = torch.tensor(feature_len_list, device=device)
+    num_processed_frames = torch.tensor(num_processed_frames_list, device=device)
+
+    # Make sure it has at least 1 frame after subsampling
+    tail_length = params.subsampling_factor + 5
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPSILON,
+        )
+
+    # Stack states of all streams
+    states = stack_states(state_list)
+
+    encoder_out, encoder_out_lens, states = model.encoder(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+    elif params.decoding_method == "fast_beam_search":
+        # feature_len is needed to get partial results.
+        # The rnnt_decoding_stream for fast_beam_search.
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            processed_lens = (
+                num_processed_frames // params.subsampling_factor + encoder_out_lens
+            )
+        fast_beam_search_one_best(
+            model=model,
+            streams=streams,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    # Update cached states of each stream
+    state_list = unstack_states(states)
+    for i, s in enumerate(state_list):
+        streams[i].states = s
+
+    finished_streams = [i for i, stream in enumerate(streams) if stream.done]
+    return finished_streams
+
+
+def create_streaming_feature_extractor() -> Fbank:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return Fbank(opts)
+
+
+def decode_dataset(
+    cuts: CutSet,
+    model: nn.Module,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+):
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The Transducer model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search.
+
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = next(model.parameters()).device
+
+    log_interval = 300
+
+    fbank = create_streaming_feature_extractor()
+
+    decode_results = []
+    streams = []
+    for num, cut in enumerate(cuts):
+        # Each utterance has a Stream.
+        stream = Stream(
+            params=params,
+            cut_id=cut.id,
+            decoding_graph=decoding_graph,
+            device=device,
+            LOG_EPS=LOG_EPSILON,
+        )
+
+        stream.states = model.encoder.get_init_states(device=device)
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+        feature = fbank(samples)
+        stream.set_feature(feature)
+        stream.ground_truth = cut.supervisions[0].text
+
+        streams.append(stream)
+
+        while len(streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                model=model,
+                streams=streams,
+                params=params,
+                decoding_graph=decoding_graph,
+            )
+
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        streams[i].id,
+                        streams[i].ground_truth.split(),
+                        sp.decode(streams[i].decoding_result()).split(),
+                    )
+                )
+                del streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    while len(streams) > 0:
+        finished_streams = decode_one_chunk(
+            model=model,
+            streams=streams,
+            params=params,
+            decoding_graph=decoding_graph,
+        )
+
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    streams[i].id,
+                    streams[i].ground_truth.split(),
+                    sp.decode(streams[i].decoding_result()).split(),
+                )
+            )
+            del streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    else:
+        key = f"beam_size_{params.beam_size}"
+
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=sorted(results))
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-streaming-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.device = device
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            model=model,
+            params=params,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220810)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py b/egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py
new file mode 100755
index 000000000..03dfe1997
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py
@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./lstm_transducer_stateless/test_model.py
+"""
+
+import os
+from pathlib import Path
+
+import torch
+from export import (
+    export_decoder_model_jit_trace,
+    export_encoder_model_jit_trace,
+    export_joiner_model_jit_trace,
+)
+from lstm import stack_states, unstack_states
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.encoder_dim = 512
+    params.rnn_hidden_size = 1024
+    params.num_encoder_layers = 12
+    params.aux_layer_period = 0
+    params.exp_dir = Path("exp_test_model")
+
+    model = get_transducer_model(params)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    if not os.path.exists(params.exp_dir):
+        os.path.mkdir(params.exp_dir)
+
+    encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+    print("The model has been successfully exported using jit.trace.")
+
+
+def test_states_stack_and_unstack():
+    layer, batch, hidden, cell = 12, 100, 512, 1024
+    states = (
+        torch.randn(layer, batch, hidden),
+        torch.randn(layer, batch, cell),
+    )
+    states2 = stack_states(unstack_states(states))
+    assert torch.allclose(states[0], states2[0])
+    assert torch.allclose(states[1], states2[1])
+
+
+def main():
+    test_model()
+    test_states_stack_and_unstack()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py
new file mode 100644
index 000000000..7567dd58c
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py
@@ -0,0 +1,257 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./lstm_transducer_stateless/test_scaling_converter.py
+"""
+
+import copy
+
+import torch
+from scaling import (
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledEmbedding,
+    ScaledLinear,
+    ScaledLSTM,
+)
+from scaling_converter import (
+    convert_scaled_to_non_scaled,
+    scaled_conv1d_to_conv1d,
+    scaled_conv2d_to_conv2d,
+    scaled_embedding_to_embedding,
+    scaled_linear_to_linear,
+    scaled_lstm_to_lstm,
+)
+from train import get_params, get_transducer_model
+
+
+def get_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.encoder_dim = 512
+    params.rnn_hidden_size = 1024
+    params.num_encoder_layers = 12
+    params.aux_layer_period = -1
+
+    model = get_transducer_model(params)
+    return model
+
+
+def test_scaled_linear_to_linear():
+    N = 5
+    in_features = 10
+    out_features = 20
+    for bias in [True, False]:
+        scaled_linear = ScaledLinear(
+            in_features=in_features,
+            out_features=out_features,
+            bias=bias,
+        )
+        linear = scaled_linear_to_linear(scaled_linear)
+        x = torch.rand(N, in_features)
+
+        y1 = scaled_linear(x)
+        y2 = linear(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_linear = torch.jit.script(scaled_linear)
+        jit_linear = torch.jit.script(linear)
+
+        y3 = jit_scaled_linear(x)
+        y4 = jit_linear(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_conv1d_to_conv1d():
+    in_channels = 3
+    for bias in [True, False]:
+        scaled_conv1d = ScaledConv1d(
+            in_channels,
+            6,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        conv1d = scaled_conv1d_to_conv1d(scaled_conv1d)
+
+        x = torch.rand(20, in_channels, 10)
+        y1 = scaled_conv1d(x)
+        y2 = conv1d(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_conv1d = torch.jit.script(scaled_conv1d)
+        jit_conv1d = torch.jit.script(conv1d)
+
+        y3 = jit_scaled_conv1d(x)
+        y4 = jit_conv1d(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_conv2d_to_conv2d():
+    in_channels = 1
+    for bias in [True, False]:
+        scaled_conv2d = ScaledConv2d(
+            in_channels=in_channels,
+            out_channels=3,
+            kernel_size=3,
+            padding=1,
+            bias=bias,
+        )
+
+        conv2d = scaled_conv2d_to_conv2d(scaled_conv2d)
+
+        x = torch.rand(20, in_channels, 10, 20)
+        y1 = scaled_conv2d(x)
+        y2 = conv2d(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_conv2d = torch.jit.script(scaled_conv2d)
+        jit_conv2d = torch.jit.script(conv2d)
+
+        y3 = jit_scaled_conv2d(x)
+        y4 = jit_conv2d(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_embedding_to_embedding():
+    scaled_embedding = ScaledEmbedding(
+        num_embeddings=500,
+        embedding_dim=10,
+        padding_idx=0,
+    )
+    embedding = scaled_embedding_to_embedding(scaled_embedding)
+
+    for s in [10, 100, 300, 500, 800, 1000]:
+        x = torch.randint(low=0, high=500, size=(s,))
+        scaled_y = scaled_embedding(x)
+        y = embedding(x)
+        assert torch.equal(scaled_y, y)
+
+
+def test_scaled_lstm_to_lstm():
+    input_size = 512
+    batch_size = 20
+    for bias in [True, False]:
+        for hidden_size in [512, 1024]:
+            scaled_lstm = ScaledLSTM(
+                input_size=input_size,
+                hidden_size=hidden_size,
+                num_layers=1,
+                bias=bias,
+                proj_size=0 if hidden_size == input_size else input_size,
+            )
+
+            lstm = scaled_lstm_to_lstm(scaled_lstm)
+
+            x = torch.rand(200, batch_size, input_size)
+            h0 = torch.randn(1, batch_size, input_size)
+            c0 = torch.randn(1, batch_size, hidden_size)
+
+            y1, (h1, c1) = scaled_lstm(x, (h0, c0))
+            y2, (h2, c2) = lstm(x, (h0, c0))
+            assert torch.allclose(y1, y2)
+            assert torch.allclose(h1, h2)
+            assert torch.allclose(c1, c2)
+
+            jit_scaled_lstm = torch.jit.trace(lstm, (x, (h0, c0)))
+            y3, (h3, c3) = jit_scaled_lstm(x, (h0, c0))
+            assert torch.allclose(y1, y3)
+            assert torch.allclose(h1, h3)
+            assert torch.allclose(c1, c3)
+
+
+def test_convert_scaled_to_non_scaled():
+    for inplace in [False, True]:
+        model = get_model()
+        model.eval()
+
+        orig_model = copy.deepcopy(model)
+
+        converted_model = convert_scaled_to_non_scaled(model, inplace=inplace)
+
+        model = orig_model
+
+        # test encoder
+        N = 2
+        T = 100
+        vocab_size = model.decoder.vocab_size
+
+        x = torch.randn(N, T, 80, dtype=torch.float32)
+        x_lens = torch.full((N,), x.size(1))
+
+        e1, e1_lens, _ = model.encoder(x, x_lens)
+        e2, e2_lens, _ = converted_model.encoder(x, x_lens)
+
+        assert torch.all(torch.eq(e1_lens, e2_lens))
+        assert torch.allclose(e1, e2), (e1 - e2).abs().max()
+
+        # test decoder
+        U = 50
+        y = torch.randint(low=1, high=vocab_size - 1, size=(N, U))
+
+        d1 = model.decoder(y)
+        d2 = model.decoder(y)
+
+        assert torch.allclose(d1, d2)
+
+        # test simple projection
+        lm1 = model.simple_lm_proj(d1)
+        am1 = model.simple_am_proj(e1)
+
+        lm2 = converted_model.simple_lm_proj(d2)
+        am2 = converted_model.simple_am_proj(e2)
+
+        assert torch.allclose(lm1, lm2)
+        assert torch.allclose(am1, am2)
+
+        # test joiner
+        e = torch.rand(2, 3, 4, 512)
+        d = torch.rand(2, 3, 4, 512)
+
+        j1 = model.joiner(e, d)
+        j2 = converted_model.joiner(e, d)
+        assert torch.allclose(j1, j2)
+
+
+@torch.no_grad()
+def main():
+    test_scaled_linear_to_linear()
+    test_scaled_conv1d_to_conv1d()
+    test_scaled_conv2d_to_conv2d()
+    test_scaled_embedding_to_embedding()
+    test_scaled_lstm_to_lstm()
+    test_convert_scaled_to_non_scaled()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220730)
+    main()
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/train.py b/egs/librispeech/ASR/lstm_transducer_stateless3/train.py
new file mode 100755
index 000000000..6ef4c9860
--- /dev/null
+++ b/egs/librispeech/ASR/lstm_transducer_stateless3/train.py
@@ -0,0 +1,1172 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./lstm_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir lstm_transducer_stateless3/exp \
+  --full-libri 1 \
+  --max-duration 500
+
+# For mix precision training:
+
+./lstm_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir lstm_transducer_stateless3/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from lstm import RNN
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of RNN encoder layers..",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Encoder output dimension.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Decoder output dimension.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="Joiner output dimension.",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Dimension of feed forward.",
+    )
+
+    parser.add_argument(
+        "--rnn-hidden-size",
+        type=int,
+        default=1024,
+        help="Hidden dim for LSTM layers.",
+    )
+
+    parser.add_argument(
+        "--aux-layer-period",
+        type=int,
+        default=0,
+        help="""Peroid of auxiliary layers used for randomly combined during training.
+        If set to 0, will not use the random combiner (Default).
+        You can set a positive integer to use the random combiner, e.g., 3.
+        """,
+    )
+
+    parser.add_argument(
+        "--grad-norm-threshold",
+        type=float,
+        default=25.0,
+        help="""For each sequence element in batch, its gradient will be
+        filtered out if the gradient norm is larger than
+        `grad_norm_threshold * median`, where `median` is the median
+        value of gradient norms of all elememts in batch.""",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=40,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=10,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+            "is_pnnx": False,
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = RNN(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        rnn_hidden_size=params.rnn_hidden_size,
+        grad_norm_threshold=params.grad_norm_threshold,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        aux_layer_period=params.aux_layer_period,
+        is_pnnx=params.is_pnnx,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0 and not params.print_diagnostics:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if (
+            batch_idx > 0
+            and batch_idx % params.valid_interval == 0
+            and not params.print_diagnostics
+        ):
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    # # overwrite it
+    # scheduler.base_lrs = [params.initial_lr for _ in scheduler.base_lrs]
+    # print(scheduler.base_lrs)
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./lstm.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 3) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/prepare.sh b/egs/librispeech/ASR/prepare.sh
new file mode 100755
index 000000000..4a5072cc0
--- /dev/null
+++ b/egs/librispeech/ASR/prepare.sh
@@ -0,0 +1,439 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/LibriSpeech
+#      You can find BOOKS.TXT, test-clean, train-clean-360, etc, inside it.
+#      You can download them from https://www.openslr.org/12
+#
+#  - $dl_dir/lm
+#      This directory contains the following files downloaded from
+#       http://www.openslr.org/resources/11
+#
+#        - 3-gram.pruned.1e-7.arpa.gz
+#        - 3-gram.pruned.1e-7.arpa
+#        - 4-gram.arpa.gz
+#        - 4-gram.arpa
+#        - librispeech-vocab.txt
+#        - librispeech-lexicon.txt
+#        - librispeech-lm-norm.txt.gz
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download LM"
+  mkdir -p $dl_dir/lm
+  if [ ! -e $dl_dir/lm/.done ]; then
+    ./local/download_lm.py --out-dir=$dl_dir/lm
+    touch $dl_dir/lm/.done
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/LibriSpeech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/LibriSpeech $dl_dir/LibriSpeech
+  #
+  if [ ! -d $dl_dir/LibriSpeech/train-other-500 ]; then
+    lhotse download librispeech --full $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LibriSpeech manifest"
+  # We assume that you have downloaded the LibriSpeech corpus
+  # to $dl_dir/LibriSpeech
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.librispeech.done ]; then
+    lhotse prepare librispeech -j $nj $dl_dir/LibriSpeech data/manifests
+    touch data/manifests/.librispeech.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for librispeech"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.librispeech.done ]; then
+    ./local/compute_fbank_librispeech.py
+    touch data/fbank/.librispeech.done
+  fi
+
+  if [ ! -f data/fbank/librispeech_cuts_train-all-shuf.jsonl.gz ]; then
+    cat <(gunzip -c data/fbank/librispeech_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/fbank/librispeech_cuts_train-clean-360.jsonl.gz) \
+      <(gunzip -c data/fbank/librispeech_cuts_train-other-500.jsonl.gz) | \
+      shuf | gzip -c > data/fbank/librispeech_cuts_train-all-shuf.jsonl.gz
+  fi
+
+  if [ ! -e data/fbank/.librispeech-validated.done ]; then
+    log "Validating data/fbank for LibriSpeech"
+    parts=(
+      train-clean-100
+      train-clean-360
+      train-other-500
+      test-clean
+      test-other
+      dev-clean
+      dev-other
+    )
+    for part in ${parts[@]}; do
+      python3 ./local/validate_manifest.py \
+        data/fbank/librispeech_cuts_${part}.jsonl.gz
+    done
+    touch data/fbank/.librispeech-validated.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/lm/librispeech-lexicon.txt |
+    sort | uniq > $lang_dir/lexicon.txt
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+
+  if [ ! -f $lang_dir/L.fst ]; then
+    log "Converting L.pt to L.fst"
+    ./shared/convert-k2-to-openfst.py \
+      --olabels aux_labels \
+      $lang_dir/L.pt \
+      $lang_dir/L.fst
+  fi
+
+  if [ ! -f $lang_dir/L_disambig.fst ]; then
+    log "Converting L_disambig.pt to L_disambig.fst"
+    ./shared/convert-k2-to-openfst.py \
+      --olabels aux_labels \
+      $lang_dir/L_disambig.pt \
+      $lang_dir/L_disambig.fst
+  fi
+fi
+
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang_phone/words.txt $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      files=$(
+        find "$dl_dir/LibriSpeech/train-clean-100" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/train-clean-360" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/train-other-500" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+      log "Converting L.pt to L.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L.pt \
+        $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+      log "Converting L_disambig.pt to L_disambig.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L_disambig.pt \
+        $lang_dir/L_disambig.fst
+    fi
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare bigram token-level P for MMI training"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+      ./local/convert_transcript_words_to_tokens.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --transcript $lang_dir/transcript_words.txt \
+        --oov "<UNK>" \
+        > $lang_dir/transcript_tokens.txt
+    fi
+
+    if [ ! -f $lang_dir/P.arpa ]; then
+      ./shared/make_kn_lm.py \
+        -ngram-order 2 \
+        -text $lang_dir/transcript_tokens.txt \
+        -lm $lang_dir/P.arpa
+    fi
+
+    if [ ! -f $lang_dir/P.fst.txt ]; then
+      python3 -m kaldilm \
+        --read-symbol-table="$lang_dir/tokens.txt" \
+        --disambig-symbol='#0' \
+        --max-order=2 \
+        $lang_dir/P.arpa > $lang_dir/P.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $dl_dir/lm/3-gram.pruned.1e-7.arpa > data/lm/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      $dl_dir/lm/4-gram.arpa > data/lm/G_4_gram.fst.txt
+  fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/HL.fst ]; then
+      ./local/prepare_lang_fst.py  \
+        --lang-dir $lang_dir \
+        --ngram-G ./data/lm/G_3_gram.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+
+  # Note If ./local/compile_hlg.py throws OOM,
+  # please switch to the following command
+  #
+  # ./local/compile_hlg_using_openfst.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+
+    # Note If ./local/compile_hlg.py throws OOM,
+    # please switch to the following command
+    #
+    # ./local/compile_hlg_using_openfst.py --lang-dir $lang_dir
+  done
+fi
+
+# Compile LG for RNN-T fast_beam_search decoding
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Compile LG"
+  ./local/compile_lg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_lg.py --lang-dir $lang_dir
+  done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Generate LM training data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    lang_dir=data/lang_bpe_${vocab_size}
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/lm/librispeech-lm-norm.txt \
+      --lm-archive $out_dir/lm_data.pt
+  done
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Generate LM validation data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/valid.txt ]; then
+      files=$(
+        find "$dl_dir/LibriSpeech/dev-clean" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/dev-other" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > $out_dir/valid.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/valid.txt \
+      --lm-archive $out_dir/lm_data-valid.pt
+  done
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: Generate LM test data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/test.txt ]; then
+      files=$(
+        find "$dl_dir/LibriSpeech/test-clean" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/test-other" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > $out_dir/test.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/test.txt \
+      --lm-archive $out_dir/lm_data-test.pt
+  done
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+  log "Stage 14: Sort LM training data"
+  # Sort LM training data by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of BPE tokens
+  # in a sentence.
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data.pt \
+      --out-lm-data $out_dir/sorted_lm_data.pt \
+      --out-statistics $out_dir/statistics.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-valid.pt \
+      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+      --out-statistics $out_dir/statistics-valid.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-test.pt \
+      --out-lm-data $out_dir/sorted_lm_data-test.pt \
+      --out-statistics $out_dir/statistics-test.txt
+  done
+fi
diff --git a/egs/librispeech/ASR/pruned2_knowledge/__init__.py b/egs/librispeech/ASR/pruned2_knowledge/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py b/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py
new file mode 100644
index 000000000..ee7556e49
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py
@@ -0,0 +1,411 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest
+from lhotse.dataset import (
+    BucketingSampler,
+    CutConcatenate,
+    CutMix,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--full-libri",
+            type=str2bool,
+            default=True,
+            help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the BucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.json.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using BucketingSampler.")
+            train_sampler = BucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                bucket_method="equal_duration",
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = BucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = BucketingSampler(
+            cuts, max_duration=self.args.max_duration, shuffle=False
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_train-clean-100.json.gz")
+
+    @lru_cache()
+    def train_clean_360_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-360 cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_train-clean-360.json.gz")
+
+    @lru_cache()
+    def train_other_500_cuts(self) -> CutSet:
+        logging.info("About to get train-other-500 cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_train-other-500.json.gz")
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_dev-clean.json.gz")
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_dev-other.json.gz")
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_test-clean.json.gz")
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_test-other.json.gz")
diff --git a/egs/librispeech/ASR/pruned2_knowledge/beam_search.py b/egs/librispeech/ASR/pruned2_knowledge/beam_search.py
new file mode 100644
index 000000000..785a8f097
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/beam_search.py
@@ -0,0 +1,760 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import k2
+import torch
+from model import Transducer
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def fast_beam_search(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(k2.RnntDecodingStream(decoding_graph))
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+    decoding_streams.terminate_and_flush_to_streams()
+    lattice = decoding_streams.format_output(encoder_out_lens.tolist())
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+    return hyps
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """Greedy search for a single utterance.
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size, device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+def greedy_search_batch(
+    model: Transducer, encoder_out: torch.Tensor
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+    Returns:
+      Return a list-of-list of token IDs containing the decoded results.
+      len(ans) equals to encoder_out.size(0).
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    device = model.device
+
+    batch_size = encoder_out.size(0)
+    T = encoder_out.size(1)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    hyps = [[blank_id] * context_size for _ in range(batch_size)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (batch_size, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    # decoder_out: (batch_size, 1, decoder_out_dim)
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t : t + 1, :].unsqueeze(2)  # noqa
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits'shape (batch_size, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (batch_size, vocab_size)
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    ans = [h[context_size:] for h in hyps]
+    return ans
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def _get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
+def modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[List[int]]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      beam:
+        Number of active paths during the beam search.
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+
+    batch_size = encoder_out.size(0)
+    T = encoder_out.size(1)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    B = [HypothesisList() for _ in range(batch_size)]
+    for i in range(batch_size):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t : t + 1, :].unsqueeze(2)  # noqa
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = _get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+            topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+    ans = [h.ys[context_size:] for h in best_hyps]
+
+    return ans
+
+
+def _deprecated_modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    It decodes only one utterance at a time. We keep it only for reference.
+    The function :func:`modified_beam_search` should be preferred as it
+    supports batch decoding.
+
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # current_encoder_out is of shape (1, 1, 1, encoder_out_dim)
+        # fmt: on
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+            dtype=torch.int64,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_output is of shape (num_hyps, 1, 1, joiner_dim)
+
+        current_encoder_out = current_encoder_out.expand(
+            decoder_out.size(0), 1, 1, -1
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        # now logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        topk_hyp_indexes = topk_hyp_indexes.tolist()
+        topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
+            new_ys = hyp.ys[:]
+            new_token = topk_token_indexes[i]
+            if new_token != blank_id:
+                new_ys.append(new_token)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+            B.add(new_hyp)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    return ys
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size,
+        device=device,
+        dtype=torch.int64,
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0))
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        # TODO(fangjun): Implement prefix search to update the `log_prob`
+        # of hypotheses in A
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            cached_key = y_star.key
+
+            if cached_key not in decoder_cache:
+                decoder_input = torch.tensor(
+                    [y_star.ys[-context_size:]],
+                    device=device,
+                    dtype=torch.int64,
+                ).reshape(1, context_size)
+
+                decoder_out = model.decoder(decoder_input, need_pad=False)
+                decoder_out = model.joiner.decoder_proj(decoder_out)
+                decoder_cache[cached_key] = decoder_out
+            else:
+                decoder_out = decoder_cache[cached_key]
+
+            cached_key += f"-t-{t}"
+            if cached_key not in joint_cache:
+                logits = model.joiner(
+                    current_encoder_out,
+                    decoder_out.unsqueeze(1),
+                    project_input=False,
+                )
+
+                # TODO(fangjun): Scale the blank posterior
+                log_prob = logits.log_softmax(dim=-1)
+                # log_prob is (1, 1, 1, vocab_size)
+                log_prob = log_prob.squeeze()
+                # Now log_prob is (vocab_size,)
+                joint_cache[cached_key] = log_prob
+            else:
+                log_prob = joint_cache[cached_key]
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for i, v in zip(indices.tolist(), values.tolist()):
+                if i == blank_id:
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys
diff --git a/egs/librispeech/ASR/pruned2_knowledge/conformer.py b/egs/librispeech/ASR/pruned2_knowledge/conformer.py
new file mode 100644
index 000000000..de367c234
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/conformer.py
@@ -0,0 +1,1041 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from sampling import KnowledgeBaseLookup, create_knowledge_base
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask
+
+
+class Conformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        knowledge_M: int = 256,
+        knowledge_N: int = 2,
+        knowledge_D: int = 512,
+        knowledge_K: int = 16,
+    ) -> None:
+        super(Conformer, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        self.knowledge_base = create_knowledge_base(
+            knowledge_M, knowledge_N, knowledge_D
+        )
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        # Pass in a lambda that creates a new ConformerEncoderLayer with these
+        # args.  Don't use deepcopy because we need the knowledge_base
+        # to be shared.
+        encoder_layer_fn = lambda: ConformerEncoderLayer(  # noqa: E731
+            self.knowledge_base,
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+            knowledge_M,
+            knowledge_N,
+            knowledge_D,
+            knowledge_K,
+        )
+        self.encoder = ConformerEncoder(encoder_layer_fn, num_encoder_layers)
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, d_model)
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            # Caution: We assume the subsampling factor is 4!
+            lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(0) == lengths.max().item()
+        mask = make_pad_mask(lengths)
+
+        x = self.encoder(
+            x, pos_emb, src_key_padding_mask=mask, warmup=warmup
+        )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+      knowledge_base: shared knowledge base parameter matrix, to be passed to constructors
+              of lookup modules
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+     knowledge_M, knowledge_N, knowledge_D, knowledge_K: parameters for knowledge-base,
+       see docs for KnowlegeBaseLookup.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        knowledge_base: nn.Parameter,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        knowledge_M: int = 256,
+        knowledge_N: int = 2,
+        knowledge_D: int = 512,
+        knowledge_K: int = 16,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.lookup = KnowledgeBaseLookup(
+            knowledge_M, knowledge_N, knowledge_D, knowledge_K, d_model, knowledge_base
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src = src + self.dropout(self.conv_module(src))
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        # knowledge-base lookup
+        src = src + self.dropout(self.lookup(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer_fn, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList([encoder_layer_fn() for i in range(num_layers)])
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/librispeech/ASR/pruned2_knowledge/decode.py b/egs/librispeech/ASR/pruned2_knowledge/decode.py
new file mode 100755
index 000000000..82fd103ea
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/decode.py
@@ -0,0 +1,528 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned2_knowledge/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned2_knowledge/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned2_knowledge/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned2_knowledge/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned2_knowledge/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned2_knowledge/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned2_knowledge/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned2_knowledge/exp \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned2_knowledge/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned2_knowledge/decoder.py b/egs/librispeech/ASR/pruned2_knowledge/decoder.py
new file mode 100644
index 000000000..0b9c886c7
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/decoder.py
@@ -0,0 +1,101 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import ScaledConv1d, ScaledEmbedding
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = ScaledEmbedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = ScaledConv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim,
+                bias=False,
+            )
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = F.relu(embedding_out)
+        return embedding_out
diff --git a/egs/librispeech/ASR/pruned2_knowledge/decoder2.py b/egs/librispeech/ASR/pruned2_knowledge/decoder2.py
new file mode 100644
index 000000000..0c9cee431
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/decoder2.py
@@ -0,0 +1,261 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from subsampling import ScaledConv1d
+from torch import Tensor
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = ScaledEmbedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        if context_size > 1:
+            self.conv = ScaledConv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, embedding_dim).
+        """
+        y = y.to(torch.int64)
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        return embedding_out
+
+
+class ScaledEmbedding(nn.Module):
+    r"""A simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
+                                    is renormalized to have norm :attr:`max_norm`.
+        norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+    """
+    __constants__ = [
+        "num_embeddings",
+        "embedding_dim",
+        "padding_idx",
+        "scale_grad_by_freq",
+        "sparse",
+    ]
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        scale_grad_by_freq: bool = False,
+        sparse: bool = False,
+        scale_speed: float = 5.0,
+    ) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert (
+                    padding_idx < self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+            elif padding_idx < 0:
+                assert (
+                    padding_idx >= -self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale_speed = scale_speed
+        self.scale = nn.Parameter(torch.zeros(()))  # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        nn.init.normal_(self.weight, std=0.05)
+        nn.init.constant_(self.scale, torch.tensor(1.0 / 0.05).log() / self.scale_speed)
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        scale = (self.scale * self.scale_speed).exp()
+        if input.numel() < self.num_embeddings:
+            return (
+                F.embedding(
+                    input,
+                    self.weight,
+                    self.padding_idx,
+                    None,
+                    2.0,  # None, 2.0 relate to normalization
+                    self.scale_grad_by_freq,
+                    self.sparse,
+                )
+                * scale
+            )
+        else:
+            return F.embedding(
+                input,
+                self.weight * scale,
+                self.padding_idx,
+                None,
+                2.0,  # None, 2.0 relates to normalization
+                self.scale_grad_by_freq,
+                self.sparse,
+            )
+
+    def extra_repr(self) -> str:
+        s = "{num_embeddings}, {embedding_dim}, scale_speed={scale_speed}, scale={scale}"
+        if self.padding_idx is not None:
+            s += ", padding_idx={padding_idx}"
+        if self.scale_grad_by_freq is not False:
+            s += ", scale_grad_by_freq={scale_grad_by_freq}"
+        if self.sparse is not False:
+            s += ", sparse=True"
+        return s.format(**self.__dict__)
diff --git a/egs/librispeech/ASR/pruned2_knowledge/encoder_interface.py b/egs/librispeech/ASR/pruned2_knowledge/encoder_interface.py
new file mode 100644
index 000000000..257facce4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/encoder_interface.py
@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")
diff --git a/egs/librispeech/ASR/pruned2_knowledge/export.py b/egs/librispeech/ASR/pruned2_knowledge/export.py
new file mode 100755
index 000000000..51020aa30
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/export.py
@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned2_knowledge/export.py \
+  --exp-dir ./pruned2_knowledge/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned2_knowledge/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned2_knowledge/decode.py \
+        --exp-dir ./pruned2_knowledge/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned2_knowledge/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert args.jit is False, "Support torchscript will be added later"
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned2_knowledge/joiner.py b/egs/librispeech/ASR/pruned2_knowledge/joiner.py
new file mode 100644
index 000000000..68c663b66
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/joiner.py
@@ -0,0 +1,65 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim)
+        self.output_linear = ScaledLinear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.shape[:-1] == decoder_out.shape[:-1]
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit
diff --git a/egs/librispeech/ASR/pruned2_knowledge/model.py b/egs/librispeech/ASR/pruned2_knowledge/model.py
new file mode 100644
index 000000000..ca8c28af1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/model.py
@@ -0,0 +1,189 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned2_knowledge/optim.py b/egs/librispeech/ASR/pruned2_knowledge/optim.py
new file mode 100644
index 000000000..9f287ce70
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/optim.py
@@ -0,0 +1,315 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Union
+
+import torch
+from torch.optim import Optimizer
+
+
+class Eve(Optimizer):
+    r"""
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam\: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+        return loss
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("initial_lr", group["lr"])
+
+        self.base_lrs = [group["initial_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            print(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+     lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                       (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
+
+     E.g. suggest initial-lr = 0.003 (passed to optimizer).
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        return [x * factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = Eve(m.parameters(), lr=0.003)
+
+    scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+    print("last lr = ", scheduler.get_last_lr())
+    print("state dict = ", scheduler.state_dict())
+
+
+if __name__ == "__main__":
+    _test_eden()
diff --git a/egs/librispeech/ASR/pruned2_knowledge/sampling.py b/egs/librispeech/ASR/pruned2_knowledge/sampling.py
new file mode 100644
index 000000000..5b595c76c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/sampling.py
@@ -0,0 +1,355 @@
+#!/usr/bin/env python3
+
+# This was copied from /ceph-dan/torch-sampling/torch_sampling/sampling_ref.py,
+# its git history is there.
+
+import random
+import timeit
+from typing import Optional, Tuple
+
+import torch
+from scaling import ScaledLinear
+from torch import Tensor, nn
+from torch.cuda.amp import GradScaler, custom_bwd, custom_fwd
+from torch_scheduled_sampling import sample_combined
+
+# The main exports of this file are the module KnowledgeBaseLookup and the
+# function create_knowledge_base.
+
+
+def create_knowledge_base(M: int, N: int, D: int) -> nn.Parameter:
+    std = 0.1
+    a = (3**0.5) * std  # this sqrt(3) thing is intended to get variance of
+    # 0.1 from uniform distribution
+    ans = nn.Parameter(torch.ones(M**N, D))
+    nn.init.uniform_(ans, -a, a)
+    return ans
+
+
+def join_indexes(indexes: Tensor, M: int) -> Tensor:
+    """
+    Combines N-tuples of indexes into single indexes that can be used for
+    lookup in the knowledge base.  Args:
+      indexes: tensor of torch.int64 of shape (*, K, N), with elements in
+         {0..M-1}
+         M: the size of the original softmaxes, is upper bound on elements
+           in indexes
+       Returns:
+          joined_indexes: of shape (*, K), joined_indexes[...,k] equals
+            joined_indexes[...,0,k] + joined_indexes[...,1,k]*(M**1) ... + joined_indexes[...,1,k]*(M**(N-1))]
+    """
+    N = indexes.shape[-1]
+    n_powers = M ** torch.arange(N, device=indexes.device)  # [ 1, M, ..., M**(N-1) ]
+    return (indexes * n_powers).sum(dim=-1)
+
+
+# Note, we don't use this, we
+def weighted_matrix_lookup(
+    weights: Tensor, indexes: Tensor, knowledge_base: Tensor
+) -> Tensor:
+    """
+    Weighted combination of specified rows of a matrix.
+         weights: Tensor of shape (*, K), can contain any value but probably in [0..1].
+         indexes: Tensor of shape (*, K), with elements in [0..C-1]
+         knowledge_base: Tensor of shape (C-1, D), whose rows we'll be looking up
+      Returns:
+         tensor of shape (*, D), containing weighted sums of rows of
+         `knowledge_base`
+    """
+    if True:
+        return WeightedMatrixLookupFunction.apply(weights, indexes, knowledge_base)
+    else:
+        # simpler but less memory-efficient implementation
+        lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten())
+        D = knowledge_base.shape[-1]
+        weights = weights.unsqueeze(-2)  # (*, 1, K)
+        lookup = lookup.reshape(*indexes.shape, D)  # (*, K, D)
+        ans = torch.matmul(weights, lookup)  # ans: (*, 1, D)
+        ans = ans.squeeze(-2)
+        assert list(ans.shape) == list(weights.shape[:-2]) + [D]
+        return ans
+
+
+class WeightedMatrixLookupFunction(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(
+        ctx, weights: Tensor, indexes: Tensor, knowledge_base: Tensor
+    ) -> Tensor:
+        """
+        Weighted combination of specified rows of a matrix.
+         weights: Tensor of shape (*, K), can contain any value but probably in [0..1].
+         indexes: Tensor of shape (*, K), with elements in [0..C-1]
+         knowledge_base: Tensor of shape (C, D), whose rows we'll be looking up
+        Returns:
+         tensor of shape (*, D), containing weighted sums of rows of
+         `knowledge_base`
+        """
+        if random.random() < 0.001:
+            print("dtype[1] = ", weights.dtype)
+        ctx.save_for_backward(
+            weights.detach(), indexes.detach(), knowledge_base.detach()
+        )
+        with torch.no_grad():
+            lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten())
+            D = knowledge_base.shape[-1]
+            weights = weights.unsqueeze(-2)  # (*, 1, K)
+            lookup = lookup.reshape(*indexes.shape, D)  # (*, K, D)
+            ans = torch.matmul(weights, lookup)  # ans: (*, 1, D)
+            ans = ans.squeeze(-2)  # (*, D)
+        return ans
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, Tensor]:
+        # ans_grad: (*, D)
+        weights, indexes, knowledge_base = ctx.saved_tensors
+        knowledge_base.requires_grad = True
+        dtype = ans_grad.dtype
+        ans_grad = ans_grad.to(weights.dtype)
+        assert weights.requires_grad is False
+        D = knowledge_base.shape[-1]
+        with torch.enable_grad():
+            # we'll use torch's autograd to differentiate this operation, which
+            # is nontrivial [and anyway we need `lookup` to compute weight grad.
+            # We don't save `lookup` because it's large, that is the reason
+            # we override Torch autograd.
+            lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten())
+            lookup = lookup.reshape(*indexes.shape, D)  # (*, K, D)
+        weights = weights.unsqueeze(-1)  # (*, K, 1)
+        # forward pass: was:
+        ## ans = torch.matmul(weights, lookup)
+        ## ans: (*, 1, D)
+        ## ans = ans.squeeze(-2) # ans, ans_grad: (*, D)
+        weights_grad = torch.matmul(
+            lookup, ans_grad.unsqueeze(-1)  # (*, K, D)
+        )  # (*, D, 1)
+        weights_grad = weights_grad.squeeze(-1)  # (*, K, 1) -> (*, K)
+        lookup_grad = weights * ans_grad.unsqueeze(
+            -2
+        )  # (*, K, 1) * (*, 1, D) = (*, K, D)
+        lookup.backward(gradient=lookup_grad)
+        return weights_grad.to(dtype), None, knowledge_base.grad.to(dtype)
+
+
+class PenalizeNegentropyFunction(torch.autograd.Function):
+    """
+    Function that does nothing in forward pass, but in backprop, it is as
+    if you had added: `- tot_entropy * alpha` to the loss function, where
+    tot_entropy is the the entropy of the average of the input distributions,
+    times the number of input distributions. (We multiply by this because
+    our overall loss function is proportional to the number of frames).
+
+    This will tend to make the entropy want to become as large as possible,
+    making (-tot_entropy * alpha) as negative as possible.
+
+    Args:
+       logprobs: Tensor of shape (*, num_classes), should be the result of
+            calling some_tensor.log_softmax(dim=-1)
+     Returns:
+       logprobs
+    """
+
+    @staticmethod
+    def forward(ctx, logprobs: Tensor, alpha: float):
+        ctx.save_for_backward(logprobs.detach())
+        ctx.alpha = alpha
+        return logprobs
+
+    @staticmethod
+    def backward(ctx, logprobs_grad: Tensor) -> Tuple[Tensor, None]:
+        (logprobs,) = ctx.saved_tensors
+        with torch.enable_grad():
+            logprobs.requires_grad = True
+            # `negentropy` is the negative entropy of the average distribution.
+            # distributions.  It will be <= 0.
+            l = logprobs.reshape(-1, logprobs.shape[-1])  # noqa: E741
+            scale = ctx.alpha * l.shape[0]
+            avg_dist = l.exp().mean(dim=0)
+            negentropy = (avg_dist * (avg_dist + 1.0e-20).log()).sum()
+            if random.random() < 0.0005:
+                negentropy_individual = (l * l.exp()).sum(dim=-1).mean()
+                print(
+                    "Negentropy[individual,combined] = ",
+                    negentropy_individual.item(),
+                    ", ",
+                    negentropy.item(),
+                )
+            loss = negentropy * scale
+            loss.backward()
+        return logprobs_grad + logprobs.grad, None
+
+
+class KnowledgeBaseLookup(nn.Module):
+    """
+    Create knowledge-base lookup module.  (The knowledge-base parameter, which is
+    large, is shared between these modules).
+    Args:
+       M: int, softmax size, e.g. in [32..128]
+       N: int, number of softmaxes, in [2..3]
+       D: int, embedding dimension in knowledge base, e.g. 256
+       K: number of samples (affects speed/accuracy tradeoff), e.g. 16.
+      embedding_dim:  the dimension to project from and to, e.g. the
+        d_model of the conformer.
+    """
+
+    def __init__(
+        self,
+        M: int,
+        N: int,
+        D: int,
+        K: int,
+        embedding_dim: int,
+        knowledge_base: nn.Parameter,
+        negentropy_penalty: float = 0.001,
+    ):
+        super(KnowledgeBaseLookup, self).__init__()
+        self.knowledge_base = knowledge_base  # shared!
+        self.in_proj = ScaledLinear(embedding_dim, M * N, initial_scale=1.0)
+        # initial_scale = 4.0 because the knowlege_base activations are
+        # quite small -- if we use our optimizer they'll have stddev <= 0.1.
+        self.out_proj = ScaledLinear(D, embedding_dim, initial_scale=4.0)
+        self.M = M
+        self.N = N
+        self.K = K
+        self.negentropy_penalty = negentropy_penalty
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        Forward function that does knowledge-base lookup.
+        Args:
+             x: input, of shape (*, E) where E is embedding_dim
+                as passed to constructor
+             y: output of knowledge-base lookup, of shape (*, E)
+
+        # TODO: later we can try multiplying by a projection of x or something like that.
+        """
+        x = self.in_proj(x)  # now (*, M*N)
+        x = x.reshape(*x.shape[:-1], self.N, self.M)  # now (*, N, M)
+        x = x.log_softmax(dim=-1)  # now normalized logprobs, dim= (*, N, M)
+        x = PenalizeNegentropyFunction.apply(x, self.negentropy_penalty)
+
+        _, indexes, weights = sample_combined(x, self.K, input_is_log=True)
+        x = weighted_matrix_lookup(weights, indexes, self.knowledge_base)  # now (*, D)
+        x = self.out_proj(x)  # now (*, self.embedding_dim)
+        return x
+
+
+def _test_knowledge_base_lookup():
+    K = 16
+    N = 2
+    M = 128
+    D = 256
+    E = 255
+
+    knowledge_base: nn.Parameter = create_knowledge_base(M, N, D)
+    m = KnowledgeBaseLookup(M, N, D, K, E, knowledge_base)
+
+    B = 30
+    T = 40
+    x = torch.randn(B, T, E)
+    x.requires_grad = True
+    y = m(x)
+    assert y.shape == x.shape
+    y.sum().backward()  # make sure backward doesn't crash..
+    print("y = ", y)
+    print("x.grad = ", x.grad)
+    print("knowlege_base.grad norm = ", knowledge_base.grad.norm())
+
+    dtype = torch.float32
+    device = torch.device("cuda")
+    train_pairs = [
+        (
+            torch.randn(B, T, E, device=device, dtype=dtype),
+            torch.randn(B, T, E, device=device, dtype=dtype),
+        )
+        for _ in range(10)
+    ]
+    from optim import Eve
+
+    optimizer = Eve(m.parameters(), lr=0.005, eps=1.0e-04)
+    m = m.to(device).to(dtype)
+
+    start = timeit.default_timer()
+
+    # Epoch 0, batch 0, loss 1.0109944343566895
+    # Epoch 10, batch 0, loss 1.0146660804748535
+    # Epoch 20, batch 0, loss 1.0119813680648804
+    # Epoch 30, batch 0, loss 1.0105408430099487
+    # Epoch 40, batch 0, loss 1.0077732801437378
+    # Epoch 50, batch 0, loss 1.0050103664398193
+    # Epoch 60, batch 0, loss 1.0033129453659058
+    # Epoch 70, batch 0, loss 1.0014232397079468
+    # Epoch 80, batch 0, loss 0.9977912306785583
+    # Epoch 90, batch 0, loss 0.8274348974227905
+    # Epoch 100, batch 0, loss 0.3368612825870514
+    # Epoch 110, batch 0, loss 0.11323091387748718
+    # Time taken:  17.591704960912466
+    for epoch in range(150):
+        for n, (x, y) in enumerate(train_pairs):
+            y_out = m(x)
+            loss = ((y_out - y) ** 2).mean() * 100.0
+            if n % 10 == 0 and epoch % 10 == 0:
+                print(f"Epoch {epoch}, batch {n}, loss {loss.item()}")
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+
+    stop = timeit.default_timer()
+    print("Time taken: ", stop - start)
+
+
+def _test_knowledge_base_lookup_autocast():
+    K = 16
+    N = 2
+    M = 128
+    D = 256
+    E = 255
+
+    knowledge_base: nn.Parameter = create_knowledge_base(M, N, D)
+    m = KnowledgeBaseLookup(M, N, D, K, E, knowledge_base)
+
+    B = 30
+    T = 40
+    x = torch.randn(B, T, E)
+    x.requires_grad = True
+    y = m(x)
+    assert y.shape == x.shape
+    y.sum().backward()  # make sure backward doesn't crash..
+    print("y = ", y)
+    print("x.grad = ", x.grad)
+    print("knowlege_base.grad norm = ", knowledge_base.grad.norm())
+
+    device = torch.device("cuda")
+    train_pairs = [
+        (torch.randn(B, T, E, device=device), torch.randn(B, T, E, device=device))
+        for _ in range(10)
+    ]
+    from optim import Eve
+
+    optimizer = Eve(m.parameters(), lr=0.005, eps=1.0e-04)
+    m = m.to(device)
+
+    scaler = GradScaler(enabled=True)
+
+    start = timeit.default_timer()
+
+    for epoch in range(150):
+        for n, (x, y) in enumerate(train_pairs):
+            y_out = m(x)
+            with torch.cuda.amp.autocast(enabled=True):
+                loss = ((y_out - y) ** 2).mean() * 100.0
+            if n % 10 == 0 and epoch % 10 == 0:
+                print(f"Epoch {epoch}, batch {n}, loss {loss.item()}")
+            scaler.scale(loss).backward()
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+
+    stop = timeit.default_timer()
+    print("Time taken: ", stop - start)
+
+
+if __name__ == "__main__":
+    _test_knowledge_base_lookup()
+    _test_knowledge_base_lookup_autocast()
diff --git a/egs/librispeech/ASR/pruned2_knowledge/scaling.py b/egs/librispeech/ASR/pruned2_knowledge/scaling.py
new file mode 100644
index 000000000..527c735eb
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/scaling.py
@@ -0,0 +1,688 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+from itertools import repeat
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+from torch.cuda.amp import custom_bwd, custom_fwd
+
+
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.Iterable):
+            return x
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+_single = _ntuple(1)
+_pair = _ntuple(2)
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(
+        ctx,
+        x: Tensor,
+        channel_dim: int,
+        min_positive: float,  # e.g. 0.05
+        max_positive: float,  # e.g. 0.95
+        max_factor: float,  # e.g. 0.01
+        min_abs: float,  # e.g. 0.2
+        max_abs: float,  # e.g. 100.0
+    ) -> Tensor:
+        if x.requires_grad:
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+            xgt0 = x > 0
+            proportion_positive = torch.mean(
+                xgt0.to(x.dtype), dim=sum_dims, keepdim=True
+            )
+            factor1 = (
+                (min_positive - proportion_positive).relu()
+                * (max_factor / min_positive)
+                if min_positive != 0.0
+                else 0.0
+            )
+            factor2 = (
+                (proportion_positive - max_positive).relu()
+                * (max_factor / (max_positive - 1.0))
+                if max_positive != 1.0
+                else 0.0
+            )
+            factor = factor1 + factor2
+            if isinstance(factor, float):
+                factor = torch.zeros_like(proportion_positive)
+
+            mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True)
+            below_threshold = mean_abs < min_abs
+            above_threshold = mean_abs > max_abs
+
+            ctx.save_for_backward(factor, xgt0, below_threshold, above_threshold)
+            ctx.max_factor = max_factor
+            ctx.sum_dims = sum_dims
+        return x
+
+    @staticmethod
+    @custom_bwd
+    def backward(
+        ctx, x_grad: Tensor
+    ) -> Tuple[Tensor, None, None, None, None, None, None]:
+        factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors
+        dtype = x_grad.dtype
+        scale_factor = (
+            (below_threshold.to(dtype) - above_threshold.to(dtype))
+            * (xgt0.to(dtype) - 0.5)
+            * (ctx.max_factor * 2.0)
+        )
+
+        neg_delta_grad = x_grad.abs() * (factor + scale_factor)
+        return x_grad - neg_delta_grad, None, None, None, None, None, None
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x**2, dim=self.channel_dim, keepdim=True) + self.eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+class ScaledLinear(nn.Linear):
+    """
+    A modified version of nn.Linear where the parameters are scaled before
+    use, via:
+         weight = self.weight * self.weight_scale.exp()
+         bias = self.bias * self.bias_scale.exp()
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+        initial_speed: this affects how fast the parameter will
+           learn near the start of training; you can set it to a
+           value less than one if you suspect that a module
+           is contributing to instability near the start of training.
+           Nnote: regardless of the use of this option, it's best to
+           use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.   Must be greater than 0.
+    """
+
+    def __init__(
+        self, *args, initial_scale: float = 1.0, initial_speed: float = 1.0, **kwargs
+    ):
+        super(ScaledLinear, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in nn.Linear
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        return torch.nn.functional.linear(input, self.get_weight(), self.get_bias())
+
+
+class ScaledConv1d(nn.Conv1d):
+    # See docs for ScaledLinear
+    def __init__(
+        self, *args, initial_scale: float = 1.0, initial_speed: float = 1.0, **kwargs
+    ):
+        super(ScaledConv1d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv1d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                self.get_weight(),
+                self.get_bias(),
+                self.stride,
+                _single(0),
+                self.dilation,
+                self.groups,
+            )
+        return F.conv1d(
+            input,
+            self.get_weight(),
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+
+class ScaledConv2d(nn.Conv2d):
+    # See docs for ScaledLinear
+    def __init__(
+        self, *args, initial_scale: float = 1.0, initial_speed: float = 1.0, **kwargs
+    ):
+        super(ScaledConv2d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def _conv_forward(self, input, weight):
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv2d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                weight,
+                self.get_bias(),
+                self.stride,
+                _pair(0),
+                self.dilation,
+                self.groups,
+            )
+        return F.conv2d(
+            input,
+            weight,
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+    def forward(self, input: Tensor) -> Tensor:
+        return self._conv_forward(input, self.get_weight())
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+
+    Args:
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           min_abs:  the minimum average-absolute-value per channel, which
+              we allow, before we start to modify the derivatives to prevent
+              this.
+           max_abs:  the maximum average-absolute-value per channel, which
+               we allow, before we start to modify the derivatives to prevent
+               this.
+    """
+
+    def __init__(
+        self,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.01,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+
+    def forward(self, x: Tensor) -> Tensor:
+        return ActivationBalancerFunction.apply(
+            x,
+            self.channel_dim,
+            self.min_positive,
+            self.max_positive,
+            self.max_factor,
+            self.min_abs,
+            self.max_abs,
+        )
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    @custom_fwd
+    def forward(ctx, x: Tensor) -> Tensor:
+        x = x.detach()
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+        ctx.save_for_backward(s, y)
+        return y
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        s, y = ctx.saved_tensors
+        return (y * (1 - s) + s) * y_grad
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        return DoubleSwishFunction.apply(x)
+
+
+class ScaledEmbedding(nn.Module):
+    r"""This is a modified version of nn.Embedding that introduces a learnable scale
+    on the parameters.  Note: due to how we initialize it, it's best used with
+    schedulers like Noam that have a warmup period.
+
+    It is a simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
+                                    is renormalized to have norm :attr:`max_norm`.
+        norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+        initial_speed (float, optional):  This affects how fast the parameter will
+           learn near the start of training; you can set it to a value less than
+           one if you suspect that a module is contributing to instability near
+           the start of training.  Nnote: regardless of the use of this option,
+           it's best to use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.  Must be greater than 0.
+
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+
+    """
+    __constants__ = [
+        "num_embeddings",
+        "embedding_dim",
+        "padding_idx",
+        "scale_grad_by_freq",
+        "sparse",
+    ]
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        scale_grad_by_freq: bool = False,
+        sparse: bool = False,
+        initial_speed: float = 1.0,
+    ) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert (
+                    padding_idx < self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+            elif padding_idx < 0:
+                assert (
+                    padding_idx >= -self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale = nn.Parameter(torch.zeros(()))  # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters(initial_speed)
+
+    def reset_parameters(self, initial_speed: float = 1.0) -> None:
+        std = 0.1 / initial_speed
+        nn.init.normal_(self.weight, std=std)
+        nn.init.constant_(self.scale, torch.tensor(1.0 / std).log())
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        scale = self.scale.exp()
+        if input.numel() < self.num_embeddings:
+            return (
+                F.embedding(
+                    input,
+                    self.weight,
+                    self.padding_idx,
+                    None,
+                    2.0,  # None, 2.0 relate to normalization
+                    self.scale_grad_by_freq,
+                    self.sparse,
+                )
+                * scale
+            )
+        else:
+            return F.embedding(
+                input,
+                self.weight * scale,
+                self.padding_idx,
+                None,
+                2.0,  # None, 2.0 relates to normalization
+                self.scale_grad_by_freq,
+                self.sparse,
+            )
+
+    def extra_repr(self) -> str:
+        s = "{num_embeddings}, {embedding_dim}, scale={scale}"
+        if self.padding_idx is not None:
+            s += ", padding_idx={padding_idx}"
+        if self.scale_grad_by_freq is not False:
+            s += ", scale_grad_by_freq={scale_grad_by_freq}"
+        if self.sparse is not False:
+            s += ", sparse=True"
+        return s.format(**self.__dict__)
+
+
+def _test_activation_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x**2).mean().sqrt()
+    y_rms = (y**2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 0.5
+    x.requires_grad = True
+    m = DoubleSwish()
+    torch.autograd.gradcheck(m, x)
+
+
+if __name__ == "__main__":
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()
diff --git a/egs/librispeech/ASR/pruned2_knowledge/scaling_tmp.py b/egs/librispeech/ASR/pruned2_knowledge/scaling_tmp.py
new file mode 100644
index 000000000..3f21133a0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/scaling_tmp.py
@@ -0,0 +1,717 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+
+def _activation_balancer_loss(
+    mean_pos: Tensor,
+    mean_neg: Tensor,
+    min_positive: float,  # e.g. 0.05
+    max_positive: float,  # e.g. 0.95
+    max_factor: float,  # e.g. 0.01
+    min_abs: float,  # e.g. 0.2
+    max_abs: float,  # e.g. 100.0
+    eps: float = 1.0e-10,
+):
+    """
+    Returns a loss-function for the ActivationBalancer module.  This loss
+    function is not exposed to the user but is used internally, and eventually
+    its derivatives are scaled by some heuristic related to derivative magnitudes,
+    and added to the backpropped deriv.
+
+    Args:
+       mean_pos:  a Tensor of arbitrary dimension, probably something like (1, num_channels, 1, 1),
+               containing the mean of only the positive parts of the input features, i.e.
+               of x.relu().
+       mean_neg:  a Tensor of arbitrary dimension, probably something like (1, num_channels, 1, 1),
+               containing the mean of only the negative parts of the input features, i.e.
+               of (-x).relu().
+       min_positive: the minimum allowed value of mean_pos / (mean_pos + mean_neg) before we
+               start penalizing.
+       max_positive: the maximum allowed value of mean_pos / (mean_pos + mean_neg) before we
+               start penalizing.
+    """
+    loss_parts = []
+
+    x_mean = mean_pos - mean_neg
+    x_mean_abs = (mean_pos + mean_neg + eps).detach()
+    x_rel_mean = x_mean / x_mean_abs
+
+    if min_positive != 0.0:
+        # e.g. x_mean_floor = -0.95 + 0.05 = -0.9
+        x_rel_mean_floor = -(1 - min_positive) + min_positive
+        min_positive_loss = (x_rel_mean_floor - x_rel_mean).relu().sum() * (
+            1.0 / (2 * min_positive)
+        )
+        # this part of the loss would be 1.0 * num_channels if all these constraints were
+        # 100% violated.
+        loss_parts.append(min_positive_loss)
+
+    if max_positive != 1.0:
+        # e.g. x_mean_floor =  -0.05 + 0.95 = 0.8
+        x_rel_mean_ceil = -(1.0 - max_positive) + max_positive
+        max_positive_loss = (x_rel_mean - x_rel_mean_ceil).relu().sum() * (
+            1.0 / (1 - x_rel_mean_ceil)
+        )
+        # this part of the loss would be 1.0 * num_channels if all these constraints were
+        # 100% violated.
+        loss_parts.append(max_positive_loss)
+
+    if min_abs != 0.0:
+        min_abs_loss = (min_abs - x_mean_abs).relu().sum() / min_abs
+        # this part of the loss would be 1.0 * num_channels if all these constraints were
+        # 100% violated.
+        loss_parts.append(min_abs_loss)
+
+    if max_abs != 0.0:
+        max_abs_loss = (x_mean_abs / max_abs).log().relu()
+        # this part of the loss would be [something logarithmic] * num_channels if all these constraints were
+        # 100% violated.
+        loss_parts.append(max_abs_loss)
+
+    # the min_positive and 1 - max_positive are "ballast" added to the
+    denom = mean_pos + mean_neg + (min_positive + (1 - max_positive))
+    # num
+
+    if min_positive != 0.0:
+        pass
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        channel_dim: int,
+        min_positive: float,  # e.g. 0.05
+        max_positive: float,  # e.g. 0.95
+        max_factor: float,  # e.g. 0.01
+        min_abs: float,  # e.g. 0.2
+        max_abs: float,  # e.g. 100.0
+    ) -> Tensor:
+        if x.requires_grad:
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+            xgt0 = x > 0
+            proportion_positive = torch.mean(
+                xgt0.to(x.dtype), dim=sum_dims, keepdim=True
+            )
+            factor1 = (
+                (min_positive - proportion_positive).relu()
+                * (max_factor / min_positive)
+                if min_positive != 0.0
+                else 0.0
+            )
+            factor2 = (
+                (proportion_positive - max_positive).relu()
+                * (max_factor / (max_positive - 1.0))
+                if max_positive != 1.0
+                else 0.0
+            )
+            factor = factor1 + factor2
+            if isinstance(factor, float):
+                factor = torch.zeros_like(proportion_positive)
+
+            mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True)
+            below_threshold = mean_abs < min_abs
+            above_threshold = mean_abs > max_abs
+
+            ctx.save_for_backward(factor, xgt0, below_threshold, above_threshold)
+            ctx.max_factor = max_factor
+            ctx.sum_dims = sum_dims
+        return x
+
+    @staticmethod
+    def backward(
+        ctx, x_grad: Tensor
+    ) -> Tuple[Tensor, None, None, None, None, None, None]:
+        factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors
+        dtype = x_grad.dtype
+        scale_factor = (
+            (below_threshold.to(dtype) - above_threshold.to(dtype))
+            * (xgt0.to(dtype) - 0.5)
+            * (ctx.max_factor * 2.0)
+        )
+
+        neg_delta_grad = x_grad.abs() * (factor + scale_factor)
+        return x_grad - neg_delta_grad, None, None, None, None, None, None
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x**2, dim=self.channel_dim, keepdim=True) + self.eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+class ScaledLinear(nn.Linear):
+    """
+    A modified version of nn.Linear where the parameters are scaled before
+    use, via:
+         weight = self.weight * self.weight_scale.exp()
+         bias = self.bias * self.bias_scale.exp()
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+
+       Note: it uses the default initialization for the weight and bias,
+       inherited from nn.Linear.  For modules with small fan-in, this
+       may be larger than optimal.
+    """
+
+    def __init__(self, *args, initial_scale: float = 1.0, **kwargs):
+        super(ScaledLinear, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+
+        self._reset_parameters()  # Overrides the reset_parameters in nn.Linear
+
+    def _reset_parameters(self):
+        std = 0.01
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+            if self.bias is not None:
+                self.bias_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        return torch.nn.functional.linear(input, self.get_weight(), self.get_bias())
+
+
+class ScaledConv1d(nn.Conv1d):
+    def __init__(self, *args, initial_scale=1.0, **kwargs):
+        super(ScaledConv1d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters()  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self):
+        std = 0.01
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+            if self.bias is not None:
+                self.bias_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv1d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                self.get_weight(),
+                self.get_bias(),
+                self.stride,
+                _single(0),  # noqa: F821
+                self.dilation,
+                self.groups,
+            )
+        return F.conv1d(
+            input,
+            self.get_weight(),
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+
+class ScaledConv2d(nn.Conv2d):
+    def __init__(self, *args, initial_scale=1.0, **kwargs):
+        super(ScaledConv2d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters()  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self):
+        std = 0.01
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+            if self.bias is not None:
+                self.bias_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def _conv_forward(self, input, weight):
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv2d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                weight,
+                self.get_bias(),
+                self.stride,
+                _pair(0),  # noqa: F821
+                self.dilation,
+                self.groups,
+            )
+        return F.conv2d(
+            input,
+            weight,
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+    def forward(self, input: Tensor) -> Tensor:
+        return self._conv_forward(input, self.get_weight())
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+
+    Args:
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           min_abs:  the minimum average-absolute-value per channel, which
+              we allow, before we start to modify the derivatives to prevent
+              this.
+           max_abs:  the maximum average-absolute-value per channel, which
+               we allow, before we start to modify the derivatives to prevent
+               this.
+    """
+
+    def __init__(
+        self,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.01,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+
+    def forward(self, x: Tensor) -> Tensor:
+        return ActivationBalancerFunction.apply(
+            x,
+            self.channel_dim,
+            self.min_positive,
+            self.max_positive,
+            self.max_factor,
+            self.min_abs,
+            self.max_abs,
+        )
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        x = x.detach()
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+        ctx.save_for_backward(s, y)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        s, y = ctx.saved_tensors
+        return (y * (1 - s) + s) * y_grad
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        return DoubleSwishFunction.apply(x)
+
+
+class ScaledEmbedding(nn.Module):
+    r"""A simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
+                                    is renormalized to have norm :attr:`max_norm`.
+        norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+    """
+    __constants__ = [
+        "num_embeddings",
+        "embedding_dim",
+        "padding_idx",
+        "scale_grad_by_freq",
+        "sparse",
+    ]
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        scale_grad_by_freq: bool = False,
+        sparse: bool = False,
+    ) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert (
+                    padding_idx < self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+            elif padding_idx < 0:
+                assert (
+                    padding_idx >= -self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale = nn.Parameter(torch.zeros(()))  # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        std = 0.01
+        nn.init.normal_(self.weight, std=std)
+        nn.init.constant_(self.scale, torch.tensor(1.0 / std).log())
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        scale = self.scale.exp()
+        if input.numel() < self.num_embeddings:
+            return (
+                F.embedding(
+                    input,
+                    self.weight,
+                    self.padding_idx,
+                    None,
+                    2.0,  # None, 2.0 relate to normalization
+                    self.scale_grad_by_freq,
+                    self.sparse,
+                )
+                * scale
+            )
+        else:
+            return F.embedding(
+                input,
+                self.weight * scale,
+                self.padding_idx,
+                None,
+                2.0,  # None, 2.0 relates to normalization
+                self.scale_grad_by_freq,
+                self.sparse,
+            )
+
+    def extra_repr(self) -> str:
+        s = "{num_embeddings}, {embedding_dim}, scale={scale}"
+        if self.padding_idx is not None:
+            s += ", padding_idx={padding_idx}"
+        if self.scale_grad_by_freq is not False:
+            s += ", scale_grad_by_freq={scale_grad_by_freq}"
+        if self.sparse is not False:
+            s += ", sparse=True"
+        return s.format(**self.__dict__)
+
+
+def _test_activation_balancer_sign():
+    channel_dim = 0
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    channel_dim = 0
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x**2).mean().sqrt()
+    y_rms = (y**2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 0.5
+    x.requires_grad = True
+    m = DoubleSwish()
+    torch.autograd.gradcheck(m, x)
+
+
+if __name__ == "__main__":
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()
diff --git a/egs/librispeech/ASR/pruned2_knowledge/train.py b/egs/librispeech/ASR/pruned2_knowledge/train.py
new file mode 100755
index 000000000..a4899f7bd
--- /dev/null
+++ b/egs/librispeech/ASR/pruned2_knowledge/train.py
@@ -0,0 +1,973 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned2_knowledge/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned2_knowledge/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned2_knowledge/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --use_fp16 1 \
+  --exp-dir pruned2_knowledge/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned2_knowledge/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 18,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/asr_datamodule.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/asr_datamodule.py
new file mode 120000
index 000000000..b4e5427e0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/beam_search.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/beam_search.py
new file mode 120000
index 000000000..227d2247c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decode.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decode.py
new file mode 100755
index 000000000..072d49d9c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decode.py
@@ -0,0 +1,632 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_stateless_emformer_rnnt/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += params.left_context_length
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.left_context_length),
+        value=LOG_EPS,
+    )
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decoder.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decoder.py
new file mode 120000
index 000000000..0d5f10dc0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/emformer.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/emformer.py
new file mode 100644
index 000000000..008f40fb1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/emformer.py
@@ -0,0 +1,311 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+try:
+    from torchaudio.models import Emformer as _Emformer
+except ImportError:
+    import torchaudio
+
+    print(
+        "Please install torchaudio >= 0.11.0. "
+        f"Current version: {torchaudio.__version__}"
+    )
+    raise
+
+
+def unstack_states(
+    states: List[List[torch.Tensor]],
+) -> List[List[List[torch.Tensor]]]:
+    """Unstack the emformer state corresponding to a batch of utterances
+    into a list of states, were the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Args:
+      states:
+        A list-of-list of tensors. ``len(states)`` equals to number of
+        layers in the emformer. ``states[i]]`` contains the states for
+        the i-th layer. ``states[i][k]`` is either a 3-D tensor of shape
+        ``(T, N, C)`` or a 2-D tensor of shape ``(C, N)``
+    """
+    batch_size = states[0][0].size(1)
+    num_layers = len(states)
+
+    ans = [None] * batch_size
+    for i in range(batch_size):
+        ans[i] = [[] for _ in range(num_layers)]
+
+    for li, layer in enumerate(states):
+        for s in layer:
+            s_list = s.unbind(dim=1)
+            # We will use stack(dim=1) later in stack_states()
+            for bi, b in enumerate(ans):
+                b[li].append(s_list[bi])
+    return ans
+
+
+def stack_states(
+    state_list: List[List[List[torch.Tensor]]],
+) -> List[List[torch.Tensor]]:
+    """Stack list of emformer states that correspond to separate utterances
+    into a single emformer state so that it can be used as an input for
+    emformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the emformer model for a single utterance.
+    Returns:
+      Return a new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+    ans = []
+    for layer in state_list[0]:
+        # layer is a list of tensors
+        if batch_size > 1:
+            ans.append([[s] for s in layer])
+            # Note: We will stack ans[layer][s][] later to get ans[layer][s]
+        else:
+            ans.append([s.unsqueeze(1) for s in layer])
+
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states):
+            for si, s in enumerate(layer):
+                ans[li][si].append(s)
+                if b == batch_size - 1:
+                    ans[li][si] = torch.stack(ans[li][si], dim=1)
+                    # We will use unbind(dim=1) later in unstack_states()
+    return ans
+
+
+class Emformer(EncoderInterface):
+    """This is just a simple wrapper around torchaudio.models.Emformer.
+    We may replace it with our own implementation some time later.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        num_encoder_layers: int,
+        segment_length: int,
+        left_context_length: int,
+        right_context_length: int,
+        max_memory_size: int = 0,
+        dropout: float = 0.1,
+        subsampling_factor: int = 4,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          segment_length:
+            Number of frames per segment before subsampling.
+          left_context_length:
+            Number of frames in the left context before subsampling.
+          right_context_length:
+            Number of frames in the right context before subsampling.
+          max_memory_size:
+            TODO.
+          dropout:
+            Dropout in encoder.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.segment_length = segment_length  # before subsampling
+        self.right_context_length = right_context_length
+
+        assert right_context_length % subsampling_factor == 0
+        assert segment_length % subsampling_factor == 0
+        assert left_context_length % subsampling_factor == 0
+
+        left_context_length = left_context_length // subsampling_factor
+        right_context_length = right_context_length // subsampling_factor
+        segment_length = segment_length // subsampling_factor
+
+        self.model = _Emformer(
+            input_dim=d_model,
+            num_heads=nhead,
+            ffn_dim=dim_feedforward,
+            num_layers=num_encoder_layers,
+            segment_length=segment_length,
+            dropout=dropout,
+            activation="relu",
+            left_context_length=left_context_length,
+            right_context_length=right_context_length,
+            max_memory_size=max_memory_size,
+            weight_init_scale_strategy="depthwise",
+            tanh_on_mem=False,
+            negative_inf=-1e8,
+        )
+
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+        self.log_eps = math.log(1e-10)
+
+        self._has_init_state = False
+        self._init_state = torch.jit.Attribute([], List[List[torch.Tensor]])
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            Input features of shape (N, T, C).
+          x_lens:
+            A int32 tensor of shape (N,) containing valid frames in `x` before
+            padding. We have `x.size(1) == x_lens.max()`
+        Returns:
+          Return a tuple containing two tensors:
+
+            - encoder_out, a tensor of shape (N, T', C)
+            - encoder_out_lens, a int32 tensor of shape (N,) containing the
+              valid frames in `encoder_out` before padding
+        """
+        x = nn.functional.pad(
+            x,
+            # (left, right, top, bottom)
+            # left/right are for the channel dimension, i.e., axis 2
+            # top/bottom are for the time dimension, i.e., axis 1
+            (0, 0, 0, self.right_context_length),
+            value=self.log_eps,
+        )  # (N, T, C) -> (N, T+right_context_length, C)
+
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+
+        emformer_out, emformer_out_lens = self.model(x, x_lens)
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[List[torch.Tensor]]] = None,
+    ):
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C). Note: x also contains right
+            context frames.
+          x_lens:
+            A 2-D tensor of shap containing the number of valid frames for each
+            element in `x` before padding. Note: It also counts right context
+            frames.
+          states:
+            Internal states of the model.
+        Returns:
+          Return a tuple containing 3 tensors:
+            - encoder_out, a 3-D tensor of shape (N, T, C)
+            - encoder_out_lens: a 1-D tensor of shape (N,)
+            - next_state, internal model states for the next invocation
+        """
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+
+        emformer_out, emformer_out_lens, states = self.model.infer(x, x_lens, states)
+
+        if x.size(1) != (self.model.segment_length + self.model.right_context_length):
+            raise ValueError(
+                "Incorrect input shape."
+                f"{x.size(1)} vs {self.model.segment_length} + "
+                f"{self.model.right_context_length}"
+            )
+
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens, states
+
+    @torch.jit.export
+    def get_init_state(self, device: torch.device) -> List[List[torch.Tensor]]:
+        """Return the initial state of each layer.
+
+        Returns:
+          Return the initial state of each layer. NOTE: the returned
+          tensors are on the given device. `len(ans) == num_emformer_layers`.
+        """
+        if self._has_init_state:
+            # Note(fangjun): It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        batch_size = 1
+
+        ans: List[List[torch.Tensor]] = []
+        for layer in self.model.emformer_layers:
+            s = layer._init_state(batch_size=batch_size, device=device)
+            ans.append(s)
+
+        self._has_init_state = True
+        self._init_state = ans
+
+        return ans
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/encoder_interface.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/encoder_interface.py
new file mode 120000
index 000000000..a478f2351
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/export.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/export.py
new file mode 100755
index 000000000..ec2c9d580
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/export.py
@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./prunted_stateless_emformer_rnnt/export.py \
+  --exp-dir ./prunted_stateless_emformer_rnnt/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `prunted_stateless_emformer_rnnt/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./prunted_stateless_emformer_rnnt/decode.py \
+        --exp-dir ./prunted_stateless_emformer_rnnt/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="prunted_stateless_emformer_rnnt/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+    for p in model.parameters():
+        p.requires_grad_(False)
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/joiner.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/joiner.py
new file mode 120000
index 000000000..81ad47c55
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/model.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/model.py
new file mode 100644
index 000000000..ed6848879
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/model.py
@@ -0,0 +1,167 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, C]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+            lm=decoder_out,
+            am=encoder_out,
+            symbols=y_padded,
+            termination_symbol=blank_id,
+            lm_only_scale=lm_scale,
+            am_only_scale=am_scale,
+            boundary=boundary,
+            reduction="sum",
+            return_grad=True,
+        )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, C]
+        # lm_pruned : [B, T, prune_range, C]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=encoder_out, lm=decoder_out, ranges=ranges
+        )
+
+        # logits : [B, T, prune_range, C]
+        logits = self.joiner(am_pruned, lm_pruned)
+
+        pruned_loss = k2.rnnt_loss_pruned(
+            logits=logits,
+            symbols=y_padded,
+            ranges=ranges,
+            termination_symbol=blank_id,
+            boundary=boundary,
+            reduction="sum",
+        )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/noam.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/noam.py
new file mode 100644
index 000000000..e46bf35fb
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/noam.py
@@ -0,0 +1,104 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/subsampling.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_emformer.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_emformer.py
new file mode 100755
index 000000000..aef506e81
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_emformer.py
@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_stateless_emformer_rnnt/test_emformer.py
+"""
+
+import torch
+from emformer import Emformer, stack_states, unstack_states
+
+
+def test_emformer():
+    N = 3
+    T = 300
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, T, C)
+    x_lens = torch.randint(100, T, (N,))
+    x_lens[0] = T
+
+    y, y_lens = encoder(x, x_lens)
+
+    y_lens = (((x_lens - 1) >> 1) - 1) >> 1
+    assert x.size(0) == x.size(0)
+    assert y.size(1) == max(y_lens)
+    assert y.size(2) == output_dim
+
+    num_param = sum([p.numel() for p in encoder.parameters()])
+    print(f"Number of encoder parameters: {num_param}")
+
+
+def test_emformer_streaming_forward():
+    N = 3
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, 23, C)
+    x_lens = torch.full((N,), 23)
+    y, y_lens, states = encoder.streaming_forward(x=x, x_lens=x_lens)
+
+    state_list = unstack_states(states)
+    states2 = stack_states(state_list)
+
+    for ss, ss2 in zip(states, states2):
+        for s, s2 in zip(ss, ss2):
+            assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+
+def test_emformer_init_state():
+    num_encoder_layers = 20
+    d_model = 512
+    encoder = Emformer(
+        num_features=80,
+        output_dim=500,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=num_encoder_layers,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+    init_state = encoder.get_init_state()
+    assert len(init_state) == num_encoder_layers
+    layer0_state = init_state[0]
+    assert len(layer0_state) == 4
+
+    assert layer0_state[0].shape == (
+        0,  # max_memory_size
+        1,  # batch_size
+        d_model,  # input_dim
+    )
+
+    assert layer0_state[1].shape == (
+        encoder.model.left_context_length,
+        1,  # batch_size
+        d_model,  # input_dim
+    )
+    assert layer0_state[2].shape == layer0_state[1].shape
+    assert layer0_state[3].shape == (
+        1,  # always 1
+        1,  # batch_size
+    )
+
+
+@torch.no_grad()
+def main():
+    test_emformer()
+    test_emformer_streaming_forward()
+    test_emformer_init_state()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220329)
+    main()
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_model.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_model.py
new file mode 100755
index 000000000..573817b85
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_model.py
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_stateless_emformer_rnnt/test_model.py
+"""
+
+import torch
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.attention_dim = 512
+    params.nhead = 8
+    params.dim_feedforward = 2048
+    params.num_encoder_layers = 18
+    params.left_context_length = 128
+    params.segment_length = 8
+    params.right_context_length = 4
+    params.memory_size = 0
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/train.py b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/train.py
new file mode 100755
index 000000000..2b872f1d5
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/train.py
@@ -0,0 +1,1048 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_stateless_emformer_rnnt/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_stateless_emformer_rnnt/exp \
+  --full-libri 1 \
+  --max-duration 300
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from emformer import Emformer
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from noam import Noam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--attention-dim",
+        type=int,
+        default=512,
+        help="Attention dim for the Emformer",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads for the Emformer",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feed-forward dimension for the Emformer",
+    )
+
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=18,
+        help="Number of encoder layers for the Emformer",
+    )
+
+    parser.add_argument(
+        "--left-context-length",
+        type=int,
+        default=128,
+        help="Number of frames for the left context in the Emformer",
+    )
+
+    parser.add_argument(
+        "--segment-length",
+        type=int,
+        default=8,
+        help="Number of frames for each segment in the Emformer",
+    )
+
+    parser.add_argument(
+        "--right-context-length",
+        type=int,
+        default=4,
+        help="Number of frames for right context in the Emformer",
+    )
+
+    parser.add_argument(
+        "--memory-size",
+        type=int,
+        default=0,
+        help="Number of entries in the memory for the Emformer",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_stateless_emformer_rnnt/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for Emformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 20000
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Emformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        left_context_length=params.left_context_length,
+        segment_length=params.segment_length,
+        right_context_length=params.right_context_length,
+        max_memory_size=params.memory_size,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute pruned RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Emformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = optimizer.rate()
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 20000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./emformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless/asr_datamodule.py
new file mode 120000
index 000000000..07f39b451
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
new file mode 100644
index 000000000..830b37cfb
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
@@ -0,0 +1,1173 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import k2
+import torch
+from model import Transducer
+
+from icefall.decode import Nbest, one_best_decoding
+from icefall.utils import get_texts
+
+
+def fast_beam_search_one_best(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+    return hyps
+
+
+def fast_beam_search_nbest_LG(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # The following code is modified from nbest.intersect()
+    word_fsa = k2.invert(nbest.fsa)
+    if hasattr(lattice, "aux_labels"):
+        # delete token IDs as it is not needed
+        del word_fsa.aux_labels
+    word_fsa.scores.zero_()
+    word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa)
+    path_to_utt_map = nbest.shape.row_ids(1)
+
+    if hasattr(lattice, "aux_labels"):
+        # lattice has token IDs as labels and word IDs as aux_labels.
+        # inv_lattice has word IDs as labels and token IDs as aux_labels
+        inv_lattice = k2.invert(lattice)
+        inv_lattice = k2.arc_sort(inv_lattice)
+    else:
+        inv_lattice = k2.arc_sort(lattice)
+
+    if inv_lattice.shape[0] == 1:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=torch.zeros_like(path_to_utt_map),
+            sorted_match_a=True,
+        )
+    else:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=path_to_utt_map,
+            sorted_match_a=True,
+        )
+
+    # path_lattice has word IDs as labels and token IDs as aux_labels
+    path_lattice = k2.top_sort(k2.connect(path_lattice))
+    tot_scores = path_lattice.get_tot_scores(
+        use_double_scores=use_double_scores,
+        log_semiring=True,  # Note: we always use True
+    )
+    # See https://github.com/k2-fsa/icefall/pull/420 for why
+    # we always use log_semiring=True
+
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+    best_hyp_indexes = ragged_tot_scores.argmax()
+    best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes)
+
+    hyps = get_texts(best_path)
+
+    return hyps
+
+
+def fast_beam_search_nbest(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # at this point, nbest.fsa.scores are all zeros.
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa.scores contains acoustic scores
+
+    max_indexes = nbest.tot_scores().argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    hyps = get_texts(best_path)
+
+    return hyps
+
+
+def fast_beam_search_nbest_oracle(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    ref_texts: List[List[int]],
+    use_double_scores: bool = True,
+    nbest_scale: float = 0.5,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    we select `num_paths` linear paths from the lattice. The path
+    that has the minimum edit distance with the given reference transcript
+    is used as the output.
+
+    This is the best result we can achieve for any nbest based rescoring
+    methods.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      ref_texts:
+        A list-of-list of integers containing the reference transcripts.
+        If the decoding_graph is a trivial_graph, the integer ID is the
+        BPE token ID.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    hyps = nbest.build_levenshtein_graphs()
+    refs = k2.levenshtein_graph(ref_texts, device=hyps.device)
+
+    levenshtein_alignment = k2.levenshtein_alignment(
+        refs=refs,
+        hyps=hyps,
+        hyp_to_ref_map=nbest.shape.row_ids(1),
+        sorted_match_ref=True,
+    )
+
+    tot_scores = levenshtein_alignment.get_tot_scores(
+        use_double_scores=False, log_semiring=False
+    )
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+
+    max_indexes = ragged_tot_scores.argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    hyps = get_texts(best_path)
+    return hyps
+
+
+def fast_beam_search(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> k2.Fsa:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    Returns:
+      Return an FsaVec with axes [utt][state][arc] containing the decoded
+      lattice. Note: When the input graph is a TrivialGraph, the returned
+      lattice is actually an acceptor.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(k2.RnntDecodingStream(decoding_graph))
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, encoder_out_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+            # in some old versions of pytorch, the type of index requires
+            # to be LongTensor. In the newest version of pytorch, the type
+            # of index can be IntTensor or LongTensor. For supporting the
+            # old and new versions of pytorch, we set the type of index
+            # to LongTensor.
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2), decoder_out.unsqueeze(1)
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+    decoding_streams.terminate_and_flush_to_streams()
+    lattice = decoding_streams.format_output(encoder_out_lens.tolist())
+
+    return lattice
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """Greedy search for a single utterance.
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    unk_id = getattr(model, "unk_id", blank_id)
+
+    device = next(model.parameters()).device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size, device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        logits = model.joiner(current_encoder_out, decoder_out.unsqueeze(1))
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y not in (blank_id, unk_id):
+            hyp.append(y)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+def greedy_search_batch(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+    Returns:
+      Return a list-of-list of token IDs containing the decoded results.
+      len(ans) equals to encoder_out.size(0).
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = next(model.parameters()).device
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    # decoder_out: (N, 1, decoder_out_dim)
+
+    encoder_out = packed_encoder_out.data
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(current_encoder_out, decoder_out.unsqueeze(1))
+        # logits'shape (batch_size, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (batch_size, vocab_size)
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v not in (blank_id, unk_id):
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis, use_max: bool = False) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+          use_max:
+            True to select the hypothesis with the larger log_prob in case there
+            already exists a hypothesis whose `ys` equals to `hyp.ys`.
+            False to use log_add.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            if use_max:
+                old_hyp.log_prob = max(old_hyp.log_prob, hyp.log_prob)
+            else:
+                torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
+def modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: int = 4,
+    use_max: bool = False,
+) -> List[List[int]]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      use_max:
+        True to use max operation to select the hypothesis with the largest
+        log_prob when there are duplicate hypotheses; False to use log-add.
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+            ),
+            use_max=use_max,
+        )
+
+    encoder_out = packed_encoder_out.data
+
+    offset = 0
+    finalized_B = []
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        # decoder_output is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+def _deprecated_modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+    use_max: bool = False,
+) -> List[int]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    It decodes only one utterance at a time. We keep it only for reference.
+    The function :func:`modified_beam_search` should be preferred as it
+    supports batch decoding.
+
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+      use_max:
+        True to use max operation to select the hypothesis with the largest
+        log_prob when there are duplicate hypotheses; False to use log-add.
+    Returns:
+      Return the decoded result.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    device = next(model.parameters()).device
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        ),
+        use_max=use_max,
+    )
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t + 1, :].unsqueeze(2)
+        # current_encoder_out is of shape (1, 1, 1, encoder_out_dim)
+        # fmt: on
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+            dtype=torch.int64,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        # decoder_output is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        current_encoder_out = current_encoder_out.expand(
+            decoder_out.size(0), 1, 1, -1
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        # now logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            topk_hyp_indexes = topk_hyp_indexes.tolist()
+            topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
+            new_ys = hyp.ys[:]
+            new_token = topk_token_indexes[i]
+            if new_token not in (blank_id, unk_id):
+                new_ys.append(new_token)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+            B.add(new_hyp, use_max=use_max)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    return ys
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+    use_max: bool = False,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+      use_max:
+        True to use max operation to select the hypothesis with the largest
+        log_prob when there are duplicate hypotheses; False to use log-add.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    device = next(model.parameters()).device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size,
+        device=device,
+        dtype=torch.int64,
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0), use_max=use_max)
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        # TODO(fangjun): Implement prefix search to update the `log_prob`
+        # of hypotheses in A
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            cached_key = y_star.key
+
+            if cached_key not in decoder_cache:
+                decoder_input = torch.tensor(
+                    [y_star.ys[-context_size:]],
+                    device=device,
+                    dtype=torch.int64,
+                ).reshape(1, context_size)
+
+                decoder_out = model.decoder(decoder_input, need_pad=False)
+                decoder_cache[cached_key] = decoder_out
+            else:
+                decoder_out = decoder_cache[cached_key]
+
+            cached_key += f"-t-{t}"
+            if cached_key not in joint_cache:
+                logits = model.joiner(current_encoder_out, decoder_out.unsqueeze(1))
+
+                # TODO(fangjun): Scale the blank posterior
+
+                log_prob = logits.log_softmax(dim=-1)
+                # log_prob is (1, 1, 1, vocab_size)
+                log_prob = log_prob.squeeze()
+                # Now log_prob is (vocab_size,)
+                joint_cache[cached_key] = log_prob
+            else:
+                log_prob = joint_cache[cached_key]
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(
+                Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob),
+                use_max=use_max,
+            )
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for i, v in zip(indices.tolist(), values.tolist()):
+                if i in (blank_id, unk_id):
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                A.add(
+                    Hypothesis(ys=new_ys, log_prob=new_log_prob),
+                    use_max=use_max,
+                )
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless/conformer.py
new file mode 120000
index 000000000..70a7ddf11
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/conformer.py
@@ -0,0 +1 @@
+../transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless/decode.py
new file mode 100755
index 000000000..3c4500087
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/decode.py
@@ -0,0 +1,790 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(6) decode in streaming mode (take greedy search as an example)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --decode-chunk-size 16 \
+    --left-context 64 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling). Set -1 to use full attention.",
+    )
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        if params.decode_chunk_size > 0:
+            # except the case of using full attention
+            feature_lens += params.left_context
+            feature = torch.nn.functional.pad(
+                feature,
+                pad=(0, 0, 0, params.left_context),
+                value=LOG_EPS,
+            )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless/decode_stream.py
new file mode 100644
index 000000000..e522943c0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/decode_stream.py
@@ -0,0 +1,146 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class DecodeStream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        initial_states: List[torch.Tensor],
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+    ) -> None:
+        """
+        Args:
+          initial_states:
+            Initial decode states of the model, e.g. the return value of
+            `get_init_state` in conformer.py
+          decoding_graph:
+            Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+            Used only when decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+        """
+        if params.decoding_method == "fast_beam_search":
+            assert decoding_graph is not None
+            assert device == decoding_graph.device
+
+        self.params = params
+        self.cut_id = cut_id
+        self.LOG_EPS = math.log(1e-10)
+
+        self.states = initial_states
+
+        # It contains a 2-D tensors representing the feature frames.
+        self.features: torch.Tensor = None
+
+        self.num_frames: int = 0
+        # how many frames have been processed. (before subsampling).
+        # we only modify this value in `func:get_feature_frames`.
+        self.num_processed_frames: int = 0
+
+        self._done: bool = False
+
+        # The transcript of current utterance.
+        self.ground_truth: str = ""
+
+        # The decoding result (partial or final) of current utterance.
+        self.hyp: List = []
+
+        # how many frames have been processed, after subsampling (i.e. a
+        # cumulative sum of the second return value of
+        # encoder.streaming_forward
+        self.done_frames: int = 0
+
+        self.pad_length = (params.right_context + 2) * params.subsampling_factor + 3
+
+        if params.decoding_method == "greedy_search":
+            self.hyp = [params.blank_id] * params.context_size
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[params.blank_id] * params.context_size,
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    @property
+    def done(self) -> bool:
+        """Return True if all the features are processed."""
+        return self._done
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    def set_features(
+        self,
+        features: torch.Tensor,
+    ) -> None:
+        """Set features tensor of current utterance."""
+        assert features.dim() == 2, features.dim()
+        self.features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, self.pad_length),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+        self.num_frames = self.features.size(0)
+
+    def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]:
+        """Consume chunk_size frames of features"""
+        chunk_length = chunk_size + self.pad_length
+
+        ret_length = min(self.num_frames - self.num_processed_frames, chunk_length)
+
+        ret_features = self.features[
+            self.num_processed_frames : self.num_processed_frames + ret_length  # noqa
+        ]
+
+        self.num_processed_frames += chunk_size
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_features, ret_length
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.params.decoding_method == "greedy_search":
+            return self.hyp[self.params.context_size :]  # noqa
+        elif self.params.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.params.context_size :]  # noqa
+        else:
+            assert self.params.decoding_method == "fast_beam_search"
+            return self.hyp
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless/decoder.py
new file mode 100644
index 000000000..03847b449
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/decoder.py
@@ -0,0 +1,106 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        unk_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          unk_id:
+            The ID of the unk symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+        )
+        self.blank_id = blank_id
+        self.unk_id = unk_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+        else:
+            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
+            # when inference with torch.jit.script and context_size == 1
+            self.conv = nn.Identity()
+        self.output_linear = nn.Linear(embedding_dim, vocab_size)
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U) with blank prepended.
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, embedding_dim).
+        """
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = self.output_linear(F.relu(embedding_out))
+        return embedding_out
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless/export-onnx.py
new file mode 100755
index 000000000..0a2132e56
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/export-onnx.py
@@ -0,0 +1,528 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import setup_logger
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer"""
+
+    def __init__(self, encoder: Conformer):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+        """
+        super().__init__()
+        self.encoder = encoder
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder"""
+
+    def __init__(self, decoder: Decoder):
+        super().__init__()
+        self.decoder = decoder
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        output = decoder_output.squeeze(1)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, inner_linear: nn.Linear, output_linear: nn.Linear):
+        super().__init__()
+        self.inner_linear = inner_linear
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.inner_linear(torch.tanh(logit))
+        output = self.output_linear(nn.functional.relu(logit))
+
+        return output
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless3",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.inner_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    encoder = OnnxEncoder(encoder=model.encoder)
+
+    decoder = OnnxDecoder(decoder=model.decoder)
+
+    joiner = OnnxJoiner(
+        inner_linear=model.joiner.inner_linear, output_linear=model.joiner.output_linear
+    )
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/export.py b/egs/librispeech/ASR/pruned_transducer_stateless/export.py
new file mode 100755
index 000000000..4b20e3a2b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/export.py
@@ -0,0 +1,197 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless/export.py \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless/decode.py \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size, <blk> is
+    # defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless/joiner.py
new file mode 100644
index 000000000..c9522df8a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/joiner.py
@@ -0,0 +1,52 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, inner_dim: int, output_dim: int):
+        super().__init__()
+
+        self.inner_linear = nn.Linear(input_dim, inner_dim)
+        self.output_linear = nn.Linear(inner_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim
+        assert encoder_out.ndim in (2, 4)
+        assert encoder_out.shape == decoder_out.shape
+
+        logit = encoder_out + decoder_out
+
+        logit = self.inner_linear(torch.tanh(logit))
+
+        output = self.output_linear(F.relu(logit))
+
+        return output
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/model.py b/egs/librispeech/ASR/pruned_transducer_stateless/model.py
new file mode 100644
index 000000000..2cca7fa27
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/model.py
@@ -0,0 +1,175 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        reduction: str = "sum",
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, C]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+            lm=decoder_out,
+            am=encoder_out,
+            symbols=y_padded,
+            termination_symbol=blank_id,
+            lm_only_scale=lm_scale,
+            am_only_scale=am_scale,
+            boundary=boundary,
+            reduction=reduction,
+            return_grad=True,
+        )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, C]
+        # lm_pruned : [B, T, prune_range, C]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=encoder_out, lm=decoder_out, ranges=ranges
+        )
+
+        # logits : [B, T, prune_range, C]
+        logits = self.joiner(am_pruned, lm_pruned)
+
+        pruned_loss = k2.rnnt_loss_pruned(
+            logits=logits,
+            symbols=y_padded,
+            ranges=ranges,
+            termination_symbol=blank_id,
+            boundary=boundary,
+            reduction=reduction,
+        )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_check.py
new file mode 120000
index 000000000..66d63b807
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_check.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_check.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py
new file mode 100755
index 000000000..8134d43f8
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py
@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+3. Run this file
+
+./pruned_transducer_stateless/onnx_decode.py \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --max-duration 600 \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+"""
+
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+
+from onnx_pretrained import greedy_search, OnnxModel
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, sp: spm.SentencePieceProcessor, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    hyps = [sp.decode(h).split() for h in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, sp=sp, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    blank_id = sp.piece_to_id("<blk>")
+    assert blank_id == 0, blank_id
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(dl=test_dl, model=model, sp=sp)
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_pretrained.py
new file mode 120000
index 000000000..7607623c8
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless/pretrained.py
new file mode 100755
index 000000000..02f9f1b03
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/pretrained.py
@@ -0,0 +1,388 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless/exp/pretrained.pt is generated by
+./pruned_transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    if params.simulate_streaming:
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/profile.py b/egs/librispeech/ASR/pruned_transducer_stateless/profile.py
new file mode 100755
index 000000000..09e4a7af4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/profile.py
@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage: ./pruned_transducer_stateless/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from icefall.profiler import get_model_profile
+from train import get_encoder_model, add_model_arguments, get_params
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = get_encoder_model(params)
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(model=model, args=(feature, feature_lens))
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless/streaming_beam_search.py
new file mode 100644
index 000000000..9e09200a1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/streaming_beam_search.py
@@ -0,0 +1,276 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from decode_stream import DecodeStream
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    T = encoder_out.size(1)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (N, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+    num_active_paths: int = 4,
+) -> None:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      num_active_paths:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(num_active_paths)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    streams: List[DecodeStream],
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first generated by Fsa-based beam search, then we get the
+    recognition by applying shortest path on the lattice.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      streams:
+        A list of stream objects.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+    best_path = one_best_decoding(lattice)
+    hyp_tokens = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyp_tokens[i]
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless/streaming_decode.py
new file mode 100755
index 000000000..f4b01fd06
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/streaming_decode.py
@@ -0,0 +1,576 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless/streaming_decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --decode-chunk-size 8 \
+        --left-context 32 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --decoding_method greedy_search \
+        --num-decode-streams 1000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num-active-paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        decode_stream.set_features(fbank(samples.to(device)))
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        # sort results so we can easily compare the difference between two
+        # recognition results
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/subsampling.py b/egs/librispeech/ASR/pruned_transducer_stateless/subsampling.py
new file mode 120000
index 000000000..73068da26
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/subsampling.py
@@ -0,0 +1 @@
+../transducer/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/test_decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless/test_decoder.py
new file mode 100755
index 000000000..36712018d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/test_decoder.py
@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    unk_id = 2
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        unk_id=unk_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, vocab_size)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, vocab_size)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless/test_model.py
new file mode 100755
index 000000000..fc82d8c69
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/test_model.py
@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.dynamic_chunk_training = False
+    params.short_chunk_size = 25
+    params.num_left_chunks = 4
+    params.causal_convolution = False
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def test_model_streaming():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.dynamic_chunk_training = True
+    params.short_chunk_size = 25
+    params.num_left_chunks = 4
+    params.causal_convolution = True
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    test_model()
+    test_model_streaming()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/train.py b/egs/librispeech/ASR/pruned_transducer_stateless/train.py
new file mode 100755
index 000000000..cf4032027
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/train.py
@@ -0,0 +1,1042 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# train a streaming model
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 25 \
+  --num-left-chunks 4 \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    measure_gradient_norms,
+    measure_weight_norms,
+    optim_step_and_measure_param_change,
+    setup_logger,
+    str2bool,
+)
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            "log_diagnostics": False,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 30000
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        sampler=sampler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            reduction="none",
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    def maybe_log_gradients(tag: str):
+        if (
+            params.log_diagnostics
+            and tb_writer is not None
+            and params.batch_idx_train % (params.log_interval * 5) == 0
+        ):
+            tb_writer.add_scalars(
+                tag,
+                measure_gradient_norms(model, norm="l2"),
+                global_step=params.batch_idx_train,
+            )
+
+    def maybe_log_weights(tag: str):
+        if (
+            params.log_diagnostics
+            and tb_writer is not None
+            and params.batch_idx_train % (params.log_interval * 5) == 0
+        ):
+            tb_writer.add_scalars(
+                tag,
+                measure_weight_norms(model, norm="l2"),
+                global_step=params.batch_idx_train,
+            )
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        loss.backward()
+
+        maybe_log_weights("train/param_norms")
+        maybe_log_gradients("train/grad_norms")
+
+        old_parameters = None
+        if (
+            params.log_diagnostics
+            and tb_writer is not None
+            and params.batch_idx_train % (params.log_interval * 5) == 0
+        ):
+            old_parameters = {
+                n: p.detach().clone() for n, p in model.named_parameters()
+            }
+
+        optimizer.step()
+
+        if old_parameters is not None:
+            deltas = optim_step_and_measure_param_change(model, old_parameters)
+            tb_writer.add_scalars(
+                "train/relative_param_change_per_minibatch",
+                deltas,
+                global_step=params.batch_idx_train,
+            )
+
+        optimizer.zero_grad()
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                sampler=train_dl.sampler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 30000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.encoder_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sampler=train_dl.sampler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/transformer.py b/egs/librispeech/ASR/pruned_transducer_stateless/transformer.py
new file mode 120000
index 000000000..e43f520f9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/transformer.py
@@ -0,0 +1 @@
+../transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 120000
index 000000000..07f39b451
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 100644
index 000000000..7fcd242fc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1,2942 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Tuple, Union
+
+import k2
+import sentencepiece as spm
+import torch
+from torch import nn
+
+from icefall import ContextGraph, ContextState, NgramLm, NgramLmStateCost
+from icefall.decode import Nbest, one_best_decoding
+from icefall.lm_wrapper import LmScorer
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.transformer_lm.model import TransformerLM
+from icefall.utils import (
+    DecodingResults,
+    add_eos,
+    add_sos,
+    get_texts,
+    get_texts_with_timestamp,
+)
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    temperature: float = 1.0,
+    ilme_scale: float = 0.0,
+    blank_penalty: float = 0.0,
+    return_timestamps: bool = False,
+    allow_partial: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using fast beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        temperature=temperature,
+        ilme_scale=ilme_scale,
+        allow_partial=allow_partial,
+        blank_penalty=blank_penalty,
+    )
+
+    best_path = one_best_decoding(lattice)
+
+    if not return_timestamps:
+        return get_texts(best_path)
+    else:
+        return get_texts_with_timestamp(best_path)
+
+
+def fast_beam_search_nbest_LG(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+    temperature: float = 1.0,
+    blank_penalty: float = 0.0,
+    ilme_scale: float = 0.0,
+    return_timestamps: bool = False,
+    allow_partial: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        temperature=temperature,
+        allow_partial=allow_partial,
+        blank_penalty=blank_penalty,
+        ilme_scale=ilme_scale,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # The following code is modified from nbest.intersect()
+    word_fsa = k2.invert(nbest.fsa)
+    if hasattr(lattice, "aux_labels"):
+        # delete token IDs as it is not needed
+        del word_fsa.aux_labels
+    word_fsa.scores.zero_()
+    word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa)
+    path_to_utt_map = nbest.shape.row_ids(1)
+
+    if hasattr(lattice, "aux_labels"):
+        # lattice has token IDs as labels and word IDs as aux_labels.
+        # inv_lattice has word IDs as labels and token IDs as aux_labels
+        inv_lattice = k2.invert(lattice)
+        inv_lattice = k2.arc_sort(inv_lattice)
+    else:
+        inv_lattice = k2.arc_sort(lattice)
+
+    if inv_lattice.shape[0] == 1:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=torch.zeros_like(path_to_utt_map),
+            sorted_match_a=True,
+        )
+    else:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=path_to_utt_map,
+            sorted_match_a=True,
+        )
+
+    # path_lattice has word IDs as labels and token IDs as aux_labels
+    path_lattice = k2.top_sort(k2.connect(path_lattice))
+    tot_scores = path_lattice.get_tot_scores(
+        use_double_scores=use_double_scores,
+        log_semiring=True,  # Note: we always use True
+    )
+    # See https://github.com/k2-fsa/icefall/pull/420 for why
+    # we always use log_semiring=True
+
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+    best_hyp_indexes = ragged_tot_scores.argmax()
+    best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes)
+
+    if not return_timestamps:
+        return get_texts(best_path)
+    else:
+        return get_texts_with_timestamp(best_path)
+
+
+def fast_beam_search_nbest(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+    temperature: float = 1.0,
+    blank_penalty: float = 0.0,
+    return_timestamps: bool = False,
+    allow_partial: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        blank_penalty=blank_penalty,
+        temperature=temperature,
+        allow_partial=allow_partial,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # at this point, nbest.fsa.scores are all zeros.
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa.scores contains acoustic scores
+
+    max_indexes = nbest.tot_scores().argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    if not return_timestamps:
+        return get_texts(best_path)
+    else:
+        return get_texts_with_timestamp(best_path)
+
+
+def fast_beam_search_nbest_oracle(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    ref_texts: List[List[int]],
+    use_double_scores: bool = True,
+    nbest_scale: float = 0.5,
+    temperature: float = 1.0,
+    blank_penalty: float = 0.0,
+    return_timestamps: bool = False,
+    allow_partial: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using fast beam search, and then
+    we select `num_paths` linear paths from the lattice. The path
+    that has the minimum edit distance with the given reference transcript
+    is used as the output.
+
+    This is the best result we can achieve for any nbest based rescoring
+    methods.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      ref_texts:
+        A list-of-list of integers containing the reference transcripts.
+        If the decoding_graph is a trivial_graph, the integer ID is the
+        BPE token ID.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        temperature=temperature,
+        allow_partial=allow_partial,
+        blank_penalty=blank_penalty,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    hyps = nbest.build_levenshtein_graphs()
+    refs = k2.levenshtein_graph(ref_texts, device=hyps.device)
+
+    levenshtein_alignment = k2.levenshtein_alignment(
+        refs=refs,
+        hyps=hyps,
+        hyp_to_ref_map=nbest.shape.row_ids(1),
+        sorted_match_ref=True,
+    )
+
+    tot_scores = levenshtein_alignment.get_tot_scores(
+        use_double_scores=False, log_semiring=False
+    )
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+
+    max_indexes = ragged_tot_scores.argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    if not return_timestamps:
+        return get_texts(best_path)
+    else:
+        return get_texts_with_timestamp(best_path)
+
+
+def fast_beam_search(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    temperature: float = 1.0,
+    subtract_ilme: bool = False,
+    ilme_scale: float = 0.1,
+    allow_partial: bool = False,
+    blank_penalty: float = 0.0,
+) -> k2.Fsa:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      temperature:
+        Softmax temperature.
+    Returns:
+      Return an FsaVec with axes [utt][state][arc] containing the decoded
+      lattice. Note: When the input graph is a TrivialGraph, the returned
+      lattice is actually an acceptor.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(k2.RnntDecodingStream(decoding_graph))
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+
+        if blank_penalty != 0:
+            logits[:, 0] -= blank_penalty
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)
+
+        if ilme_scale != 0:
+            ilme_logits = model.joiner(
+                torch.zeros_like(
+                    current_encoder_out, device=current_encoder_out.device
+                ).unsqueeze(2),
+                decoder_out.unsqueeze(1),
+                project_input=False,
+            )
+            ilme_logits = ilme_logits.squeeze(1).squeeze(1)
+            if blank_penalty != 0:
+                ilme_logits[:, 0] -= blank_penalty
+            ilme_log_probs = (ilme_logits / temperature).log_softmax(dim=-1)
+            log_probs -= ilme_scale * ilme_log_probs
+
+        decoding_streams.advance(log_probs)
+    decoding_streams.terminate_and_flush_to_streams()
+    lattice = decoding_streams.format_output(
+        encoder_out_lens.tolist(), allow_partial=allow_partial
+    )
+
+    return lattice
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    max_sym_per_frame: int,
+    blank_penalty: float = 0.0,
+    return_timestamps: bool = False,
+) -> Union[List[int], DecodingResults]:
+    """Greedy search for a single utterance.
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    unk_id = getattr(model, "unk_id", blank_id)
+
+    device = next(model.parameters()).device
+
+    decoder_input = torch.tensor(
+        [-1] * (context_size - 1) + [blank_id], device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # timestamp[i] is the frame index after subsampling
+    # on which hyp[i] is decoded
+    timestamp = []
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits is (1, 1, 1, vocab_size)
+
+        if blank_penalty != 0:
+            logits[:, :, :, 0] -= blank_penalty
+
+        y = logits.argmax().item()
+        if y not in (blank_id, unk_id):
+            hyp.append(y)
+            timestamp.append(t)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    if not return_timestamps:
+        return hyp
+    else:
+        return DecodingResults(
+            hyps=[hyp],
+            timestamps=[timestamp],
+        )
+
+
+def greedy_search_batch(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    blank_penalty: float = 0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = next(model.parameters()).device
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[-1] * (context_size - 1) + [blank_id] for _ in range(N)]
+
+    # timestamp[n][i] is the frame index after subsampling
+    # on which hyp[n][i] is decoded
+    timestamps = [[] for _ in range(N)]
+    # scores[n][i] is the logits on which hyp[n][i] is decoded
+    scores = [[] for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+    # decoder_out: (N, 1, decoder_out_dim)
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits'shape (batch_size, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (batch_size, vocab_size)
+        assert logits.ndim == 2, logits.shape
+
+        if blank_penalty != 0:
+            logits[:, 0] -= blank_penalty
+
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v not in (blank_id, unk_id):
+                hyps[i].append(v)
+                timestamps[i].append(t)
+                scores[i].append(logits[i, v].item())
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    ans_timestamps = []
+    ans_scores = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(timestamps[unsorted_indices[i]])
+        ans_scores.append(scores[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+            scores=ans_scores,
+        )
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    # timestamp[i] is the frame index after subsampling
+    # on which ys[i] is decoded
+    timestamp: List[int] = field(default_factory=list)
+
+    # the lm score for next token given the current ys
+    lm_score: Optional[torch.Tensor] = None
+
+    # the RNNLM states (h and c in LSTM)
+    state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
+
+    # N-gram LM state
+    state_cost: Optional[NgramLmStateCost] = None
+
+    # Context graph state
+    context_state: Optional[ContextState] = None
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int, length_norm: bool = False) -> "HypothesisList":
+        """Return the top-k hypothesis.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        """
+        hyps = list(self._data.items())
+
+        if length_norm:
+            hyps = sorted(
+                hyps, key=lambda h: h[1].log_prob / len(h[1].ys), reverse=True
+            )[:k]
+        else:
+            hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_graph: Optional[ContextGraph] = None,
+    beam: int = 4,
+    temperature: float = 1.0,
+    blank_penalty: float = 0.0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                context_state=None if context_graph is None else context_graph.root,
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        if blank_penalty != 0:
+            logits[:, 0] -= blank_penalty
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                context_score = 0
+                new_context_state = None if context_graph is None else hyp.context_state
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    new_timestamp.append(t)
+                    if context_graph is not None:
+                        (
+                            context_score,
+                            new_context_state,
+                        ) = context_graph.forward_one_step(hyp.context_state, new_token)
+
+                new_log_prob = topk_log_probs[k] + context_score
+
+                new_hyp = Hypothesis(
+                    ys=new_ys,
+                    log_prob=new_log_prob,
+                    timestamp=new_timestamp,
+                    context_state=new_context_state,
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+
+    # finalize context_state, if the matched contexts do not reach final state
+    # we need to add the score on the corresponding backoff arc
+    if context_graph is not None:
+        finalized_B = [HypothesisList() for _ in range(len(B))]
+        for i, hyps in enumerate(B):
+            for hyp in list(hyps):
+                context_score, new_context_state = context_graph.finalize(
+                    hyp.context_state
+                )
+                finalized_B[i].add(
+                    Hypothesis(
+                        ys=hyp.ys,
+                        log_prob=hyp.log_prob + context_score,
+                        timestamp=hyp.timestamp,
+                        context_state=new_context_state,
+                    )
+                )
+        B = finalized_B
+
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    sorted_timestamps = [h.timestamp for h in best_hyps]
+    ans = []
+    ans_timestamps = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(sorted_timestamps[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+        )
+
+
+def modified_beam_search_lm_rescore(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    LM: LmScorer,
+    lm_scale_list: List[int],
+    beam: int = 4,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+    Rescore the final results with RNNLM and return the one with the highest score
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      temperature:
+        Softmax temperature.
+      LM:
+        A neural network language model
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    new_timestamp.append(t)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(
+                    ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+
+    # get the am_scores for n-best list
+    hyps_shape = get_hyps_shape(B)
+    am_scores = torch.tensor([hyp.log_prob.item() for b in B for hyp in b])
+    am_scores = k2.RaggedTensor(value=am_scores, shape=hyps_shape).to(device)
+
+    # now LM rescore
+    # prepare input data to LM
+    candidate_seqs = [hyp.ys[context_size:] for b in B for hyp in b]
+    possible_seqs = k2.RaggedTensor(candidate_seqs)
+    row_splits = possible_seqs.shape.row_splits(1)
+    sentence_token_lengths = row_splits[1:] - row_splits[:-1]
+    possible_seqs_with_sos = add_sos(possible_seqs, sos_id=1)
+    possible_seqs_with_eos = add_eos(possible_seqs, eos_id=1)
+    sentence_token_lengths += 1
+
+    x = possible_seqs_with_sos.pad(mode="constant", padding_value=blank_id)
+    y = possible_seqs_with_eos.pad(mode="constant", padding_value=blank_id)
+    x = x.to(device).to(torch.int64)
+    y = y.to(device).to(torch.int64)
+    sentence_token_lengths = sentence_token_lengths.to(device).to(torch.int64)
+
+    lm_scores = LM.lm(x=x, y=y, lengths=sentence_token_lengths)
+    assert lm_scores.ndim == 2
+    lm_scores = -1 * lm_scores.sum(dim=1)
+
+    ans = {}
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+
+    # get the best hyp with different lm_scale
+    for lm_scale in lm_scale_list:
+        key = f"nnlm_scale_{lm_scale:.2f}"
+        tot_scores = am_scores.values + lm_scores * lm_scale
+        ragged_tot_scores = k2.RaggedTensor(shape=am_scores.shape, value=tot_scores)
+        max_indexes = ragged_tot_scores.argmax().tolist()
+        unsorted_hyps = [candidate_seqs[idx] for idx in max_indexes]
+        hyps = []
+        for idx in unsorted_indices:
+            hyps.append(unsorted_hyps[idx])
+
+        ans[key] = hyps
+    return ans
+
+
+def modified_beam_search_lm_rescore_LODR(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    LM: LmScorer,
+    LODR_lm: NgramLm,
+    sp: spm.SentencePieceProcessor,
+    lm_scale_list: List[int],
+    beam: int = 4,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Union[List[List[int]], DecodingResults]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+    Rescore the final results with RNNLM and return the one with the highest score
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      temperature:
+        Softmax temperature.
+      LM:
+        A neural network language model
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    new_timestamp.append(t)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(
+                    ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+
+    # get the am_scores for n-best list
+    hyps_shape = get_hyps_shape(B)
+    am_scores = torch.tensor([hyp.log_prob.item() for b in B for hyp in b])
+    am_scores = k2.RaggedTensor(value=am_scores, shape=hyps_shape).to(device)
+
+    # now LM rescore
+    # prepare input data to LM
+    candidate_seqs = [hyp.ys[context_size:] for b in B for hyp in b]
+    possible_seqs = k2.RaggedTensor(candidate_seqs)
+    row_splits = possible_seqs.shape.row_splits(1)
+    sentence_token_lengths = row_splits[1:] - row_splits[:-1]
+    possible_seqs_with_sos = add_sos(possible_seqs, sos_id=1)
+    possible_seqs_with_eos = add_eos(possible_seqs, eos_id=1)
+    sentence_token_lengths += 1
+
+    x = possible_seqs_with_sos.pad(mode="constant", padding_value=blank_id)
+    y = possible_seqs_with_eos.pad(mode="constant", padding_value=blank_id)
+    x = x.to(device).to(torch.int64)
+    y = y.to(device).to(torch.int64)
+    sentence_token_lengths = sentence_token_lengths.to(device).to(torch.int64)
+
+    lm_scores = LM.lm(x=x, y=y, lengths=sentence_token_lengths)
+    assert lm_scores.ndim == 2
+    lm_scores = -1 * lm_scores.sum(dim=1)
+
+    # now LODR scores
+    import math
+
+    LODR_scores = []
+    for seq in candidate_seqs:
+        tokens = " ".join(sp.id_to_piece(seq))
+        LODR_scores.append(LODR_lm.score(tokens))
+    LODR_scores = torch.tensor(LODR_scores).to(device) * math.log(
+        10
+    )  # arpa scores are 10-based
+    assert lm_scores.shape == LODR_scores.shape
+
+    ans = {}
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+
+    LODR_scale_list = [0.05 * i for i in range(1, 20)]
+    # get the best hyp with different lm_scale and lodr_scale
+    for lm_scale in lm_scale_list:
+        for lodr_scale in LODR_scale_list:
+            key = f"nnlm_scale_{lm_scale:.2f}_lodr_scale_{lodr_scale:.2f}"
+            tot_scores = (
+                am_scores.values / lm_scale + lm_scores - LODR_scores * lodr_scale
+            )
+            ragged_tot_scores = k2.RaggedTensor(shape=am_scores.shape, value=tot_scores)
+            max_indexes = ragged_tot_scores.argmax().tolist()
+            unsorted_hyps = [candidate_seqs[idx] for idx in max_indexes]
+            hyps = []
+            for idx in unsorted_indices:
+                hyps.append(unsorted_hyps[idx])
+
+            ans[key] = hyps
+    return ans
+
+
+def _deprecated_modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+    return_timestamps: bool = False,
+) -> Union[List[int], DecodingResults]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    It decodes only one utterance at a time. We keep it only for reference.
+    The function :func:`modified_beam_search` should be preferred as it
+    supports batch decoding.
+
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+      return_timestamps:
+        Whether to return timestamps.
+
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    device = next(model.parameters()).device
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[-1] * (context_size - 1) + [blank_id],
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+            timestamp=[],
+        )
+    )
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # current_encoder_out is of shape (1, 1, 1, encoder_out_dim)
+        # fmt: on
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+            dtype=torch.int64,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_output is of shape (num_hyps, 1, 1, joiner_dim)
+
+        current_encoder_out = current_encoder_out.expand(
+            decoder_out.size(0), 1, 1, -1
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        # now logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            topk_hyp_indexes = topk_hyp_indexes.tolist()
+            topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
+            new_ys = hyp.ys[:]
+            new_timestamp = hyp.timestamp[:]
+            new_token = topk_token_indexes[i]
+            if new_token not in (blank_id, unk_id):
+                new_ys.append(new_token)
+                new_timestamp.append(t)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(
+                ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
+            )
+            B.add(new_hyp)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    if not return_timestamps:
+        return ys
+    else:
+        return DecodingResults(hyps=[ys], timestamps=[best_hyp.timestamp])
+
+
+def beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+    temperature: float = 1.0,
+    blank_penalty: float = 0.0,
+    return_timestamps: bool = False,
+) -> Union[List[int], DecodingResults]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+
+    Returns:
+      If return_timestamps is False, return the decoded result.
+      Else, return a DecodingResults object containing
+      decoded result and corresponding timestamps.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+
+    device = next(model.parameters()).device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size,
+        device=device,
+        dtype=torch.int64,
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[-1] * (context_size - 1) + [blank_id], log_prob=0.0, timestamp=[]
+        )
+    )
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        # TODO(fangjun): Implement prefix search to update the `log_prob`
+        # of hypotheses in A
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            cached_key = y_star.key
+
+            if cached_key not in decoder_cache:
+                decoder_input = torch.tensor(
+                    [y_star.ys[-context_size:]],
+                    device=device,
+                    dtype=torch.int64,
+                ).reshape(1, context_size)
+
+                decoder_out = model.decoder(decoder_input, need_pad=False)
+                decoder_out = model.joiner.decoder_proj(decoder_out)
+                decoder_cache[cached_key] = decoder_out
+            else:
+                decoder_out = decoder_cache[cached_key]
+
+            cached_key += f"-t-{t}"
+            if cached_key not in joint_cache:
+                logits = model.joiner(
+                    current_encoder_out,
+                    decoder_out.unsqueeze(1),
+                    project_input=False,
+                )
+
+                if blank_penalty != 0:
+                    logits[:, :, :, 0] -= blank_penalty
+
+                # TODO(fangjun): Scale the blank posterior
+                log_prob = (logits / temperature).log_softmax(dim=-1)
+                # log_prob is (1, 1, 1, vocab_size)
+                log_prob = log_prob.squeeze()
+                # Now log_prob is (vocab_size,)
+                joint_cache[cached_key] = log_prob
+            else:
+                log_prob = joint_cache[cached_key]
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(
+                Hypothesis(
+                    ys=y_star.ys[:],
+                    log_prob=new_y_star_log_prob,
+                    timestamp=y_star.timestamp[:],
+                )
+            )
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for i, v in zip(indices.tolist(), values.tolist()):
+                if i in (blank_id, unk_id):
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                new_timestamp = y_star.timestamp + [t]
+                A.add(
+                    Hypothesis(
+                        ys=new_ys,
+                        log_prob=new_log_prob,
+                        timestamp=new_timestamp,
+                    )
+                )
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    if not return_timestamps:
+        return ys
+    else:
+        return DecodingResults(hyps=[ys], timestamps=[best_hyp.timestamp])
+
+
+def fast_beam_search_with_nbest_rescoring(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    ngram_lm_scale_list: List[float],
+    num_paths: int,
+    G: k2.Fsa,
+    sp: spm.SentencePieceProcessor,
+    word_table: k2.SymbolTable,
+    oov_word: str = "<UNK>",
+    use_double_scores: bool = True,
+    nbest_scale: float = 0.5,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Dict[str, Union[List[List[int]], DecodingResults]]:
+    """It limits the maximum number of symbols per frame to 1.
+    A lattice is first obtained using fast beam search, num_path are selected
+    and rescored using a given language model. The shortest path within the
+    lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi.
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      ngram_lm_scale_list:
+        A list of floats representing LM score scales.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      G:
+        An FsaVec containing only a single FSA. It is an n-gram LM.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      oov_word:
+        OOV words are replaced with this word.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      Return the decoded result in a dict, where the key has the form
+      'ngram_lm_scale_xx' and the value is the decoded results
+      optionally with timestamps. `xx` is the ngram LM scale value
+      used during decoding, i.e., 0.1.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        temperature=temperature,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # at this point, nbest.fsa.scores are all zeros.
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa.scores contains acoustic scores
+
+    am_scores = nbest.tot_scores()
+
+    # Now we need to compute the LM scores of each path.
+    # (1) Get the token IDs of each Path. We assume the decoding_graph
+    # is an acceptor, i.e., lattice is also an acceptor
+    tokens_shape = nbest.fsa.arcs.shape().remove_axis(1)  # [path][arc]
+
+    tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.labels.contiguous())
+    tokens = tokens.remove_values_leq(0)  # remove -1 and 0
+
+    token_list: List[List[int]] = tokens.tolist()
+    word_list: List[List[str]] = sp.decode(token_list)
+
+    assert isinstance(oov_word, str), oov_word
+    assert oov_word in word_table, oov_word
+    oov_word_id = word_table[oov_word]
+
+    word_ids_list: List[List[int]] = []
+
+    for words in word_list:
+        this_word_ids = []
+        for w in words.split():
+            if w in word_table:
+                this_word_ids.append(word_table[w])
+            else:
+                this_word_ids.append(oov_word_id)
+        word_ids_list.append(this_word_ids)
+
+    word_fsas = k2.linear_fsa(word_ids_list, device=lattice.device)
+    word_fsas_with_self_loops = k2.add_epsilon_self_loops(word_fsas)
+
+    num_unique_paths = len(word_ids_list)
+
+    b_to_a_map = torch.zeros(
+        num_unique_paths,
+        dtype=torch.int32,
+        device=lattice.device,
+    )
+
+    rescored_word_fsas = k2.intersect_device(
+        a_fsas=G,
+        b_fsas=word_fsas_with_self_loops,
+        b_to_a_map=b_to_a_map,
+        sorted_match_a=True,
+        ret_arc_maps=False,
+    )
+
+    rescored_word_fsas = k2.remove_epsilon_self_loops(rescored_word_fsas)
+    rescored_word_fsas = k2.top_sort(k2.connect(rescored_word_fsas))
+    ngram_lm_scores = rescored_word_fsas.get_tot_scores(
+        use_double_scores=True,
+        log_semiring=False,
+    )
+
+    ans: Dict[str, Union[List[List[int]], DecodingResults]] = {}
+    for s in ngram_lm_scale_list:
+        key = f"ngram_lm_scale_{s}"
+        tot_scores = am_scores.values + s * ngram_lm_scores
+        ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+        max_indexes = ragged_tot_scores.argmax()
+        best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+        if not return_timestamps:
+            ans[key] = get_texts(best_path)
+        else:
+            ans[key] = get_texts_with_timestamp(best_path)
+
+    return ans
+
+
+def fast_beam_search_with_nbest_rnn_rescoring(
+    model: nn.Module,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    ngram_lm_scale_list: List[float],
+    num_paths: int,
+    G: k2.Fsa,
+    sp: spm.SentencePieceProcessor,
+    word_table: k2.SymbolTable,
+    rnn_lm_model: torch.nn.Module,
+    rnn_lm_scale_list: List[float],
+    oov_word: str = "<UNK>",
+    use_double_scores: bool = True,
+    nbest_scale: float = 0.5,
+    temperature: float = 1.0,
+    return_timestamps: bool = False,
+) -> Dict[str, Union[List[List[int]], DecodingResults]]:
+    """It limits the maximum number of symbols per frame to 1.
+    A lattice is first obtained using fast beam search, num_path are selected
+    and rescored using a given language model and a rnn-lm.
+    The shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a LG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi.
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      ngram_lm_scale_list:
+        A list of floats representing LM score scales.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      G:
+        An FsaVec containing only a single FSA. It is an n-gram LM.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      rnn_lm_model:
+        A rnn-lm model used for LM rescoring
+      rnn_lm_scale_list:
+        A list of floats representing RNN score scales.
+      oov_word:
+        OOV words are replaced with this word.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      temperature:
+        Softmax temperature.
+      return_timestamps:
+        Whether to return timestamps.
+    Returns:
+      Return the decoded result in a dict, where the key has the form
+      'ngram_lm_scale_xx' and the value is the decoded results
+      optionally with timestamps. `xx` is the ngram LM scale value
+      used during decoding, i.e., 0.1.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+        temperature=temperature,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # at this point, nbest.fsa.scores are all zeros.
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa.scores contains acoustic scores
+
+    am_scores = nbest.tot_scores()
+
+    # Now we need to compute the LM scores of each path.
+    # (1) Get the token IDs of each Path. We assume the decoding_graph
+    # is an acceptor, i.e., lattice is also an acceptor
+    tokens_shape = nbest.fsa.arcs.shape().remove_axis(1)  # [path][arc]
+
+    tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.labels.contiguous())
+    tokens = tokens.remove_values_leq(0)  # remove -1 and 0
+
+    token_list: List[List[int]] = tokens.tolist()
+    word_list: List[List[str]] = sp.decode(token_list)
+
+    assert isinstance(oov_word, str), oov_word
+    assert oov_word in word_table, oov_word
+    oov_word_id = word_table[oov_word]
+
+    word_ids_list: List[List[int]] = []
+
+    for words in word_list:
+        this_word_ids = []
+        for w in words.split():
+            if w in word_table:
+                this_word_ids.append(word_table[w])
+            else:
+                this_word_ids.append(oov_word_id)
+        word_ids_list.append(this_word_ids)
+
+    word_fsas = k2.linear_fsa(word_ids_list, device=lattice.device)
+    word_fsas_with_self_loops = k2.add_epsilon_self_loops(word_fsas)
+
+    num_unique_paths = len(word_ids_list)
+
+    b_to_a_map = torch.zeros(
+        num_unique_paths,
+        dtype=torch.int32,
+        device=lattice.device,
+    )
+
+    rescored_word_fsas = k2.intersect_device(
+        a_fsas=G,
+        b_fsas=word_fsas_with_self_loops,
+        b_to_a_map=b_to_a_map,
+        sorted_match_a=True,
+        ret_arc_maps=False,
+    )
+
+    rescored_word_fsas = k2.remove_epsilon_self_loops(rescored_word_fsas)
+    rescored_word_fsas = k2.top_sort(k2.connect(rescored_word_fsas))
+    ngram_lm_scores = rescored_word_fsas.get_tot_scores(
+        use_double_scores=True,
+        log_semiring=False,
+    )
+
+    # Now RNN-LM
+    blank_id = model.decoder.blank_id
+    sos_id = sp.piece_to_id("sos_id")
+    eos_id = sp.piece_to_id("eos_id")
+
+    sos_tokens = add_sos(tokens, sos_id)
+    tokens_eos = add_eos(tokens, eos_id)
+    sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+    sentence_lengths = sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+
+    x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+    y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+    x_tokens = x_tokens.to(torch.int64)
+    y_tokens = y_tokens.to(torch.int64)
+    sentence_lengths = sentence_lengths.to(torch.int64)
+
+    rnn_lm_nll = rnn_lm_model(x=x_tokens, y=y_tokens, lengths=sentence_lengths)
+    assert rnn_lm_nll.ndim == 2
+    assert rnn_lm_nll.shape[0] == len(token_list)
+    rnn_lm_scores = -1 * rnn_lm_nll.sum(dim=1)
+
+    ans: Dict[str, List[List[int]]] = {}
+    for n_scale in ngram_lm_scale_list:
+        for rnn_scale in rnn_lm_scale_list:
+            key = f"ngram_lm_scale_{n_scale}_rnn_lm_scale_{rnn_scale}"
+            tot_scores = (
+                am_scores.values + n_scale * ngram_lm_scores + rnn_scale * rnn_lm_scores
+            )
+            ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+            max_indexes = ragged_tot_scores.argmax()
+            best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+            if not return_timestamps:
+                ans[key] = get_texts(best_path)
+            else:
+                ans[key] = get_texts_with_timestamp(best_path)
+
+    return ans
+
+
+def modified_beam_search_ngram_rescoring(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    ngram_lm: NgramLm,
+    ngram_lm_scale: float,
+    beam: int = 4,
+    temperature: float = 1.0,
+) -> List[List[int]]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+      temperature:
+        Softmax temperature.
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    lm_scale = ngram_lm_scale
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                state_cost=NgramLmStateCost(ngram_lm),
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [
+                hyp.log_prob.reshape(1, 1) + hyp.state_cost.lm_score * lm_scale
+                for hyps in A
+                for hyp in hyps
+            ]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = (logits / temperature).log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+        vocab_size = log_probs.size(-1)
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    state_cost = hyp.state_cost.forward_one_step(new_token)
+                else:
+                    state_cost = hyp.state_cost
+
+                # We only keep AM scores in new_hyp.log_prob
+                new_log_prob = topk_log_probs[k] - hyp.state_cost.lm_score * lm_scale
+
+                new_hyp = Hypothesis(
+                    ys=new_ys, log_prob=new_log_prob, state_cost=state_cost
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+def modified_beam_search_LODR(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    LODR_lm: NgramLm,
+    LODR_lm_scale: float,
+    LM: LmScorer,
+    beam: int = 4,
+    context_graph: Optional[ContextGraph] = None,
+) -> List[List[int]]:
+    """This function implements LODR (https://arxiv.org/abs/2203.16776) with
+    `modified_beam_search`. It uses a bi-gram language model as the estimate
+    of the internal language model and subtracts its score during shallow fusion
+    with an external language model. This implementation uses a RNNLM as the
+    external language model.
+
+    Args:
+        model (Transducer):
+            The transducer model
+        encoder_out (torch.Tensor):
+            Encoder output in (N,T,C)
+        encoder_out_lens (torch.Tensor):
+            A 1-D tensor of shape (N,), containing the number of
+            valid frames in encoder_out before padding.
+        LODR_lm:
+            A low order n-gram LM, whose score will be subtracted during shallow fusion
+        LODR_lm_scale:
+            The scale of the LODR_lm
+        LM:
+            A neural net LM, e.g an RNNLM or transformer LM
+        beam (int, optional):
+            Beam size. Defaults to 4.
+
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+    assert LM is not None
+    lm_scale = LM.lm_scale
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    sos_id = getattr(LM, "sos_id", 1)
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    # get initial lm score and lm state by scoring the "sos" token
+    sos_token = torch.tensor([[sos_id]]).to(torch.int64).to(device)
+    lens = torch.tensor([1]).to(device)
+    init_score, init_states = LM.score_token(sos_token, lens)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                state=init_states,  # state of the NN LM
+                lm_score=init_score.reshape(-1),
+                state_cost=NgramLmStateCost(
+                    LODR_lm
+                ),  # state of the source domain ngram
+                context_state=None if context_graph is None else context_graph.root,
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]  # get batch
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+        """
+        for all hyps with a non-blank new token, score this token.
+        It is a little confusing here because this for-loop
+        looks very similar to the one below. Here, we go through all
+        top-k tokens and only add the non-blanks ones to the token_list.
+        LM will score those tokens given the LM states. Note that
+        the variable `scores` is the LM score after seeing the new
+        non-blank token.
+        """
+        token_list = []
+        hs = []
+        cs = []
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_token = topk_token_indexes[k]
+                if new_token not in (blank_id, unk_id):
+                    if LM.lm_type == "rnn":
+                        token_list.append([new_token])
+                        # store the LSTM states
+                        hs.append(hyp.state[0])
+                        cs.append(hyp.state[1])
+                    else:
+                        # for transformer LM
+                        token_list.append(
+                            [sos_id] + hyp.ys[context_size:] + [new_token]
+                        )
+
+        # forward NN LM to get new states and scores
+        if len(token_list) != 0:
+            x_lens = torch.tensor([len(tokens) for tokens in token_list]).to(device)
+            if LM.lm_type == "rnn":
+                tokens_to_score = (
+                    torch.tensor(token_list).to(torch.int64).to(device).reshape(-1, 1)
+                )
+                hs = torch.cat(hs, dim=1).to(device)
+                cs = torch.cat(cs, dim=1).to(device)
+                state = (hs, cs)
+            else:
+                # for transformer LM
+                tokens_list = [torch.tensor(tokens) for tokens in token_list]
+                tokens_to_score = (
+                    torch.nn.utils.rnn.pad_sequence(
+                        tokens_list, batch_first=True, padding_value=0.0
+                    )
+                    .to(device)
+                    .to(torch.int64)
+                )
+
+                state = None
+
+            scores, lm_states = LM.score_token(tokens_to_score, x_lens, state)
+
+        count = 0  # index, used to locate score and lm states
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                ys = hyp.ys[:]
+
+                # current score of hyp
+                lm_score = hyp.lm_score
+                state = hyp.state
+
+                hyp_log_prob = topk_log_probs[k]  # get score of current hyp
+                new_token = topk_token_indexes[k]
+
+                context_score = 0
+                new_context_state = None if context_graph is None else hyp.context_state
+                if new_token not in (blank_id, unk_id):
+                    if context_graph is not None:
+                        (
+                            context_score,
+                            new_context_state,
+                        ) = context_graph.forward_one_step(hyp.context_state, new_token)
+
+                    ys.append(new_token)
+                    state_cost = hyp.state_cost.forward_one_step(new_token)
+
+                    # calculate the score of the latest token
+                    current_ngram_score = state_cost.lm_score - hyp.state_cost.lm_score
+
+                    assert current_ngram_score <= 0.0, (
+                        state_cost.lm_score,
+                        hyp.state_cost.lm_score,
+                    )
+                    # score = score + TDLM_score - LODR_score
+                    # LODR_LM_scale should be a negative number here
+                    hyp_log_prob += (
+                        lm_score[new_token] * lm_scale
+                        + LODR_lm_scale * current_ngram_score
+                        + context_score
+                    )  # add the lm score
+
+                    lm_score = scores[count]
+                    if LM.lm_type == "rnn":
+                        state = (
+                            lm_states[0][:, count, :].unsqueeze(1),
+                            lm_states[1][:, count, :].unsqueeze(1),
+                        )
+                    count += 1
+                else:
+                    state_cost = hyp.state_cost
+
+                new_hyp = Hypothesis(
+                    ys=ys,
+                    log_prob=hyp_log_prob,
+                    state=state,
+                    lm_score=lm_score,
+                    state_cost=state_cost,
+                    context_state=new_context_state,
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+
+    # finalize context_state, if the matched contexts do not reach final state
+    # we need to add the score on the corresponding backoff arc
+    if context_graph is not None:
+        finalized_B = [HypothesisList() for _ in range(len(B))]
+        for i, hyps in enumerate(B):
+            for hyp in list(hyps):
+                context_score, new_context_state = context_graph.finalize(
+                    hyp.context_state
+                )
+                finalized_B[i].add(
+                    Hypothesis(
+                        ys=hyp.ys,
+                        log_prob=hyp.log_prob + context_score,
+                        timestamp=hyp.timestamp,
+                        context_state=new_context_state,
+                    )
+                )
+        B = finalized_B
+
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+def modified_beam_search_lm_shallow_fusion(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    LM: LmScorer,
+    beam: int = 4,
+    return_timestamps: bool = False,
+) -> List[List[int]]:
+    """Modified_beam_search + NN LM shallow fusion
+
+    Args:
+        model (Transducer):
+            The transducer model
+        encoder_out (torch.Tensor):
+            Encoder output in (N,T,C)
+        encoder_out_lens (torch.Tensor):
+            A 1-D tensor of shape (N,), containing the number of
+            valid frames in encoder_out before padding.
+        sp:
+            Sentence piece generator.
+        LM (LmScorer):
+            A neural net LM, e.g RNN or Transformer
+        beam (int, optional):
+            Beam size. Defaults to 4.
+
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+    assert LM is not None
+    lm_scale = LM.lm_scale
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    sos_id = getattr(LM, "sos_id", 1)
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    # get initial lm score and lm state by scoring the "sos" token
+    sos_token = torch.tensor([[sos_id]]).to(torch.int64).to(device)
+    lens = torch.tensor([1]).to(device)
+    init_score, init_states = LM.score_token(sos_token, lens)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                state=init_states,
+                lm_score=init_score.reshape(-1),
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]  # get batch
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )
+
+        lm_scores = torch.cat(
+            [hyp.lm_score.reshape(1, -1) for hyps in A for hyp in hyps]
+        )
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+        """
+        for all hyps with a non-blank new token, score this token.
+        It is a little confusing here because this for-loop
+        looks very similar to the one below. Here, we go through all
+        top-k tokens and only add the non-blanks ones to the token_list.
+        `LM` will score those tokens given the LM states. Note that
+        the variable `scores` is the LM score after seeing the new
+        non-blank token.
+        """
+        token_list = []  # a list of list
+        hs = []
+        cs = []
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_token = topk_token_indexes[k]
+                if new_token not in (blank_id, unk_id):
+                    if LM.lm_type == "rnn":
+                        token_list.append([new_token])
+                        # store the LSTM states
+                        hs.append(hyp.state[0])
+                        cs.append(hyp.state[1])
+                    else:
+                        # for transformer LM
+                        token_list.append(
+                            [sos_id] + hyp.ys[context_size:] + [new_token]
+                        )
+
+        if len(token_list) != 0:
+            x_lens = torch.tensor([len(tokens) for tokens in token_list]).to(device)
+            if LM.lm_type == "rnn":
+                tokens_to_score = (
+                    torch.tensor(token_list).to(torch.int64).to(device).reshape(-1, 1)
+                )
+                hs = torch.cat(hs, dim=1).to(device)
+                cs = torch.cat(cs, dim=1).to(device)
+                state = (hs, cs)
+            else:
+                # for transformer LM
+                tokens_list = [torch.tensor(tokens) for tokens in token_list]
+                tokens_to_score = (
+                    torch.nn.utils.rnn.pad_sequence(
+                        tokens_list, batch_first=True, padding_value=0.0
+                    )
+                    .to(device)
+                    .to(torch.int64)
+                )
+
+                state = None
+
+            scores, lm_states = LM.score_token(tokens_to_score, x_lens, state)
+
+        count = 0  # index, used to locate score and lm states
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                ys = hyp.ys[:]
+
+                lm_score = hyp.lm_score
+                state = hyp.state
+
+                hyp_log_prob = topk_log_probs[k]  # get score of current hyp
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                if new_token not in (blank_id, unk_id):
+                    ys.append(new_token)
+                    new_timestamp.append(t)
+
+                    hyp_log_prob += lm_score[new_token] * lm_scale  # add the lm score
+
+                    lm_score = scores[count]
+                    if LM.lm_type == "rnn":
+                        state = (
+                            lm_states[0][:, count, :].unsqueeze(1),
+                            lm_states[1][:, count, :].unsqueeze(1),
+                        )
+                    count += 1
+
+                new_hyp = Hypothesis(
+                    ys=ys,
+                    log_prob=hyp_log_prob,
+                    state=state,
+                    lm_score=lm_score,
+                    timestamp=new_timestamp,
+                )
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    sorted_timestamps = [h.timestamp for h in best_hyps]
+    ans = []
+    ans_timestamps = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+        ans_timestamps.append(sorted_timestamps[unsorted_indices[i]])
+
+    if not return_timestamps:
+        return ans
+    else:
+        return DecodingResults(
+            hyps=ans,
+            timestamps=ans_timestamps,
+        )
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py
new file mode 100644
index 000000000..ab46e233b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1,1600 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from torch import Tensor, nn
+
+from icefall.utils import is_jit_tracing, make_pad_mask, subsequent_chunk_mask
+
+
+class Conformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+        dynamic_chunk_training (bool): whether to use dynamic chunk training, if
+            you want to train a streaming model, this is expected to be True.
+            When setting True, it will use a masking strategy to make the attention
+            see only limited left and right context.
+        short_chunk_threshold (float): a threshold to determinize the chunk size
+            to be used in masking training, if the randomly generated chunk size
+            is greater than ``max_len * short_chunk_threshold`` (max_len is the
+            max sequence length of current batch) then it will use
+            full context in training (i.e. with chunk size equals to max_len).
+            This will be used only when dynamic_chunk_training is True.
+        short_chunk_size (int): see docs above, if the randomly generated chunk
+            size equals to or less than ``max_len * short_chunk_threshold``, the
+            chunk size will be sampled uniformly from 1 to short_chunk_size.
+            This also will be used only when dynamic_chunk_training is True.
+        num_left_chunks (int): the left context (in chunks) attention can see, the
+            chunk size is decided by short_chunk_threshold and short_chunk_size.
+            A minus value means seeing full left context.
+            This also will be used only when dynamic_chunk_training is True.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        dynamic_chunk_training: bool = False,
+        short_chunk_threshold: float = 0.75,
+        short_chunk_size: int = 25,
+        num_left_chunks: int = -1,
+        causal: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.cnn_module_kernel = cnn_module_kernel
+        self.causal = causal
+        self.dynamic_chunk_training = dynamic_chunk_training
+        self.short_chunk_threshold = short_chunk_threshold
+        self.short_chunk_size = short_chunk_size
+        self.num_left_chunks = num_left_chunks
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+            causal,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self._init_state: List[torch.Tensor] = [torch.empty(0)]
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, d_model)
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        if not is_jit_tracing():
+            assert x.size(0) == lengths.max().item()
+
+        src_key_padding_mask = make_pad_mask(lengths, x.size(0))
+
+        if self.dynamic_chunk_training:
+            assert (
+                self.causal
+            ), "Causal convolution is required for streaming conformer."
+            max_len = x.size(0)
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * self.short_chunk_threshold):
+                chunk_size = max_len
+            else:
+                chunk_size = chunk_size % self.short_chunk_size + 1
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=self.num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+        else:
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=None,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        return x, lengths
+
+    @torch.jit.export
+    def get_init_state(
+        self, left_context: int, device: torch.device
+    ) -> List[torch.Tensor]:
+        """Return the initial cache state of the model.
+
+        Args:
+          left_context: The left context size (in frames after subsampling).
+
+        Returns:
+          Return the initial state of the model, it is a list containing two
+          tensors, the first one is the cache for attentions which has a shape
+          of (num_encoder_layers, left_context, encoder_dim), the second one
+          is the cache of conv_modules which has a shape of
+          (num_encoder_layers, cnn_module_kernel - 1, encoder_dim).
+
+          NOTE: the returned tensors are on the given device.
+        """
+        if len(self._init_state) == 2 and self._init_state[0].size(1) == left_context:
+            # Note: It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        init_states: List[torch.Tensor] = [
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    left_context,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+        ]
+
+        self._init_state = init_states
+
+        return init_states
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[Tensor]] = None,
+        processed_lens: Optional[Tensor] = None,
+        left_context: int = 64,
+        right_context: int = 4,
+        chunk_size: int = 16,
+        simulate_streaming: bool = False,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states:
+            The decode states for previous frames which contains the cached data.
+            It has two elements, the first element is the attn_cache which has
+            a shape of (encoder_layers, left_context, batch, attention_dim),
+            the second element is the conv_cache which has a shape of
+            (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+            Note: states will be modified in this function.
+          processed_lens:
+            How many frames (after subsampling) have been processed for each sequence.
+          left_context:
+            How many previous frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `left_context` frames
+            of left context, some have more.
+          right_context:
+            How many future frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `right_context` frames
+            of right context, some have more.
+          chunk_size:
+            The chunk size for decoding, this will be used to simulate streaming
+            decoding using masking.
+          simulate_streaming:
+            If setting True, it will use a masking strategy to simulate streaming
+            fashion (i.e. every chunk data only see limited left context and
+            right context). The whole sequence is supposed to be send at a time
+            When using simulate_streaming.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+            - decode_states, the updated states including the information
+              of current chunk.
+        """
+
+        # x: [N, T, C]
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        if not simulate_streaming:
+            assert states is not None
+            assert processed_lens is not None
+            assert (
+                len(states) == 2
+                and states[0].shape
+                == (self.encoder_layers, left_context, x.size(0), self.d_model)
+                and states[1].shape
+                == (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    x.size(0),
+                    self.d_model,
+                )
+            ), f"""The length of states MUST be equal to 2, and the shape of
+             first element should be {(self.encoder_layers, left_context, x.size(0), self.d_model)},
+             given {states[0].shape}. the shape of second element should be
+             {(self.encoder_layers, self.cnn_module_kernel - 1, x.size(0), self.d_model)},
+             given {states[1].shape}."""
+
+            lengths -= 2  # we will cut off 1 frame on each side of encoder_embed output
+
+            src_key_padding_mask = make_pad_mask(lengths)
+
+            processed_mask = torch.arange(left_context, device=x.device).expand(
+                x.size(0), left_context
+            )
+            processed_lens = processed_lens.view(x.size(0), 1)
+            processed_mask = (processed_lens <= processed_mask).flip(1)
+
+            src_key_padding_mask = torch.cat(
+                [processed_mask, src_key_padding_mask], dim=1
+            )
+
+            embed = self.encoder_embed(x)
+
+            # cut off 1 frame on each size of embed as they see the padding
+            # value which causes a training and decoding mismatch.
+            embed = embed[:, 1:-1, :]
+
+            embed, pos_enc = self.encoder_pos(embed, left_context)
+            embed = embed.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+
+            x, states = self.encoder.chunk_forward(
+                embed,
+                pos_enc,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+                states=states,
+                left_context=left_context,
+                right_context=right_context,
+            )  # (T, B, F)
+            if right_context > 0:
+                x = x[0:-right_context, ...]
+                lengths -= right_context
+        else:
+            assert states is None
+            states = []  # just to make torch.script.jit happy
+            # this branch simulates streaming decoding using mask as we are
+            # using in training time.
+            src_key_padding_mask = make_pad_mask(lengths)
+            x = self.encoder_embed(x)
+            x, pos_emb = self.encoder_pos(x)
+            x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+            assert x.size(0) == lengths.max().item()
+
+            if chunk_size < 0:
+                # use full attention
+                chunk_size = x.size(0)
+                left_context = -1
+
+            num_left_chunks = -1
+            if left_context >= 0:
+                assert left_context % chunk_size == 0
+                num_left_chunks = left_context // chunk_size
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths, states
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training and streaming decoding.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        causal: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel, causal=causal)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+        src_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            src_mask: the mask for the src sequence (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        conv, _ = self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        src = src + self.dropout(conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        assert not self.training
+        assert len(states) == 2
+        assert states[0].shape == (left_context, src.size(1), src.size(2))
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # We put the attention cache this level (i.e. before linear transformation)
+        # to save memory consumption, when decoding in streaming fashion, the
+        # batch size would be thousands (for 32GB machine), if we cache key & val
+        # separately, it needs extra several GB memory.
+        # TODO(WeiKang): Move cache to self_attn level (i.e. cache key & val
+        # separately) if needed.
+        key = torch.cat([states[0], src], dim=0)
+        val = key
+        if right_context > 0:
+            states[0] = key[
+                -(left_context + right_context) : -right_context, ...  # noqa
+            ]
+        else:
+            states[0] = key[-left_context:, ...]
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            key,
+            val,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+            left_context=left_context,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        conv, conv_cache = self.conv_module(src, states[1], right_context)
+        states[1] = conv_cache
+
+        src = src + self.dropout(conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        return src, states
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            mask: the mask for the src sequence (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (encoder_layers, left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        assert not self.training
+        assert len(states) == 2
+        assert states[0].shape == (
+            self.num_layers,
+            left_context,
+            src.size(1),
+            src.size(2),
+        )
+        assert states[1].size(0) == self.num_layers
+
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            cache = [states[0][layer_index], states[1][layer_index]]
+            output, cache = mod.chunk_forward(
+                output,
+                pos_emb,
+                states=cache,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+                left_context=left_context,
+                right_context=right_context,
+            )
+            states[0][layer_index] = cache[0]
+            states[1][layer_index] = cache[1]
+
+        return output, states
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        if is_jit_tracing():
+            # 10k frames correspond to ~100k ms, e.g., 100 seconds, i.e.,
+            # It assumes that the maximum input won't have more than
+            # 10k frames.
+            #
+            # TODO(fangjun): Use torch.jit.script() for this module
+            max_len = 10000
+
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor, left_context: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_1 = x.size(1) + left_context
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_1 * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_1, self.d_model)
+        pe_negative = torch.zeros(x_size_1, self.d_model)
+        position = torch.arange(0, x_size_1, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        if isinstance(left_context, torch.Tensor):
+            left_context = left_context.item()
+        self.extend_pe(x, left_context)
+        x_size_1 = x.size(1) + left_context
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_1
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = False,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            left_context=left_context,
+        )
+
+    def rel_shift(self, x: Tensor, left_context: int = 0) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1+left_context).
+                time1 means the length of query vector.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+
+        time2 = time1 + left_context
+        if not is_jit_tracing():
+            assert (
+                n == left_context + 2 * time1 - 1
+            ), f"{n} == {left_context} + 2 * {time1} - 1"
+
+        if is_jit_tracing():
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(time2)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, time1, time2)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time1_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, time1, time2),
+                (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+                storage_offset=n_stride * (time1 - 1),
+            )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = False,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        if not is_jit_tracing():
+            assert embed_dim == embed_dim_to_check
+            assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        if not is_jit_tracing():
+            assert (
+                head_dim * num_heads == embed_dim
+            ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None and not is_jit_tracing():
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        if not is_jit_tracing():
+            assert pos_emb_bsz in (1, bsz)  # actually it is 1
+
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1)
+        p = p.permute(0, 2, 3, 1)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(q_with_bias_v, p)  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd, left_context)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        if not is_jit_tracing():
+            assert list(attn_output_weights.size()) == [
+                bsz * num_heads,
+                tgt_len,
+                src_len,
+            ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+
+        # If we are using dynamic_chunk_training and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`, at this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax), so, we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, tgt_len, src_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+
+        if not is_jit_tracing():
+            assert list(attn_output.size()) == [
+                bsz * num_heads,
+                tgt_len,
+                head_dim,
+            ]
+
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+        causal (bool): Whether to use causal convolution.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        causal: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+        self.causal = causal
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.lorder = kernel_size - 1
+        padding = (kernel_size - 1) // 2
+        if self.causal:
+            padding = 0
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        cache: Optional[Tensor] = None,
+        right_context: int = 0,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: The cache of depthwise_conv, only used in real streaming
+                decoding.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+              of right context, some have more.
+
+        Returns:
+            If cache is None return the output tensor (#time, batch, channels).
+            If cache is not None, return a tuple of Tensor, the first one is
+            the output tensor (#time, batch, channels), the second one is the
+            new cache for next chunk (#kernel_size - 1, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        if self.causal and self.lorder > 0:
+            if cache is None:
+                # Make depthwise_conv causal by
+                # manualy padding self.lorder zeros to the left
+                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+            else:
+                assert not self.training, "Cache should be None in training time"
+                assert cache.size(0) == self.lorder
+                x = torch.cat([cache.permute(1, 2, 0), x], dim=2)
+                if right_context > 0:
+                    cache = x.permute(2, 0, 1)[
+                        -(self.lorder + right_context) : (-right_context),  # noqa
+                        ...,
+                    ]
+                else:
+                    cache = x.permute(2, 0, 1)[-self.lorder :, ...]  # noqa
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        # torch.jit.script requires return types be the same as annotated above
+        if cache is None:
+            cache = torch.empty(0)
+
+        return x.permute(2, 0, 1), cache
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+if __name__ == "__main__":
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..c57514193
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,803 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) decode in streaming mode (take greedy search as an example)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --decode-chunk-size 16 \
+    --left-context 64 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG and
+        fast_beam_search_nbest_LG. It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is fast_beam_search_LG
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling). Set -1 to use full attention.",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_oracle, and
+        fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        if params.decode_chunk_size > 0:
+            # except the case of using full attention
+            feature_lens += params.left_context
+            feature = torch.nn.functional.pad(
+                feature,
+                pad=(0, 0, 0, params.left_context),
+                value=LOG_EPS,
+            )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_oracle, and
+        fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decode_stream.py
new file mode 120000
index 000000000..30f264813
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decode_stream.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/decode_stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
new file mode 100644
index 000000000..d44ed6f81
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1,122 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import ScaledConv1d, ScaledEmbedding
+
+from icefall.utils import is_jit_tracing
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = ScaledEmbedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = ScaledConv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim,
+                bias=False,
+            )
+        else:
+            # It is to support torch script
+            self.conv = nn.Identity()
+
+    def forward(
+        self,
+        y: torch.Tensor,
+        need_pad: bool = True  # Annotation should be Union[bool, torch.Tensor]
+        # but, torch.jit.script does not support Union.
+    ) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        if isinstance(need_pad, torch.Tensor):
+            # This is for torch.jit.trace(), which cannot handle the case
+            # when the input argument is not a tensor.
+            need_pad = bool(need_pad)
+
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        if torch.jit.is_tracing():
+            # This is for exporting to PNNX via ONNX
+            embedding_out = self.embedding(y)
+        else:
+            embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                if not is_jit_tracing():
+                    assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = F.relu(embedding_out)
+        return embedding_out
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/export.py b/egs/librispeech/ASR/pruned_transducer_stateless2/export.py
new file mode 100755
index 000000000..e2db98f73
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,226 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py
new file mode 100644
index 000000000..9f88bd029
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1,67 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+from icefall.utils import is_jit_tracing
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim)
+        self.output_linear = ScaledLinear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        if not is_jit_tracing():
+            assert encoder_out.ndim == decoder_out.ndim
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/lstmp.py b/egs/librispeech/ASR/pruned_transducer_stateless2/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/model.py b/egs/librispeech/ASR/pruned_transducer_stateless2/model.py
new file mode 100644
index 000000000..272d06c37
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1,207 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output
+            contains unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        reduction: str = "sum",
+        delay_penalty: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          delay_penalty:
+            A constant value used to penalize symbol delay, to encourage
+            streaming models to emit symbols earlier.
+            See https://github.com/k2-fsa/k2/issues/955 and
+            https://arxiv.org/pdf/2211.00490.pdf for more details.
+        Returns:
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                delay_penalty=delay_penalty,
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                delay_penalty=delay_penalty,
+                reduction=reduction,
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
new file mode 100644
index 000000000..f54bc2709
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1,319 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Union
+
+import torch
+from torch.optim import Optimizer
+
+
+class Eve(Optimizer):
+    r"""
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam\: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                # Constrain the range of scalar weights
+                if p.numel() == 1:
+                    p.clamp_(min=-10, max=2)
+
+        return loss
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("initial_lr", group["lr"])
+
+        self.base_lrs = [group["initial_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            print(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+     lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                       (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
+
+     E.g. suggest initial-lr = 0.003 (passed to optimizer).
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        return [x * factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = Eve(m.parameters(), lr=0.003)
+
+    scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+    print("last lr = ", scheduler.get_last_lr())
+    print("state dict = ", scheduler.state_dict())
+
+
+if __name__ == "__main__":
+    _test_eden()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py
new file mode 100755
index 000000000..029f55ba0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py
@@ -0,0 +1,389 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
+./pruned_transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    if params.simulate_streaming:
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
new file mode 100644
index 000000000..91d64c1df
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1,1014 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+import random
+from itertools import repeat
+from typing import Optional, Tuple
+
+import torch
+import torch.backends.cudnn.rnn as rnn
+import torch.nn as nn
+from torch import _VF, Tensor
+
+from icefall.utils import is_jit_tracing
+
+
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.Iterable):
+            return x
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+_single = _ntuple(1)
+_pair = _ntuple(2)
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        channel_dim: int,
+        min_positive: float,  # e.g. 0.05
+        max_positive: float,  # e.g. 0.95
+        max_factor: float,  # e.g. 0.01
+        min_abs: float,  # e.g. 0.2
+        max_abs: float,  # e.g. 100.0
+    ) -> Tensor:
+        if x.requires_grad:
+            if channel_dim < 0:
+                channel_dim += x.ndim
+
+            #  sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+            # The above line is not torch scriptable for torch 1.6.0
+            # torch.jit.frontend.NotSupportedError: comprehension ifs not supported yet:  # noqa
+            sum_dims = []
+            for d in range(x.ndim):
+                if d != channel_dim:
+                    sum_dims.append(d)
+
+            xgt0 = x > 0
+            proportion_positive = torch.mean(
+                xgt0.to(x.dtype), dim=sum_dims, keepdim=True
+            )
+            factor1 = (
+                (min_positive - proportion_positive).relu()
+                * (max_factor / min_positive)
+                if min_positive != 0.0
+                else 0.0
+            )
+            factor2 = (
+                (proportion_positive - max_positive).relu()
+                * (max_factor / (max_positive - 1.0))
+                if max_positive != 1.0
+                else 0.0
+            )
+            factor = factor1 + factor2
+            if isinstance(factor, float):
+                factor = torch.zeros_like(proportion_positive)
+
+            mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True)
+            below_threshold = mean_abs < min_abs
+            above_threshold = mean_abs > max_abs
+
+            ctx.save_for_backward(factor, xgt0, below_threshold, above_threshold)
+            ctx.max_factor = max_factor
+            ctx.sum_dims = sum_dims
+        return x
+
+    @staticmethod
+    def backward(
+        ctx, x_grad: Tensor
+    ) -> Tuple[Tensor, None, None, None, None, None, None]:
+        factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors
+        dtype = x_grad.dtype
+        scale_factor = (
+            (below_threshold.to(dtype) - above_threshold.to(dtype))
+            * (xgt0.to(dtype) - 0.5)
+            * (ctx.max_factor * 2.0)
+        )
+
+        neg_delta_grad = x_grad.abs() * (factor + scale_factor)
+        return x_grad - neg_delta_grad, None, None, None, None, None, None
+
+
+class GradientFilterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        batch_dim: int,  # e.g., 1
+        threshold: float,  # e.g., 10.0
+        *params: Tensor,  # module parameters
+    ) -> Tuple[Tensor, ...]:
+        if x.requires_grad:
+            if batch_dim < 0:
+                batch_dim += x.ndim
+            ctx.batch_dim = batch_dim
+            ctx.threshold = threshold
+        return (x,) + params
+
+    @staticmethod
+    def backward(
+        ctx,
+        x_grad: Tensor,
+        *param_grads: Tensor,
+    ) -> Tuple[Tensor, ...]:
+        eps = 1.0e-20
+        dim = ctx.batch_dim
+        norm_dims = [d for d in range(x_grad.ndim) if d != dim]
+        norm_of_batch = (x_grad**2).mean(dim=norm_dims, keepdim=True).sqrt()
+        median_norm = norm_of_batch.median()
+
+        cutoff = median_norm * ctx.threshold
+        inv_mask = (cutoff + norm_of_batch) / (cutoff + eps)
+        mask = 1.0 / (inv_mask + eps)
+        x_grad = x_grad * mask
+
+        avg_mask = 1.0 / (inv_mask.mean() + eps)
+        param_grads = [avg_mask * g for g in param_grads]
+
+        return (x_grad, None, None) + tuple(param_grads)
+
+
+class GradientFilter(torch.nn.Module):
+    """This is used to filter out elements that have extremely large gradients
+    in batch and the module parameters with soft masks.
+
+    Args:
+      batch_dim (int):
+        The batch dimension.
+      threshold (float):
+        For each element in batch, its gradient will be
+        filtered out if the gradient norm is larger than
+        `grad_norm_threshold * median`, where `median` is the median
+        value of gradient norms of all elememts in batch.
+    """
+
+    def __init__(self, batch_dim: int = 1, threshold: float = 10.0):
+        super(GradientFilter, self).__init__()
+        self.batch_dim = batch_dim
+        self.threshold = threshold
+
+    def forward(self, x: Tensor, *params: Tensor) -> Tuple[Tensor, ...]:
+        if torch.jit.is_scripting() or is_jit_tracing():
+            return (x,) + params
+        else:
+            return GradientFilterFunction.apply(
+                x,
+                self.batch_dim,
+                self.threshold,
+                *params,
+            )
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+
+    def forward(self, x: Tensor) -> Tensor:
+        if not is_jit_tracing():
+            assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x**2, dim=self.channel_dim, keepdim=True) + self.eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+class ScaledLinear(nn.Linear):
+    """
+    A modified version of nn.Linear where the parameters are scaled before
+    use, via:
+         weight = self.weight * self.weight_scale.exp()
+         bias = self.bias * self.bias_scale.exp()
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+        initial_speed: this affects how fast the parameter will
+           learn near the start of training; you can set it to a
+           value less than one if you suspect that a module
+           is contributing to instability near the start of training.
+           Nnote: regardless of the use of this option, it's best to
+           use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.   Must be greater than 0.
+    """
+
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        **kwargs,
+    ):
+        super(ScaledLinear, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in nn.Linear
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        if self.bias is None or self.bias_scale is None:
+            return None
+        else:
+            return self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        return torch.nn.functional.linear(input, self.get_weight(), self.get_bias())
+
+
+class ScaledConv1d(nn.Conv1d):
+    # See docs for ScaledLinear
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        **kwargs,
+    ):
+        super(ScaledConv1d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+
+        self.bias_scale: Optional[nn.Parameter]  # for torchscript
+
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        bias = self.bias
+        bias_scale = self.bias_scale
+        if bias is None or bias_scale is None:
+            return None
+        else:
+            return bias * bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv1d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                self.get_weight(),
+                self.get_bias(),
+                self.stride,
+                (0,),
+                self.dilation,
+                self.groups,
+            )
+        return F.conv1d(
+            input,
+            self.get_weight(),
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+
+class ScaledConv2d(nn.Conv2d):
+    # See docs for ScaledLinear
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        **kwargs,
+    ):
+        super(ScaledConv2d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in**-0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        # see https://github.com/pytorch/pytorch/issues/24135
+        bias = self.bias
+        bias_scale = self.bias_scale
+        if bias is None or bias_scale is None:
+            return None
+        else:
+            return bias * bias_scale.exp()
+
+    def _conv_forward(self, input, weight):
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv2d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                weight,
+                self.get_bias(),
+                self.stride,
+                (0, 0),
+                self.dilation,
+                self.groups,
+            )
+        return F.conv2d(
+            input,
+            weight,
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+    def forward(self, input: Tensor) -> Tensor:
+        return self._conv_forward(input, self.get_weight())
+
+
+class ScaledLSTM(nn.LSTM):
+    # See docs for ScaledLinear.
+    # This class implements LSTM with scaling mechanism, using `torch._VF.lstm`
+    # Please refer to https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        grad_norm_threshold: float = 10.0,
+        **kwargs,
+    ):
+        if "bidirectional" in kwargs:
+            assert kwargs["bidirectional"] is False
+        super(ScaledLSTM, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self._scales_names = []
+        self._scales = []
+        for name in self._flat_weights_names:
+            scale_name = name + "_scale"
+            self._scales_names.append(scale_name)
+            param = nn.Parameter(initial_scale.clone().detach())
+            setattr(self, scale_name, param)
+            self._scales.append(param)
+
+        self.grad_filter = GradientFilter(batch_dim=1, threshold=grad_norm_threshold)
+
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3**0.5) * std
+        scale = self.hidden_size**-0.5
+        v = scale / std
+        for idx, name in enumerate(self._flat_weights_names):
+            if "weight" in name:
+                nn.init.uniform_(self._flat_weights[idx], -a, a)
+                with torch.no_grad():
+                    self._scales[idx] += torch.tensor(v).log()
+            elif "bias" in name:
+                nn.init.constant_(self._flat_weights[idx], 0.0)
+
+    def _flatten_parameters(self, flat_weights) -> None:
+        """Resets parameter data pointer so that they can use faster code paths.
+
+        Right now, this works only if the module is on the GPU and cuDNN is enabled.
+        Otherwise, it's a no-op.
+
+        This function is modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py  # noqa
+        """
+        # Short-circuits if _flat_weights is only partially instantiated
+        if len(flat_weights) != len(self._flat_weights_names):
+            return
+
+        for w in flat_weights:
+            if not isinstance(w, Tensor):
+                return
+        # Short-circuits if any tensor in flat_weights is not acceptable to cuDNN
+        # or the tensors in flat_weights are of different dtypes
+
+        first_fw = flat_weights[0]
+        dtype = first_fw.dtype
+        for fw in flat_weights:
+            if (
+                not isinstance(fw.data, Tensor)
+                or not (fw.data.dtype == dtype)
+                or not fw.data.is_cuda
+                or not torch.backends.cudnn.is_acceptable(fw.data)
+            ):
+                return
+
+        # If any parameters alias, we fall back to the slower, copying code path. This is
+        # a sufficient check, because overlapping parameter buffers that don't completely
+        # alias would break the assumptions of the uniqueness check in
+        # Module.named_parameters().
+        unique_data_ptrs = set(p.data_ptr() for p in flat_weights)
+        if len(unique_data_ptrs) != len(flat_weights):
+            return
+
+        with torch.cuda.device_of(first_fw):
+            # Note: no_grad() is necessary since _cudnn_rnn_flatten_weight is
+            # an inplace operation on self._flat_weights
+            with torch.no_grad():
+                if torch._use_cudnn_rnn_flatten_weight():
+                    num_weights = 4 if self.bias else 2
+                    if self.proj_size > 0:
+                        num_weights += 1
+                    torch._cudnn_rnn_flatten_weight(
+                        flat_weights,
+                        num_weights,
+                        self.input_size,
+                        rnn.get_cudnn_mode(self.mode),
+                        self.hidden_size,
+                        self.proj_size,
+                        self.num_layers,
+                        self.batch_first,
+                        bool(self.bidirectional),
+                    )
+
+    def _get_flat_weights(self):
+        """Get scaled weights, and resets their data pointer."""
+        flat_weights = []
+        for idx in range(len(self._flat_weights_names)):
+            flat_weights.append(self._flat_weights[idx] * self._scales[idx].exp())
+        self._flatten_parameters(flat_weights)
+        return flat_weights
+
+    def forward(self, input: Tensor, hx: Optional[Tuple[Tensor, Tensor]] = None):
+        # This function is modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py  # noqa
+        # The change for calling `_VF.lstm()` is:
+        # self._flat_weights -> self._get_flat_weights()
+        if hx is None:
+            h_zeros = torch.zeros(
+                self.num_layers,
+                input.size(1),
+                self.proj_size if self.proj_size > 0 else self.hidden_size,
+                dtype=input.dtype,
+                device=input.device,
+            )
+            c_zeros = torch.zeros(
+                self.num_layers,
+                input.size(1),
+                self.hidden_size,
+                dtype=input.dtype,
+                device=input.device,
+            )
+            hx = (h_zeros, c_zeros)
+
+        self.check_forward_args(input, hx, None)
+
+        flat_weights = self._get_flat_weights()
+        input, *flat_weights = self.grad_filter(input, *flat_weights)
+
+        result = _VF.lstm(
+            input,
+            hx,
+            flat_weights,
+            self.bias,
+            self.num_layers,
+            self.dropout,
+            self.training,
+            self.bidirectional,
+            self.batch_first,
+        )
+
+        output = result[0]
+        hidden = result[1:]
+        return output, hidden
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+
+    Args:
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           min_abs:  the minimum average-absolute-value per channel, which
+              we allow, before we start to modify the derivatives to prevent
+              this.
+           max_abs:  the maximum average-absolute-value per channel, which
+               we allow, before we start to modify the derivatives to prevent
+               this.
+           balance_prob: the probability to apply the ActivationBalancer.
+    """
+
+    def __init__(
+        self,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.01,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+        balance_prob: float = 0.25,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        assert 0 < balance_prob <= 1, balance_prob
+        self.balance_prob = balance_prob
+
+    def forward(self, x: Tensor) -> Tensor:
+        if random.random() >= self.balance_prob:
+            return x
+
+        return ActivationBalancerFunction.apply(
+            x,
+            self.channel_dim,
+            self.min_positive,
+            self.max_positive,
+            self.max_factor / self.balance_prob,
+            self.min_abs,
+            self.max_abs,
+        )
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        x = x.detach()
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+        ctx.save_for_backward(s, y)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        s, y = ctx.saved_tensors
+        return (y * (1 - s) + s) * y_grad
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting() or is_jit_tracing():
+            return x * torch.sigmoid(x - 1.0)
+        else:
+            return DoubleSwishFunction.apply(x)
+
+
+class ScaledEmbedding(nn.Module):
+    r"""This is a modified version of nn.Embedding that introduces a learnable scale
+    on the parameters.  Note: due to how we initialize it, it's best used with
+    schedulers like Noam that have a warmup period.
+
+    It is a simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+        initial_speed (float, optional):  This affects how fast the parameter will
+           learn near the start of training; you can set it to a value less than
+           one if you suspect that a module is contributing to instability near
+           the start of training.  Note: regardless of the use of this option,
+           it's best to use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.  Must be greater than 0.
+
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+
+    """
+    __constants__ = [
+        "num_embeddings",
+        "embedding_dim",
+        "padding_idx",
+        "scale_grad_by_freq",
+        "sparse",
+    ]
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        scale_grad_by_freq: bool = False,
+        sparse: bool = False,
+        initial_speed: float = 1.0,
+    ) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert (
+                    padding_idx < self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+            elif padding_idx < 0:
+                assert (
+                    padding_idx >= -self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale = nn.Parameter(torch.zeros(()))  # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters(initial_speed)
+
+    def reset_parameters(self, initial_speed: float = 1.0) -> None:
+        std = 0.1 / initial_speed
+        nn.init.normal_(self.weight, std=std)
+        nn.init.constant_(self.scale, torch.tensor(1.0 / std).log())
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        scale = self.scale.exp()
+        if input.numel() < self.num_embeddings:
+            return (
+                F.embedding(
+                    input,
+                    self.weight,
+                    self.padding_idx,
+                    None,
+                    2.0,  # None, 2.0 relate to normalization
+                    self.scale_grad_by_freq,
+                    self.sparse,
+                )
+                * scale
+            )
+        else:
+            return F.embedding(
+                input,
+                self.weight * scale,
+                self.padding_idx,
+                None,
+                2.0,  # None, 2.0 relates to normalization
+                self.scale_grad_by_freq,
+                self.sparse,
+            )
+
+    def extra_repr(self) -> str:
+        # s = "{num_embeddings}, {embedding_dim}, scale={scale}"
+        s = "{num_embeddings}, {embedding_dim}"
+        if self.padding_idx is not None:
+            s += ", padding_idx={padding_idx}"
+        if self.scale_grad_by_freq is not False:
+            s += ", scale_grad_by_freq={scale_grad_by_freq}"
+        if self.sparse is not False:
+            s += ", sparse=True"
+        return s.format(**self.__dict__)
+
+
+def _test_activation_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x**2).mean().sqrt()
+    y_rms = (y**2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 0.5
+    x.requires_grad = True
+    m = DoubleSwish()
+    torch.autograd.gradcheck(m, x)
+
+
+def _test_scaled_lstm():
+    N, L = 2, 30
+    dim_in, dim_hidden = 10, 20
+    m = ScaledLSTM(input_size=dim_in, hidden_size=dim_hidden, bias=True)
+    x = torch.randn(L, N, dim_in)
+    h0 = torch.randn(1, N, dim_hidden)
+    c0 = torch.randn(1, N, dim_hidden)
+    y, (h, c) = m(x, (h0, c0))
+    assert y.shape == (L, N, dim_hidden)
+    assert h.shape == (1, N, dim_hidden)
+    assert c.shape == (1, N, dim_hidden)
+
+
+def _test_grad_filter():
+    threshold = 50.0
+    time, batch, channel = 200, 5, 128
+    grad_filter = GradientFilter(batch_dim=1, threshold=threshold)
+
+    for i in range(2):
+        x = torch.randn(time, batch, channel, requires_grad=True)
+        w = nn.Parameter(torch.ones(5))
+        b = nn.Parameter(torch.zeros(5))
+
+        x_out, w_out, b_out = grad_filter(x, w, b)
+
+        w_out_grad = torch.randn_like(w)
+        b_out_grad = torch.randn_like(b)
+        x_out_grad = torch.rand_like(x)
+        if i % 2 == 1:
+            # The gradient norm of the first element must be larger than
+            # `threshold * median`, where `median` is the median value
+            # of gradient norms of all elements in batch.
+            x_out_grad[:, 0, :] = torch.full((time, channel), threshold)
+
+        torch.autograd.backward(
+            [x_out, w_out, b_out], [x_out_grad, w_out_grad, b_out_grad]
+        )
+
+        print(
+            "_test_grad_filter: for gradient norms, the first element > median * threshold ",  # noqa
+            i % 2 == 1,
+        )
+
+        print(
+            "_test_grad_filter: x_out_grad norm = ",
+            (x_out_grad**2).mean(dim=(0, 2)).sqrt(),
+        )
+        print(
+            "_test_grad_filter: x.grad norm = ",
+            (x.grad**2).mean(dim=(0, 2)).sqrt(),
+        )
+        print("_test_grad_filter: w_out_grad = ", w_out_grad)
+        print("_test_grad_filter: w.grad = ", w.grad)
+        print("_test_grad_filter: b_out_grad = ", b_out_grad)
+        print("_test_grad_filter: b.grad = ", b.grad)
+
+
+if __name__ == "__main__":
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()
+    _test_scaled_lstm()
+    _test_grad_filter()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py
new file mode 100644
index 000000000..e6e0fb1c8
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py
@@ -0,0 +1,282 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from decode_stream import DecodeStream
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    T = encoder_out.size(1)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (N, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+    num_active_paths: int = 4,
+) -> None:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      num_active_paths:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(num_active_paths)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    streams: List[DecodeStream],
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first generated by Fsa-based beam search, then we get the
+    recognition by applying shortest path on the lattice.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      streams:
+        A list of stream objects.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+    best_path = one_best_decoding(lattice)
+    hyp_tokens = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyp_tokens[i]
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_decode.py
new file mode 100755
index 000000000..9c4a13606
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/streaming_decode.py
@@ -0,0 +1,579 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless2/streaming_decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --left-context 32 \
+        --decode-chunk-size 8 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --decoding_method greedy_search \
+        --num-decode-streams 1000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        # sort results so we can easily compare the difference between two
+        # recognition results
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+    # Decoding in streaming requires causal convolution
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless2/test_model.py
new file mode 120000
index 000000000..4196e587c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/test_model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/test_model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
new file mode 100755
index 000000000..6c19f2cb0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,1113 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+# train a streaming model
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 25 \
+  --num-left-chunks 4 \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 0 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
new file mode 100644
index 000000000..057624272
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
@@ -0,0 +1,293 @@
+# Copyright      2021  Piotr Żelasko
+#                2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Optional
+
+from lhotse import CutSet, Fbank, FbankConfig
+from lhotse.dataset import (
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures, PrecomputedFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class AsrDataModule:
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler. "
+            "(you might want to increase it for larger datasets).",
+        )
+
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--on-the-fly-num-workers",
+            type=int,
+            default=0,
+            help="The number of workers for on-the-fly feature extraction",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available. Used only in dev/test CutSet",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        on_the_fly_feats: bool,
+        cuts_musan: Optional[CutSet] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            Cuts for training.
+          cuts_musan:
+            If not None, it is the cuts for mixing.
+          on_the_fly_feats:
+            True to use OnTheFlyFeatures;
+            False to use PrecomputedFeatures.
+        """
+        transforms = []
+        if cuts_musan is not None:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        input_transforms = []
+
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=2,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        # NOTE: the PerturbSpeed transform should be added only if we
+        # remove it from data prep stage.
+        # Add on-the-fly speed perturbation; since originally it would
+        # have increased epoch size by 3, we will apply prob 2/3 and use
+        # 3x more epochs.
+        # Speed perturbation probably should come first before
+        # concatenation, but in principle the transforms order doesn't have
+        # to be strict (e.g. could be randomized)
+        # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+        # Drop feats to be on the safe side.
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_strategy=(
+                OnTheFlyFeatures(
+                    extractor=Fbank(FbankConfig(num_mel_bins=80)),
+                    num_workers=self.args.on_the_fly_num_workers,
+                )
+                if on_the_fly_feats
+                else PrecomputedFeatures()
+            ),
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        logging.info("Using DynamicBucketingSampler.")
+        train_sampler = DynamicBucketingSampler(
+            cuts_train,
+            max_duration=self.args.max_duration,
+            shuffle=self.args.shuffle,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
+        )
+
+        logging.info("About to create train dataloader")
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless3/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py
new file mode 120000
index 000000000..3b84b9573
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py
new file mode 100755
index 000000000..b4804ecde
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py
@@ -0,0 +1,631 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless3/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --max-duration 600 \
+        --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless3/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --max-duration 600 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless3/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --max-duration 600 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless3/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --max-duration 600 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from gigaspeech import GigaSpeech
+from gigaspeech_scoring import asr_text_post_processing
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for computed nbest oracle WER
+        when the decoding method is fast_beam_search_nbest_oracle.
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding_method is fast_beam_search_nbest_oracle.
+        """,
+    )
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[List[str], List[str]]],
+) -> List[Tuple[List[str], List[str]]]:
+    new_results = []
+    for ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is
+        fast_beam_search or fast_beam_search_nbest_oracle.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}_"
+                f"num_paths_{params.num_paths}_"
+                f"nbest_scale_{params.nbest_scale}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = post_processing(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / "giga" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        params.suffix += f"-num-paths-{params.num_paths}"
+        params.suffix += f"-nbest-scale-{params.nbest_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.unk_id()
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+    model.unk_id = params.unk_id
+
+    # In beam_search.py, we are using model.decoder() and model.joiner(),
+    # so we have to switch to the branch for the GigaSpeech dataset.
+    model.decoder = model.decoder_giga
+    model.joiner = model.joiner_giga
+
+    if params.decoding_method in (
+        "fast_beam_search",
+        "fast_beam_search_nbest_oracle",
+    ):
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+
+    test_cuts = gigaspeech.test_cuts()
+    dev_cuts = gigaspeech.dev_cuts()
+
+    test_dl = asr_datamodule.test_dataloaders(test_cuts)
+    dev_dl = asr_datamodule.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_sets_dl = [test_dl, dev_dl]
+
+    for test_set, dl in zip(test_sets, test_sets_dl):
+        results_dict = decode_dataset(
+            dl=dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py
new file mode 100755
index 000000000..7c62bfa58
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py
@@ -0,0 +1,1225 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search (with LM shallow fusion)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+(9) modified beam search with LM shallow fusion + LODR
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --max-duration 600 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --decoding-method modified_beam_search_LODR \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.4 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16 \
+"""
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    fast_beam_search_with_nbest_rescoring,
+    fast_beam_search_with_nbest_rnn_rescoring,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+    modified_beam_search_ngram_rescoring,
+)
+from librispeech import LibriSpeech
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_ngram_rescoring
+          - modified_beam_search_lm_shallow_fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is, fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling). Set -1 to use full attention.",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=1.0,
+        help="""Softmax temperature.
+         The output of the model is (logits / temperature).log_softmax().
+         """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=Path,
+        default=Path("./data/lm"),
+        help="""Used only when --decoding-method is
+         fast_beam_search_with_nbest_rescoring.
+         It should contain either G_4_gram.pt or G_4_gram.fst.txt
+         """,
+    )
+
+    parser.add_argument(
+        "--words-txt",
+        type=Path,
+        default=Path("./data/lang_bpe_500/words.txt"),
+        help="""Used only when --decoding-method is
+         fast_beam_search_with_nbest_rescoring.
+         It is the word table.
+         """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-scale",
+        type=float,
+        default=0.0,
+        help="""Used only when --method is modified-beam-search_rnnlm_shallow_fusion.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=3,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring""",
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    G: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    rnn_lm_model: Optional[RnnLmModel] = None,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      G:
+        Optional. Used only when decoding method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_oracle,
+        or fast_beam_search_with_nbest_rescoring.
+        It an FsaVec containing an acceptor.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        if params.decode_chunk_size > 0:
+            # except the case of using full attention
+            feature_lens += params.left_context
+            feature = torch.nn.functional.pad(
+                feature,
+                pad=(0, 0, 0, params.left_context),
+                value=LOG_EPS,
+            )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            temperature=params.temperature,
+        )
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            temperature=params.temperature,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            temperature=params.temperature,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            temperature=params.temperature,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            temperature=params.temperature,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_with_nbest_rescoring":
+        ngram_lm_scale_list = [-0.5, -0.2, -0.1, -0.05, -0.02, 0]
+        ngram_lm_scale_list += [0.01, 0.02, 0.05]
+        ngram_lm_scale_list += [0.1, 0.3, 0.5, 0.8]
+        ngram_lm_scale_list += [1.0, 1.5, 2.5, 3]
+        hyp_tokens = fast_beam_search_with_nbest_rescoring(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+            ngram_lm_scale_list=ngram_lm_scale_list,
+            num_paths=params.num_paths,
+            G=G,
+            sp=sp,
+            word_table=word_table,
+            use_double_scores=True,
+            nbest_scale=params.nbest_scale,
+            temperature=params.temperature,
+        )
+    elif params.decoding_method == "fast_beam_search_with_nbest_rnn_rescoring":
+        ngram_lm_scale_list = [-0.5, -0.2, -0.1, -0.05, -0.02, 0]
+        ngram_lm_scale_list += [0.01, 0.02, 0.05]
+        ngram_lm_scale_list += [0.1, 0.3, 0.5, 0.8]
+        ngram_lm_scale_list += [1.0, 1.5, 2.5, 3]
+        hyp_tokens = fast_beam_search_with_nbest_rnn_rescoring(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+            ngram_lm_scale_list=ngram_lm_scale_list,
+            num_paths=params.num_paths,
+            G=G,
+            sp=sp,
+            word_table=word_table,
+            rnn_lm_model=rnn_lm_model,
+            rnn_lm_scale_list=ngram_lm_scale_list,
+            use_double_scores=True,
+            nbest_scale=params.nbest_scale,
+            temperature=params.temperature,
+        )
+    elif params.decoding_method == "modified_beam_search_ngram_rescoring":
+        hyp_tokens = modified_beam_search_ngram_rescoring(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+                f"temperature_{params.temperature}"
+            ): hyps
+        }
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+                f"temperature_{params.temperature}"
+            ): hyps
+        }
+    elif params.decoding_method in [
+        "fast_beam_search_with_nbest_rescoring",
+        "fast_beam_search_with_nbest_rnn_rescoring",
+    ]:
+        prefix = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}_"
+            f"num_paths_{params.num_paths}_"
+            f"nbest_scale_{params.nbest_scale}_"
+            f"temperature_{params.temperature}_"
+        )
+        ans: Dict[str, List[List[str]]] = {}
+        for key, hyp in hyp_tokens.items():
+            t: List[str] = sp.decode(hyp)
+            ans[prefix + key] = [s.split() for s in t]
+        return ans
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+        return {key: hyps}
+    else:
+        return {
+            (f"beam_size_{params.beam_size}_temperature_{params.temperature}"): hyps
+        }
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    G: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    rnn_lm_model: Optional[RnnLmModel] = None,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      G:
+        Optional. Used only when decoding method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_oracle,
+        or fast_beam_search_with_nbest_rescoring.
+        It's an FsaVec containing an acceptor.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+            G=G,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            rnn_lm_model=rnn_lm_model,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+def load_ngram_LM(
+    lm_dir: Path, word_table: k2.SymbolTable, device: torch.device
+) -> k2.Fsa:
+    """Read a ngram model from the given directory.
+    Args:
+      lm_dir:
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+      word_table:
+        The word table mapping words to IDs and vice versa.
+      device:
+        The resulting FSA will be moved to this device.
+    Returns:
+      Return an FsaVec containing a single acceptor.
+    """
+    lm_dir = Path(lm_dir)
+    assert lm_dir.is_dir(), f"{lm_dir} does not exist"
+
+    pt_file = lm_dir / "G_4_gram.pt"
+
+    if pt_file.is_file():
+        logging.info(f"Loading pre-compiled {pt_file}")
+        d = torch.load(pt_file, map_location=device)
+        G = k2.Fsa.from_dict(d)
+        G = k2.add_epsilon_self_loops(G)
+        G = k2.arc_sort(G)
+        return G
+
+    txt_file = lm_dir / "G_4_gram.fst.txt"
+
+    assert txt_file.is_file(), f"{txt_file} does not exist"
+    logging.info(f"Loading {txt_file}")
+    logging.warning("It may take 8 minutes (Will be cached for later use).")
+    with open(txt_file) as f:
+        G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+
+        # G.aux_labels is not needed in later computations, so
+        # remove it here.
+        del G.aux_labels
+        # Now G is an acceptor
+
+        first_word_disambig_id = word_table["#0"]
+        # CAUTION: The following line is crucial.
+        # Arcs entering the back-off state have label equal to #0.
+        # We have to change it to 0 here.
+        G.labels[G.labels >= first_word_disambig_id] = 0
+
+        # See https://github.com/k2-fsa/k2/issues/874
+        # for why we need to set G.properties to None
+        G.__dict__["_properties"] = None
+
+        G = k2.Fsa.from_fsas([G]).to(device)
+
+        # Save a dummy value so that it can be loaded in C++.
+        # See https://github.com/pytorch/pytorch/issues/67902
+        # for why we need to do this.
+        G.dummy = 1
+
+        logging.info(f"Saving to {pt_file} for later use")
+        torch.save(G.as_dict(), pt_file)
+
+        G = k2.add_epsilon_self_loops(G)
+        G = k2.arc_sort(G)
+        return G
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "fast_beam_search_with_nbest_rescoring",
+        "fast_beam_search_with_nbest_rnn_rescoring",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_ngram_rescoring",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        params.suffix += f"-temperature-{params.temperature}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        params.suffix += f"-temperature-{params.temperature}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+        params.suffix += f"-temperature-{params.temperature}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+    model.unk_id = params.unk_id
+
+    G = None
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        elif params.decoding_method in [
+            "fast_beam_search_with_nbest_rescoring",
+            "fast_beam_search_with_nbest_rnn_rescoring",
+        ]:
+            logging.info(f"Loading word symbol table from {params.words_txt}")
+            word_table = k2.SymbolTable.from_file(params.words_txt)
+
+            G = load_ngram_LM(
+                lm_dir=params.lm_dir,
+                word_table=word_table,
+                device=device,
+            )
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+            logging.info(f"G properties_str: {G.properties_str}")
+            rnn_lm_model = None
+            if params.decoding_method == "fast_beam_search_with_nbest_rnn_rescoring":
+                rnn_lm_model = RnnLmModel(
+                    vocab_size=params.vocab_size,
+                    embedding_dim=params.rnn_lm_embedding_dim,
+                    hidden_dim=params.rnn_lm_hidden_dim,
+                    num_layers=params.rnn_lm_num_layers,
+                    tie_weights=params.rnn_lm_tie_weights,
+                )
+                if params.rnn_lm_avg == 1:
+                    load_checkpoint(
+                        f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                        rnn_lm_model,
+                    )
+                    rnn_lm_model.to(device)
+                else:
+                    rnn_lm_model = load_averaged_model(
+                        params.rnn_lm_exp_dir,
+                        rnn_lm_model,
+                        params.rnn_lm_epoch,
+                        params.rnn_lm_avg,
+                        device,
+                    )
+                rnn_lm_model.eval()
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+            rnn_lm_model = None
+    else:
+        decoding_graph = None
+        word_table = None
+        rnn_lm_model = None
+
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+
+    else:
+        LM = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            G=G,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            rnn_lm_model=rnn_lm_model,
+            LM=LM,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode_stream.py
new file mode 120000
index 000000000..30f264813
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode_stream.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/decode_stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless3/export-onnx.py
new file mode 100755
index 000000000..2685ea95a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/export-onnx.py
@@ -0,0 +1,540 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-1224000-avg-14.pt"
+
+cd exp
+ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless3/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import num_tokens, setup_logger
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless3",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/export.py b/egs/librispeech/ASR/pruned_transducer_stateless3/export.py
new file mode 100755
index 000000000..925b15646
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/export.py
@@ -0,0 +1,447 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+It will also generate 3 other files: `encoder_jit_script.pt`,
+`decoder_jit_script.pt`, and `joiner_jit_script.pt`.
+
+(2) Export to torchscript model using torch.jit.trace()
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10 \
+  --jit-trace 1
+
+It will generates 3 files: `encoder_jit_trace.pt`,
+`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`.
+
+(3) Export `model.state_dict()`
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless3/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless3/decode.py \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate 4 files:
+         - encoder_jit_script.pt
+         - decoder_jit_script.pt
+         - joiner_jit_script.pt
+         - cpu_jit.pt (which combines the above 3 files)
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace.
+        It will generate 3 files:
+         - encoder_jit_trace.pt
+         - decoder_jit_trace.pt
+         - joiner_jit_trace.pt
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_script(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.script()
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    script_model = torch.jit.script(encoder_model)
+    script_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_script(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.script()
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    script_model = torch.jit.script(decoder_model)
+    script_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_script(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+    """
+    script_model = torch.jit.script(joiner_model)
+    script_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+
+        # Also export encoder/decoder/joiner separately
+        encoder_filename = params.exp_dir / "encoder_jit_script.pt"
+        export_encoder_model_jit_script(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_script.pt"
+        export_decoder_model_jit_script(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_script.pt"
+        export_joiner_model_jit_script(model.joiner, joiner_filename)
+
+    elif params.jit_trace is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    else:
+        logging.info("Not using torchscript")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py
new file mode 100644
index 000000000..f3bd6284e
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py
@@ -0,0 +1,92 @@
+# Copyright      2021  Piotr Żelasko
+#                2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import glob
+import logging
+import re
+from pathlib import Path
+
+import lhotse
+from lhotse import CutSet, load_manifest_lazy
+
+
+class GigaSpeech:
+    def __init__(self, manifest_dir: str):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files:
+
+                - gigaspeech_XL_split_2000/gigaspeech_cuts_XL.*.jsonl.gz
+                - gigaspeech_cuts_L_raw.jsonl.gz
+                - gigaspeech_cuts_M_raw.jsonl.gz
+                - gigaspeech_cuts_S_raw.jsonl.gz
+                - gigaspeech_cuts_XS_raw.jsonl.gz
+                - gigaspeech_cuts_DEV_raw.jsonl.gz
+                - gigaspeech_cuts_TEST_raw.jsonl.gz
+        """
+        self.manifest_dir = Path(manifest_dir)
+
+    def train_XL_cuts(self) -> CutSet:
+        logging.info("About to get train-XL cuts")
+
+        filenames = list(
+            glob.glob(
+                f"{self.manifest_dir}/gigaspeech_XL_split_2000/gigaspeech_cuts_XL.*.jsonl.gz"  # noqa
+            )
+        )
+
+        pattern = re.compile(r"gigaspeech_cuts_XL.([0-9]+).jsonl.gz")
+        idx_filenames = [(int(pattern.search(f).group(1)), f) for f in filenames]
+        idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
+
+        sorted_filenames = [f[1] for f in idx_filenames]
+
+        logging.info(f"Loading {len(sorted_filenames)} splits")
+
+        return lhotse.combine(lhotse.load_manifest_lazy(p) for p in sorted_filenames)
+
+    def train_L_cuts(self) -> CutSet:
+        f = self.manifest_dir / "gigaspeech_cuts_L_raw.jsonl.gz"
+        logging.info(f"About to get train-L cuts from {f}")
+        return CutSet.from_jsonl_lazy(f)
+
+    def train_M_cuts(self) -> CutSet:
+        f = self.manifest_dir / "gigaspeech_cuts_M_raw.jsonl.gz"
+        logging.info(f"About to get train-M cuts from {f}")
+        return CutSet.from_jsonl_lazy(f)
+
+    def train_S_cuts(self) -> CutSet:
+        f = self.manifest_dir / "gigaspeech_cuts_S_raw.jsonl.gz"
+        logging.info(f"About to get train-S cuts from {f}")
+        return CutSet.from_jsonl_lazy(f)
+
+    def train_XS_cuts(self) -> CutSet:
+        f = self.manifest_dir / "gigaspeech_cuts_XS_raw.jsonl.gz"
+        logging.info(f"About to get train-XS cuts from {f}")
+        return CutSet.from_jsonl_lazy(f)
+
+    def test_cuts(self) -> CutSet:
+        f = self.manifest_dir / "gigaspeech_cuts_TEST.jsonl.gz"
+        logging.info(f"About to get TEST cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def dev_cuts(self) -> CutSet:
+        f = self.manifest_dir / "gigaspeech_cuts_DEV.jsonl.gz"
+        logging.info(f"About to get DEV cuts from {f}")
+        return load_manifest_lazy(f)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py
new file mode 120000
index 000000000..b89dd0a38
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py
@@ -0,0 +1 @@
+../../../gigaspeech/ASR/conformer_ctc/gigaspeech_scoring.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless3/jit_pretrained.py
new file mode 100755
index 000000000..0669284b3
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/jit_pretrained.py
@@ -0,0 +1,335 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, either exported by `torch.jit.trace()`
+or by `torch.jit.script()`, and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit-trace 1
+
+or
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless3/jit_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless3/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./pruned_transducer_stateless3/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./pruned_transducer_stateless3/exp/joiner_jit_trace.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+or
+
+./pruned_transducer_stateless3/jit_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless3/exp/encoder_jit_script.pt \
+  --decoder-model-filename ./pruned_transducer_stateless3/exp/decoder_jit_script.pt \
+  --joiner-model-filename ./pruned_transducer_stateless3/exp/joiner_jit_script.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_size: int,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      decoder:
+        The decoder model.
+      joiner:
+        The joiner model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+      context_size:
+        The context size of the decoder model.
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        decoder=decoder,
+        joiner=joiner,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        context_size=args.context_size,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py b/egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py
new file mode 100644
index 000000000..9f2cb6225
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py
@@ -0,0 +1,83 @@
+# Copyright      2021  Piotr Żelasko
+#                2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+
+
+class LibriSpeech:
+    def __init__(self, manifest_dir: str):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files::
+
+                - librispeech_cuts_dev-clean.jsonl.gz
+                - librispeech_cuts_dev-other.jsonl.gz
+                - librispeech_cuts_test-clean.jsonl.gz
+                - librispeech_cuts_test-other.jsonl.gz
+                - librispeech_cuts_train-clean-100.jsonl.gz
+                - librispeech_cuts_train-clean-360.jsonl.gz
+                - librispeech_cuts_train-other-500.jsonl.gz
+        """
+        self.manifest_dir = Path(manifest_dir)
+
+    def train_clean_100_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_train-clean-100.jsonl.gz"
+        logging.info(f"About to get train-clean-100 cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def train_clean_360_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_train-clean-360.jsonl.gz"
+        logging.info(f"About to get train-clean-360 cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def train_other_500_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_train-other-500.jsonl.gz"
+        logging.info(f"About to get train-other-500 cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def test_clean_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        logging.info(f"About to get test-clean cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def test_other_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        logging.info(f"About to get test-other cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def dev_clean_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        logging.info(f"About to get dev-clean cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def dev_other_cuts(self) -> CutSet:
+        f = self.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        logging.info(f"About to get dev-other cuts from {f}")
+        return load_manifest_lazy(f)
+
+    def train_all_shuf_cuts(self) -> CutSet:
+        logging.info(
+            "About to get the shuffled train-clean-100, \
+            train-clean-360 and train-other-500 cuts"
+        )
+        return load_manifest_lazy(
+            self.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
+        )
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/lstmp.py b/egs/librispeech/ASR/pruned_transducer_stateless3/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/model.py b/egs/librispeech/ASR/pruned_transducer_stateless3/model.py
new file mode 100644
index 000000000..d45f6dadc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/model.py
@@ -0,0 +1,245 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        decoder_giga: Optional[nn.Module] = None,
+        joiner_giga: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim). Its output shape is (N, T, U, vocab_size).
+            Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          encoder_dim:
+            Output dimension of the encoder network.
+          decoder_dim:
+            Output dimension of the decoder network.
+          joiner_dim:
+            Input dimension of the joiner network.
+          vocab_size:
+            Output dimension of the joiner network.
+          decoder_giga:
+            Optional. The decoder network for the GigaSpeech dataset.
+          joiner_giga:
+            Optional. The joiner network for the GigaSpeech dataset.
+        """
+        super().__init__()
+
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_giga = decoder_giga
+        self.joiner_giga = joiner_giga
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+        if decoder_giga is not None:
+            self.simple_am_proj_giga = ScaledLinear(
+                encoder_dim, vocab_size, initial_speed=0.5
+            )
+            self.simple_lm_proj_giga = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        libri: bool = True,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        reduction: str = "sum",
+        delay_penalty: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          libri:
+            True to use the decoder and joiner for the LibriSpeech dataset.
+            False to use the decoder and joiner for the GigaSpeech dataset.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          delay_penalty:
+            A constant value used to penalize symbol delay, to encourage
+            streaming models to emit symbols earlier.
+            See https://github.com/k2-fsa/k2/issues/955 and
+            https://arxiv.org/pdf/2211.00490.pdf for more details.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(encoder_out_lens > 0)
+
+        if libri:
+            decoder = self.decoder
+            simple_lm_proj = self.simple_lm_proj
+            simple_am_proj = self.simple_am_proj
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_giga
+            simple_lm_proj = self.simple_lm_proj_giga
+            simple_am_proj = self.simple_am_proj_giga
+            joiner = self.joiner_giga
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = encoder_out_lens
+
+        lm = simple_lm_proj(decoder_out)
+        am = simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                delay_penalty=delay_penalty,
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=joiner.encoder_proj(encoder_out),
+            lm=joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                delay_penalty=delay_penalty,
+                reduction=reduction,
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py
new file mode 100755
index 000000000..5ca4173c1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py
@@ -0,0 +1,239 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script checks that exported onnx models produce the same output
+with the given torchscript model for the same input.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-1224000-avg-14.pt"
+
+cd exp
+ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+popd
+
+2. Export the model via torchscript (torch.jit.script())
+
+./pruned_transducer_stateless3/export.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --jit 1
+
+It will generate the following file in $repo/exp:
+    - cpu_jit.pt
+
+3. Export the model to ONNX
+
+./pruned_transducer_stateless3/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+4. Run this file
+
+./pruned_transducer_stateless3/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-9999-avg-1.onnx
+"""
+
+import argparse
+import logging
+
+from icefall import is_module_available
+from onnx_pretrained import OnnxModel
+
+import torch
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript model",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    C = 80
+    for i in range(3):
+        N = torch.randint(low=1, high=20, size=(1,)).item()
+        T = torch.randint(low=30, high=50, size=(1,)).item()
+        logging.info(f"test_encoder: iter {i}, N={N}, T={T}")
+
+        x = torch.rand(N, T, C)
+        x_lens = torch.randint(low=30, high=T + 1, size=(N,))
+        x_lens[0] = T
+
+        torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens)
+        torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out)
+
+        onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens)
+
+        assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), (
+            (torch_encoder_out - onnx_encoder_out).abs().max()
+        )
+
+
+def test_decoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    context_size = onnx_model.context_size
+    vocab_size = onnx_model.vocab_size
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_decoder: iter {i}, N={N}")
+        x = torch.randint(
+            low=1,
+            high=vocab_size,
+            size=(N, context_size),
+            dtype=torch.int64,
+        )
+        torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False]))
+        torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out)
+        torch_decoder_out = torch_decoder_out.squeeze(1)
+
+        onnx_decoder_out = onnx_model.run_decoder(x)
+        assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
+            (torch_decoder_out - onnx_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1]
+    decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1]
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_joiner: iter {i}, N={N}")
+        encoder_out = torch.rand(N, encoder_dim)
+        decoder_out = torch.rand(N, decoder_dim)
+
+        projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out)
+        projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out)
+
+        torch_joiner_out = torch_model.joiner(encoder_out, decoder_out)
+        onnx_joiner_out = onnx_model.run_joiner(
+            projected_encoder_out, projected_decoder_out
+        )
+
+        assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
+            (torch_joiner_out - onnx_joiner_out).abs().max()
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_model = torch.jit.load(args.jit_filename)
+
+    onnx_model = OnnxModel(
+        encoder_model_filename=args.onnx_encoder_filename,
+        decoder_model_filename=args.onnx_decoder_filename,
+        joiner_model_filename=args.onnx_joiner_filename,
+    )
+
+    logging.info("Test encoder")
+    test_encoder(torch_model, onnx_model)
+
+    logging.info("Test decoder")
+    test_decoder(torch_model, onnx_model)
+
+    logging.info("Test joiner")
+    test_joiner(torch_model, onnx_model)
+    logging.info("Finished checking ONNX models")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# See https://github.com/pytorch/pytorch/issues/38342
+# and https://github.com/pytorch/pytorch/issues/33354
+#
+# If we don't do this, the delay increases whenever there is
+# a new request that changes the actual batch size.
+# If you use `py-spy dump --pid <server-pid> --native`, you will
+# see a lot of time is spent in re-compiling the torch script model.
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    torch.manual_seed(20220727)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py
new file mode 100755
index 000000000..3b1c72cf1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py
@@ -0,0 +1,321 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-1224000-avg-14.pt"
+
+cd exp
+ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless3/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+2. Run this file
+
+./pruned_transducer_stateless3/onnx_decode.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --max-duration 600 \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+"""
+
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from librispeech import LibriSpeech
+
+from onnx_pretrained import greedy_search, OnnxModel
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, sp: spm.SentencePieceProcessor, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    hyps = [sp.decode(h).split() for h in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, sp=sp, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    blank_id = sp.piece_to_id("<blk>")
+    assert blank_id == 0, blank_id
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(dl=test_dl, model=model, sp=sp)
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_pretrained.py
new file mode 100755
index 000000000..de3e03da6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_pretrained.py
@@ -0,0 +1,420 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-iter-1224000-avg-14.pt"
+
+cd exp
+ln -s pretrained-iter-1224000-avg-14.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless3/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+3. Run this file
+
+./pruned_transducer_stateless3/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 4
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 2-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, its shape is (N, T', joiner_dim)
+            - encoder_out_lens, its shape is (N,)
+        """
+        out = self.encoder.run(
+            [
+                self.encoder.get_outputs()[0].name,
+                self.encoder.get_outputs()[1].name,
+            ],
+            {
+                self.encoder.get_inputs()[0].name: x.numpy(),
+                self.encoder.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, joiner_dim)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.run_decoder(decoder_input)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        # current_encoder_out's shape: (batch_size, joiner_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+        logits = model.run_joiner(current_encoder_out, decoder_out)
+
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                dtype=torch.int64,
+            )
+            decoder_out = model.run_decoder(decoder_input)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += symbol_table[i]
+        return text.replace("▁", " ").strip()
+
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = token_ids_to_words(hyp)
+        s += f"{filename}:\n{words}\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless3/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py
new file mode 100755
index 000000000..abda4e2d4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py
@@ -0,0 +1,398 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless3/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless3/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless3/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless3/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless3/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless3/exp/pretrained.pt is generated by
+./pruned_transducer_stateless3/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    if params.simulate_streaming:
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py
new file mode 100644
index 000000000..a6540c584
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py
@@ -0,0 +1,320 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file provides functions to convert `ScaledLinear`, `ScaledConv1d`,
+`ScaledConv2d`, and `ScaledEmbedding` to their non-scaled counterparts:
+`nn.Linear`, `nn.Conv1d`, `nn.Conv2d`, and `nn.Embedding`.
+
+The scaled version are required only in the training time. It simplifies our
+life by converting them to their non-scaled version during inference.
+"""
+
+import copy
+import re
+from typing import List
+
+import torch
+import torch.nn as nn
+from lstmp import LSTMP
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledEmbedding,
+    ScaledLinear,
+    ScaledLSTM,
+)
+
+
+class NonScaledNorm(nn.Module):
+    """See BasicNorm for doc"""
+
+    def __init__(
+        self,
+        num_channels: int,
+        eps_exp: float,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+    ):
+        super().__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.eps_exp = eps_exp
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not torch.jit.is_tracing():
+            assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp
+        ).pow(-0.5)
+        return x * scales
+
+
+def scaled_linear_to_linear(scaled_linear: ScaledLinear) -> nn.Linear:
+    """Convert an instance of ScaledLinear to nn.Linear.
+
+    Args:
+      scaled_linear:
+        The layer to be converted.
+    Returns:
+      Return a linear layer. It satisfies:
+
+        scaled_linear(x) == linear(x)
+
+      for any given input tensor `x`.
+    """
+    assert isinstance(scaled_linear, ScaledLinear), type(scaled_linear)
+
+    weight = scaled_linear.get_weight()
+    bias = scaled_linear.get_bias()
+    has_bias = bias is not None
+
+    linear = torch.nn.Linear(
+        in_features=scaled_linear.in_features,
+        out_features=scaled_linear.out_features,
+        bias=True,  # otherwise, it throws errors when converting to PNNX format
+        # device=weight.device,  # Pytorch version before v1.9.0 does not have
+        # this argument. Comment out for now, we will
+        # see if it will raise error for versions
+        # after v1.9.0
+    )
+    linear.weight.data.copy_(weight)
+
+    if has_bias:
+        linear.bias.data.copy_(bias)
+    else:
+        linear.bias.data.zero_()
+
+    return linear
+
+
+def scaled_conv1d_to_conv1d(scaled_conv1d: ScaledConv1d) -> nn.Conv1d:
+    """Convert an instance of ScaledConv1d to nn.Conv1d.
+
+    Args:
+      scaled_conv1d:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Conv1d that has the same `forward()` behavior
+      of the given `scaled_conv1d`.
+    """
+    assert isinstance(scaled_conv1d, ScaledConv1d), type(scaled_conv1d)
+
+    weight = scaled_conv1d.get_weight()
+    bias = scaled_conv1d.get_bias()
+    has_bias = bias is not None
+
+    conv1d = nn.Conv1d(
+        in_channels=scaled_conv1d.in_channels,
+        out_channels=scaled_conv1d.out_channels,
+        kernel_size=scaled_conv1d.kernel_size,
+        stride=scaled_conv1d.stride,
+        padding=scaled_conv1d.padding,
+        dilation=scaled_conv1d.dilation,
+        groups=scaled_conv1d.groups,
+        bias=scaled_conv1d.bias is not None,
+        padding_mode=scaled_conv1d.padding_mode,
+    )
+
+    conv1d.weight.data.copy_(weight)
+    if has_bias:
+        conv1d.bias.data.copy_(bias)
+
+    return conv1d
+
+
+def scaled_conv2d_to_conv2d(scaled_conv2d: ScaledConv2d) -> nn.Conv2d:
+    """Convert an instance of ScaledConv2d to nn.Conv2d.
+
+    Args:
+      scaled_conv2d:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Conv2d that has the same `forward()` behavior
+      of the given `scaled_conv2d`.
+    """
+    assert isinstance(scaled_conv2d, ScaledConv2d), type(scaled_conv2d)
+
+    weight = scaled_conv2d.get_weight()
+    bias = scaled_conv2d.get_bias()
+    has_bias = bias is not None
+
+    conv2d = nn.Conv2d(
+        in_channels=scaled_conv2d.in_channels,
+        out_channels=scaled_conv2d.out_channels,
+        kernel_size=scaled_conv2d.kernel_size,
+        stride=scaled_conv2d.stride,
+        padding=scaled_conv2d.padding,
+        dilation=scaled_conv2d.dilation,
+        groups=scaled_conv2d.groups,
+        bias=scaled_conv2d.bias is not None,
+        padding_mode=scaled_conv2d.padding_mode,
+    )
+
+    conv2d.weight.data.copy_(weight)
+    if has_bias:
+        conv2d.bias.data.copy_(bias)
+
+    return conv2d
+
+
+def scaled_embedding_to_embedding(
+    scaled_embedding: ScaledEmbedding,
+) -> nn.Embedding:
+    """Convert an instance of ScaledEmbedding to nn.Embedding.
+
+    Args:
+      scaled_embedding:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Embedding that has the same `forward()` behavior
+      of the given `scaled_embedding`.
+    """
+    assert isinstance(scaled_embedding, ScaledEmbedding), type(scaled_embedding)
+    embedding = nn.Embedding(
+        num_embeddings=scaled_embedding.num_embeddings,
+        embedding_dim=scaled_embedding.embedding_dim,
+        padding_idx=scaled_embedding.padding_idx,
+        scale_grad_by_freq=scaled_embedding.scale_grad_by_freq,
+        sparse=scaled_embedding.sparse,
+    )
+    weight = scaled_embedding.weight
+    scale = scaled_embedding.scale
+
+    embedding.weight.data.copy_(weight * scale.exp())
+
+    return embedding
+
+
+def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm:
+    assert isinstance(basic_norm, BasicNorm), type(BasicNorm)
+    norm = NonScaledNorm(
+        num_channels=basic_norm.num_channels,
+        eps_exp=basic_norm.eps.data.exp().item(),
+        channel_dim=basic_norm.channel_dim,
+    )
+    return norm
+
+
+def scaled_lstm_to_lstm(scaled_lstm: ScaledLSTM) -> nn.LSTM:
+    """Convert an instance of ScaledLSTM to nn.LSTM.
+
+    Args:
+      scaled_lstm:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.LSTM that has the same `forward()` behavior
+      of the given `scaled_lstm`.
+    """
+    assert isinstance(scaled_lstm, ScaledLSTM), type(scaled_lstm)
+    lstm = nn.LSTM(
+        input_size=scaled_lstm.input_size,
+        hidden_size=scaled_lstm.hidden_size,
+        num_layers=scaled_lstm.num_layers,
+        bias=scaled_lstm.bias,
+        batch_first=scaled_lstm.batch_first,
+        dropout=scaled_lstm.dropout,
+        bidirectional=scaled_lstm.bidirectional,
+        proj_size=scaled_lstm.proj_size,
+    )
+
+    assert lstm._flat_weights_names == scaled_lstm._flat_weights_names
+    for idx in range(len(scaled_lstm._flat_weights_names)):
+        scaled_weight = scaled_lstm._flat_weights[idx] * scaled_lstm._scales[idx].exp()
+        lstm._flat_weights[idx].data.copy_(scaled_weight)
+
+    return lstm
+
+
+# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule  # noqa
+# get_submodule was added to nn.Module at v1.9.0
+def get_submodule(model, target):
+    if target == "":
+        return model
+    atoms: List[str] = target.split(".")
+    mod: torch.nn.Module = model
+    for item in atoms:
+        if not hasattr(mod, item):
+            raise AttributeError(
+                mod._get_name() + " has no " "attribute `" + item + "`"
+            )
+        mod = getattr(mod, item)
+        if not isinstance(mod, torch.nn.Module):
+            raise AttributeError("`" + item + "` is not " "an nn.Module")
+    return mod
+
+
+def convert_scaled_to_non_scaled(
+    model: nn.Module,
+    inplace: bool = False,
+    is_onnx: bool = False,
+):
+    """Convert `ScaledLinear`, `ScaledConv1d`, and `ScaledConv2d`
+    in the given modle to their unscaled version `nn.Linear`, `nn.Conv1d`,
+    and `nn.Conv2d`.
+
+    Args:
+      model:
+        The model to be converted.
+      inplace:
+        If True, the input model is modified inplace.
+        If False, the input model is copied and we modify the copied version.
+      is_onnx:
+        If True, we are going to export the model to ONNX. In this case,
+        we will convert nn.LSTM with proj_size to LSTMP.
+    Return:
+      Return a model without scaled layers.
+    """
+    if not inplace:
+        model = copy.deepcopy(model)
+
+    excluded_patterns = r"(self|src)_attn\.(in|out)_proj"
+    p = re.compile(excluded_patterns)
+
+    d = {}
+    for name, m in model.named_modules():
+        if isinstance(m, ScaledLinear):
+            if p.search(name) is not None:
+                continue
+            d[name] = scaled_linear_to_linear(m)
+        elif isinstance(m, ScaledConv1d):
+            d[name] = scaled_conv1d_to_conv1d(m)
+        elif isinstance(m, ScaledConv2d):
+            d[name] = scaled_conv2d_to_conv2d(m)
+        elif isinstance(m, ScaledEmbedding):
+            d[name] = scaled_embedding_to_embedding(m)
+        elif isinstance(m, BasicNorm):
+            d[name] = convert_basic_norm(m)
+        elif isinstance(m, ScaledLSTM):
+            if is_onnx:
+                d[name] = LSTMP(scaled_lstm_to_lstm(m))
+                # See
+                # https://github.com/pytorch/pytorch/issues/47887
+                #  d[name] = torch.jit.script(LSTMP(scaled_lstm_to_lstm(m)))
+            else:
+                d[name] = scaled_lstm_to_lstm(m)
+        elif isinstance(m, ActivationBalancer):
+            d[name] = nn.Identity()
+
+    for k, v in d.items():
+        if "." in k:
+            parent, child = k.rsplit(".", maxsplit=1)
+            setattr(get_submodule(model, parent), child, v)
+        else:
+            setattr(model, k, v)
+
+    return model
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless3/streaming_beam_search.py
new file mode 120000
index 000000000..3a5f89833
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/streaming_beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/streaming_decode.py
new file mode 100755
index 000000000..e7c1affc2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/streaming_decode.py
@@ -0,0 +1,578 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless3/streaming_decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --left-context 32 \
+        --decode-chunk-size 8 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --decoding_method greedy_search \
+        --num-decode-streams 1000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from librispeech import LibriSpeech
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+    # Decoding in streaming requires causal convolution
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeech(params.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless3/test_model.py
new file mode 120000
index 000000000..4196e587c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/test_model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/test_model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py
new file mode 100755
index 000000000..b98248128
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/test_onnx.py
@@ -0,0 +1,401 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file is to test that models can be exported to onnx.
+"""
+import os
+
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
+import onnxruntime as ort
+import torch
+from conformer import (
+    Conformer,
+    ConformerEncoder,
+    ConformerEncoderLayer,
+    Conv2dSubsampling,
+    RelPositionalEncoding,
+)
+from scaling_converter import convert_scaled_to_non_scaled
+
+from icefall.utils import make_pad_mask
+
+ort.set_default_logger_severity(3)
+
+
+def test_conv2d_subsampling():
+    filename = "conv2d_subsampling.onnx"
+    opset_version = 11
+    N = 30
+    T = 50
+    num_features = 80
+    d_model = 512
+    x = torch.rand(N, T, num_features)
+
+    encoder_embed = Conv2dSubsampling(num_features, d_model)
+    encoder_embed.eval()
+    encoder_embed = convert_scaled_to_non_scaled(encoder_embed, inplace=True)
+
+    jit_model = torch.jit.trace(encoder_embed, x)
+
+    torch.onnx.export(
+        encoder_embed,
+        x,
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x"],
+        output_names=["y"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "y": {0: "N", 1: "T"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    assert input_nodes[0].name == "x"
+    assert input_nodes[0].shape == ["N", "T", num_features]
+
+    inputs = {input_nodes[0].name: x.numpy()}
+
+    onnx_y = session.run(["y"], inputs)[0]
+
+    onnx_y = torch.from_numpy(onnx_y)
+    torch_y = jit_model(x)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    os.remove(filename)
+
+
+def test_rel_pos():
+    filename = "rel_pos.onnx"
+
+    opset_version = 11
+    N = 30
+    T = 50
+    num_features = 80
+    d_model = 512
+    x = torch.rand(N, T, num_features)
+
+    encoder_pos = RelPositionalEncoding(d_model, dropout_rate=0.1)
+    encoder_pos.eval()
+    encoder_pos = convert_scaled_to_non_scaled(encoder_pos, inplace=True)
+
+    jit_model = torch.jit.trace(encoder_pos, x)
+
+    torch.onnx.export(
+        encoder_pos,
+        (x, torch.zeros(1, dtype=torch.int64)),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x"],
+        output_names=["y", "pos_emb"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "y": {0: "N", 1: "T"},
+            "pos_emb": {0: "N", 1: "T"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    assert input_nodes[0].name == "x"
+    assert input_nodes[0].shape == ["N", "T", num_features]
+
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+    }
+    onnx_y, onnx_pos_emb = session.run(["y", "pos_emb"], inputs)
+    onnx_y = torch.from_numpy(onnx_y)
+    onnx_pos_emb = torch.from_numpy(onnx_pos_emb)
+
+    torch_y, torch_pos_emb = jit_model(x)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    assert torch.allclose(onnx_pos_emb, torch_pos_emb, atol=1e-05), (
+        (onnx_pos_emb - torch_pos_emb).abs().max()
+    )
+    print(onnx_y.abs().sum(), torch_y.abs().sum())
+    print(onnx_pos_emb.abs().sum(), torch_pos_emb.abs().sum())
+
+    os.remove(filename)
+
+
+def test_conformer_encoder_layer():
+    filename = "conformer_encoder_layer.onnx"
+    opset_version = 11
+    N = 30
+    T = 50
+
+    d_model = 512
+    nhead = 8
+    dim_feedforward = 2048
+    dropout = 0.1
+    layer_dropout = 0.075
+    cnn_module_kernel = 31
+    causal = False
+
+    x = torch.rand(N, T, d_model)
+    x_lens = torch.full((N,), fill_value=T, dtype=torch.int64)
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    encoder_pos = RelPositionalEncoding(d_model, dropout)
+    encoder_pos.eval()
+    encoder_pos = convert_scaled_to_non_scaled(encoder_pos, inplace=True)
+
+    x, pos_emb = encoder_pos(x)
+    x = x.permute(1, 0, 2)
+
+    encoder_layer = ConformerEncoderLayer(
+        d_model,
+        nhead,
+        dim_feedforward,
+        dropout,
+        layer_dropout,
+        cnn_module_kernel,
+        causal,
+    )
+    encoder_layer.eval()
+    encoder_layer = convert_scaled_to_non_scaled(encoder_layer, inplace=True)
+
+    jit_model = torch.jit.trace(encoder_layer, (x, pos_emb, src_key_padding_mask))
+
+    torch.onnx.export(
+        encoder_layer,
+        (x, pos_emb, src_key_padding_mask),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "pos_emb", "src_key_padding_mask"],
+        output_names=["y"],
+        dynamic_axes={
+            "x": {0: "T", 1: "N"},
+            "pos_emb": {0: "N", 1: "T"},
+            "src_key_padding_mask": {0: "N", 1: "T"},
+            "y": {0: "T", 1: "N"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+        input_nodes[1].name: pos_emb.numpy(),
+        input_nodes[2].name: src_key_padding_mask.numpy(),
+    }
+    onnx_y = session.run(["y"], inputs)[0]
+    onnx_y = torch.from_numpy(onnx_y)
+
+    torch_y = jit_model(x, pos_emb, src_key_padding_mask)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    print(onnx_y.abs().sum(), torch_y.abs().sum(), onnx_y.shape, torch_y.shape)
+
+    os.remove(filename)
+
+
+def test_conformer_encoder():
+    filename = "conformer_encoder.onnx"
+
+    opset_version = 11
+    N = 3
+    T = 15
+
+    d_model = 512
+    nhead = 8
+    dim_feedforward = 2048
+    dropout = 0.1
+    layer_dropout = 0.075
+    cnn_module_kernel = 31
+    causal = False
+    num_encoder_layers = 12
+
+    x = torch.rand(N, T, d_model)
+    x_lens = torch.full((N,), fill_value=T, dtype=torch.int64)
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    encoder_pos = RelPositionalEncoding(d_model, dropout)
+    encoder_pos.eval()
+    encoder_pos = convert_scaled_to_non_scaled(encoder_pos, inplace=True)
+
+    x, pos_emb = encoder_pos(x)
+    x = x.permute(1, 0, 2)
+
+    encoder_layer = ConformerEncoderLayer(
+        d_model,
+        nhead,
+        dim_feedforward,
+        dropout,
+        layer_dropout,
+        cnn_module_kernel,
+        causal,
+    )
+    encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+    encoder.eval()
+    encoder = convert_scaled_to_non_scaled(encoder, inplace=True)
+
+    jit_model = torch.jit.trace(encoder, (x, pos_emb, src_key_padding_mask))
+
+    torch.onnx.export(
+        encoder,
+        (x, pos_emb, src_key_padding_mask),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "pos_emb", "src_key_padding_mask"],
+        output_names=["y"],
+        dynamic_axes={
+            "x": {0: "T", 1: "N"},
+            "pos_emb": {0: "N", 1: "T"},
+            "src_key_padding_mask": {0: "N", 1: "T"},
+            "y": {0: "T", 1: "N"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+        input_nodes[1].name: pos_emb.numpy(),
+        input_nodes[2].name: src_key_padding_mask.numpy(),
+    }
+    onnx_y = session.run(["y"], inputs)[0]
+    onnx_y = torch.from_numpy(onnx_y)
+
+    torch_y = jit_model(x, pos_emb, src_key_padding_mask)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    print(onnx_y.abs().sum(), torch_y.abs().sum(), onnx_y.shape, torch_y.shape)
+
+    os.remove(filename)
+
+
+def test_conformer():
+    filename = "conformer.onnx"
+    opset_version = 11
+    N = 3
+    T = 15
+    num_features = 80
+    x = torch.rand(N, T, num_features)
+    x_lens = torch.full((N,), fill_value=T, dtype=torch.int64)
+
+    conformer = Conformer(num_features=num_features)
+    conformer.eval()
+    conformer = convert_scaled_to_non_scaled(conformer, inplace=True)
+
+    jit_model = torch.jit.trace(conformer, (x, x_lens))
+    torch.onnx.export(
+        conformer,
+        (x, x_lens),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["y", "y_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "y": {0: "N", 1: "T"},
+            "y_lens": {0: "N"},
+        },
+    )
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+        input_nodes[1].name: x_lens.numpy(),
+    }
+    onnx_y, onnx_y_lens = session.run(["y", "y_lens"], inputs)
+    onnx_y = torch.from_numpy(onnx_y)
+    onnx_y_lens = torch.from_numpy(onnx_y_lens)
+
+    torch_y, torch_y_lens = jit_model(x, x_lens)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    assert torch.allclose(onnx_y_lens, torch_y_lens, atol=1e-05), (
+        (onnx_y_lens - torch_y_lens).abs().max()
+    )
+    print(onnx_y.abs().sum(), torch_y.abs().sum(), onnx_y.shape, torch_y.shape)
+    print(onnx_y_lens, torch_y_lens)
+
+    os.remove(filename)
+
+
+@torch.no_grad()
+def main():
+    test_conv2d_subsampling()
+    test_rel_pos()
+    test_conformer_encoder_layer()
+    test_conformer_encoder()
+    test_conformer()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20221011)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/test_scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless3/test_scaling.py
new file mode 100755
index 000000000..42de2410a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/test_scaling.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless3/test_scaling.py
+"""
+
+import torch
+from scaling import ActivationBalancer, ScaledConv1d, ScaledConv2d
+
+
+def test_scaled_conv1d():
+    for bias in [True, False]:
+        conv1d = ScaledConv1d(
+            3,
+            6,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        torch.jit.script(conv1d)
+
+
+def test_scaled_conv2d():
+    for bias in [True, False]:
+        conv2d = ScaledConv2d(
+            in_channels=1,
+            out_channels=3,
+            kernel_size=3,
+            padding=1,
+            bias=bias,
+        )
+        torch.jit.script(conv2d)
+
+
+def main():
+    test_scaled_conv1d()
+    test_scaled_conv2d()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/test_scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless3/test_scaling_converter.py
new file mode 100644
index 000000000..2e131158f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/test_scaling_converter.py
@@ -0,0 +1,218 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless3/test_scaling_converter.py
+"""
+
+import copy
+
+import torch
+from scaling import ScaledConv1d, ScaledConv2d, ScaledEmbedding, ScaledLinear
+from scaling_converter import (
+    convert_scaled_to_non_scaled,
+    scaled_conv1d_to_conv1d,
+    scaled_conv2d_to_conv2d,
+    scaled_embedding_to_embedding,
+    scaled_linear_to_linear,
+)
+from train import get_params, get_transducer_model
+
+
+def get_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.dynamic_chunk_training = False
+    params.short_chunk_size = 25
+    params.num_left_chunks = 4
+    params.causal_convolution = False
+
+    model = get_transducer_model(params, enable_giga=False)
+    return model
+
+
+def test_scaled_linear_to_linear():
+    N = 5
+    in_features = 10
+    out_features = 20
+    for bias in [True, False]:
+        scaled_linear = ScaledLinear(
+            in_features=in_features,
+            out_features=out_features,
+            bias=bias,
+        )
+        linear = scaled_linear_to_linear(scaled_linear)
+        x = torch.rand(N, in_features)
+
+        y1 = scaled_linear(x)
+        y2 = linear(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_linear = torch.jit.script(scaled_linear)
+        jit_linear = torch.jit.script(linear)
+
+        y3 = jit_scaled_linear(x)
+        y4 = jit_linear(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_conv1d_to_conv1d():
+    in_channels = 3
+    for bias in [True, False]:
+        scaled_conv1d = ScaledConv1d(
+            in_channels,
+            6,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        conv1d = scaled_conv1d_to_conv1d(scaled_conv1d)
+
+        x = torch.rand(20, in_channels, 10)
+        y1 = scaled_conv1d(x)
+        y2 = conv1d(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_conv1d = torch.jit.script(scaled_conv1d)
+        jit_conv1d = torch.jit.script(conv1d)
+
+        y3 = jit_scaled_conv1d(x)
+        y4 = jit_conv1d(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_conv2d_to_conv2d():
+    in_channels = 1
+    for bias in [True, False]:
+        scaled_conv2d = ScaledConv2d(
+            in_channels=in_channels,
+            out_channels=3,
+            kernel_size=3,
+            padding=1,
+            bias=bias,
+        )
+
+        conv2d = scaled_conv2d_to_conv2d(scaled_conv2d)
+
+        x = torch.rand(20, in_channels, 10, 20)
+        y1 = scaled_conv2d(x)
+        y2 = conv2d(x)
+        assert torch.allclose(y1, y2)
+
+        jit_scaled_conv2d = torch.jit.script(scaled_conv2d)
+        jit_conv2d = torch.jit.script(conv2d)
+
+        y3 = jit_scaled_conv2d(x)
+        y4 = jit_conv2d(x)
+
+        assert torch.allclose(y3, y4)
+        assert torch.allclose(y1, y4)
+
+
+def test_scaled_embedding_to_embedding():
+    scaled_embedding = ScaledEmbedding(
+        num_embeddings=500,
+        embedding_dim=10,
+        padding_idx=0,
+    )
+    embedding = scaled_embedding_to_embedding(scaled_embedding)
+
+    for s in [10, 100, 300, 500, 800, 1000]:
+        x = torch.randint(low=0, high=500, size=(s,))
+        scaled_y = scaled_embedding(x)
+        y = embedding(x)
+        assert torch.equal(scaled_y, y)
+
+
+def test_convert_scaled_to_non_scaled():
+    for inplace in [False, True]:
+        model = get_model()
+        model.eval()
+
+        orig_model = copy.deepcopy(model)
+
+        converted_model = convert_scaled_to_non_scaled(model, inplace=inplace)
+
+        model = orig_model
+
+        # test encoder
+        N = 2
+        T = 100
+        vocab_size = model.decoder.vocab_size
+
+        x = torch.randn(N, T, 80, dtype=torch.float32)
+        x_lens = torch.full((N,), x.size(1))
+
+        e1, e1_lens = model.encoder(x, x_lens)
+        e2, e2_lens = converted_model.encoder(x, x_lens)
+
+        assert torch.all(torch.eq(e1_lens, e2_lens))
+        assert torch.allclose(e1, e2), (e1 - e2).abs().max()
+
+        # test decoder
+        U = 50
+        y = torch.randint(low=1, high=vocab_size - 1, size=(N, U))
+
+        d1 = model.decoder(y)
+        d2 = converted_model.decoder(y)
+
+        assert torch.allclose(d1, d2)
+
+        # test simple projection
+        lm1 = model.simple_lm_proj(d1)
+        am1 = model.simple_am_proj(e1)
+
+        lm2 = converted_model.simple_lm_proj(d2)
+        am2 = converted_model.simple_am_proj(e2)
+
+        assert torch.allclose(lm1, lm2)
+        assert torch.allclose(am1, am2)
+
+        # test joiner
+        e = torch.rand(2, 3, 4, 512)
+        d = torch.rand(2, 3, 4, 512)
+
+        j1 = model.joiner(e, d)
+        j2 = converted_model.joiner(e, d)
+        assert torch.allclose(j1, j2)
+
+
+@torch.no_grad()
+def main():
+    test_scaled_linear_to_linear()
+    test_scaled_conv1d_to_conv1d()
+    test_scaled_conv2d_to_conv2d()
+    test_scaled_embedding_to_embedding()
+    test_convert_scaled_to_non_scaled()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220730)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/train.py b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
new file mode 100755
index 000000000..fdafa5a87
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
@@ -0,0 +1,1242 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+cd egs/librispeech/ASR/
+./prepare.sh
+./prepare_giga_speech.sh
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from gigaspeech import GigaSpeech
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from librispeech import LibriSpeech
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--full-libri",
+        type=str2bool,
+        default=True,
+        help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless3/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=4,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--giga-prob",
+        type=float,
+        default=0.5,
+        help="The probability to select a batch from the GigaSpeech dataset",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(
+    params: AttributeDict,
+    enable_giga: bool = True,
+) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    if enable_giga:
+        logging.info("Use giga")
+        decoder_giga = get_decoder_model(params)
+        joiner_giga = get_joiner_model(params)
+    else:
+        logging.info("Disable giga")
+        decoder_giga = None
+        joiner_giga = None
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        decoder_giga=decoder_giga,
+        joiner_giga=joiner_giga,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_libri(c: Cut) -> bool:
+    """Return True if this cut is from the LibriSpeech dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of GigaSpeech to dict(origin='giga')
+      See ../local/preprocess_gigaspeech.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    libri = is_libri(supervisions["cut"][0])
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            libri=libri,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    giga_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      giga_train_dl:
+        Dataloader for the GigaSpeech training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      rng:
+        For selecting which dataset to use.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    libri_tot_loss = MetricsTracker()
+    giga_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.giga_prob, params.giga_prob]
+
+    iter_libri = iter(train_dl)
+    iter_giga = iter(giga_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_libri if idx == 0 else iter_giga
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            name = "libri" if idx == 0 else "giga"
+            logging.info(f"{name} reaches end of dataloader")
+            break
+
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        libri = is_libri(batch["supervisions"]["cut"][0])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        if libri:
+            libri_tot_loss = (
+                libri_tot_loss * (1 - 1 / params.reset_interval)
+            ) + loss_info
+            prefix = "libri"  # for logging only
+        else:
+            giga_tot_loss = (
+                giga_tot_loss * (1 - 1 / params.reset_interval)
+            ) + loss_info
+            prefix = "giga"
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], "
+                f"libri_tot_loss[{libri_tot_loss}], "
+                f"giga_tot_loss[{giga_tot_loss}], "
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                libri_tot_loss.write_summary(
+                    tb_writer, "train/libri_tot_", params.batch_idx_train
+                )
+                giga_tot_loss.write_summary(
+                    tb_writer, "train/giga_tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(
+    cuts: CutSet,
+    sp: spm.SentencePieceProcessor,
+) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    cuts = cuts.filter(remove_short_and_long_utt)
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    train_cuts = filter_short_and_long_utterances(train_cuts, sp)
+
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+    # XL 10k hours
+    # L  2.5k hours
+    # M  1k hours
+    # S  250 hours
+    # XS 10 hours
+    # DEV 12 hours
+    # Test 40 hours
+    if params.full_libri:
+        logging.info("Using the XL subset of GigaSpeech (10k hours)")
+        train_giga_cuts = gigaspeech.train_XL_cuts()
+    else:
+        logging.info("Using the S subset of GigaSpeech (250 hours)")
+        train_giga_cuts = gigaspeech.train_S_cuts()
+
+    train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts, sp)
+    train_giga_cuts = train_giga_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    giga_train_dl = asr_datamodule.train_dataloaders(
+        train_giga_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    # It's time consuming to include `giga_train_dl` here
+    #  for dl in [train_dl, giga_train_dl]:
+    for dl in [train_dl]:
+        if params.start_batch <= 0:
+            scan_pessimistic_batches_for_oom(
+                model=model,
+                train_dl=dl,
+                optimizer=optimizer,
+                sp=sp,
+                params=params,
+                warmup=0.0 if params.start_epoch == 0 else 1.0,
+            )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            giga_train_dl=giga_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert 0 <= args.giga_prob < 1, args.giga_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless4/__init__.py
new file mode 120000
index 000000000..b24e5e357
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/__init__.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless4/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless4/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless4/conformer.py
new file mode 120000
index 000000000..3b84b9573
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/conformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py
new file mode 100755
index 000000000..5195a4ef6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py
@@ -0,0 +1,975 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) decode in streaming mode (take greedy search as an example)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --decode-chunk-size 16 \
+    --left-context 64 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+To evaluate symbol delay, you should:
+(1) Generate cuts with word-time alignments:
+./add_alignments.sh
+(2) Set the argument "--manifest-dir data/fbank_ali" while decoding.
+For example:
+./pruned_transducer_stateless4/decode.py \
+    --epoch 40 \
+    --avg 20 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search \
+    --manifest-dir data/fbank_ali
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import ContextGraph
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    DecodingResults,
+    parse_hyp_and_timestamp,
+    setup_logger,
+    store_transcripts_and_timestamps,
+    str2bool,
+    write_error_stats_with_timestamps,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling). Set -1 to use full attention.",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="""Left context can be seen during decoding (in frames after subsampling). """,  # noqa
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, Tuple[List[List[str]], List[List[float]]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It is a tuple. `len(value[0])` and `len(value[1])` are both
+                 equal to the batch size. `value[0][i]` and `value[1][i]`
+                 are the decoding result and timestamps for the i-th utterance
+                 in the given batch respectively.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result and timestamps. See above description for the
+      format of the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        if params.decode_chunk_size > 0:
+            # except the case of using full attention
+            feature_lens += params.left_context
+            feature = torch.nn.functional.pad(
+                feature,
+                pad=(0, 0, 0, params.left_context),
+                value=LOG_EPS,
+            )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        res = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        res = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "fast_beam_search_nbest":
+        res = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        res = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        res = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            return_timestamps=True,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        res = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+            return_timestamps=True,
+        )
+    else:
+        batch_size = encoder_out.size(0)
+        tokens = []
+        timestamps = []
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                res = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                    return_timestamps=True,
+                )
+            elif params.decoding_method == "beam_search":
+                res = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                    return_timestamps=True,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            tokens.extend(res.tokens)
+            timestamps.extend(res.timestamps)
+        res = DecodingResults(hyps=tokens, timestamps=timestamps)
+
+    hyps, timestamps = parse_hyp_and_timestamp(
+        res=res,
+        sp=sp,
+        subsampling_factor=params.subsampling_factor,
+        frame_shift_ms=params.frame_shift_ms,
+        word_table=word_table,
+    )
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": (hyps, timestamps)}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: (hyps, timestamps)}
+    else:
+        key = f"beam_size_{params.beam_size}"
+        if params.has_contexts:
+            key += f"-context-score-{params.context_score}"
+        else:
+            key += "-no-context-words"
+        return {key: (hyps, timestamps)}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str], List[float], List[float]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains five elements:
+      - cut_id
+      - reference transcript
+      - predicted result
+      - timestamp of reference transcript
+      - timestamp of predicted result
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        timestamps_ref = []
+        for cut in batch["supervisions"]["cut"]:
+            for s in cut.supervisions:
+                time = []
+                if s.alignment is not None and "word" in s.alignment:
+                    time = [
+                        aliword.start
+                        for aliword in s.alignment["word"]
+                        if aliword.symbol != ""
+                    ]
+                timestamps_ref.append(time)
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            context_graph=context_graph,
+        )
+
+        for name, (hyps, timestamps_hyp) in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts) and len(timestamps_hyp) == len(
+                timestamps_ref
+            )
+            for cut_id, hyp_words, ref_text, time_hyp, time_ref in zip(
+                cut_ids, hyps, texts, timestamps_hyp, timestamps_ref
+            ):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words, time_ref, time_hyp))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[
+        str,
+        List[Tuple[List[str], List[str], List[str], List[float], List[float]]],
+    ],
+):
+    test_set_wers = dict()
+    test_set_delays = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts_and_timestamps(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer, mean_delay, var_delay = write_error_stats_with_timestamps(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+            test_set_delays[key] = (mean_delay, var_delay)
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    test_set_delays = sorted(test_set_delays.items(), key=lambda x: x[1][0])
+    delays_info = (
+        params.res_dir / f"symbol-delay-summary-{test_set_name}-{params.suffix}.txt"
+    )
+    with open(delays_info, "w") as f:
+        print("settings\tsymbol-delay", file=f)
+        for key, val in test_set_delays:
+            print(
+                "{}\tmean: {}s, variance: {}".format(key, val[0], val[1]),
+                file=f,
+            )
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    s = "\nFor {}, symbol-delay of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_delays:
+        s += "{}\tmean: {}s, variance: {}{}\n".format(key, val[0], val[1], note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.has_contexts:
+            params.suffix += f"-context-score-{params.context_score}"
+        else:
+            params.suffix += "-no-context-words"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts = []
+            for line in open(params.context_file).readlines():
+                contexts.append((sp.encode(line.strip()), 0.0))
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless4/decode_stream.py
new file mode 120000
index 000000000..30f264813
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/decode_stream.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/decode_stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless4/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless4/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/export.py b/egs/librispeech/ASR/pruned_transducer_stateless4/export.py
new file mode 100755
index 000000000..08d736f52
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/export.py
@@ -0,0 +1,291 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless4/export.py \
+  --exp-dir ./pruned_transducer_stateless4/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless4/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless4/decode.py \
+        --exp-dir ./pruned_transducer_stateless4/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model \
+        --use-averaged-model True
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless4/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/lstmp.py b/egs/librispeech/ASR/pruned_transducer_stateless4/lstmp.py
new file mode 120000
index 000000000..9aa06f82f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/lstmp.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/model.py b/egs/librispeech/ASR/pruned_transducer_stateless4/model.py
new file mode 120000
index 000000000..ebb6d774d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless4/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/profile.py b/egs/librispeech/ASR/pruned_transducer_stateless4/profile.py
new file mode 100755
index 000000000..252bdf060
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/profile.py
@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage: ./pruned_transducer_stateless4/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from typing import Tuple
+from torch import Tensor, nn
+
+from icefall.profiler import get_model_profile
+from scaling import BasicNorm, DoubleSwish
+from train import get_encoder_model, get_joiner_model, add_model_arguments, get_params
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _basic_norm_flops_compute(module, input, output):
+    assert len(input) == 1, len(input)
+    # estimate as layer_norm, see icefall/profiler.py
+    flops = input[0].numel() * 5
+    module.__flops__ += int(flops)
+
+
+def _doubleswish_module_flops_compute(module, input, output):
+    # For DoubleSwish
+    assert len(input) == 1, len(input)
+    # estimate as swish/silu, see icefall/profiler.py
+    flops = input[0].numel()
+    module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    BasicNorm: _basic_norm_flops_compute,
+    DoubleSwish: _doubleswish_module_flops_compute,
+}
+
+
+class Model(nn.Module):
+    """A Wrapper for encoder and encoder_proj"""
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        encoder_proj: nn.Module,
+    ) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(self, feature: Tensor, feature_lens: Tensor) -> Tuple[Tensor, Tensor]:
+        encoder_out, encoder_out_lens = self.encoder(feature, feature_lens)
+
+        logits = self.encoder_proj(encoder_out)
+
+        return logits, encoder_out_lens
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = Model(
+        encoder=get_encoder_model(params),
+        encoder_proj=get_joiner_model(params).encoder_proj,
+    )
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(
+        model=model,
+        args=(feature, feature_lens),
+        module_hoop_mapping=MODULE_HOOK_MAPPING,
+    )
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless4/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless4/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless4/streaming_beam_search.py
new file mode 120000
index 000000000..3a5f89833
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/streaming_beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless4/streaming_decode.py
new file mode 100755
index 000000000..e966aa4b1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/streaming_decode.py
@@ -0,0 +1,647 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless4/streaming_decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --left-context 32 \
+        --decode-chunk-size 8 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless4/exp \
+        --decoding_method greedy_search \
+        --num-decode-streams 200
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    # Decoding in streaming requires causal convolution
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless4/test_model.py
new file mode 120000
index 000000000..4196e587c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/test_model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/test_model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
new file mode 100755
index 000000000..875b03f7f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
@@ -0,0 +1,1161 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless4/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless4/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+# train a streaming model
+./pruned_transducer_stateless4/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 1 \
+  --dynamic-chunk-training 1 \
+  --causal-convolution 1 \
+  --short-chunk-size 25 \
+  --num-left-chunks 4 \
+  --max-duration 300
+
+"""
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless4/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
new file mode 100644
index 000000000..8bbceec61
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1,1822 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from torch import Tensor, nn
+
+from icefall.utils import is_jit_tracing, make_pad_mask, subsequent_chunk_mask
+
+
+class Conformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module.
+        dynamic_chunk_training (bool): whether to use dynamic chunk training, if
+            you want to train a streaming model, this is expected to be True.
+            When setting True, it will use a masking strategy to make the attention
+            see only limited left and right context.
+        short_chunk_threshold (float): a threshold to determinize the chunk size
+            to be used in masking training, if the randomly generated chunk size
+            is greater than ``max_len * short_chunk_threshold`` (max_len is the
+            max sequence length of current batch) then it will use
+            full context in training (i.e. with chunk size equals to max_len).
+            This will be used only when dynamic_chunk_training is True.
+        short_chunk_size (int): see docs above, if the randomly generated chunk
+            size equals to or less than ``max_len * short_chunk_threshold``, the
+            chunk size will be sampled uniformly from 1 to short_chunk_size.
+            This also will be used only when dynamic_chunk_training is True.
+        num_left_chunks (int): the left context (in chunks) attention can see, the
+            chunk size is decided by short_chunk_threshold and short_chunk_size.
+            A minus value means seeing full left context.
+            This also will be used only when dynamic_chunk_training is True.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        aux_layer_period: int = 3,
+        dynamic_chunk_training: bool = False,
+        short_chunk_threshold: float = 0.75,
+        short_chunk_size: int = 25,
+        num_left_chunks: int = -1,
+        causal: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        self.encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.cnn_module_kernel = cnn_module_kernel
+        self.causal = causal
+        self.dynamic_chunk_training = dynamic_chunk_training
+        self.short_chunk_threshold = short_chunk_threshold
+        self.short_chunk_size = short_chunk_size
+        self.num_left_chunks = num_left_chunks
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+            cnn_module_kernel=cnn_module_kernel,
+            causal=causal,
+        )
+        # aux_layers from 1/3
+        self.encoder = ConformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            aux_layers=list(
+                range(
+                    num_encoder_layers // 3,
+                    num_encoder_layers - 1,
+                    aux_layer_period,
+                )
+            ),
+        )
+        self._init_state: List[torch.Tensor] = [torch.empty(0)]
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, d_model)
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+        assert x.size(0) == lengths.max().item()
+        src_key_padding_mask = make_pad_mask(lengths)
+
+        if self.dynamic_chunk_training:
+            assert (
+                self.causal
+            ), "Causal convolution is required for streaming conformer."
+            max_len = x.size(0)
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * self.short_chunk_threshold):
+                chunk_size = max_len
+            else:
+                chunk_size = chunk_size % self.short_chunk_size + 1
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=self.num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+        else:
+            x = self.encoder(
+                x,
+                pos_emb,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths
+
+    @torch.jit.export
+    def get_init_state(
+        self, left_context: int, device: torch.device
+    ) -> List[torch.Tensor]:
+        """Return the initial cache state of the model.
+        Args:
+          left_context: The left context size (in frames after subsampling).
+        Returns:
+          Return the initial state of the model, it is a list containing two
+          tensors, the first one is the cache for attentions which has a shape
+          of (num_encoder_layers, left_context, encoder_dim), the second one
+          is the cache of conv_modules which has a shape of
+          (num_encoder_layers, cnn_module_kernel - 1, encoder_dim).
+          NOTE: the returned tensors are on the given device.
+        """
+        if len(self._init_state) == 2 and self._init_state[0].size(1) == left_context:
+            # Note: It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        init_states: List[torch.Tensor] = [
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    left_context,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+        ]
+
+        self._init_state = init_states
+
+        return init_states
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[Tensor]] = None,
+        processed_lens: Optional[Tensor] = None,
+        left_context: int = 64,
+        right_context: int = 4,
+        chunk_size: int = 16,
+        simulate_streaming: bool = False,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states:
+            The decode states for previous frames which contains the cached data.
+            It has two elements, the first element is the attn_cache which has
+            a shape of (encoder_layers, left_context, batch, attention_dim),
+            the second element is the conv_cache which has a shape of
+            (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+            Note: states will be modified in this function.
+          processed_lens:
+            How many frames (after subsampling) have been processed for each sequence.
+          left_context:
+            How many previous frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `left_context` frames
+            of left context, some have more.
+          right_context:
+            How many future frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `right_context` frames
+            of right context, some have more.
+          chunk_size:
+            The chunk size for decoding, this will be used to simulate streaming
+            decoding using masking.
+          simulate_streaming:
+            If setting True, it will use a masking strategy to simulate streaming
+            fashion (i.e. every chunk data only see limited left context and
+            right context). The whole sequence is supposed to be send at a time
+            When using simulate_streaming.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+            - decode_states, the updated states including the information
+              of current chunk.
+        """
+
+        # x: [N, T, C]
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        if not simulate_streaming:
+            assert states is not None
+            assert processed_lens is not None
+            assert (
+                len(states) == 2
+                and states[0].shape
+                == (self.encoder_layers, left_context, x.size(0), self.d_model)
+                and states[1].shape
+                == (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    x.size(0),
+                    self.d_model,
+                )
+            ), f"""The length of states MUST be equal to 2, and the shape of
+             first element should be {(self.encoder_layers, left_context, x.size(0), self.d_model)},
+             given {states[0].shape}. the shape of second element should be
+             {(self.encoder_layers, self.cnn_module_kernel - 1, x.size(0), self.d_model)},
+             given {states[1].shape}."""
+
+            lengths -= 2  # we will cut off 1 frame on each side of encoder_embed output
+
+            src_key_padding_mask = make_pad_mask(lengths)
+
+            processed_mask = torch.arange(left_context, device=x.device).expand(
+                x.size(0), left_context
+            )
+            processed_lens = processed_lens.view(x.size(0), 1)
+            processed_mask = (processed_lens <= processed_mask).flip(1)
+
+            src_key_padding_mask = torch.cat(
+                [processed_mask, src_key_padding_mask], dim=1
+            )
+
+            embed = self.encoder_embed(x)
+
+            # cut off 1 frame on each size of embed as they see the padding
+            # value which causes a training and decoding mismatch.
+            embed = embed[:, 1:-1, :]
+
+            embed, pos_enc = self.encoder_pos(embed, left_context)
+            embed = embed.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+
+            x, states = self.encoder.chunk_forward(
+                embed,
+                pos_enc,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+                states=states,
+                left_context=left_context,
+                right_context=right_context,
+            )  # (T, B, F)
+            if right_context > 0:
+                x = x[0:-right_context, ...]
+                lengths -= right_context
+        else:
+            assert states is None
+            states = []  # just to make torch.script.jit happy
+            # this branch simulates streaming decoding using mask as we are
+            # using in training time.
+            src_key_padding_mask = make_pad_mask(lengths)
+            x = self.encoder_embed(x)
+            x, pos_emb = self.encoder_pos(x)
+            x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+            assert x.size(0) == lengths.max().item()
+
+            num_left_chunks = -1
+            if left_context >= 0:
+                assert left_context % chunk_size == 0
+                num_left_chunks = left_context // chunk_size
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths, states
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training and streaming decoding.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        causal: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel, causal=causal)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        conv, _ = self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        src = src + self.dropout(conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        """
+        Pass the input through the encoder layer.
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        assert not self.training
+        assert len(states) == 2
+        assert states[0].shape == (left_context, src.size(1), src.size(2))
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # We put the attention cache this level (i.e. before linear transformation)
+        # to save memory consumption, when decoding in streaming fashion, the
+        # batch size would be thousands (for 32GB machine), if we cache key & val
+        # separately, it needs extra several GB memory.
+        # TODO(WeiKang): Move cache to self_attn level (i.e. cache key & val
+        # separately) if needed.
+        key = torch.cat([states[0], src], dim=0)
+        val = key
+        if right_context > 0:
+            states[0] = key[
+                -(left_context + right_context) : -right_context, ...  # noqa
+            ]
+        else:
+            states[0] = key[-left_context:, ...]
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            key,
+            val,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+            left_context=left_context,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        conv, conv_cache = self.conv_module(src, states[1], right_context)
+        states[1] = conv_cache
+
+        src = src + self.dropout(conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        return src, states
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        aux_layers: List[int],
+    ) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert len(set(aux_layers)) == len(aux_layers)
+
+        assert num_layers - 1 not in aux_layers
+        self.aux_layers = aux_layers + [num_layers - 1]
+
+        self.combiner = RandomCombine(
+            num_inputs=len(self.aux_layers),
+            final_weight=0.5,
+            pure_prob=0.333,
+            stddev=2.0,
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        outputs = []
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+            if i in self.aux_layers:
+                outputs.append(output)
+
+        output = self.combiner(outputs)
+
+        return output
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        r"""Pass the input through the encoder layers in turn.
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (encoder_layers, left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+        """
+        assert not self.training
+        assert len(states) == 2
+        assert states[0].shape == (
+            self.num_layers,
+            left_context,
+            src.size(1),
+            src.size(2),
+        )
+        assert states[1].size(0) == self.num_layers
+
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            cache = [states[0][layer_index], states[1][layer_index]]
+            output, cache = mod.chunk_forward(
+                output,
+                pos_emb,
+                states=cache,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+                left_context=left_context,
+                right_context=right_context,
+            )
+            states[0][layer_index] = cache[0]
+            states[1][layer_index] = cache[1]
+
+        return output, states
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor, left_context: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_1 = x.size(1) + left_context
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_1 * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_1, self.d_model)
+        pe_negative = torch.zeros(x_size_1, self.d_model)
+        position = torch.arange(0, x_size_1, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor, left_context: int = 0) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x, left_context)
+        x_size_1 = x.size(1) + left_context
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_1
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            left_context=left_context,
+        )
+
+    def rel_shift(self, x: Tensor, left_context: int = 0) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        time2 = time1 + left_context
+        assert (
+            n == left_context + 2 * time1 - 1
+        ), f"{n} == {left_context} + 2 * {time1} - 1"
+        # Note: TorchScript requires explicit arg for stride()
+
+        if is_jit_tracing():
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(time2)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, time1, time2)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time1_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, time1, time2),
+                (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+                storage_offset=n_stride * (time1 - 1),
+            )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1)
+        p = p.permute(0, 2, 3, 1)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(q_with_bias_v, p)  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd, left_context)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+
+        # If we are using dynamic_chunk_training and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`, at this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax), so, we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, tgt_len, src_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+        causal (bool): Whether to use causal convolution.
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        causal: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.causal = causal
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.lorder = kernel_size - 1
+        padding = (kernel_size - 1) // 2
+        if self.causal:
+            padding = 0
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        cache: Optional[Tensor] = None,
+        right_context: int = 0,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: The cache of depthwise_conv, only used in real streaming
+                decoding.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+            If cache is None return the output tensor (#time, batch, channels).
+            If cache is not None, return a tuple of Tensor, the first one is
+            the output tensor (#time, batch, channels), the second one is the
+            new cache for next chunk (#kernel_size - 1, batch, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        if self.causal and self.lorder > 0:
+            if cache is None:
+                # Make depthwise_conv causal by
+                # manualy padding self.lorder zeros to the left
+                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+            else:
+                assert not self.training, "Cache should be None in training time"
+                assert cache.size(0) == self.lorder
+                x = torch.cat([cache.permute(1, 2, 0), x], dim=2)
+                if right_context > 0:
+                    cache = x.permute(2, 0, 1)[
+                        -(self.lorder + right_context) : (-right_context),  # noqa
+                        ...,
+                    ]
+                else:
+                    cache = x.permute(2, 0, 1)[-self.lorder :, ...]  # noqa
+
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        # torch.jit.script requires return types be the same as annotated above
+        if cache is None:
+            cache = torch.empty(0)
+
+        return x.permute(2, 0, 1), cache
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+class RandomCombine(nn.Module):
+    """
+    This module combines a list of Tensors, all with the same shape, to
+    produce a single output of that same shape which, in training time,
+    is a random combination of all the inputs; but which in test time
+    will be just the last input.
+
+    The idea is that the list of Tensors will be a list of outputs of multiple
+    conformer layers.  This has a similar effect as iterated loss. (See:
+    DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER
+    NETWORKS).
+    """
+
+    def __init__(
+        self,
+        num_inputs: int,
+        final_weight: float = 0.5,
+        pure_prob: float = 0.5,
+        stddev: float = 2.0,
+    ) -> None:
+        """
+        Args:
+          num_inputs:
+            The number of tensor inputs, which equals the number of layers'
+            outputs that are fed into this module.  E.g. in an 18-layer neural
+            net if we output layers 16, 12, 18, num_inputs would be 3.
+          final_weight:
+            The amount of weight or probability we assign to the
+            final layer when randomly choosing layers or when choosing
+            continuous layer weights.
+          pure_prob:
+            The probability, on each frame, with which we choose
+            only a single layer to output (rather than an interpolation)
+          stddev:
+            A standard deviation that we add to log-probs for computing
+            randomized weights.
+
+        The method of choosing which layers, or combinations of layers, to use,
+        is conceptually as follows::
+
+            With probability `pure_prob`::
+               With probability `final_weight`: choose final layer,
+               Else: choose random non-final layer.
+            Else::
+               Choose initial log-weights that correspond to assigning
+               weight `final_weight` to the final layer and equal
+               weights to other layers; then add Gaussian noise
+               with variance `stddev` to these log-weights, and normalize
+               to weights (note: the average weight assigned to the
+               final layer here will not be `final_weight` if stddev>0).
+        """
+        super().__init__()
+        assert 0 <= pure_prob <= 1, pure_prob
+        assert 0 < final_weight < 1, final_weight
+        assert num_inputs >= 1
+
+        self.num_inputs = num_inputs
+        self.final_weight = final_weight
+        self.pure_prob = pure_prob
+        self.stddev = stddev
+
+        self.final_log_weight = (
+            torch.tensor((final_weight / (1 - final_weight)) * (self.num_inputs - 1))
+            .log()
+            .item()
+        )
+
+    def forward(self, inputs: List[Tensor]) -> Tensor:
+        """Forward function.
+        Args:
+          inputs:
+            A list of Tensor, e.g. from various layers of a transformer.
+            All must be the same shape, of (*, num_channels)
+        Returns:
+          A Tensor of shape (*, num_channels).  In test mode
+          this is just the final input.
+        """
+        num_inputs = self.num_inputs
+        assert len(inputs) == num_inputs
+        if not self.training or torch.jit.is_scripting():
+            return inputs[-1]
+
+        # Shape of weights: (*, num_inputs)
+        num_channels = inputs[0].shape[-1]
+        num_frames = inputs[0].numel() // num_channels
+
+        ndim = inputs[0].ndim
+        # stacked_inputs: (num_frames, num_channels, num_inputs)
+        stacked_inputs = torch.stack(inputs, dim=ndim).reshape(
+            (num_frames, num_channels, num_inputs)
+        )
+
+        # weights: (num_frames, num_inputs)
+        weights = self._get_random_weights(
+            inputs[0].dtype, inputs[0].device, num_frames
+        )
+
+        weights = weights.reshape(num_frames, num_inputs, 1)
+        # ans: (num_frames, num_channels, 1)
+        ans = torch.matmul(stacked_inputs, weights)
+        # ans: (*, num_channels)
+
+        ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,))
+
+        # The following if causes errors for torch script in torch 1.6.0
+        #  if __name__ == "__main__":
+        #      # for testing only...
+        #      print("Weights = ", weights.reshape(num_frames, num_inputs))
+        return ans
+
+    def _get_random_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> Tensor:
+        """Return a tensor of random weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), such that
+          `ans.sum(dim=1)` is all ones.
+        """
+        pure_prob = self.pure_prob
+        if pure_prob == 0.0:
+            return self._get_random_mixed_weights(dtype, device, num_frames)
+        elif pure_prob == 1.0:
+            return self._get_random_pure_weights(dtype, device, num_frames)
+        else:
+            p = self._get_random_pure_weights(dtype, device, num_frames)
+            m = self._get_random_mixed_weights(dtype, device, num_frames)
+            return torch.where(
+                torch.rand(num_frames, 1, device=device) < self.pure_prob, p, m
+            )
+
+    def _get_random_pure_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ):
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A one-hot tensor of shape `(num_frames, self.num_inputs)`, with
+          exactly one weight equal to 1.0 on each frame.
+        """
+        final_prob = self.final_weight
+
+        # final contains self.num_inputs - 1 in all elements
+        final = torch.full((num_frames,), self.num_inputs - 1, device=device)
+        # nonfinal contains random integers in [0..num_inputs - 2], these are for non-final weights.
+        nonfinal = torch.randint(self.num_inputs - 1, (num_frames,), device=device)
+
+        indexes = torch.where(
+            torch.rand(num_frames, device=device) < final_prob, final, nonfinal
+        )
+        ans = torch.nn.functional.one_hot(indexes, num_classes=self.num_inputs).to(
+            dtype=dtype
+        )
+        return ans
+
+    def _get_random_mixed_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ):
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), which elements
+          in [0..1] that sum to one over the second axis, i.e.
+          `ans.sum(dim=1)` is all ones.
+        """
+        logprobs = (
+            torch.randn(num_frames, self.num_inputs, dtype=dtype, device=device)
+            * self.stddev
+        )
+        logprobs[:, -1] += self.final_log_weight
+        return logprobs.softmax(dim=1)
+
+
+def _test_random_combine(final_weight: float, pure_prob: float, stddev: float):
+    print(
+        f"_test_random_combine: final_weight={final_weight}, pure_prob={pure_prob}, stddev={stddev}"
+    )
+    num_inputs = 3
+    num_channels = 50
+    m = RandomCombine(
+        num_inputs=num_inputs,
+        final_weight=final_weight,
+        pure_prob=pure_prob,
+        stddev=stddev,
+    )
+
+    x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)]
+
+    y = m(x)
+    assert y.shape == x[0].shape
+    assert torch.allclose(y, x[0])  # .. since actually all ones.
+
+
+def _test_random_combine_main():
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.0)
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.3)
+    _test_random_combine(0.5, 1, 0.3)
+    _test_random_combine(0.5, 0.5, 0.3)
+
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
+
+    _test_random_combine_main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..7a3e63218
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,1024 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+(2) beam search (not recommended)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+(4) fast beam search (one best)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+(5) fast beam search (nbest)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+(7) fast beam search (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search with RNNLM shallow fusion
+./pruned_transducer_stateless5/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+(9) modified beam search with LM shallow fusion + LODR
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --max-duration 600 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --decoding-method modified_beam_search_LODR \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.4 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16 \
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+    modified_beam_search_ngram_rescoring,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling). Set -1 to use full attention.",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=3,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring, or LODR
+            """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        if params.decode_chunk_size > 0:
+            # except the case of using full attention
+            feature_lens += params.left_context
+            feature = torch.nn.functional.pad(
+                feature,
+                pad=(0, 0, 0, params.left_context),
+                value=LOG_EPS,
+            )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_ngram_rescoring":
+        hyp_tokens = modified_beam_search_ngram_rescoring(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+
+    else:
+        LM = None
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode_stream.py
new file mode 120000
index 000000000..30f264813
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode_stream.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless/decode_stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py b/egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py
new file mode 100755
index 000000000..b90d81dcf
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py
@@ -0,0 +1,684 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-25-avg-5.pt"
+
+cd exp
+ln -s pretrained-epoch-25-avg-5.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx-streaming.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+
+You can find the exported models in
+https://huggingface.co/csukuangfj/sherpa-onnx-streaming-conformer-en-2023-05-09
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+        self.num_encoder_layers = encoder.encoder_layers
+        self.encoder_dim = encoder.d_model
+        self.cnn_module_kernel = encoder.cnn_module_kernel
+
+        # Note you can tune these values
+        self.left_context = 64  # after subsampling
+        self.chunk_size = 16  # after subsampling
+        self.right_context = 0  # after subsampling
+
+        subsampling_factor = 4
+        self.pad_length = (self.right_context + 2) * subsampling_factor + 3
+
+        self.T = (self.chunk_size * subsampling_factor) + self.pad_length
+        self.decode_chunk_len = self.chunk_size * subsampling_factor
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cached_attn: torch.Tensor,
+        cached_conv: torch.Tensor,
+        processed_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, self.T, C)
+          cached_attn:
+            A 3-D tensor of shape
+            (num_encoder_layers, self.left_context, N, self.encoder_dim)
+          cached_conv:
+            A 3-D tensor of shape
+            (num_encoder_layers, self.cnn_module_kernel-1, N, self.encoder_dim)
+          processed_lens:
+            A 1-D tensor of shape (N,). It contains number of processed frames
+            after subsampling. Its dtype is torch.int64.
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, self.chunk_size, joiner_dim)
+            - new_cached_attn, it has the same shape as cached_attn
+            - new_cached_conv, it has the same shape as cached_conv
+        """
+        assert x.size(1) == self.T, (x.shape, self.T)
+        N = x.size(0)
+        x_lens = torch.full((N,), fill_value=self.T, device=x.device, dtype=torch.int64)
+
+        (
+            encoder_out,
+            _,
+            [new_cached_attn, new_cached_conv],
+        ) = self.encoder.streaming_forward(
+            x,
+            x_lens,
+            states=[cached_attn, cached_conv],
+            processed_lens=processed_lens,
+            left_context=self.left_context,
+            right_context=self.right_context,
+            chunk_size=self.chunk_size,
+        )
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, new_cached_attn, new_cached_conv
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    N = 1
+    x = torch.zeros(N, encoder_model.T, 80, dtype=torch.float32)
+    cached_attn = torch.zeros(
+        encoder_model.num_encoder_layers,
+        encoder_model.left_context,
+        N,
+        encoder_model.encoder_dim,
+    )
+    cached_conv = torch.zeros(
+        encoder_model.num_encoder_layers,
+        encoder_model.cnn_module_kernel - 1,
+        N,
+        encoder_model.encoder_dim,
+    )
+    processed_lens = torch.zeros((N,), dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, cached_attn, cached_conv, processed_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "cached_attn", "cached_conv", "processed_lens"],
+        output_names=["encoder_out", "new_cached_attn", "new_cached_conv"],
+        dynamic_axes={
+            "x": {0: "N"},
+            "cached_attn": {2: "N"},
+            "cached_conv": {2: "N"},
+            "processed_lens": {0: "N"},
+            "encoder_out": {0: "N"},
+            "new_cached_attn": {2: "N"},
+            "new_cached_conv": {2: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless5",
+        "pad_length": str(encoder_model.pad_length),
+        "decode_chunk_len": str(encoder_model.decode_chunk_len),
+        "encoder_dim": str(encoder_model.encoder_dim),
+        "num_encoder_layers": str(encoder_model.num_encoder_layers),
+        "cnn_module_kernel": str(encoder_model.cnn_module_kernel),
+        "left_context": str(encoder_model.left_context),
+        "right_context": str(encoder_model.right_context),
+        "chunk_size": str(encoder_model.chunk_size),
+        "T": str(encoder_model.T),
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if not params.causal_convolution:
+        logging.info("Seting causal_convolution to True for exporting streaming models")
+        params.causal_convolution = True
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum(p.numel() for p in model.parameters())
+    logging.info(f"Number of model parameters: {num_param}")
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx.py
new file mode 100755
index 000000000..02aa24f2c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx.py
@@ -0,0 +1,608 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-39-avg-7.pt"
+
+cd exp
+ln -s pretrained-epoch-39-avg-7.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless5",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/export.py b/egs/librispeech/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..e5223be26
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,289 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        convert_scaled_to_non_scaled(model, inplace=True)
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/lstmp.py b/egs/librispeech/ASR/pruned_transducer_stateless5/lstmp.py
new file mode 120000
index 000000000..4f377cd01
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/lstmp.py
@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/model.py b/egs/librispeech/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..ebb6d774d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_check.py
new file mode 120000
index 000000000..66d63b807
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_check.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_check.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py
new file mode 100755
index 000000000..6f26e34b5
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py
@@ -0,0 +1,326 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-2022-07-07
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-2022-07-07
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-30-avg-10.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-10.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp/ \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+2. Run this file
+
+./pruned_transducer_stateless5/onnx_decode.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --max-duration 600 \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+"""
+
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+
+from onnx_pretrained import greedy_search, OnnxModel
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, sp: spm.SentencePieceProcessor, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    hyps = [sp.decode(h).split() for h in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, sp=sp, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    blank_id = sp.piece_to_id("<blk>")
+    assert blank_id == 0, blank_id
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(dl=test_dl, model=model, sp=sp)
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained-streaming.py b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained-streaming.py
new file mode 100755
index 000000000..6e290e799
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained-streaming.py
@@ -0,0 +1,459 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This script loads ONNX models exported by ./export-onnx.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-25-avg-5.pt"
+cd exp
+ln -s pretrained-epoch-25-avg-5.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx-streaming.py \
+  --bpe-model ./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir ./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files in $repo/exp
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file with the exported ONNX models
+
+./pruned_transducer_stateless5/onnx_pretrained-streaming.py \
+  --encoder-model-filename ./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename ./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename ./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens=./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/data/lang_bpe_500/tokens.txt \
+  ./icefall_librispeech_streaming_pruned_transducer_stateless5_20220729/test_waves/1221-135766-0001.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+
+You can find the exported models in
+https://huggingface.co/csukuangfj/sherpa-onnx-streaming-conformer-en-2023-05-09
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_encoder_states()
+
+    def init_encoder_states(self, batch_size: int = 1):
+        encoder_meta = self.encoder.get_modelmeta().custom_metadata_map
+        print(encoder_meta)
+
+        model_type = encoder_meta["model_type"]
+        assert model_type == "conformer", model_type
+
+        decode_chunk_len = int(encoder_meta["decode_chunk_len"])
+        T = int(encoder_meta["T"])
+        pad_length = int(encoder_meta["pad_length"])
+
+        encoder_dim = int(encoder_meta["encoder_dim"])
+        cnn_module_kernel = int(encoder_meta["cnn_module_kernel"])
+        left_context = int(encoder_meta["left_context"])
+        num_encoder_layers = int(encoder_meta["num_encoder_layers"])
+
+        self.cached_attn = torch.zeros(
+            num_encoder_layers,
+            left_context,
+            batch_size,
+            encoder_dim,
+        ).numpy()
+        self.cached_conv = torch.zeros(
+            num_encoder_layers,
+            cnn_module_kernel - 1,
+            batch_size,
+            encoder_dim,
+        ).numpy()
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"pad_length: {pad_length}")
+        logging.info(f"encoder_dim: {encoder_dim}")
+        logging.info(f"cnn_module_kernel: {cnn_module_kernel}")
+        logging.info(f"left_context: {left_context}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+
+        self.segment = T
+        self.offset = decode_chunk_len
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def _build_encoder_input_output(
+        self, x: torch.Tensor, processed_lens: int
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        assert x.size(0) == 1
+        encoder_input = {
+            "x": x.numpy(),
+            "cached_attn": self.cached_attn,
+            "cached_conv": self.cached_conv,
+            "processed_lens": torch.full(
+                (1,), fill_value=processed_lens, dtype=torch.int64
+            ).numpy(),
+        }
+        encoder_output = ["encoder_out", "new_cached_attn", "new_cached_conv"]
+
+        return encoder_input, encoder_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        self.cached_attn = states[0]
+        self.cached_conv = states[1]
+
+    def run_encoder(self, x: torch.Tensor, num_processed_frames: int) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, self.T, C). It only implements N == 1
+          num_processed_frames:
+            Number of processed frames before subsampling.
+        Returns:
+          Return a 3-D tensor of shape (N, chunk_size, joiner_dim)
+        """
+        # assume subsampling_factor is 4
+        num_processed_frames = num_processed_frames // 4
+        encoder_input, encoder_output_names = self._build_encoder_input_output(
+            x, num_processed_frames
+        )
+        out = self.encoder.run(encoder_output_names, encoder_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    context_size: int,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+) -> List[int]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (1, T, joiner_dim)
+      context_size:
+        The context size of the decoder model.
+      decoder_out:
+        Optional. Decoder output of the previous chunk.
+      hyp:
+        Decoding results for previous chunks.
+    Returns:
+      Return the decoded results so far.
+    """
+
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert hyp is not None, hyp
+
+    encoder_out = encoder_out.squeeze(0)
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t : t + 1]
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+            decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(1.0 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    context_size = model.context_size
+    hyp = None
+    decoder_out = None
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            encoder_out = model.run_encoder(frames, num_processed_frames)
+            hyp, decoder_out = greedy_search(
+                model,
+                encoder_out,
+                context_size,
+                decoder_out,
+                hyp,
+            )
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py
new file mode 120000
index 000000000..7607623c8
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100755
index 000000000..304fa8693
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,351 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless5/scaling_converter.py
new file mode 120000
index 000000000..3b667058d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless5/streaming_beam_search.py
new file mode 120000
index 000000000..3a5f89833
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/streaming_beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/streaming_decode.py
new file mode 100755
index 000000000..f65f47fc2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/streaming_decode.py
@@ -0,0 +1,647 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless5/streaming_decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --left-context 32 \
+        --decode-chunk-size 8 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --decoding_method greedy_search \
+        --num-decode-streams 200
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    # Decoding in streaming requires causal convolution
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py
new file mode 100755
index 000000000..71b36e029
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless5/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 18
+    params.dim_feedforward = 1024
+    params.encoder_dim = 256
+    params.nhead = 4
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..66dc5f991
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1193 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If the batch contains more than 10 utterances AND
+            # if either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless6/__init__.py
new file mode 120000
index 000000000..b24e5e357
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/__init__.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless6/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless6/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless6/conformer.py
new file mode 100644
index 000000000..0667e7f61
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/conformer.py
@@ -0,0 +1,1043 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask
+
+
+class Conformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        middle_output_layer: int = None,  # 0-based layer index
+    ) -> None:
+        super(Conformer, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+        )
+
+        output_layers = []
+        if middle_output_layer is not None:
+            assert middle_output_layer >= 0 and middle_output_layer < num_encoder_layers
+            output_layers.append(middle_output_layer)
+
+        # The last layer is always needed.
+        output_layers.append(num_encoder_layers - 1)
+
+        self.encoder = ConformerEncoder(
+            encoder_layer, num_encoder_layers, output_layers=output_layers
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[List[torch.Tensor], torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, d_model)
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        assert x.size(0) == lengths.max().item()
+        mask = make_pad_mask(lengths)
+
+        layer_results = self.encoder(
+            x, pos_emb, src_key_padding_mask=mask, warmup=warmup
+        )  # (T, N, C)
+
+        return layer_results, lengths
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src = src + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        output_layers: List[int],
+    ) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+        self.output_layers = output_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> List[Tensor]:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        layer_results = []
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+            if i in self.output_layers:
+                # (T, N, C) --> (N, T, C)
+                layer_results.append(output.permute(1, 0, 2))
+
+        return layer_results
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py
new file mode 100755
index 000000000..95534efef
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py
@@ -0,0 +1,628 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless6/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--enable-distillation",
+        type=str2bool,
+        default=True,
+        help="Whether to eanble distillation.",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    layer_results, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    encoder_out = layer_results[-1]
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless6/decoder.py
new file mode 120000
index 000000000..0793c5709
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless6/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/export.py b/egs/librispeech/ASR/pruned_transducer_stateless6/export.py
new file mode 100755
index 000000000..38f48b2ed
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/export.py
@@ -0,0 +1,211 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless6/export.py \
+  --exp-dir ./pruned_transducer_stateless6/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless6/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless6/decode.py \
+        --exp-dir ./pruned_transducer_stateless6/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless6/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py b/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py
new file mode 100755
index 000000000..86cf34877
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py
@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import os
+from pathlib import Path
+
+import torch
+from asr_datamodule import LibriSpeechAsrDataModule
+from hubert_xlarge import HubertXlargeFineTuned
+from vq_utils import CodebookIndexExtractor
+
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless6/exp/",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--use-extracted-codebook",
+        type=str2bool,
+        default=False,
+        help="Whether to use the extracted codebook indexes.",
+    )
+
+    return parser
+
+
+def get_world_size():
+    warn_message = (
+        "It's better to use GPU to extrac codebook indices"
+        "Please set with commonds like: export CUDA_VISIBLE_DEVICES=0,1,2,3"
+    )
+    assert (
+        torch.cuda.is_available() and "CUDA_VISIBLE_DEVICES" in os.environ
+    ), warn_message
+    world_size = len(os.environ["CUDA_VISIBLE_DEVICES"].split(","))
+    assert world_size > 0, warn_message
+    return world_size
+
+
+def main():
+    world_size = get_world_size()
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    HubertXlargeFineTuned.add_arguments(parser)
+    CodebookIndexExtractor.add_arguments(parser)
+
+    args = parser.parse_args()
+    params = AttributeDict()
+    params.update(vars(args))
+
+    # reset some parameters needed by hubert.
+    params.update(HubertXlargeFineTuned.get_params())
+    params.device = torch.device("cuda", 0)
+    params.world_size = world_size
+
+    extractor = CodebookIndexExtractor(params=params)
+    if not params.use_extracted_codebook:
+        extractor.extract_and_save_embedding()
+        extractor.train_quantizer()
+        extractor.extract_codebook_indexes()
+
+    extractor.reuse_manifests()
+    extractor.join_manifests()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_decode.py
new file mode 100755
index 000000000..b8440f90a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_decode.py
@@ -0,0 +1,195 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import torch
+from asr_datamodule import LibriSpeechAsrDataModule
+from hubert_xlarge import HubertXlargeFineTuned
+
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless6/exp/",
+        help="The experiment dir",
+    )
+
+    return parser
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    hubert_model: HubertXlargeFineTuned,
+    params: AttributeDict,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+
+    Returns:
+      Return a dict, whose key is decoding method "ctc_greedy_search".
+      Its value is a list of tuples.
+      Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        # hyps is a list, every element is decode result of a sentence.
+        hyps = hubert_model.ctc_greedy_search(batch)
+
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_text, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            hyp_words = hyp_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+        results["ctc_greedy_search"].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 20 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{key}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    HubertXlargeFineTuned.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = AttributeDict()
+    params.update(vars(args))
+    # reset some parameters needed by hubert.
+    params.update(HubertXlargeFineTuned.get_params())
+
+    params.res_dir = params.exp_dir / f"ctc_greedy_search-{params.teacher_model_id}"
+
+    setup_logger(f"{params.res_dir}/log/log-ctc_greedy_search")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+    params.device = device
+
+    hubert_model = HubertXlargeFineTuned(params)
+
+    librispeech = LibriSpeechAsrDataModule(params)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            hubert_model=hubert_model,
+            params=params,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py
new file mode 100644
index 000000000..4f9417c9f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py
@@ -0,0 +1,208 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import torch
+from fairseq import checkpoint_utils, tasks, utils
+from fairseq.data.data_utils import post_process
+from omegaconf import OmegaConf
+
+from icefall.utils import AttributeDict
+
+
+def _load_hubert_model(params: AttributeDict):
+    """
+    Load the hubert model.
+
+    The model loaded is specified by params.hubert_model_dir
+    and params.teacher_model_id.
+
+    Returned model carries hubert,
+    while processor is responsible to map model's output to human readable transcripts.
+    """
+    cfg_task = OmegaConf.create(
+        {
+            "_name": "hubert_pretraining",
+            "single_target": True,
+            "fine_tuning": True,
+            "data": str(params.hubert_model_dir),
+        }
+    )
+    model_path = Path(params.hubert_model_dir) / (params.teacher_model_id + ".pt")
+    task = tasks.setup_task(cfg_task)
+    processor = task.target_dictionary
+    models, saved_cfg = checkpoint_utils.load_model_ensemble(
+        utils.split_paths(str(model_path), separator="\\"),
+        arg_overrides={},
+        strict=True,
+        suffix="",
+        num_shards=1,
+    )
+    model = models[0]
+    model.to(params.device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    return model, processor
+
+
+class HubertXlargeFineTuned:
+    """
+    A wrapper of hubert extra large fine-tuned model.
+
+    A teacher model is responsible for:
+        1. load teacher model
+        2. extracting embeddings to train quantizer.
+        3. extract codebook indices
+        4. verify its performance with ctc_greedy_search method.
+    """
+
+    def __init__(self, params: AttributeDict):
+        self.model, self.processor = _load_hubert_model(params)
+        self.w2v_model = self.model.w2v_encoder.w2v_model
+        self.params = params
+
+    @staticmethod
+    def get_params() -> AttributeDict:
+        """Return a dict containing parameters defined in other modules.
+
+        Their default value conflits to hubert's requirements so they are reset as following.
+        """
+        params = AttributeDict(
+            {
+                # parameters defined in asr_datamodule.py
+                "input_strategy": "AudioSamples",
+                "enable_musan": False,
+                "enable_spec_aug": False,
+                "return_cuts": True,
+                "drop_last": False,
+                # parameters used by quantizer
+                "embedding_dim": 1280,
+            }
+        )
+        return params
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        # Options about model loading.
+        parser.add_argument(
+            "--hubert-model-dir",
+            type=Path,
+            default="./pruned_transducer_stateless6/exp/hubert_models/",
+            help="path to save downloaded hubert models.",
+        )
+
+        parser.add_argument(
+            "--teacher-model-id",
+            type=str,
+            default="hubert_xtralarge_ll60k_finetune_ls960",
+            help="""could be one of:
+            [
+                "hubert_xtralarge_ll60k_finetune_ls960",  # fine-tuned model.
+                "hubert_xtralarge_ll60k.pt",  # pretrained model without fintuing.
+            ]""",
+        )
+        parser.add_argument(
+            "--total-layers",
+            type=int,
+            default=48,
+        )
+
+    # Modified from HubertModel.forward to extract all middle layers output
+    def extract_layers_result(
+        self,
+        batch: Dict,
+    ) -> List[torch.Tensor]:
+        """
+        Extract activations from all layers.
+        """
+        features = batch["inputs"]
+
+        # corresponding task.normalize in fairseq
+        features = torch.nn.functional.layer_norm(features, features.shape)
+
+        supervisions = batch["supervisions"]
+        num_samples = supervisions["num_samples"]
+        B, T = features.shape
+        padding_mask = torch.arange(0, T).expand(B, T) > num_samples.reshape([-1, 1])
+
+        padding_mask = padding_mask.to(self.params.device)
+        features = features.to(self.params.device)
+
+        features = self.w2v_model.forward_features(features)
+
+        features = features.transpose(1, 2)
+        features = self.w2v_model.layer_norm(features)
+
+        padding_mask = self.w2v_model.forward_padding_mask(features, padding_mask)
+
+        if self.w2v_model.post_extract_proj is not None:
+            features = self.w2v_model.post_extract_proj(features)
+
+        _, layer_results = self.w2v_model.encoder(
+            features,
+            padding_mask=padding_mask,
+        )
+        return layer_results
+
+    def extract_embedding(self, batch) -> Tuple[torch.tensor, List[int]]:
+        """
+        Eextract embeddings specified by self.params.embedding_layer.
+
+        These embeddings could be used to train quantizer
+        or to extract codebook indexes.
+
+        The returned List[int] is valid length of each embedding.
+        We only want to store codebook indexes related to
+        these valid embeddings.
+        """
+        supervisions = batch["supervisions"]
+        cut_list = supervisions["cut"]
+        assert all(c.start == 0 for c in cut_list)
+        layer_results = self.extract_layers_result(batch)
+        embeddings = layer_results[self.params.embedding_layer - 1][0]
+        encoder_embedding = embeddings.transpose(0, 1)  # N, T, C
+        N = encoder_embedding.shape[0]
+        assert len(cut_list) == N
+        # 320 is from: 16,000 / 50 = sample_rate / hbuert output frame rate
+        num_frames = (supervisions["num_samples"] // 320).tolist()
+        return encoder_embedding, num_frames
+
+    def ctc_greedy_search(self, batch):
+        """
+        Mainly used to verify hubert model is used correctly.
+        """
+        layer_results = self.extract_layers_result(batch=batch)
+        encoder_out = self.w2v_model.encoder.layer_norm(
+            layer_results[self.params.total_layers - 1][0]
+        )
+        encoder_out = self.model.w2v_encoder.proj(encoder_out.transpose(0, 1))
+
+        toks = encoder_out.argmax(dim=-1)
+        blank = 0
+        toks = [tok.unique_consecutive() for tok in toks]
+        hyps = [self.processor.string(tok[tok != blank].int().cpu()) for tok in toks]
+        hyps = [post_process(hyp, "letter") for hyp in hyps]
+
+        return hyps
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless6/joiner.py
new file mode 120000
index 000000000..815fd4bb6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/model.py b/egs/librispeech/ASR/pruned_transducer_stateless6/model.py
new file mode 100644
index 000000000..daadb70c9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/model.py
@@ -0,0 +1,259 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        num_codebooks: int = 0,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output
+            contains unnormalized probs, i.e., not processed by log-softmax.
+          num_codebooks:
+            Used by distillation loss.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+        from icefall import is_module_available
+
+        if not is_module_available("multi_quantization"):
+            raise ValueError("Please 'pip install multi_quantization' first.")
+
+        from multi_quantization.prediction import JointCodebookLoss
+
+        if num_codebooks > 0:
+            self.codebook_loss_net = JointCodebookLoss(
+                predictor_channels=encoder_dim,
+                num_codebooks=num_codebooks,
+                is_joint=False,
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        reduction: str = "sum",
+        codebook_indexes: torch.Tensor = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          codebook_indexes:
+            codebook_indexes extracted from a teacher model.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        layer_results, x_lens = self.encoder(x, x_lens, warmup=warmup)
+        encoder_out = layer_results[-1]
+        middle_layer_output = layer_results[0]
+        if self.training and codebook_indexes is not None:
+            assert hasattr(self, "codebook_loss_net")
+            if codebook_indexes.shape[1] != middle_layer_output.shape[1]:
+                codebook_indexes = self.concat_successive_codebook_indexes(
+                    middle_layer_output, codebook_indexes
+                )
+            codebook_loss = self.codebook_loss_net(
+                middle_layer_output, codebook_indexes
+            )
+        else:
+            # when codebook index is not available.
+            codebook_loss = None
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction=reduction,
+            )
+
+        return (simple_loss, pruned_loss, codebook_loss)
+
+    @staticmethod
+    def concat_successive_codebook_indexes(middle_layer_output, codebook_indexes):
+        # Output rate of hubert is 50 frames per second,
+        # while that of current encoder is 25.
+        # Following code handling two issues:
+        # 1.
+        #   Roughly speaking, to generate another frame output,
+        #   hubert needes extra two frames,
+        #   while current encoder needs extra four frames.
+        #   Suppose there are only extra three frames provided,
+        #   hubert will generate another frame while current encoder does nothing.
+        # 2.
+        #   codebook loss is a frame-wise loss, to enalbe 25 frames studnet output
+        #   learns from 50 frames teacher output, two successive frames of teacher model
+        #   output is concatenated together.
+        t_expected = middle_layer_output.shape[1]
+        N, T, C = codebook_indexes.shape
+
+        # Handling issue 1.
+        if T >= t_expected * 2:
+            codebook_indexes = codebook_indexes[:, : t_expected * 2, :]
+        # Handling issue 2.
+        codebook_indexes = codebook_indexes.reshape(N, t_expected, C * 2)
+        assert middle_layer_output.shape[1] == codebook_indexes.shape[1]
+        return codebook_indexes
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless6/optim.py
new file mode 120000
index 000000000..e2deb4492
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless6/scaling.py
new file mode 120000
index 000000000..09d802cc4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless6/test_model.py
new file mode 100755
index 000000000..9bd75ba21
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/test_model.py
@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless6/test_model.py
+"""
+
+import torch
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.enable_distiallation = False
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
new file mode 100755
index 000000000..8f033cb9a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
@@ -0,0 +1,1153 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless6/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless6/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless6/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless6/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+# For distillation with codebook_indexes:
+
+./pruned_transducer_stateless6/train.py \
+  --manifest-dir ./data/vq_fbank \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless6/exp \
+  --full-libri 0 \
+  --max-duration 300
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut, MonoCut
+from lhotse.dataset.collation import collate_custom_field
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless6/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--codebook-loss-scale",
+        type=float,
+        default=0.1,
+        help="The scale of codebook loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--enable-distillation",
+        type=str2bool,
+        default=True,
+        help="Whether to eanble distillation.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+            # parameters for distillation with codebook indexes.
+            "distillation_layer": 5,  # 0-based index
+            # Since output rate of hubert is 50, while that of encoder is 8,
+            # two successive codebook_index are concatenated together.
+            # Detailed in function Transducer::concat_sucessive_codebook_indexes
+            "num_codebooks": 16,  # used to construct distillation loss
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        middle_output_layer=params.distillation_layer
+        if params.enable_distillation
+        else None,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        num_codebooks=params.num_codebooks if params.enable_distillation else 0,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def extract_codebook_indexes(batch):
+    cuts = batch["supervisions"]["cut"]
+    # -100 is identical to ignore_value in CE loss computation.
+    cuts_pre_mixed = [c if isinstance(c, MonoCut) else c.tracks[0].cut for c in cuts]
+    codebook_indexes, codebook_indexes_lens = collate_custom_field(
+        cuts_pre_mixed, "codebook_indexes", pad_value=-100
+    )
+    return codebook_indexes, codebook_indexes_lens
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    info = MetricsTracker()
+    if is_training and params.enable_distillation:
+        codebook_indexes, _ = extract_codebook_indexes(batch)
+        codebook_indexes = codebook_indexes.to(device)
+    else:
+        codebook_indexes = None
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, codebook_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            codebook_indexes=codebook_indexes,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+            # If the batch contains more than 10 utterances AND
+            # if either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+        if is_training and params.enable_distillation:
+            assert codebook_loss is not None
+            loss += params.codebook_loss_scale * codebook_loss
+
+    assert loss.requires_grad == is_training
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if is_training and params.enable_distillation:
+        info["codebook_loss"] = codebook_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+
+    # Note: it's better to set --spec-aug-time-warpi-factor=-1
+    # when doing distillation with vq.
+    assert args.spec_aug_time_warp_factor < 1
+
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py b/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
new file mode 100644
index 000000000..3bca7db2c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
@@ -0,0 +1,468 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import copy
+import glob
+import logging
+import os
+from functools import cached_property
+from pathlib import Path
+from typing import List, Tuple
+
+import numpy as np
+import torch
+import torch.multiprocessing as mp
+
+from icefall import is_module_available
+
+if not is_module_available("multi_quantization"):
+    raise ValueError("Please 'pip install multi_quantization' first.")
+
+import multi_quantization as quantization
+from asr_datamodule import LibriSpeechAsrDataModule
+from hubert_xlarge import HubertXlargeFineTuned
+from lhotse import CutSet, load_manifest
+from lhotse.cut import MonoCut
+from lhotse.features.io import NumpyHdf5Writer
+
+from icefall.utils import AttributeDict, setup_logger
+
+
+class CodebookIndexExtractor:
+    """
+    A wrapper of quantiation.Quantizer.
+
+    It's responsible for:
+        1. extract and save activations from a teacher model.
+        2. train quantizer from previous activations.
+        3. extract codebook indexes for whole training set.
+           Normally this step needs multi GPUs.
+    """
+
+    def __init__(self, params: AttributeDict):
+        self.params = params
+        params.subsets = ["clean-100"]
+        if self.params.full_libri:
+            self.params.subsets += ["clean-360", "other-500"]
+
+        self.init_dirs()
+        setup_logger(f"{self.vq_dir}/log-vq_extraction")
+
+    def init_dirs(self):
+        # vq_dir is the root dir for quantization, containing:
+        # training data, trained quantizer, and extracted codebook indexes
+        self.vq_dir = (
+            self.params.exp_dir
+            / f"vq/{self.params.teacher_model_id}_layer{self.params.embedding_layer}_cb{self.params.num_codebooks}/"
+        )
+        self.vq_dir.mkdir(parents=True, exist_ok=True)
+
+        # manifest_dir contains:
+        # splited original manifests, extracted codebook indexes with related manifests # noqa
+        self.manifest_dir = self.vq_dir / f"splits{self.params.world_size}"
+        self.manifest_dir.mkdir(parents=True, exist_ok=True)
+
+        # It's doesn't matter whether ori_manifest_dir is str or Path.
+        # Set it to Path to be consistent.
+        self.ori_manifest_dir = Path("./data/fbank/")
+        self.dst_manifest_dir = Path(
+            f"./data/vq_fbank_layer"
+            + f"{self.params.embedding_layer}_cb{self.params.num_codebooks}/"
+        )
+
+        self.dst_manifest_dir.mkdir(parents=True, exist_ok=True)
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        # Options about teacher embeddings eatraction.
+        parser.add_argument(
+            "--embedding-layer",
+            type=int,
+            help="layer to extract teacher embeddings, 1-based.",
+            default=36,
+        )
+
+        parser.add_argument(
+            "--num-utts",
+            type=int,
+            default=1000,
+            help="num utts to train quantizer",
+        )
+
+        parser.add_argument(
+            "--num-codebooks",
+            type=int,
+            default=8,
+            help="""number of codebooks,
+            i.e. number of codebook indexes each teacher embedding is compressed.
+            """,
+        )
+
+    @property
+    def embedding_file_path(self):
+        """
+        The saved embedding is used to train quantizer.
+        """
+        embedding_file_id = (
+            f"num_utts_{self.params.num_utts}"
+            + f"-layer_{self.params.embedding_layer}"
+            + "-embedding_embeddings.h5"
+        )
+
+        embedding_file_path = self.vq_dir / embedding_file_id
+        return embedding_file_path
+
+    @torch.no_grad()
+    def extract_and_save_embedding(self):
+        """
+        The extract embedding is used to train quantizer.
+        """
+        if self.embedding_file_path.exists():
+            warn_message = (
+                f"{self.embedding_file_path} already exists."
+                + " Skip extracting embeddings from teacher model"
+            )
+            logging.warn(warn_message)
+            return
+
+        logging.info("Start to extract embeddings for training the quantizer.")
+        total_cuts = 0
+        with NumpyHdf5Writer(self.embedding_file_path) as writer:
+            for batch_idx, batch in enumerate(self.quantizer_train_dl):
+                cut_list = batch["supervisions"]["cut"]
+                (
+                    encoder_embedding,
+                    num_frames,
+                ) = self.teacher_model.extract_embedding(batch)
+                encoder_embedding = encoder_embedding.cpu().numpy()
+                for idx, cut in enumerate(cut_list):
+                    cut.encoder_embedding = writer.store_array(
+                        key=cut.id,
+                        value=encoder_embedding[idx][: num_frames[idx]],
+                    )
+                total_cuts += len(cut_list)
+                logging.info(
+                    f"Processed {total_cuts} output of {self.params.num_utts} cuts."
+                )
+
+        logging.info(f"Processed all {total_cuts} cuts.")
+
+    @property
+    def quantizer_train_dl(self):
+        # used to train quantizer.
+        librispeech = LibriSpeechAsrDataModule(self.params)
+        quantizer_trian_cuts = librispeech.train_clean_100_cuts().subset(
+            first=self.params.num_utts
+        )
+        return librispeech.train_dataloaders(quantizer_trian_cuts)
+
+    @cached_property
+    def quantizer_file_path(self):
+        quantizer_file_id = (
+            f"num_utts-{self.params.num_utts}"
+            + f"-layer-{self.params.embedding_layer}"
+            + f"-num_codebooks_{self.params.num_codebooks}"
+            + "-quantizer.pt"
+        )
+        quantizer_file_path = Path(self.vq_dir) / quantizer_file_id
+
+        return quantizer_file_path
+
+    def train_quantizer(self):
+        if self.quantizer_file_path.exists():
+            warn_message = (
+                f"{self.quantizer_file_path} already exists."
+                + " Skip trainning quantizer."
+            )
+            logging.warn(warn_message)
+            return
+
+        assert self.embedding_file_path.exists()
+        logging.info("Start to train quantizer.")
+        trainer = quantization.QuantizerTrainer(
+            dim=self.params.embedding_dim,
+            bytes_per_frame=self.params.num_codebooks,
+            device=self.params.device,
+        )
+        train, valid = quantization.read_hdf5_data(self.embedding_file_path)
+        B = 512  # Minibatch size, this is very arbitrary,
+        # it's close to what we used when we tuned this method.
+
+        def minibatch_generator(data: torch.Tensor, repeat: bool):
+            assert 3 * B < data.shape[0]
+            cur_offset = 0
+            while True if repeat else cur_offset + B <= data.shape[0]:
+                start = cur_offset % (data.shape[0] + 1 - B)
+                end = start + B
+                cur_offset += B
+                yield data[start:end, :].to(self.params.device).to(dtype=torch.float)
+
+        for x in minibatch_generator(train, repeat=True):
+            trainer.step(x)
+            if trainer.done():
+                break
+
+        quantizer = trainer.get_quantizer()
+        torch.save(quantizer.state_dict(), self.quantizer_file_path)
+
+    def split_ori_manifests(self):
+        """
+        When multi gpus are available, split original manifests
+        and extract codebook indexes in a prallel way.
+        """
+        for subset in self.params.subsets:
+            logging.info(f"About to split {subset}.")
+            ori_manifest = f"./data/fbank/librispeech_cuts_train-{subset}.jsonl.gz"
+            split_cmd = f"lhotse split {self.params.world_size} {ori_manifest} {self.manifest_dir}"
+            os.system(f"{split_cmd}")
+
+    def join_manifests(self):
+        """
+        Join the vq manifest to the original manifest according to cut id.
+        """
+        logging.info("Start to join manifest files.")
+        for subset in self.params.subsets:
+            vq_manifest_path = (
+                self.dst_manifest_dir / f"librispeech_cuts_train-{subset}-vq.jsonl.gz"
+            )
+            ori_manifest_path = (
+                self.ori_manifest_dir / f"librispeech_cuts_train-{subset}.jsonl.gz"
+            )
+            dst_vq_manifest_path = (
+                self.dst_manifest_dir / f"librispeech_cuts_train-{subset}.jsonl.gz"
+            )
+            cuts_vq = load_manifest(vq_manifest_path)
+            cuts_ori = load_manifest(ori_manifest_path)
+            assert len(cuts_vq) == len(cuts_ori), "Cuts should have the same length!"
+
+            if set(cuts_vq.ids) == set(cuts_ori.ids):
+                # IDs match exactly
+                cuts_vq = cuts_vq.sort_like(cuts_ori)
+                for cut_idx, (cut_vq, cut_ori) in enumerate(zip(cuts_vq, cuts_ori)):
+                    assert cut_vq.id == cut_ori.id, (cut_vq.id, cut_ori.id)
+                    cut_ori.codebook_indexes = cut_vq.codebook_indexes
+            else:
+                # in case of ID mismatch, remap them
+                # get the mapping between audio and cut ID
+                logging
+                ori_id_map = {}
+                for id in cuts_ori.ids:
+                    # some text normalization
+                    if "sp" in id:
+                        clean_id = "-".join(id.split("-")[:3]) + "_" + id.split("_")[-1]
+                    else:
+                        clean_id = "-".join(id.split("-")[:3])
+                    ori_id_map[clean_id] = id
+
+                for id in cuts_vq.ids:
+                    if "sp" in id:
+                        clean_id = "-".join(id.split("-")[:3]) + "_" + id.split("_")[-1]
+                    else:
+                        clean_id = "-".join(id.split("-")[:3])
+                    assert clean_id in ori_id_map, clean_id
+                    cuts_ori[ori_id_map[clean_id]].codebook_indexes = cuts_vq[
+                        id
+                    ].codebook_indexes
+
+            CutSet.from_cuts(cuts_ori).to_jsonl(dst_vq_manifest_path)
+            logging.info(f"Processed {subset}.")
+            logging.info(f"Saved to {dst_vq_manifest_path}.")
+
+    def merge_vq_manifests(self):
+        """
+        Merge generated vq included manfiests and storage to self.dst_manifest_dir.
+        """
+        for subset in self.params.subsets:
+            vq_manifests = (
+                f"{self.manifest_dir}/"
+                + f"with_codebook_indexes-librispeech-cuts_train-{subset}*.jsonl.gz"
+            )
+            dst_vq_manifest = (
+                self.dst_manifest_dir / f"librispeech_cuts_train-{subset}-vq.jsonl.gz"
+            )
+            if 1 == self.params.world_size:
+                merge_cmd = f"cp {vq_manifests} {dst_vq_manifest}"
+            else:
+                merge_cmd = f"lhotse combine {vq_manifests} {dst_vq_manifest}"
+            os.system(f"{merge_cmd}")
+
+    def reuse_manifests(self):
+        """
+        Only train-* subsets are extracted codebook indexes from.
+        The reset subsets are just a link from ./data/fbank.
+        """
+
+        def is_train(manifest: str) -> bool:
+            for train_subset in ["clean-100", "clean-360", "other-500"]:
+                if train_subset in manifest:
+                    return True
+            return False
+
+        # Type of self.ori_nanifest_dir is Path
+        # and result type of glob.glob is str.
+        reusable_manifests = [
+            manifest
+            for manifest in glob.glob(f"{self.ori_manifest_dir}/*.gz")
+            if not is_train(manifest)
+        ]
+        for manifest_path in reusable_manifests:
+            ori_manifest_path = Path(manifest_path).resolve()
+            # Path cannot used as a parameter of str.replace.
+            # Cast them to str.
+            dst_manifest_path = Path(
+                manifest_path.replace(
+                    str(self.ori_manifest_dir), str(self.dst_manifest_dir)
+                )
+            ).resolve()
+            if not dst_manifest_path.exists():
+                os.symlink(ori_manifest_path, dst_manifest_path)
+
+    def create_vq_fbank(self):
+        self.merge_vq_manifests()
+
+    @cached_property
+    def teacher_model(self):
+        return HubertXlargeFineTuned(self.params)
+
+    @cached_property
+    def quantizer(self):
+        assert self.quantizer_file_path.exists()
+        quantizer = quantization.Quantizer(
+            dim=self.params.embedding_dim,
+            num_codebooks=self.params.num_codebooks,
+            codebook_size=256,
+        )
+        quantizer.load_state_dict(torch.load(self.quantizer_file_path))
+        quantizer.to(self.params.device)
+        return quantizer
+
+    def load_ori_dl(self, subset):
+        if self.params.world_size == 1:
+            ori_manifest_path = f"./data/fbank/librispeech_cuts_train-{subset}.jsonl.gz"
+        else:
+            ori_manifest_path = (
+                self.manifest_dir
+                / f"librispeech_cuts_train-{subset}.{self.params.manifest_index}.jsonl.gz"  # noqa
+            )
+
+        cuts = load_manifest(ori_manifest_path)
+        dl = LibriSpeechAsrDataModule(self.params).train_dataloaders(cuts)
+        return dl
+
+    def _release_gpu_memory(self):
+        self.__dict__.pop("teacher_model", None)
+        self.__dict__.pop("quantizer", None)
+        torch.cuda.empty_cache()
+
+    def extract_codebook_indexes(self):
+        logging.info("Start to extract codebook indexes.")
+        if self.params.world_size == 1:
+            self.extract_codebook_indexes_imp()
+        else:
+            # Since a new extractor will be created for each rank in
+            # compute_codebook_indexes_parallel, it's better to
+            # release the GPU memory occupied by current extractor.
+            self._release_gpu_memory()
+
+            # Prepare split manifests for each job.
+            self.split_ori_manifests()
+            mp.spawn(
+                compute_codebook_indexes_parallel,
+                args=(self.params,),
+                nprocs=self.params.world_size,
+                join=True,
+            )
+        self.create_vq_fbank()
+
+    @torch.no_grad()
+    def extract_codebook_indexes_imp(self):
+        for subset in self.params.subsets:
+            num_cuts = 0
+            new_cuts = []
+            if self.params.world_size == 1:
+                manifest_file_id = f"{subset}"
+            else:
+                manifest_file_id = f"{subset}-{self.params.manifest_index}"
+
+            manifest_file_path = self.manifest_dir / manifest_file_id
+            with NumpyHdf5Writer(manifest_file_path) as writer:
+                for batch_idx, batch in enumerate(self.load_ori_dl(subset)):
+                    (
+                        encoder_embedding,
+                        num_frames,
+                    ) = self.teacher_model.extract_embedding(batch)
+                    codebook_indexes = self.quantizer.encode(encoder_embedding)
+                    # [N, T, C]
+                    codebook_indexes = codebook_indexes.to("cpu").numpy()
+                    assert np.min(codebook_indexes) >= 0
+                    assert np.max(codebook_indexes) < 256
+                    supervisions = batch["supervisions"]
+                    cut_list = supervisions["cut"]
+                    assert len(cut_list) == codebook_indexes.shape[0]
+                    assert all(c.start == 0 for c in supervisions["cut"])
+
+                    new_cut_list = []
+                    for idx, cut in enumerate(cut_list):
+                        new_cut = MonoCut(
+                            id=cut.id,
+                            start=cut.start,
+                            duration=cut.duration,
+                            channel=cut.channel,
+                        )
+                        new_cut.codebook_indexes = writer.store_array(
+                            key=cut.id,
+                            value=codebook_indexes[idx][: num_frames[idx]],
+                            frame_shift=0.02,
+                            temporal_dim=0,
+                            start=0,
+                        )
+                        new_cut_list.append(new_cut)
+                    new_cuts += new_cut_list
+                    num_cuts += len(cut_list)
+                    message = f"Processed {num_cuts} cuts from {subset}"
+                    if self.params.world_size > 1:
+                        message += f" by job {self.params.manifest_index}"
+                    logging.info(f"{message}.")
+
+                json_file_path = (
+                    self.manifest_dir
+                    / f"with_codebook_indexes-librispeech-cuts_train-{manifest_file_id}.jsonl.gz"  # noqa
+                )
+                CutSet.from_cuts(new_cuts).to_jsonl(json_file_path)
+
+
+@torch.no_grad()
+def compute_codebook_indexes_parallel(
+    rank: int,
+    params,
+) -> List[Tuple[str, List[int]]]:
+    """Create an extractor for each rank and extract codebook indexes parallelly.
+
+    Normally, this function is called by torch.multiprocessing
+    when multi GPUs are available.
+    """
+    params = copy.deepcopy(params)
+    device = torch.device("cuda", rank)
+    params.device = device
+
+    # rank is 0-based while split manifests by "lhotse split" is 1-based.
+    params.manifest_index = rank + 1
+
+    extractor = CodebookIndexExtractor(params=params)
+    extractor.extract_codebook_indexes_imp()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless7/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py b/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py
new file mode 100644
index 000000000..bfb5fe609
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py
@@ -0,0 +1,206 @@
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List
+
+import k2
+import torch
+
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+
+# The force alignment problem can be formulated as finding
+# a path in a rectangular lattice, where the path starts
+# from the lower left corner and ends at the upper right
+# corner. The horizontal axis of the lattice is `t` (representing
+# acoustic frame indexes) and the vertical axis is `u` (representing
+# BPE tokens of the transcript).
+#
+# The notations `t` and `u` are from the paper
+# https://arxiv.org/pdf/1211.3711.pdf
+#
+# Beam search is used to find the path with the highest log probabilities.
+#
+# It assumes the maximum number of symbols that can be
+# emitted per frame is 1.
+
+
+def batch_force_alignment(
+    model: torch.nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    ys_list: List[List[int]],
+    beam_size: int = 4,
+) -> List[int]:
+    """Compute the force alignment of a batch of utterances given their transcripts
+    in BPE tokens and the corresponding acoustic output from the encoder.
+
+    Caution:
+      This function is modified from `modified_beam_search` in beam_search.py.
+      We assume that the maximum number of sybmols per frame is 1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A tensor of shape (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      ys_list:
+        A list of BPE token IDs list. We require that for each utterance i,
+        len(ys_list[i]) <= encoder_out_lens[i].
+      beam_size:
+        Size of the beam used in beam search.
+
+    Returns:
+      Return a list of frame indexes list for each utterance i,
+      where len(ans[i]) == len(ys_list[i]).
+    """
+    assert encoder_out.ndim == 3, encoder_out.ndim
+    assert encoder_out.size(0) == len(ys_list), (encoder_out.size(0), len(ys_list))
+    assert encoder_out.size(0) > 0, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    sorted_indices = packed_encoder_out.sorted_indices.tolist()
+    encoder_out_lens = encoder_out_lens.tolist()
+    ys_lens = [len(ys) for ys in ys_list]
+    sorted_encoder_out_lens = [encoder_out_lens[i] for i in sorted_indices]
+    sorted_ys_lens = [ys_lens[i] for i in sorted_indices]
+    sorted_ys_list = [ys_list[i] for i in sorted_indices]
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                timestamp=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+        sorted_encoder_out_lens = sorted_encoder_out_lens[:batch_size]
+        sorted_ys_lens = sorted_ys_lens[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out, decoder_out, project_input=False
+        )  # (num_hyps, 1, 1, vocab_size)
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(
+            shape=log_probs_shape, value=log_probs.reshape(-1)
+        )  # [batch][num_hyps*vocab_size]
+
+        for i in range(batch_size):
+            for h, hyp in enumerate(A[i]):
+                pos_u = len(hyp.timestamp)
+                idx_offset = h * vocab_size
+                if (sorted_encoder_out_lens[i] - 1 - t) >= (sorted_ys_lens[i] - pos_u):
+                    # emit blank token
+                    new_hyp = Hypothesis(
+                        log_prob=ragged_log_probs[i][idx_offset + blank_id],
+                        ys=hyp.ys[:],
+                        timestamp=hyp.timestamp[:],
+                    )
+                    B[i].add(new_hyp)
+                if pos_u < sorted_ys_lens[i]:
+                    # emit non-blank token
+                    new_token = sorted_ys_list[i][pos_u]
+                    new_hyp = Hypothesis(
+                        log_prob=ragged_log_probs[i][idx_offset + new_token],
+                        ys=hyp.ys + [new_token],
+                        timestamp=hyp.timestamp + [t],
+                    )
+                    B[i].add(new_hyp)
+
+            if len(B[i]) > beam_size:
+                B[i] = B[i].topk(beam_size, length_norm=True)
+
+    B = B + finalized_B
+    sorted_hyps = [b.get_most_probable() for b in B]
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    hyps = [sorted_hyps[i] for i in unsorted_indices]
+    ans = []
+    for i, hyp in enumerate(hyps):
+        assert hyp.ys[context_size:] == ys_list[i], (hyp.ys[context_size:], ys_list[i])
+        ans.append(hyp.timestamp)
+
+    return ans
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py b/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py
new file mode 100755
index 000000000..27ef0a244
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py
@@ -0,0 +1,345 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+The script gets forced-alignments based on the modified_beam_search decoding method.
+Both token-level alignments and word-level alignments are saved to the new cuts manifests.
+
+It loads a checkpoint and uses it to get the forced-alignments.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+Usage of this script:
+
+./pruned_transducer_stateless7/compute_ali.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --dataset test-clean \
+    --max-duration 300 \
+    --beam-size 4 \
+    --cuts-out-dir data/fbank_ali_beam_search
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from alignment import batch_force_alignment
+from asr_datamodule import LibriSpeechAsrDataModule
+from lhotse import CutSet
+from lhotse.serialization import SequentialJsonlWriter
+from lhotse.supervision import AlignmentItem
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import AttributeDict, convert_timestamp, parse_timestamp
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="""The name of the dataset to compute alignments for.
+        Possible values are:
+        - test-clean
+        - test-other
+        - train-clean-100
+        - train-clean-360
+        - train-other-500
+        - dev-clean
+        - dev-other
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--cuts-out-dir",
+        type=str,
+        default="data/fbank_ali_beam_search",
+        help="The dir to save the new cuts manifests with alignments",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def align_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Tuple[List[List[str]], List[List[str]], List[List[float]], List[List[float]]]:
+    """Get forced-alignments for one batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      token_list:
+        A list of token list.
+      word_list:
+        A list of word list.
+      token_time_list:
+        A list of timestamps list for tokens.
+      word_time_list.
+        A list of timestamps list for words.
+
+      where len(token_list) == len(word_list) == len(token_time_list) == len(word_time_list),
+      len(token_list[i]) == len(token_time_list[i]),
+      and len(word_list[i]) == len(word_time_list[i])
+
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    texts = supervisions["text"]
+    ys_list: List[List[int]] = sp.encode(texts, out_type=int)
+
+    frame_indexes = batch_force_alignment(
+        model, encoder_out, encoder_out_lens, ys_list, params.beam_size
+    )
+
+    token_list = []
+    word_list = []
+    token_time_list = []
+    word_time_list = []
+    for i in range(encoder_out.size(0)):
+        tokens = sp.id_to_piece(ys_list[i])
+        words = texts[i].split()
+        token_time = convert_timestamp(
+            frame_indexes[i], params.subsampling_factor, params.frame_shift_ms
+        )
+        word_time = parse_timestamp(tokens, token_time)
+        assert len(word_time) == len(words), (len(word_time), len(words))
+
+        token_list.append(tokens)
+        word_list.append(words)
+        token_time_list.append(token_time)
+        word_time_list.append(word_time)
+
+    return token_list, word_list, token_time_list, word_time_list
+
+
+def align_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    writer: SequentialJsonlWriter,
+) -> None:
+    """Get forced-alignments for the dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      writer:
+        Writer to save the cuts with alignments.
+    """
+    log_interval = 20
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    for batch_idx, batch in enumerate(dl):
+        token_list, word_list, token_time_list, word_time_list = align_one_batch(
+            params=params, model=model, sp=sp, batch=batch
+        )
+
+        cut_list = batch["supervisions"]["cut"]
+        for cut, token, word, token_time, word_time in zip(
+            cut_list, token_list, word_list, token_time_list, word_time_list
+        ):
+            assert len(cut.supervisions) == 1, f"{len(cut.supervisions)}"
+            token_ali = [
+                AlignmentItem(
+                    symbol=token[i],
+                    start=round(token_time[i], ndigits=3),
+                    duration=None,
+                )
+                for i in range(len(token))
+            ]
+            word_ali = [
+                AlignmentItem(
+                    symbol=word[i], start=round(word_time[i], ndigits=3), duration=None
+                )
+                for i in range(len(word))
+            ]
+            cut.supervisions[0].alignment = {"word": word_ali, "token": token_ali}
+            writer.write(cut, flush=True)
+
+        num_cuts += len(cut_list)
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.dataset == "test-clean":
+        test_clean_cuts = librispeech.test_clean_cuts()
+        dl = librispeech.test_dataloaders(test_clean_cuts)
+    elif params.dataset == "test-other":
+        test_other_cuts = librispeech.test_other_cuts()
+        dl = librispeech.test_dataloaders(test_other_cuts)
+    elif params.dataset == "train-clean-100":
+        train_clean_100_cuts = librispeech.train_clean_100_cuts()
+        dl = librispeech.train_dataloaders(train_clean_100_cuts)
+    elif params.dataset == "train-clean-360":
+        train_clean_360_cuts = librispeech.train_clean_360_cuts()
+        dl = librispeech.train_dataloaders(train_clean_360_cuts)
+    elif params.dataset == "train-other-500":
+        train_other_500_cuts = librispeech.train_other_500_cuts()
+        dl = librispeech.train_dataloaders(train_other_500_cuts)
+    elif params.dataset == "dev-clean":
+        dev_clean_cuts = librispeech.dev_clean_cuts()
+        dl = librispeech.valid_dataloaders(dev_clean_cuts)
+    else:
+        assert params.dataset == "dev-other", f"{params.dataset}"
+        dev_other_cuts = librispeech.dev_other_cuts()
+        dl = librispeech.valid_dataloaders(dev_other_cuts)
+
+    cuts_out_dir = Path(params.cuts_out_dir)
+    cuts_out_dir.mkdir(parents=True, exist_ok=True)
+    cuts_out_path = cuts_out_dir / f"librispeech_cuts_{params.dataset}.jsonl.gz"
+
+    with CutSet.open_writer(cuts_out_path) as writer:
+        align_dataset(dl=dl, params=params, model=model, sp=sp, writer=writer)
+
+    logging.info(
+        f"For dataset {params.dataset}, the cut manifest with framewise token alignments "
+        f"and word alignments are saved to {cuts_out_path}"
+    )
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decode.py
new file mode 100755
index 000000000..eb8841cc4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decode.py
@@ -0,0 +1,962 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search with RNNLM shallow fusion
+./pruned_transducer_stateless7/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+(9) modified beam search with LM shallow fusion + LODR
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --max-duration 600 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --decoding-method modified_beam_search_LODR \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.4 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16 \
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+    modified_beam_search_ngram_rescoring,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=3,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring, or LODR
+            """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+
+    else:
+        LM = None
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py
new file mode 100644
index 000000000..b0e4be0d1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py
@@ -0,0 +1,862 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang,
+#                                                 Yifan Yang,)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7/decode_gigaspeech.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp_giga_finetune \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+
+# from asr_datamodule import LibriSpeechAsrDataModule
+from gigaspeech import GigaSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from gigaspeech_scoring import asr_text_post_processing
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp_giga_finetune",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = post_processing(results)
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    """
+    This scripts test a libri model with libri BPE
+    on Gigaspeech.
+    """
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / (params.decoding_method + "_gigaspeech")
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py
new file mode 100644
index 000000000..bfd019ff5
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py
@@ -0,0 +1,109 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim // 4,  # group size == 4
+                bias=False,
+            )
+        else:
+            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
+            # when inference with torch.jit.script and context_size == 1
+            self.conv = nn.Identity()
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        if torch.jit.is_tracing():
+            # This is for exporting to PNNX via ONNX
+            embedding_out = self.embedding(y)
+        else:
+            embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = F.relu(embedding_out)
+        return embedding_out
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless7/export-onnx.py
new file mode 100755
index 000000000..b75548f8b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/export-onnx.py
@@ -0,0 +1,600 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Zengrui Jin)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained-epoch-30-avg-9.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-9.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless7",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7/export.py
new file mode 100755
index 000000000..eb4c4d282
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/export.py
@@ -0,0 +1,320 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless7/decode.py \
+        --exp-dir ./pruned_transducer_stateless7/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --tokens data/lang_bpe_500/tokens.txt \
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py b/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py
new file mode 100755
index 000000000..a7a8ef149
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py
@@ -0,0 +1,1369 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Xiaoyu Yang,
+#                                                       Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless7/finetune.py \
+  --world-size 2 \
+  --num-epochs 20 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp_giga_finetune \
+  --subset S \
+  --use-fp16 1 \
+  --base-lr 0.005 \
+  --lr-epochs 100 \
+  --lr-batches 100000 \
+  --bpe-model icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/data/lang_bpe_500/bpe.model \
+  --do-finetune True \
+  --use-mux True \
+  --finetune-ckpt icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp/pretrained.pt \
+  --max-duration 500
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from gigaspeech import GigaSpeechAsrDataModule
+from joiner import Joiner
+from lhotse.cut import Cut, CutSet
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_finetune_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--do-finetune",
+        type=str2bool,
+        default=True,
+        help="Whether to fine-tune.",
+    )
+    parser.add_argument(
+        "--use-mux",
+        type=str2bool,
+        default=False,
+        help="""
+        Whether to adapt. If true, we will mix 5% of the new data
+        with 95% of the original data to fine-tune.
+        """,
+    )
+
+    parser.add_argument(
+        "--init-modules",
+        type=str,
+        default=None,
+        help="""
+        Modules to be initialized. It matches all parameters starting with
+        a specific key. The keys are given with Comma seperated. If None,
+        all modules will be initialised. For example, if you only want to
+        initialise all parameters staring with "encoder", use "encoder";
+        if you want to initialise parameters starting with encoder or decoder,
+        use "encoder,joiner".
+        """,
+    )
+
+    parser.add_argument(
+        "--finetune-ckpt",
+        type=str,
+        default=None,
+        help="Fine-tuning from which checkpoint (a path to a .pt file)",
+    )
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="""Embedding dimension in the 2 blocks of zipformer encoder
+        layers, comma separated
+        """,
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers,\
+        comma separated; not the same as embedding dimension.
+        """,
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="""Unmasked dimensions in the encoders, relates to augmentation
+        during training. Must be <= each of encoder_dims. Empirically, less
+        than 256 seems to make performance worse.
+        """,
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="""Path to the BPE model.
+        This should be the bpe model of the original model
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.005, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=100000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. During fine-tuning, we set this very large so that the
+        learning rate slowly decays with number of batches. You may tune
+        its value by yourself.
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=100,
+        help="""Number of epochs that affects how rapidly the learning rate
+        decreases. During fine-tuning, we set this very large so that the
+        learning rate slowly decays with number of batches. You may tune
+        its value by yourself.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+    add_finetune_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def load_model_params(
+    ckpt: str, model: nn.Module, init_modules: List[str] = None, strict: bool = True
+):
+    """Load model params from checkpoint
+
+    Args:
+        ckpt (str): Path to the checkpoint
+        model (nn.Module): model to be loaded
+
+    """
+    logging.info(f"Loading checkpoint from {ckpt}")
+    checkpoint = torch.load(ckpt, map_location="cpu")
+
+    # if module list is empty, load the whole model from ckpt
+    if not init_modules:
+        if next(iter(checkpoint["model"])).startswith("module."):
+            logging.info("Loading checkpoint saved by DDP")
+
+            dst_state_dict = model.state_dict()
+            src_state_dict = checkpoint["model"]
+            for key in dst_state_dict.keys():
+                src_key = "{}.{}".format("module", key)
+                dst_state_dict[key] = src_state_dict.pop(src_key)
+            assert len(src_state_dict) == 0
+            model.load_state_dict(dst_state_dict, strict=strict)
+        else:
+            model.load_state_dict(checkpoint["model"], strict=strict)
+    else:
+        src_state_dict = checkpoint["model"]
+        dst_state_dict = model.state_dict()
+        for module in init_modules:
+            logging.info(f"Loading parameters starting with prefix {module}")
+            src_keys = [
+                k for k in src_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
+            dst_keys = [
+                k for k in dst_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
+            assert set(src_keys) == set(dst_keys)  # two sets should match exactly
+            for key in src_keys:
+                dst_state_dict[key] = src_state_dict.pop(key)
+
+        model.load_state_dict(dst_state_dict, strict=strict)
+
+    return None
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    if is_training:
+        batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            # Skip the warmup by adding a huge number to batch_count
+            if params.do_finetune:
+                set_batch_count(model, params.batch_idx_train + 100000)
+            else:
+                set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it. The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have
+            # different behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    # load model parameters for model fine-tuning
+    if params.do_finetune:
+        modules = params.init_modules.split(",") if params.init_modules else None
+        checkpoints = load_model_params(
+            ckpt=params.finetune_ckpt, model=model, init_modules=modules
+        )
+        if rank == 0:
+            # model_avg is only used with rank 0
+            model_avg = copy.deepcopy(model).to(torch.float64)
+    else:
+        assert params.start_epoch > 0, params.start_epoch
+        checkpoints = load_checkpoint_if_available(
+            params=params, model=model, model_avg=model_avg
+        )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(
+        optimizer=optimizer,
+        lr_batches=params.lr_batches,
+        lr_epochs=params.lr_epochs,
+        warmup_batches=0,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    if params.use_mux:
+        librispeech = LibriSpeechAsrDataModule(args)
+        train_cuts = CutSet.mux(
+            librispeech.train_all_shuf_cuts(),
+            gigaspeech.train_cuts(),
+            weights=[0.95, 0.05],
+        )
+    else:
+        train_cuts = gigaspeech.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            # logging.warning(
+            #    f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = gigaspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = gigaspeech.dev_cuts()
+    valid_dl = gigaspeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(
+        parser
+    )  # you may replace this with your own dataset
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py b/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py
new file mode 100755
index 000000000..37edc0390
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py
@@ -0,0 +1,282 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) use the averaged model with checkpoint exp_dir/epoch-xxx.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --epoch 28 \
+    --avg 15 \
+    --use-averaged-model True \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `epoch-28-avg-15-use-averaged-model.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15-use-averaged-model.pt")`.
+
+(2) use the averaged model with checkpoint exp_dir/checkpoint-iter.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --iter 22000 \
+    --avg 5 \
+    --use-averaged-model True \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `iter-22000-avg-5-use-averaged-model.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5-use-averaged-model.pt")`.
+
+(3) use the original model with checkpoint exp_dir/epoch-xxx.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --epoch 28 \
+    --avg 15 \
+    --use-averaged-model False \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `epoch-28-avg-15.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15.pt")`.
+
+(4) use the original model with checkpoint exp_dir/checkpoint-iter.pt
+./pruned_transducer_stateless7/generate_model_from_checkpoint.py \
+    --iter 22000 \
+    --avg 5 \
+    --use-averaged-model False \
+    --exp-dir ./pruned_transducer_stateless7/exp
+
+It will generate a file `iter-22000-avg-5.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5.pt")`.
+"""
+
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+import sentencepiece as spm
+import torch
+
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import str2bool
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model."
+        "If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    print("Script started")
+
+    device = torch.device("cpu")
+    print(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    print("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            print(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+            filename = params.exp_dir / f"iter-{params.iter}-avg-{params.avg}.pt"
+            torch.save({"model": model.state_dict()}, filename)
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+            filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+            torch.save({"model": model.state_dict()}, filename)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            print(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+            filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+            torch.save({"model": model.state_dict()}, filename)
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            print(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+            filename = (
+                params.exp_dir
+                / f"iter-{params.iter}-avg-{params.avg}-use-averaged-model.pt"
+            )
+            torch.save({"model": model.state_dict()}, filename)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            print(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+            filename = (
+                params.exp_dir
+                / f"epoch-{params.epoch}-avg-{params.avg}-use-averaged-model.pt"
+            )
+            torch.save({"model": model.state_dict()}, filename)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    print("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py
new file mode 100644
index 000000000..cd432fd6f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py
@@ -0,0 +1,406 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class GigaSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+
+        # GigaSpeech specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="XL",
+            help="Select the GigaSpeech subset (XS|S|M|L|XL)",
+        )
+        group.add_argument(
+            "--small-dev",
+            type=str2bool,
+            default=False,
+            help="Should we use only 1000 utterances for dev (speeds up training)",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get train_{self.args.subset} cuts")
+        path = self.args.manifest_dir / f"cuts_{self.args.subset}.jsonl.gz"
+        cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(self.args.manifest_dir / "cuts_DEV.jsonl.gz")
+        if self.args.small_dev:
+            return cuts_valid.subset(first=1000)
+        else:
+            return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST.jsonl.gz")
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech_scoring.py b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech_scoring.py
new file mode 120000
index 000000000..fdfa6ce4b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech_scoring.py
@@ -0,0 +1 @@
+../../../gigaspeech/ASR/pruned_transducer_stateless2/gigaspeech_scoring.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7/jit_pretrained.py
new file mode 100755
index 000000000..5af6dae25
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/jit_pretrained.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py
new file mode 100644
index 000000000..62a4d22d6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py
@@ -0,0 +1,64 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = nn.Linear(encoder_dim, joiner_dim)
+        self.decoder_proj = nn.Linear(decoder_dim, joiner_dim)
+        self.output_linear = nn.Linear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim
+        assert encoder_out.ndim in (2, 4)
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7/model.py
new file mode 100644
index 000000000..add0e6a18
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/model.py
@@ -0,0 +1,198 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import penalize_abs_values_gt
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = nn.Linear(
+            encoder_dim,
+            vocab_size,
+        )
+        self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        # x.T_dim == max(x_len)
+        assert x.size(1) == x_lens.max().item(), (x.shape, x_lens, x_lens.max())
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        # if self.training and random.random() < 0.25:
+        #    lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04)
+        # if self.training and random.random() < 0.25:
+        #    am = penalize_abs_values_gt(am, 30.0, 1.0e-04)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_check.py
new file mode 120000
index 000000000..20e334271
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_check.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless5/onnx_check.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py
new file mode 100755
index 000000000..67585ee47
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py
@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-30-avg-9.pt"
+
+cd exp
+ln -s pretrained-epoch-30-avg-9.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+2. Run this file
+
+./pruned_transducer_stateless7/onnx_decode.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --max-duration 600 \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+"""
+
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+
+from onnx_pretrained import greedy_search, OnnxModel
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, sp: spm.SentencePieceProcessor, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    hyps = [sp.decode(h).split() for h in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, sp=sp, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    blank_id = sp.piece_to_id("<blk>")
+    assert blank_id == 0, blank_id
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(dl=test_dl, model=model, sp=sp)
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_pretrained.py
new file mode 120000
index 000000000..7607623c8
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
new file mode 100644
index 000000000..8ab3589da
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
@@ -0,0 +1,1098 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import contextlib
+import logging
+import random
+from collections import defaultdict
+from typing import List, Optional, Tuple, Union
+
+import torch
+from lhotse.utils import fix_random_seed
+from scaling import ActivationBalancer
+from torch import Tensor
+from torch.optim import Optimizer
+
+
+class BatchedOptimizer(Optimizer):
+    """
+    This class adds to class Optimizer the capability to optimize parameters in batches:
+    it will stack the parameters and their grads for you so the optimizer can work
+    on tensors with an extra leading dimension.  This is intended for speed with GPUs,
+    as it reduces the number of kernels launched in the optimizer.
+
+    Args:
+      params:
+    """
+
+    def __init__(self, params, defaults):
+        super(BatchedOptimizer, self).__init__(params, defaults)
+
+    @contextlib.contextmanager
+    def batched_params(self, param_group, group_params_names):
+        """
+        This function returns (technically, yields) a list of
+          of tuples (p, state), where
+        p is a `fake` parameter that is stacked (over axis 0) from real parameters
+        that share the same shape, and its gradient is also stacked;
+        `state` is the state corresponding to this batch of parameters
+        (it will be physically located in the "state" for one of the real
+        parameters, the last one that has any particular shape and dtype).
+
+        This function is decorated as a context manager so that it can
+        write parameters back to their "real" locations.
+
+        The idea is, instead of doing:
+        <code>
+          for p in group["params"]:
+             state = self.state[p]
+             ...
+        </code>
+        you can do:
+        <code>
+          with self.batched_params(group["params"]) as batches:
+             for p, state, p_names in batches:
+                 ...
+        </code>
+
+        Args:
+          group: a parameter group, which is a list of parameters; should be
+                one of self.param_groups.
+          group_params_names: name for each parameter in group,
+                which is List[str].
+        """
+        batches = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
+        batches_names = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of str
+
+        assert len(param_group) == len(group_params_names)
+        for p, named_p in zip(param_group, group_params_names):
+            key = (str(p.dtype), *p.shape)
+            batches[key].append(p)
+            batches_names[key].append(named_p)
+
+        batches_names_keys = list(batches_names.keys())
+        sorted_idx = sorted(
+            range(len(batches_names)), key=lambda i: batches_names_keys[i]
+        )
+        batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx]
+        batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
+
+        stacked_params_dict = dict()
+
+        # turn batches into a list, in deterministic order.
+        # tuples will contain tuples of (stacked_param, state, stacked_params_names),
+        # one for each batch in `batches`.
+        tuples = []
+
+        for batch, batch_names in zip(batches, batches_names):
+            p = batch[0]
+            # we arbitrarily store the state in the
+            # state corresponding to the 1st parameter in the
+            # group.  class Optimizer will take care of saving/loading state.
+            state = self.state[p]
+            p_stacked = torch.stack(batch)
+            grad = torch.stack(
+                [torch.zeros_like(p) if p.grad is None else p.grad for p in batch]
+            )
+            p_stacked.grad = grad
+            stacked_params_dict[key] = p_stacked
+            tuples.append((p_stacked, state, batch_names))
+
+        yield tuples  # <-- calling code will do the actual optimization here!
+
+        for (stacked_params, _state, _names), batch in zip(tuples, batches):
+            for i, p in enumerate(batch):  # batch is list of Parameter
+                p.copy_(stacked_params[i])
+
+
+class ScaledAdam(BatchedOptimizer):
+    """
+     Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
+     proportional to the norm of that parameter; and also learn the scale of the parameter,
+     in log space, subject to upper and lower limits (as if we had factored each parameter as
+     param = underlying_param * log_scale.exp())
+
+
+     Args:
+          params:  The parameters or param_groups to optimize (like other Optimizer subclasses)
+              lr:  The learning rate.  We will typically use a learning rate schedule that starts
+                   at 0.03 and decreases over time, i.e. much higher than other common
+                   optimizers.
+     clipping_scale: (e.g. 2.0)
+                   A scale for gradient-clipping: if specified, the normalized gradients
+                   over the whole model will be clipped to have 2-norm equal to
+                   `clipping_scale` times the median 2-norm over the most recent period
+                   of `clipping_update_period` minibatches.  By "normalized gradients",
+                   we mean after multiplying by the rms parameter value for this tensor
+                   [for non-scalars]; this is appropriate because our update is scaled
+                   by this quantity.
+            betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
+                   Must satisfy 0 < beta <= beta2 < 1.
+     scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
+                   scale of each parameter tensor and scalar parameters of the mode..
+                   If each parameter were decomposed
+                   as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
+                   would be a the scaling factor on the learning rate of p_scale.
+              eps:  A general-purpose epsilon to prevent division by zero
+    param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be >= this value)
+    param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be <= this value)
+       scalar_max: Maximum absolute value for scalar parameters (applicable if your
+                   model has any parameters with numel() == 1).
+    size_update_period: The periodicity, in steps, with which we update the size (scale)
+                   of the parameter tensor.  This is provided to save a little time
+                   in the update.
+     clipping_update_period: if clipping_scale is specified, this is the period
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=3e-02,
+        clipping_scale=None,
+        betas=(0.9, 0.98),
+        scalar_lr_scale=0.1,
+        eps=1.0e-08,
+        param_min_rms=1.0e-05,
+        param_max_rms=3.0,
+        scalar_max=10.0,
+        size_update_period=4,
+        clipping_update_period=100,
+        parameters_names=None,
+        show_dominant_parameters=True,
+    ):
+        assert parameters_names is not None, (
+            "Please prepare parameters_names,"
+            "which is a List[List[str]]. Each List[str] is for a group"
+            "and each str is for a parameter"
+        )
+        defaults = dict(
+            lr=lr,
+            clipping_scale=clipping_scale,
+            betas=betas,
+            scalar_lr_scale=scalar_lr_scale,
+            eps=eps,
+            param_min_rms=param_min_rms,
+            param_max_rms=param_max_rms,
+            scalar_max=scalar_max,
+            size_update_period=size_update_period,
+            clipping_update_period=clipping_update_period,
+        )
+
+        super(ScaledAdam, self).__init__(params, defaults)
+        assert len(self.param_groups) == len(parameters_names)
+        self.parameters_names = parameters_names
+        self.show_dominant_parameters = show_dominant_parameters
+
+    def __setstate__(self, state):
+        super(ScaledAdam, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        batch = True
+
+        for group, group_params_names in zip(self.param_groups, self.parameters_names):
+            with self.batched_params(group["params"], group_params_names) as batches:
+                # batches is list of pairs (stacked_param, state).  stacked_param is like
+                # a regular parameter, and will have a .grad, but the 1st dim corresponds to
+                # a stacking dim, it is not a real dim.
+
+                if (
+                    len(batches[0][1]) == 0
+                ):  # if len(first state) == 0: not yet initialized
+                    clipping_scale = 1
+                else:
+                    clipping_scale = self._get_clipping_scale(group, batches)
+
+                for p, state, _ in batches:
+                    # Perform optimization step.
+                    # grad is not going to be None, we handled that when creating the batches.
+                    grad = p.grad
+                    if grad.is_sparse:
+                        raise RuntimeError(
+                            "ScaledAdam optimizer does not support sparse gradients"
+                        )
+                    # State initialization
+                    if len(state) == 0:
+                        self._init_state(group, p, state)
+
+                    self._step_one_batch(group, p, state, clipping_scale)
+
+        return loss
+
+    def _init_state(self, group: dict, p: Tensor, state: dict):
+        """
+        Initializes state dict for parameter 'p'.  Assumes that dim 0 of tensor p
+        is actually the batch dimension, corresponding to batched-together
+        parameters of a given shape.
+
+
+        Args:
+           group:   Dict to look up configuration values.
+               p: The parameter that we are initializing the state for
+           state: Dict from string to whatever state we are initializing
+        """
+        size_update_period = group["size_update_period"]
+
+        state["step"] = 0
+
+        kwargs = {"device": p.device, "dtype": p.dtype}
+
+        # 'delta' implements conventional momentum.  There are
+        # several different kinds of update going on, so rather than
+        # compute "exp_avg" like in Adam, we store and decay a
+        # parameter-change "delta", which combines all forms of
+        # update.  this is equivalent to how it's done in Adam,
+        # except for the first few steps.
+        state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+
+        if numel > 1:
+            # "param_rms" just periodically records the scalar root-mean-square value of
+            # the parameter tensor.
+            # it has a shape like (batch_size, 1, 1, 1, 1)
+            param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+            state["param_rms"] = param_rms
+
+            state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
+            state["scale_grads"] = torch.zeros(
+                size_update_period, *param_rms.shape, **kwargs
+            )
+
+        # exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
+        state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+    def _get_clipping_scale(
+        self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]]
+    ) -> float:
+        """
+        Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
+        by this amount before applying the rest of the update.
+
+        Args:
+           group: the parameter group, an item in self.param_groups
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+        """
+        assert len(tuples) >= 1
+        clipping_scale = group["clipping_scale"]
+        (first_p, first_state, _) = tuples[0]
+        step = first_state["step"]
+        if clipping_scale is None or step == 0:
+            # no clipping.  return early on step == 0 because the other
+            # parameters' state won't have been initialized yet.
+            return 1.0
+        clipping_update_period = group["clipping_update_period"]
+
+        tot_sumsq = torch.tensor(0.0, device=first_p.device)
+        for p, state, param_names in tuples:
+            grad = p.grad
+            if grad.is_sparse:
+                raise RuntimeError(
+                    "ScaledAdam optimizer does not support sparse gradients"
+                )
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                tot_sumsq += (grad**2).sum()  # sum() to change shape [1] to []
+            else:
+                tot_sumsq += ((grad * state["param_rms"]) ** 2).sum()
+
+        tot_norm = tot_sumsq.sqrt()
+        if "model_norms" not in first_state:
+            first_state["model_norms"] = torch.zeros(
+                clipping_update_period, device=p.device
+            )
+        first_state["model_norms"][step % clipping_update_period] = tot_norm
+
+        if step % clipping_update_period == 0:
+            # Print some stats.
+            # We don't reach here if step == 0 because we would have returned
+            # above.
+            sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
+            quartiles = []
+            for n in range(0, 5):
+                index = min(
+                    clipping_update_period - 1, (clipping_update_period // 4) * n
+                )
+                quartiles.append(sorted_norms[index].item())
+
+            median = quartiles[2]
+            threshold = clipping_scale * median
+            first_state["model_norm_threshold"] = threshold
+            percent_clipped = (
+                first_state["num_clipped"] * 100.0 / clipping_update_period
+                if "num_clipped" in first_state
+                else 0.0
+            )
+            first_state["num_clipped"] = 0
+            quartiles = " ".join(["%.3e" % x for x in quartiles])
+            logging.info(
+                f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
+                f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
+            )
+
+        if step < clipping_update_period:
+            return 1.0  # We have not yet estimated a norm to clip to.
+        else:
+            try:
+                model_norm_threshold = first_state["model_norm_threshold"]
+            except KeyError:
+                logging.info(
+                    "Warning: model_norm_threshold not in state: possibly "
+                    "you changed config when restarting, adding clipping_scale option?"
+                )
+                return 1.0
+            ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
+            if ans < 1.0:
+                first_state["num_clipped"] += 1
+            if ans < 0.1:
+                logging.warn(
+                    f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
+                )
+                if self.show_dominant_parameters:
+                    assert p.shape[0] == len(param_names)
+                    self._show_gradient_dominating_parameter(tuples, tot_sumsq)
+            return ans
+
+    def _show_gradient_dominating_parameter(
+        self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor
+    ):
+        """
+        Show information of parameter wihch dominanting tot_sumsq.
+
+        Args:
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+            tot_sumsq: sumsq of all parameters. Though it's could be calculated
+                from tuples, we still pass it to save some time.
+        """
+        all_sumsq_orig = {}
+        for p, state, batch_param_names in tuples:
+            # p is a stacked batch parameters.
+            batch_grad = p.grad
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                batch_sumsq_orig = batch_grad**2
+                # Dummpy values used by following `zip` statement.
+                batch_rms_orig = torch.ones(p.shape[0])
+            else:
+                batch_rms_orig = state["param_rms"]
+                batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum(
+                    dim=list(range(1, batch_grad.ndim))
+                )
+
+            for name, sumsq_orig, rms, grad in zip(
+                batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
+            ):
+                proportion_orig = sumsq_orig / tot_sumsq
+                all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
+
+        assert torch.isclose(
+            sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
+            torch.tensor(1.0),
+        )
+        sorted_by_proportion = {
+            k: v
+            for k, v in sorted(
+                all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True
+            )
+        }
+        dominant_param_name = next(iter(sorted_by_proportion))
+        (
+            dominant_proportion,
+            dominant_sumsq,
+            dominant_rms,
+            dominant_grad,
+        ) = sorted_by_proportion[dominant_param_name]
+        logging.info(
+            f"Parameter Dominanting tot_sumsq {dominant_param_name}"
+            f" with proportion {dominant_proportion:.2f},"
+            f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
+            f"={dominant_sumsq:.3e},"
+            f" grad_sumsq = {(dominant_grad**2).sum():.3e},"
+            f" orig_rms_sq={(dominant_rms**2).item():.3e}"
+        )
+
+    def _step_one_batch(
+        self, group: dict, p: Tensor, state: dict, clipping_scale: float
+    ):
+        """
+        Do the step for one parameter, which is actually going to be a batch of
+        `real` parameters, with dim 0 as the batch dim.
+        Args:
+                  group:  dict to look up configuration values
+                    p: parameter to update (actually multiple parameters stacked together
+                       as a batch)
+                  state: state-dict for p, to look up the optimizer state
+        """
+        lr = group["lr"]
+        size_update_period = group["size_update_period"]
+        beta1 = group["betas"][0]
+
+        grad = p.grad
+        if clipping_scale != 1.0:
+            grad = grad * clipping_scale
+        step = state["step"]
+        delta = state["delta"]
+
+        delta.mul_(beta1)
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        if numel > 1:
+            # Update the size/scale of p, and set param_rms
+            scale_grads = state["scale_grads"]
+            scale_grads[step % size_update_period] = (p * grad).sum(
+                dim=list(range(1, p.ndim)), keepdim=True
+            )
+            if step % size_update_period == size_update_period - 1:
+                param_rms = state["param_rms"]  # shape: (batch_size, 1, 1, ..)
+                param_rms.copy_(
+                    (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+                )
+                if step > 0:
+                    # self._size_update() learns the overall scale on the
+                    # parameter, by shrinking or expanding it.
+                    self._size_update(group, scale_grads, p, state)
+
+        if numel == 1:
+            # For parameters with 1 element we just use regular Adam.
+            # Updates delta.
+            self._step_scalar(group, p, state)
+        else:
+            self._step(group, p, state)
+
+        state["step"] = step + 1
+
+    def _size_update(
+        self, group: dict, scale_grads: Tensor, p: Tensor, state: dict
+    ) -> None:
+        """
+               Called only where p.numel() > 1, this updates the scale of the parameter.
+               If we imagine: p =  underlying_param * scale.exp(), and we are doing
+               gradient descent on underlying param and on scale, this function does the update
+               on `scale`.
+
+               Args:
+              group: dict to look up configuration values
+        scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
+                      grads w.r.t. the scales.
+                  p:  The parameter to update
+               state: The state-dict of p
+        """
+
+        param_rms = state["param_rms"]
+        beta1, beta2 = group["betas"]
+        size_lr = group["lr"] * group["scalar_lr_scale"]
+        param_min_rms = group["param_min_rms"]
+        param_max_rms = group["param_max_rms"]
+        eps = group["eps"]
+        step = state["step"]
+        batch_size = p.shape[0]
+
+        size_update_period = scale_grads.shape[0]
+        # correct beta2 for the size update period: we will have
+        # faster decay at this level.
+        beta2_corr = beta2**size_update_period
+
+        scale_exp_avg_sq = state["scale_exp_avg_sq"]  # shape: (batch_size, 1, 1, ..)
+        scale_exp_avg_sq.mul_(beta2_corr).add_(
+            (scale_grads**2).mean(dim=0),  # mean over dim `size_update_period`
+            alpha=1 - beta2_corr,
+        )  # shape is (batch_size, 1, 1, ...)
+
+        # The 1st time we reach here is when size_step == 1.
+        size_step = (step + 1) // size_update_period
+        bias_correction2 = 1 - beta2_corr**size_step
+        # we don't bother with bias_correction1; this will help prevent divergence
+        # at the start of training.
+
+        denom = scale_exp_avg_sq.sqrt() + eps
+
+        scale_step = (
+            -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom
+        )
+
+        is_too_small = param_rms < param_min_rms
+        is_too_large = param_rms > param_max_rms
+
+        # when the param gets too small, just don't shrink it any further.
+        scale_step.masked_fill_(is_too_small, 0.0)
+        # when it gets too large, stop it from getting any larger.
+        scale_step.masked_fill_(is_too_large, -size_lr * size_update_period)
+        delta = state["delta"]
+        # the factor of (1-beta1) relates to momentum.
+        delta.add_(p * scale_step, alpha=(1 - beta1))
+
+    def _step(self, group: dict, p: Tensor, state: dict):
+        """
+        This function does the core update of self.step(), in the case where the members of
+        the batch have more than 1 element.
+
+        Args:
+            group: A dict which will be used to look up configuration values
+                p: The parameter to be updated
+             grad: The grad of p
+            state: The state-dict corresponding to parameter p
+
+        This function modifies p.
+        """
+        grad = p.grad
+        lr = group["lr"]
+        beta1, beta2 = group["betas"]
+        eps = group["eps"]
+        param_min_rms = group["param_min_rms"]
+        step = state["step"]
+
+        exp_avg_sq = state["exp_avg_sq"]
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
+
+        this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0)
+        bias_correction2 = 1 - beta2 ** (this_step + 1)
+        if bias_correction2 < 0.99:
+            # note: not in-place.
+            exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
+
+        denom = exp_avg_sq.sqrt()
+        denom += eps
+        grad = grad / denom
+
+        alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
+
+        delta = state["delta"]
+        delta.add_(grad * alpha)
+        p.add_(delta)
+
+    def _step_scalar(self, group: dict, p: Tensor, state: dict):
+        """
+        A simplified form of the core update for scalar tensors, where we cannot get a good
+        estimate of the parameter rms.
+        """
+        beta1, beta2 = group["betas"]
+        scalar_max = group["scalar_max"]
+        eps = group["eps"]
+        lr = group["lr"] * group["scalar_lr_scale"]
+        grad = p.grad
+
+        exp_avg_sq = state["exp_avg_sq"]  # shape: (batch_size,)
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+        # bias_correction2 is like in Adam.  Don't bother with bias_correction1;
+        # slower update at the start will help stability anyway.
+        bias_correction2 = 1 - beta2 ** (state["step"] + 1)
+        denom = (exp_avg_sq / bias_correction2).sqrt() + eps
+
+        delta = state["delta"]
+        delta.add_(grad / denom, alpha=-lr * (1 - beta1))
+        p.clamp_(min=-scalar_max, max=scalar_max)
+        p.add_(delta)
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("base_lr", group["lr"])
+
+        self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            logging.info(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+    The basic formula (before warmup) is:
+      lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                     (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup
+    where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
+    and then stays constant at 1.
+
+
+     E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        warmup_batches: Union[int, float] = 500.0,
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+        self.warmup_batches = warmup_batches
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        warmup_factor = (
+            1.0
+            if self.batch >= self.warmup_batches
+            else 0.5 + 0.5 * (self.batch / self.warmup_batches)
+        )
+
+        return [x * factor * warmup_factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = ScaledAdam(m.parameters(), lr=0.03)
+
+    scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+
+    logging.info(f"last lr = {scheduler.get_last_lr()}")
+    logging.info(f"state dict = {scheduler.state_dict()}")
+
+
+def _plot_eden_lr():
+    import matplotlib.pyplot as plt
+
+    m = torch.nn.Linear(100, 100)
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in m.named_parameters()]
+    )
+
+    for lr_epoch in [4, 10, 100]:
+        for lr_batch in [100, 400]:
+            optim = ScaledAdam(
+                m.parameters(), lr=0.03, parameters_names=parameters_names
+            )
+            scheduler = Eden(
+                optim, lr_batches=lr_batch, lr_epochs=lr_epoch, verbose=True
+            )
+            lr = []
+
+            for epoch in range(10):
+                scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+                for step in range(500):
+                    lr.append(scheduler.get_lr())
+
+                    x = torch.randn(200, 100).detach()
+                    x.requires_grad = True
+                    y = m(x)
+                    dy = torch.randn(200, 100).detach()
+                    f = (y * dy).sum()
+                    f.backward()
+
+                    optim.step()
+                    scheduler.step_batch()
+                    optim.zero_grad()
+            plt.plot(lr, label=f"lr_epoch:{lr_epoch}, lr_batch:{lr_batch}")
+
+    plt.legend()
+    plt.savefig("lr.png")
+
+
+# This is included mostly as a baseline for ScaledAdam.
+class Eve(Optimizer):
+    """
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                if random.random() < 0.0005:
+                    step = (exp_avg / denom) * step_size
+                    logging.info(
+                        f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}"
+                    )
+
+        return loss
+
+
+def _test_scaled_adam(hidden_dim: int):
+    import timeit
+
+    from scaling import ScaledLinear
+
+    E = 100
+    B = 4
+    T = 2
+    logging.info("in test_eve_cain")
+    # device = torch.device('cuda')
+    device = torch.device("cpu")
+    dtype = torch.float32
+
+    fix_random_seed(42)
+    # these input_magnitudes and output_magnitudes are to test that
+    # Abel is working as we expect and is able to adjust scales of
+    # different dims differently.
+    input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+    output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+
+    for iter in [1, 0]:
+        fix_random_seed(42)
+        Linear = torch.nn.Linear if iter == 0 else ScaledLinear
+
+        m = torch.nn.Sequential(
+            Linear(E, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, E),
+        ).to(device)
+
+        train_pairs = [
+            (
+                100.0
+                * torch.randn(B, T, E, device=device, dtype=dtype)
+                * input_magnitudes,
+                torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes,
+            )
+            for _ in range(20)
+        ]
+
+        if iter == 0:
+            optim = Eve(m.parameters(), lr=0.003)
+        elif iter == 1:
+            optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0)
+        scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False)
+
+        start = timeit.default_timer()
+        avg_loss = 0.0
+        for epoch in range(180):
+            scheduler.step_epoch()
+            # if epoch == 100 and iter in [2,3]:
+            #    optim.reset_speedup()  # check it doesn't crash.
+
+            # if epoch == 130:
+            #    opts = diagnostics.TensorDiagnosticOptions(
+            #        512
+            #    )  # allow 4 megabytes per sub-module
+            #    diagnostic = diagnostics.attach_diagnostics(m, opts)
+
+            for n, (x, y) in enumerate(train_pairs):
+                y_out = m(x)
+                loss = ((y_out - y) ** 2).mean() * 100.0
+                if epoch == 0 and n == 0:
+                    avg_loss = loss.item()
+                else:
+                    avg_loss = 0.98 * avg_loss + 0.02 * loss.item()
+                if n == 0 and epoch % 5 == 0:
+                    # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item()
+                    # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item()
+                    # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item()
+                    # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item()
+                    # scale1 = '%.2e' % (m[0].weight_scale.exp().item())
+                    # scale1b = '%.2e' % (m[0].bias_scale.exp().item())
+                    # scale2 = '%.2e' % (m[2].weight_scale.exp().item())
+                    # scale2b = '%.2e' % (m[2].bias_scale.exp().item())
+                    lr = scheduler.get_last_lr()[0]
+                    logging.info(
+                        f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}"
+                    )  # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
+                loss.log().backward()
+                optim.step()
+                optim.zero_grad()
+                scheduler.step_batch()
+
+        # diagnostic.print_diagnostics()
+
+        stop = timeit.default_timer()
+        logging.info(f"Iter={iter}, Time taken: {stop - start}")
+
+        logging.info(f"last lr = {scheduler.get_last_lr()}")
+        # logging.info("state dict = ", scheduler.state_dict())
+        # logging.info("optim state_dict = ", optim.state_dict())
+        logging.info(f"input_magnitudes = {input_magnitudes}")
+        logging.info(f"output_magnitudes = {output_magnitudes}")
+
+
+if __name__ == "__main__":
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    logging.getLogger().setLevel(logging.INFO)
+    import subprocess
+
+    s = subprocess.check_output(
+        "git status -uno .; git log -1; git diff HEAD .", shell=True
+    )
+    logging.info(s)
+    import sys
+
+    if len(sys.argv) > 1:
+        hidden_dim = int(sys.argv[1])
+    else:
+        hidden_dim = 200
+
+    # _test_scaled_adam(hidden_dim)
+    # _test_eden()
+    _plot_eden_lr()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7/pretrained.py
new file mode 100755
index 000000000..86c922cda
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/pretrained.py
@@ -0,0 +1,360 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --tokens ./data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/profile.py b/egs/librispeech/ASR/pruned_transducer_stateless7/profile.py
new file mode 100755
index 000000000..0d308e966
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/profile.py
@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage: ./pruned_transducer_stateless7/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from typing import Tuple
+from torch import Tensor, nn
+
+from icefall.profiler import get_model_profile
+from scaling import BasicNorm, DoubleSwish
+from train import get_encoder_model, get_joiner_model, add_model_arguments, get_params
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _basic_norm_flops_compute(module, input, output):
+    assert len(input) == 1, len(input)
+    # estimate as layer_norm, see icefall/profiler.py
+    flops = input[0].numel() * 5
+    module.__flops__ += int(flops)
+
+
+def _doubleswish_module_flops_compute(module, input, output):
+    # For DoubleSwish
+    assert len(input) == 1, len(input)
+    # estimate as swish/silu, see icefall/profiler.py
+    flops = input[0].numel()
+    module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    BasicNorm: _basic_norm_flops_compute,
+    DoubleSwish: _doubleswish_module_flops_compute,
+}
+
+
+class Model(nn.Module):
+    """A Wrapper for encoder and encoder_proj"""
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        encoder_proj: nn.Module,
+    ) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(self, feature: Tensor, feature_lens: Tensor) -> Tuple[Tensor, Tensor]:
+        encoder_out, encoder_out_lens = self.encoder(feature, feature_lens)
+
+        logits = self.encoder_proj(encoder_out)
+
+        return logits, encoder_out_lens
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = Model(
+        encoder=get_encoder_model(params),
+        encoder_proj=get_joiner_model(params).encoder_proj,
+    )
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(
+        model=model,
+        args=(feature, feature_lens),
+        module_hoop_mapping=MODULE_HOOK_MAPPING,
+    )
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py
new file mode 100644
index 000000000..30a737061
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py
@@ -0,0 +1,1180 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+import logging
+import random
+from functools import reduce
+from itertools import repeat
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from torch.nn import Embedding as ScaledEmbedding
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        scale_factor: Tensor,
+        sign_factor: Optional[Tensor],
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        if sign_factor is None:
+            ctx.save_for_backward(xgt0, scale_factor)
+        else:
+            ctx.save_for_backward(xgt0, scale_factor, sign_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        if len(ctx.saved_tensors) == 3:
+            xgt0, scale_factor, sign_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+                sign_factor = sign_factor.unsqueeze(-1)
+            factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        else:
+            xgt0, scale_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+            factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+def _compute_scale_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_abs: float,
+    max_abs: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32)
+
+    if min_abs == 0.0:
+        below_threshold = 0.0
+    else:
+        # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if
+        # x_abs)_mean , min_abs.
+        below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp(
+            min=0, max=max_factor
+        )
+
+    above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp(
+        min=0, max=max_factor
+    )
+
+    return below_threshold - above_threshold
+
+
+def _compute_sign_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_positive: float,
+    max_positive: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims)
+    if min_positive == 0.0:
+        factor1 = 0.0
+    else:
+        # 0 if proportion_positive >= min_positive, else can be
+        # as large as max_factor.
+        factor1 = (
+            (min_positive - proportion_positive) * (gain_factor / min_positive)
+        ).clamp_(min=0, max=max_factor)
+
+    if max_positive == 1.0:
+        factor2 = 0.0
+    else:
+        # 0 if self.proportion_positive <= max_positive, else can be
+        # as large as -max_factor.
+        factor2 = (
+            (proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive))
+        ).clamp_(min=0, max=max_factor)
+    sign_factor = factor1 - factor2
+    # require min_positive != 0 or max_positive != 1:
+    assert not isinstance(sign_factor, float)
+    return sign_factor
+
+
+class ActivationScaleBalancerFunction(torch.autograd.Function):
+    """
+    This object is used in class ActivationBalancer when the user specified
+    min_positive=0, max_positive=1, so there are no constraints on the signs
+    of the activations and only the absolute value has a constraint.
+    """
+
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        sign_factor: Tensor,
+        scale_factor: Tensor,
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        ctx.save_for_backward(xgt0, sign_factor, scale_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        xgt0, sign_factor, scale_factor = ctx.saved_tensors
+        for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+            sign_factor = sign_factor.unsqueeze(-1)
+            scale_factor = scale_factor.unsqueeze(-1)
+
+        factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+class RandomClampFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        min: Optional[float],
+        max: Optional[float],
+        prob: float,
+        reflect: float,
+    ) -> Tensor:
+        x_clamped = torch.clamp(x, min=min, max=max)
+        mask = torch.rand_like(x) < prob
+        ans = torch.where(mask, x_clamped, x)
+        if x.requires_grad:
+            ctx.save_for_backward(ans == x)
+            ctx.reflect = reflect
+        if reflect != 0.0:
+            ans = ans * (1.0 + reflect) - (x * reflect)
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, None, None, None]:
+        (is_same,) = ctx.saved_tensors
+        x_grad = ans_grad * is_same.to(ans_grad.dtype)
+        reflect = ctx.reflect
+        if reflect != 0.0:
+            x_grad = x_grad * (1.0 + reflect) - (ans_grad * reflect)
+        return x_grad, None, None, None, None
+
+
+def random_clamp(
+    x: Tensor,
+    min: Optional[float] = None,
+    max: Optional[float] = None,
+    prob: float = 0.5,
+    reflect: float = 0.0,
+):
+    return RandomClampFunction.apply(x, min, max, prob, reflect)
+
+
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
+    """
+    A randomized way of casting a floating point value to half precision.
+    """
+    if x.dtype == torch.float16:
+        return x
+    x_abs = x.abs()
+    is_too_small = x_abs < min_abs
+    # for elements where is_too_small is true, random_val will contain +-min_abs with
+    # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
+    # for those elements].
+    random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs)
+    return torch.where(is_too_small, random_val, x).to(torch.float16)
+
+
+class RandomGradFunction(torch.autograd.Function):
+    """
+    Does nothing in forward pass; in backward pass, gets rid of very small grads using
+    randomized approach that preserves expectations (intended to reduce roundoff).
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, min_abs: float) -> Tensor:
+        ctx.min_abs = min_abs
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None]:
+        if ans_grad.dtype == torch.float16:
+            return (
+                random_cast_to_half(ans_grad.to(torch.float32), min_abs=ctx.min_abs),
+                None,
+            )
+        else:
+            return ans_grad, None
+
+
+class RandomGrad(torch.nn.Module):
+    """
+    Gets rid of very small gradients using an expectation-preserving method, intended to increase
+    accuracy of training when using amp (automatic mixed precision)
+    """
+
+    def __init__(self, min_abs: float = 5.0e-06):
+        super(RandomGrad, self).__init__()
+        self.min_abs = min_abs
+
+    def forward(self, x: Tensor):
+        if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing():
+            return x
+        else:
+            return RandomGradFunction.apply(x, self.min_abs)
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    """
+    Tries to handle half-precision derivatives in a randomized way that should
+    be more accurate for training than the default behavior.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, dim: int):
+        ans = x.softmax(dim=dim)
+        # if x dtype is float16, x.softmax() returns a float32 because
+        # (presumably) that op does not support float16, and autocast
+        # is enabled.
+        if torch.is_autocast_enabled():
+            ans = ans.to(torch.float16)
+        ctx.save_for_backward(ans)
+        ctx.x_dtype = x.dtype
+        ctx.dim = dim
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        (ans,) = ctx.saved_tensors
+        with torch.cuda.amp.autocast(enabled=False):
+            ans_grad = ans_grad.to(torch.float32)
+            ans = ans.to(torch.float32)
+            x_grad = ans_grad * ans
+            x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True)
+            return x_grad, None
+
+
+def softmax(x: Tensor, dim: int):
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x.softmax(dim)
+
+    return SoftmaxFunction.apply(x, dim)
+
+
+class MaxEigLimiterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.channel_dim = channel_dim
+        ctx.grad_scale = grad_scale
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        with torch.enable_grad():
+            (x_orig, coeffs, new_direction) = ctx.saved_tensors
+            x_orig.requires_grad = True
+            num_channels = x_orig.shape[ctx.channel_dim]
+            x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
+            new_direction.requires_grad = False
+            x = x - x.mean(dim=0)
+            x_var = (x**2).mean()
+            x_residual = x - coeffs * new_direction
+            x_residual_var = (x_residual**2).mean()
+            # `variance_proportion` is the proportion of the variance accounted for
+            # by the top eigen-direction.  This is to be minimized.
+            variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+            variance_proportion.backward()
+        x_orig_grad = x_orig.grad
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
+        return x_grad + x_extra_grad.detach(), None, None, None, None
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    eps_min: float
+    eps_max: float
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+        eps_min: float = -3.0,
+        eps_max: float = 3.0,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+        self.eps_min = eps_min
+        self.eps_max = eps_max
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        eps = self.eps
+        if self.training and random.random() < 0.25:
+            # with probability 0.25, in training mode, clamp eps between the min
+            # and max; this will encourage it to learn parameters within the
+            # allowed range by making parameters that are outside the allowed
+            # range noisy.
+
+            # gradients to allow the parameter to get back into the allowed
+            # region if it happens to exit it.
+            eps = eps.clamp(min=self.eps_min, max=self.eps_max)
+        scales = (
+            torch.mean(x**2, dim=self.channel_dim, keepdim=True) + eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
+    """
+    Behaves like a constructor of a modified version of nn.Linear
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Linear(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv1d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv1d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           sign_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_positive and max_positive
+              are violated.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+          min_prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.  Early in training we may use
+             higher probabilities than this; it will decay to this value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.04,
+        sign_gain_factor: float = 0.01,
+        scale_gain_factor: float = 0.02,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+        min_prob: float = 0.1,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.min_prob = min_prob
+        self.sign_gain_factor = sign_gain_factor
+        self.scale_gain_factor = scale_gain_factor
+
+        # count measures how many times the forward() function has been called.
+        # We occasionally sync this to a tensor called `count`, that exists to
+        # make sure it is synced to disk when we load and save the model.
+        self.cpu_count = 0
+        self.register_buffer("count", torch.tensor(0, dtype=torch.int64))
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or not x.requires_grad or torch.jit.is_tracing():
+            return _no_op(x)
+
+        count = self.cpu_count
+        self.cpu_count += 1
+
+        if random.random() < 0.01:
+            # Occasionally sync self.cpu_count with self.count.
+            # count affects the decay of 'prob'.  don't do this on every iter,
+            # because syncing with the GPU is slow.
+            self.cpu_count = max(self.cpu_count, self.count.item())
+            self.count.fill_(self.cpu_count)
+
+        # the prob of doing some work exponentially decreases from 0.5 till it hits
+        # a floor at min_prob (==0.1, by default)
+        prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0)))
+
+        if random.random() < prob:
+            sign_gain_factor = 0.5
+            if self.min_positive != 0.0 or self.max_positive != 1.0:
+                sign_factor = _compute_sign_factor(
+                    x,
+                    self.channel_dim,
+                    self.min_positive,
+                    self.max_positive,
+                    gain_factor=self.sign_gain_factor / prob,
+                    max_factor=self.max_factor,
+                )
+            else:
+                sign_factor = None
+
+            scale_factor = _compute_scale_factor(
+                x.detach(),
+                self.channel_dim,
+                min_abs=self.min_abs,
+                max_abs=self.max_abs,
+                gain_factor=self.scale_gain_factor / prob,
+                max_factor=self.max_factor,
+            )
+            return ActivationBalancerFunction.apply(
+                x,
+                scale_factor,
+                sign_factor,
+                self.channel_dim,
+            )
+        else:
+            return _no_op(x)
+
+
+def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float) -> Tensor:
+    """
+    Returns x unmodified, but in backprop will put a penalty for the excess of
+    the absolute values of elements of x over the limit "limit".  E.g. if
+    limit == 10.0, then if x has any values over 10 it will get a penalty.
+
+    Caution: the value of this penalty will be affected by grad scaling used
+    in automatic mixed precision training.  For this reasons we use this,
+    it shouldn't really matter, or may even be helpful; we just use this
+    to disallow really implausible values of scores to be given to softmax.
+    """
+    x_sign = x.sign()
+    over_limit = (x.abs() - limit) > 0
+    # The following is a memory efficient way to penalize the absolute values of
+    # x that's over the limit.  (The memory efficiency comes when you think
+    # about which items torch needs to cache for the autograd, and which ones it
+    # can throw away).  The numerical value of aux_loss as computed here will
+    # actually be larger than it should be, by limit * over_limit.sum(), but it
+    # has the same derivative as the real aux_loss which is penalty * (x.abs() -
+    # limit).relu().
+    aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x)
+    # note: we don't do sum() here on aux)_loss, but it's as if we had done
+    # sum() due to how with_loss() works.
+    x = with_loss(x, aux_loss)
+    # you must use x for something, or this will be ineffective.
+    return x
+
+
+def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
+    if x.ndim == 2:
+        return x.diag()
+    else:
+        (batch, dim, dim) = x.shape
+        x = x.reshape(batch, dim * dim)
+        x = x[:, :: dim + 1]
+        assert x.shape == (batch, dim)
+        return x
+
+
+def _whitening_metric(x: Tensor, num_groups: int):
+    """
+    Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
+    of the centered feature covariance are the same within each group's covariance matrix
+    and also between groups.
+    Args:
+        x: a Tensor of shape (*, num_channels)
+     num_groups:  the number of groups of channels, a number >=1 that divides num_channels
+    Returns:
+        Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and
+    greater than 1.0 otherwise.
+    """
+    assert x.dtype != torch.float16
+    x = x.reshape(-1, x.shape[-1])
+    (num_frames, num_channels) = x.shape
+    assert num_channels % num_groups == 0
+    channels_per_group = num_channels // num_groups
+    x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1)
+    # x now has shape (num_groups, num_frames, channels_per_group)
+    # subtract the mean so we use the centered, not uncentered, covariance.
+    # My experience has been that when we "mess with the gradients" like this,
+    # it's better not do anything that tries to move the mean around, because
+    # that can easily cause instability.
+    x = x - x.mean(dim=1, keepdim=True)
+    # x_covar: (num_groups, channels_per_group, channels_per_group)
+    x_covar = torch.matmul(x.transpose(1, 2), x)
+    x_covar_mean_diag = _diag(x_covar).mean()
+    # the following expression is what we'd get if we took the matrix product
+    # of each covariance and measured the mean of its trace, i.e.
+    # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
+    x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group)
+    # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20)
+    return metric
+
+
+class WhiteningPenaltyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx, x: Tensor, num_groups: int, whitening_limit: float, grad_scale: float
+    ) -> Tensor:
+        ctx.save_for_backward(x)
+        ctx.num_groups = num_groups
+        ctx.whitening_limit = whitening_limit
+        ctx.grad_scale = grad_scale
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
+        with torch.enable_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                x_detached = x_orig.to(torch.float32).detach()
+                x_detached.requires_grad = True
+
+                metric = _whitening_metric(x_detached, ctx.num_groups)
+
+                if random.random() < 0.005 or __name__ == "__main__":
+                    logging.info(
+                        f"Whitening: num_groups={ctx.num_groups}, num_channels={x_orig.shape[-1]}, "
+                        f"metric={metric.item():.2f} vs. limit={ctx.whitening_limit}"
+                    )
+
+                (metric - ctx.whitening_limit).relu().backward()
+                penalty_grad = x_detached.grad
+                scale = ctx.grad_scale * (
+                    x_grad.to(torch.float32).norm() / (penalty_grad.norm() + 1.0e-20)
+                )
+                penalty_grad = penalty_grad * scale
+        return x_grad + penalty_grad.to(x_grad.dtype), None, None, None
+
+
+class Whiten(nn.Module):
+    def __init__(
+        self,
+        num_groups: int,
+        whitening_limit: float,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: float,
+    ):
+        """
+        Args:
+          num_groups: the number of groups to divide the channel dim into before
+            whitening.  We will attempt to make the feature covariance
+            within each group, after mean subtraction, as "white" as possible,
+            while having the same trace across all groups.
+         whitening_limit: a value greater than 1.0, that dictates how much
+           freedom we have to violate the constraints.  1.0 would mean perfectly
+           white, with exactly the same trace across groups; larger values
+           give more freedom.  E.g. 2.0.
+         prob: the probability with which we apply the gradient modification
+           (also affects the grad scale).  May be supplied as a float,
+           or as a pair (min_prob, max_prob)
+
+          grad_scale: determines the scale on the gradient term from this object,
+            relative to the rest of the gradient on the attention weights.
+            E.g. 0.02 (you may want to use smaller values than this if prob is large)
+        """
+        super(Whiten, self).__init__()
+        assert num_groups >= 1
+        assert whitening_limit >= 1
+        assert grad_scale >= 0
+        self.num_groups = num_groups
+        self.whitening_limit = whitening_limit
+        if isinstance(prob, float):
+            assert 0 < prob <= 1
+            self.prob = prob
+        else:
+            (self.min_prob, self.max_prob) = prob
+            assert 0 < self.min_prob < self.max_prob <= 1
+            self.prob = self.max_prob
+
+        self.grad_scale = grad_scale
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        In the forward pass, this function just returns the input unmodified.
+        In the backward pass, it will modify the gradients to ensure that the
+        distribution in each group has close to (lambda times I) as the covariance
+        after mean subtraction, with the same lambda across groups.
+        For whitening_limit > 1, there will be more freedom to violate this
+        constraint.
+
+        Args:
+           x: the input of shape (*, num_channels)
+
+        Returns:
+            x, unmodified.   You should make sure
+        you use the returned value, or the graph will be freed
+        and nothing will happen in backprop.
+        """
+        if not x.requires_grad or random.random() > self.prob or self.grad_scale == 0:
+            return _no_op(x)
+        else:
+            if hasattr(self, "min_prob") and random.random() < 0.25:
+                # occasionally switch between min_prob and max_prob, based on whether
+                # we are above or below the threshold.
+                if (
+                    _whitening_metric(x.to(torch.float32), self.num_groups)
+                    > self.whitening_limit
+                ):
+                    # there would be a change to the grad.
+                    self.prob = self.max_prob
+                else:
+                    self.prob = self.min_prob
+
+            return WhiteningPenaltyFunction.apply(
+                x, self.num_groups, self.whitening_limit, self.grad_scale
+            )
+
+
+class WithLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, y: Tensor):
+        ctx.y_shape = y.shape
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        return (
+            ans_grad,
+            torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
+        )
+
+
+def with_loss(x, y):
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    # returns x but adds y.sum() to the loss function.
+    return WithLoss.apply(x, y)
+
+
+def _no_op(x: Tensor) -> Tensor:
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    else:
+        # a no-op function that will have a node in the autograd graph,
+        # to avoid certain bugs relating to backward hooks
+        return x.chunk(1, dim=-1)[0]
+
+
+class Identity(torch.nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, x):
+        return _no_op(x)
+
+
+class MaxEig(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to discourage
+    that any given direction in activation space accounts for more than
+    a specified proportion of the covariance (e.g. 0.2).
+
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           max_var_per_eig:  the maximum proportion of the variance of the
+               features/channels, after mean subtraction, that can come from
+               any given eigenvalue.
+           min_prob: the minimum probability with which we apply this during any invocation
+               of forward(), assuming last time we applied the constraint it was
+               not active; supplied for speed.
+           scale: determines the scale with which we modify the gradients, relative
+               to the existing / unmodified gradients
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        max_var_per_eig: float = 0.2,
+        min_prob: float = 0.01,
+        scale: float = 0.01,
+    ):
+        super(MaxEig, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.scale = scale
+        assert max_var_per_eig == 0.0 or max_var_per_eig > 1.0 / num_channels
+        self.max_var_per_eig = max_var_per_eig
+
+        # we figure out the dominant direction using the power method: starting with
+        # a random vector, keep multiplying by the covariance and renormalizing.
+        with torch.no_grad():
+            # arbitrary.. would use randn() but want to leave the rest of the model's
+            # random parameters unchanged for comparison
+            direction = torch.arange(num_channels).to(torch.float)
+            direction = direction / direction.norm()
+            self.register_buffer("max_eig_direction", direction)
+
+        self.min_prob = min_prob
+        # cur_prob is the current probability we'll use to apply the ActivationBalancer.
+        # We'll regress this towards prob, each tiem we try to apply it and it is not
+        # active.
+        self.cur_prob = 1.0
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or self.max_var_per_eig <= 0
+            or random.random() > self.cur_prob
+            or torch.jit.is_tracing()
+        ):
+            return _no_op(x)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            eps = 1.0e-20
+            orig_x = x
+            x = x.to(torch.float32)
+            with torch.no_grad():
+                x = x.transpose(self.channel_dim, -1).reshape(-1, self.num_channels)
+                x = x - x.mean(dim=0)
+                new_direction, coeffs = self._find_direction_coeffs(
+                    x, self.max_eig_direction
+                )
+                x_var = (x**2).mean()
+                x_residual = x - coeffs * new_direction
+                x_residual_var = (x_residual**2).mean()
+
+                # `variance_proportion` is the proportion of the variance accounted for
+                # by the top eigen-direction.
+                variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+
+                # ensure new direction is nonzero even if x == 0, by including `direction`.
+                self._set_direction(0.1 * self.max_eig_direction + new_direction)
+
+            if random.random() < 0.01 or __name__ == "__main__":
+                logging.info(
+                    f"variance_proportion = {variance_proportion.item()}, shape={tuple(orig_x.shape)}, cur_prob={self.cur_prob}"
+                )
+
+            if variance_proportion >= self.max_var_per_eig:
+                # The constraint is active.  Note, we should quite rarely
+                # reach here, only near the beginning of training if we are
+                # starting to diverge, should this constraint be active.
+                cur_prob = self.cur_prob
+                self.cur_prob = 1.0  # next time, do the update with probability 1.0.
+                return MaxEigLimiterFunction.apply(
+                    orig_x, coeffs, new_direction, self.channel_dim, self.scale
+                )
+            else:
+                # let self.cur_prob exponentially approach self.min_prob, as
+                # long as the constraint is inactive.
+                self.cur_prob = 0.75 * self.cur_prob + 0.25 * self.min_prob
+                return orig_x
+
+    def _set_direction(self, direction: Tensor):
+        """
+        Sets self.max_eig_direction to a normalized version of `direction`
+        """
+        direction = direction.detach()
+        direction = direction / direction.norm()
+        direction_sum = direction.sum().item()
+        if direction_sum - direction_sum == 0:  # no inf/nan
+            self.max_eig_direction[:] = direction
+        else:
+            logging.info(
+                f"Warning: sum of direction in MaxEig is {direction_sum}, "
+                "num_channels={self.num_channels}, channel_dim={self.channel_dim}"
+            )
+
+    def _find_direction_coeffs(
+        self, x: Tensor, prev_direction: Tensor
+    ) -> Tuple[Tensor, Tensor, Tensor]:
+        """
+            Figure out (an approximation to) the proportion of the variance of a set of
+            feature vectors that can be attributed to the top eigen-direction.
+            Args:
+             x: a Tensor of shape (num_frames, num_channels), with num_frames > 1.
+          prev_direction:  a Tensor of shape (num_channels,), that is our previous estimate
+                   of the top eigen-direction, or a random direction if this is the first
+                   iteration.  Does not have to be normalized, but should be nonzero.
+
+        Returns: (cur_direction, coeffs), where:
+             cur_direction: a Tensor of shape (num_channels,) that is the current
+                estimate of the top eigen-direction.
+             coeffs: a Tensor of shape (num_frames, 1) that minimizes, or
+                approximately minimizes, (x - coeffs * cur_direction).norm()
+        """
+        (num_frames, num_channels) = x.shape
+        assert num_channels > 1 and num_frames > 1
+        assert prev_direction.shape == (num_channels,)
+        # `coeffs` are the coefficients of `prev_direction` in x.
+        # actually represent the coeffs up to a constant positive factor.
+        coeffs = (x * prev_direction).sum(dim=1, keepdim=True) + 1.0e-10
+        cur_direction = (x * coeffs).sum(dim=0) / ((coeffs**2).sum() + 1.0e-20)
+        return cur_direction, coeffs
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.043637 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.043637
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+def _test_max_eig():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = MaxEig(
+            num_channels, 1, 0.5, scale=0.1  # channel_dim  # max_var_per_eig
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad, atol=1.0e-02)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_whiten():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"_test_whiten(): proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_activation_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        probs.numel(),
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        magnitudes.numel(),
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+        min_prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x**2).mean().sqrt()
+    y_rms = (y**2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = DoubleSwish()
+
+    tol = (1.2 - (-0.043637)) / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_softmax():
+    a = torch.randn(2, 10, dtype=torch.float64)
+    b = a.clone()
+    a.requires_grad = True
+    b.requires_grad = True
+    a.softmax(dim=1)[:, 0].sum().backward()
+    print("a grad = ", a.grad)
+    softmax(b, dim=1)[:, 0].sum().backward()
+    print("b grad = ", b.grad)
+    assert torch.allclose(a.grad, b.grad)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_softmax()
+    _test_whiten()
+    _test_max_eig()
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
new file mode 100644
index 000000000..86067b04f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
@@ -0,0 +1,214 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file replaces various modules in a model.
+Specifically, ActivationBalancer is replaced with an identity operator;
+Whiten is also replaced with an identity operator;
+BasicNorm is replaced by a module with `exp` removed.
+"""
+
+import copy
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+from scaling import ActivationBalancer, BasicNorm, Whiten
+from zipformer import PoolingModule
+
+
+class PoolingModuleNoProj(nn.Module):
+    def forward(
+        self,
+        x: torch.Tensor,
+        cached_len: torch.Tensor,
+        cached_avg: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (T, N, C)
+          cached_len:
+            A tensor of shape (N,)
+          cached_avg:
+            A tensor of shape (N, C)
+        Returns:
+          Return a tuple containing:
+            - new_x
+            - new_cached_len
+            - new_cached_avg
+        """
+        x = x.cumsum(dim=0)  # (T, N, C)
+        x = x + (cached_avg * cached_len.unsqueeze(1)).unsqueeze(0)
+        # Cumulated numbers of frames from start
+        cum_mask = torch.arange(1, x.size(0) + 1, device=x.device)
+        cum_mask = cum_mask.unsqueeze(1) + cached_len.unsqueeze(0)  # (T, N)
+        pooling_mask = (1.0 / cum_mask).unsqueeze(2)
+        # now pooling_mask: (T, N, 1)
+        x = x * pooling_mask  # (T, N, C)
+
+        cached_len = cached_len + x.size(0)
+        cached_avg = x[-1]
+
+        return x, cached_len, cached_avg
+
+
+class PoolingModuleWithProj(nn.Module):
+    def __init__(self, proj: torch.nn.Module):
+        super().__init__()
+        self.proj = proj
+        self.pooling = PoolingModuleNoProj()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cached_len: torch.Tensor,
+        cached_avg: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (T, N, C)
+          cached_len:
+            A tensor of shape (N,)
+          cached_avg:
+            A tensor of shape (N, C)
+        Returns:
+          Return a tuple containing:
+            - new_x
+            - new_cached_len
+            - new_cached_avg
+        """
+        x, cached_len, cached_avg = self.pooling(x, cached_len, cached_avg)
+        return self.proj(x), cached_len, cached_avg
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        cached_len: torch.Tensor,
+        cached_avg: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (T, N, C)
+          cached_len:
+            A tensor of shape (N,)
+          cached_avg:
+            A tensor of shape (N, C)
+        Returns:
+          Return a tuple containing:
+            - new_x
+            - new_cached_len
+            - new_cached_avg
+        """
+        x, cached_len, cached_avg = self.pooling(x, cached_len, cached_avg)
+        return self.proj(x), cached_len, cached_avg
+
+
+class NonScaledNorm(nn.Module):
+    """See BasicNorm for doc"""
+
+    def __init__(
+        self,
+        num_channels: int,
+        eps_exp: float,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+    ):
+        super().__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.eps_exp = eps_exp
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not torch.jit.is_tracing():
+            assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp
+        ).pow(-0.5)
+        return x * scales
+
+
+def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm:
+    assert isinstance(basic_norm, BasicNorm), type(basic_norm)
+    norm = NonScaledNorm(
+        num_channels=basic_norm.num_channels,
+        eps_exp=basic_norm.eps.data.exp().item(),
+        channel_dim=basic_norm.channel_dim,
+    )
+    return norm
+
+
+def convert_pooling_module(pooling: PoolingModule) -> PoolingModuleWithProj:
+    assert isinstance(pooling, PoolingModule), type(pooling)
+    return PoolingModuleWithProj(proj=pooling.proj)
+
+
+# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule  # noqa
+# get_submodule was added to nn.Module at v1.9.0
+def get_submodule(model, target):
+    if target == "":
+        return model
+    atoms: List[str] = target.split(".")
+    mod: torch.nn.Module = model
+    for item in atoms:
+        if not hasattr(mod, item):
+            raise AttributeError(
+                mod._get_name() + " has no " "attribute `" + item + "`"
+            )
+        mod = getattr(mod, item)
+        if not isinstance(mod, torch.nn.Module):
+            raise AttributeError("`" + item + "` is not " "an nn.Module")
+    return mod
+
+
+def convert_scaled_to_non_scaled(
+    model: nn.Module,
+    inplace: bool = False,
+    is_pnnx: bool = False,
+):
+    """
+    Args:
+      model:
+        The model to be converted.
+      inplace:
+        If True, the input model is modified inplace.
+        If False, the input model is copied and we modify the copied version.
+      is_pnnx:
+        True if we are going to export the model for PNNX.
+    Return:
+      Return a model without scaled layers.
+    """
+    if not inplace:
+        model = copy.deepcopy(model)
+
+    d = {}
+    for name, m in model.named_modules():
+        if isinstance(m, BasicNorm):
+            d[name] = convert_basic_norm(m)
+        elif isinstance(m, (ActivationBalancer, Whiten)):
+            d[name] = nn.Identity()
+        elif isinstance(m, PoolingModule) and is_pnnx:
+            d[name] = convert_pooling_module(m)
+
+    for k, v in d.items():
+        if "." in k:
+            parent, child = k.rsplit(".", maxsplit=1)
+            setattr(get_submodule(model, parent), child, v)
+        else:
+            setattr(model, k, v)
+
+    return model
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/test_compute_ali.py b/egs/librispeech/ASR/pruned_transducer_stateless7/test_compute_ali.py
new file mode 100755
index 000000000..081f7ba1a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/test_compute_ali.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script compares the word-level alignments generated based on modified_beam_search decoding
+(in ./pruned_transducer_stateless7/compute_ali.py) to the reference alignments generated
+by torchaudio framework (in ./add_alignments.sh).
+
+Usage:
+
+./pruned_transducer_stateless7/compute_ali.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --dataset test-clean \
+    --max-duration 300 \
+    --beam-size 4 \
+    --cuts-out-dir data/fbank_ali_beam_search
+
+And the you can run:
+
+./pruned_transducer_stateless7/test_compute_ali.py \
+  --cuts-out-dir ./data/fbank_ali_test \
+  --cuts-ref-dir ./data/fbank_ali_torch \
+  --dataset train-clean-100
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import load_manifest
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--cuts-out-dir",
+        type=Path,
+        default="./data/fbank_ali",
+        help="The dir that saves the generated cuts manifests with alignments",
+    )
+
+    parser.add_argument(
+        "--cuts-ref-dir",
+        type=Path,
+        default="./data/fbank_ali_torch",
+        help="The dir that saves the reference cuts manifests with alignments",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="""The name of the dataset:
+        Possible values are:
+            - test-clean
+            - test-other
+            - train-clean-100
+            - train-clean-360
+            - train-other-500
+            - dev-clean
+            - dev-other
+        """,
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+
+    cuts_out_jsonl = args.cuts_out_dir / f"librispeech_cuts_{args.dataset}.jsonl.gz"
+    cuts_ref_jsonl = args.cuts_ref_dir / f"librispeech_cuts_{args.dataset}.jsonl.gz"
+
+    logging.info(f"Loading {cuts_out_jsonl} and {cuts_ref_jsonl}")
+    cuts_out = load_manifest(cuts_out_jsonl)
+    cuts_ref = load_manifest(cuts_ref_jsonl)
+    cuts_ref = cuts_ref.sort_like(cuts_out)
+
+    all_time_diffs = []
+    for cut_out, cut_ref in zip(cuts_out, cuts_ref):
+        time_out = [
+            ali.start
+            for ali in cut_out.supervisions[0].alignment["word"]
+            if ali.symbol != ""
+        ]
+        time_ref = [
+            ali.start
+            for ali in cut_ref.supervisions[0].alignment["word"]
+            if ali.symbol != ""
+        ]
+        assert len(time_out) == len(time_ref), (len(time_out), len(time_ref))
+        diff = [
+            round(abs(out - ref), ndigits=3) for out, ref in zip(time_out, time_ref)
+        ]
+        all_time_diffs += diff
+
+    all_time_diffs = torch.tensor(all_time_diffs)
+    logging.info(
+        f"For the word-level alignments abs difference on dataset {args.dataset}, "
+        f"mean: {'%.2f' % all_time_diffs.mean()}s, std: {'%.2f' % all_time_diffs.std()}s"
+    )
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py
new file mode 100755
index 000000000..cdf914df3
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py
@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless7/test_model.py
+"""
+
+import torch
+
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    # Test jit script
+    convert_scaled_to_non_scaled(model, inplace=True)
+    # We won't use the forward() method of the model in C++, so just ignore
+    # it here.
+    # Otherwise, one of its arguments is a ragged tensor and is not
+    # torch scriptabe.
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    print("Using torch.jit.script")
+    model = torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py
new file mode 100644
index 000000000..f3f7b1ea9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/test_onnx.py
@@ -0,0 +1,380 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file is to test that models can be exported to onnx.
+"""
+import os
+
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
+import onnxruntime as ort
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from zipformer import (
+    Conv2dSubsampling,
+    RelPositionalEncoding,
+    Zipformer,
+    ZipformerEncoder,
+    ZipformerEncoderLayer,
+)
+
+ort.set_default_logger_severity(3)
+
+
+def test_conv2d_subsampling():
+    filename = "conv2d_subsampling.onnx"
+    opset_version = 13
+    N = 30
+    T = 50
+    num_features = 80
+    d_model = 512
+    x = torch.rand(N, T, num_features)
+
+    encoder_embed = Conv2dSubsampling(num_features, d_model)
+    encoder_embed.eval()
+    encoder_embed = convert_scaled_to_non_scaled(encoder_embed, inplace=True)
+
+    torch.onnx.export(
+        encoder_embed,
+        x,
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x"],
+        output_names=["y"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "y": {0: "N", 1: "T"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    assert input_nodes[0].name == "x"
+    assert input_nodes[0].shape == ["N", "T", num_features]
+
+    inputs = {input_nodes[0].name: x.numpy()}
+
+    onnx_y = session.run(["y"], inputs)[0]
+
+    onnx_y = torch.from_numpy(onnx_y)
+    torch_y = encoder_embed(x)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    os.remove(filename)
+
+
+def test_rel_pos():
+    filename = "rel_pos.onnx"
+
+    opset_version = 13
+    N = 30
+    T = 50
+    num_features = 80
+    d_model = 512
+    x = torch.rand(N, T, num_features)
+
+    encoder_pos = RelPositionalEncoding(d_model, dropout_rate=0.1)
+    encoder_pos.eval()
+    encoder_pos = convert_scaled_to_non_scaled(encoder_pos, inplace=True)
+
+    x = x.permute(1, 0, 2)
+
+    torch.onnx.export(
+        encoder_pos,
+        x,
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x"],
+        output_names=["pos_emb"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "pos_emb": {0: "N", 1: "T"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    assert input_nodes[0].name == "x"
+    assert input_nodes[0].shape == ["N", "T", num_features]
+
+    inputs = {input_nodes[0].name: x.numpy()}
+    onnx_pos_emb = session.run(["pos_emb"], inputs)
+    onnx_pos_emb = torch.from_numpy(onnx_pos_emb[0])
+
+    torch_pos_emb = encoder_pos(x)
+    assert torch.allclose(onnx_pos_emb, torch_pos_emb, atol=1e-05), (
+        (onnx_pos_emb - torch_pos_emb).abs().max()
+    )
+    print(onnx_pos_emb.abs().sum(), torch_pos_emb.abs().sum())
+
+    os.remove(filename)
+
+
+def test_zipformer_encoder_layer():
+    filename = "zipformer_encoder_layer.onnx"
+    opset_version = 13
+    N = 30
+    T = 50
+
+    d_model = 384
+    attention_dim = 192
+    nhead = 8
+    feedforward_dim = 1024
+    dropout = 0.1
+    cnn_module_kernel = 31
+    pos_dim = 4
+
+    x = torch.rand(N, T, d_model)
+
+    encoder_pos = RelPositionalEncoding(d_model, dropout)
+    encoder_pos.eval()
+    encoder_pos = convert_scaled_to_non_scaled(encoder_pos, inplace=True)
+
+    x = x.permute(1, 0, 2)
+    pos_emb = encoder_pos(x)
+
+    encoder_layer = ZipformerEncoderLayer(
+        d_model,
+        attention_dim,
+        nhead,
+        feedforward_dim,
+        dropout,
+        cnn_module_kernel,
+        pos_dim,
+    )
+    encoder_layer.eval()
+    encoder_layer = convert_scaled_to_non_scaled(encoder_layer, inplace=True)
+
+    torch.onnx.export(
+        encoder_layer,
+        (x, pos_emb),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "pos_emb"],
+        output_names=["y"],
+        dynamic_axes={
+            "x": {0: "T", 1: "N"},
+            "pos_emb": {0: "N", 1: "T"},
+            "y": {0: "T", 1: "N"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+        input_nodes[1].name: pos_emb.numpy(),
+    }
+    onnx_y = session.run(["y"], inputs)[0]
+    onnx_y = torch.from_numpy(onnx_y)
+
+    torch_y = encoder_layer(x, pos_emb)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    print(onnx_y.abs().sum(), torch_y.abs().sum(), onnx_y.shape, torch_y.shape)
+
+    os.remove(filename)
+
+
+def test_zipformer_encoder():
+    filename = "zipformer_encoder.onnx"
+
+    opset_version = 13
+    N = 3
+    T = 15
+
+    d_model = 512
+    attention_dim = 192
+    nhead = 8
+    feedforward_dim = 1024
+    dropout = 0.1
+    cnn_module_kernel = 31
+    pos_dim = 4
+    num_encoder_layers = 12
+
+    warmup_batches = 4000.0
+    warmup_begin = warmup_batches / (num_encoder_layers + 1)
+    warmup_end = warmup_batches / (num_encoder_layers + 1)
+
+    x = torch.rand(N, T, d_model)
+
+    encoder_layer = ZipformerEncoderLayer(
+        d_model,
+        attention_dim,
+        nhead,
+        feedforward_dim,
+        dropout,
+        cnn_module_kernel,
+        pos_dim,
+    )
+    encoder = ZipformerEncoder(
+        encoder_layer, num_encoder_layers, dropout, warmup_begin, warmup_end
+    )
+    encoder.eval()
+    encoder = convert_scaled_to_non_scaled(encoder, inplace=True)
+
+    # jit_model = torch.jit.trace(encoder, (pos_emb))
+
+    torch_y = encoder(x)
+
+    torch.onnx.export(
+        encoder,
+        (x, torch.ones(1, dtype=torch.float32)),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x"],
+        output_names=["y"],
+        dynamic_axes={
+            "x": {0: "T", 1: "N"},
+            "y": {0: "T", 1: "N"},
+        },
+    )
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+        input_nodes[1].name: torch.ones(1, dtype=torch.float32).numpy(),
+    }
+    onnx_y = session.run(["y"], inputs)[0]
+    onnx_y = torch.from_numpy(onnx_y)
+
+    torch_y = encoder(x)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    print(onnx_y.abs().sum(), torch_y.abs().sum(), onnx_y.shape, torch_y.shape)
+
+    os.remove(filename)
+
+
+def test_zipformer():
+    filename = "zipformer.onnx"
+    opset_version = 11
+    N = 3
+    T = 15
+    num_features = 80
+    x = torch.rand(N, T, num_features)
+    x_lens = torch.full((N,), fill_value=T, dtype=torch.int64)
+
+    zipformer = Zipformer(num_features=num_features)
+    zipformer.eval()
+    zipformer = convert_scaled_to_non_scaled(zipformer, inplace=True)
+
+    # jit_model = torch.jit.trace(zipformer, (x, x_lens))
+    torch.onnx.export(
+        zipformer,
+        (x, x_lens),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["y", "y_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "y": {0: "N", 1: "T"},
+            "y_lens": {0: "N"},
+        },
+    )
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    session = ort.InferenceSession(
+        filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+
+    input_nodes = session.get_inputs()
+    inputs = {
+        input_nodes[0].name: x.numpy(),
+        input_nodes[1].name: x_lens.numpy(),
+    }
+    onnx_y, onnx_y_lens = session.run(["y", "y_lens"], inputs)
+    onnx_y = torch.from_numpy(onnx_y)
+    onnx_y_lens = torch.from_numpy(onnx_y_lens)
+
+    torch_y, torch_y_lens = zipformer(x, x_lens)
+    assert torch.allclose(onnx_y, torch_y, atol=1e-05), (onnx_y - torch_y).abs().max()
+
+    assert torch.allclose(onnx_y_lens, torch_y_lens, atol=1e-05), (
+        (onnx_y_lens - torch_y_lens).abs().max()
+    )
+    print(onnx_y.abs().sum(), torch_y.abs().sum(), onnx_y.shape, torch_y.shape)
+    print(onnx_y_lens, torch_y_lens)
+
+    os.remove(filename)
+
+
+@torch.no_grad()
+def main():
+    test_conv2d_subsampling()
+    test_rel_pos()
+    test_zipformer_encoder_layer()
+    test_zipformer_encoder()
+    test_zipformer()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20221011)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
new file mode 100755
index 000000000..fac3706d2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
@@ -0,0 +1,1254 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Yifan   Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+    symlink_or_copy,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    epoch_basename = f"epoch-{params.cur_epoch}.pt"
+    filename = params.exp_dir / epoch_basename
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        symlink_or_copy(
+            exp_dir=params.exp_dir, src=epoch_basename, dst="best-train-loss.pt"
+        )
+
+    if params.best_valid_epoch == params.cur_epoch:
+        symlink_or_copy(
+            exp_dir=params.exp_dir, src=epoch_basename, dst="best-valid-loss.pt"
+        )
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted in descending order by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    if is_training:
+        batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.mini_libri:
+        train_cuts = librispeech.train_clean_5_cuts()
+    elif params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if params.mini_libri:
+        valid_cuts = librispeech.dev_clean_2_cuts()
+    else:
+        valid_cuts = librispeech.dev_clean_cuts()
+        valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
new file mode 100644
index 000000000..cbde2a2e4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
@@ -0,0 +1,1909 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import itertools
+import logging
+import math
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    Identity,
+    MaxEig,
+    ScaledConv1d,
+    Whiten,
+    _diag,
+    penalize_abs_values_gt,
+    random_clamp,
+    softmax,
+)
+from torch import Tensor, nn
+
+from icefall.dist import get_rank
+from icefall.utils import is_jit_tracing, make_pad_mask
+
+
+class Zipformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        d_model: (int,int): embedding dimension of 2 encoder stacks
+        attention_dim: (int,int): attention dimension of 2 encoder stacks
+        nhead (int, int): number of heads
+        dim_feedforward (int, int): feedforward dimension in 2 encoder stacks
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+        warmup_batches (float): number of batches to warm up over
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_downsampling_factor: int = 2,
+        encoder_dims: Tuple[int] = (384, 384),
+        attention_dim: Tuple[int] = (256, 256),
+        encoder_unmasked_dims: Tuple[int] = (256, 256),
+        zipformer_downsampling_factors: Tuple[int] = (2, 4),
+        nhead: Tuple[int] = (8, 8),
+        feedforward_dim: Tuple[int] = (1536, 2048),
+        num_encoder_layers: Tuple[int] = (12, 12),
+        dropout: float = 0.1,
+        cnn_module_kernels: Tuple[int] = (31, 31),
+        pos_dim: int = 4,
+        warmup_batches: float = 4000.0,
+    ) -> None:
+        super(Zipformer, self).__init__()
+
+        self.num_features = num_features
+        assert 0 < encoder_dims[0] <= encoder_dims[1]
+        self.encoder_dims = encoder_dims
+        self.encoder_unmasked_dims = encoder_unmasked_dims
+        self.zipformer_downsampling_factors = zipformer_downsampling_factors
+        self.output_downsampling_factor = output_downsampling_factor
+
+        # will be written to, see set_batch_count()
+        self.batch_count = 0
+        self.warmup_end = warmup_batches
+
+        for u, d in zip(encoder_unmasked_dims, encoder_dims):
+            assert u <= d, (u, d)
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, (T - 7)//2, encoder_dims).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> (T - 7)//2
+        #   (2) embedding: num_features -> encoder_dims
+        self.encoder_embed = Conv2dSubsampling(
+            num_features, encoder_dims[0], dropout=dropout
+        )
+
+        # each one will be ZipformerEncoder or DownsampledZipformerEncoder
+        encoders = []
+
+        num_encoders = len(encoder_dims)
+        for i in range(num_encoders):
+            encoder_layer = ZipformerEncoderLayer(
+                encoder_dims[i],
+                attention_dim[i],
+                nhead[i],
+                feedforward_dim[i],
+                dropout,
+                cnn_module_kernels[i],
+                pos_dim,
+            )
+
+            # For the segment of the warmup period, we let the Conv2dSubsampling
+            # layer learn something.  Then we start to warm up the other encoders.
+            encoder = ZipformerEncoder(
+                encoder_layer,
+                num_encoder_layers[i],
+                dropout,
+                warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1),
+                warmup_end=warmup_batches * (i + 2) / (num_encoders + 1),
+            )
+
+            if zipformer_downsampling_factors[i] != 1:
+                encoder = DownsampledZipformerEncoder(
+                    encoder,
+                    input_dim=encoder_dims[i - 1] if i > 0 else encoder_dims[0],
+                    output_dim=encoder_dims[i],
+                    downsample=zipformer_downsampling_factors[i],
+                )
+            encoders.append(encoder)
+        self.encoders = nn.ModuleList(encoders)
+
+        # initializes self.skip_layers and self.skip_modules
+        self._init_skip_modules()
+
+        self.downsample_output = AttentionDownsample(
+            encoder_dims[-1], encoder_dims[-1], downsample=output_downsampling_factor
+        )
+
+    def _get_layer_skip_dropout_prob(self):
+        if not self.training:
+            return 0.0
+        batch_count = self.batch_count
+        min_dropout_prob = 0.025
+
+        if batch_count > self.warmup_end:
+            return min_dropout_prob
+        else:
+            return 0.5 - (batch_count / self.warmup_end) * (0.5 - min_dropout_prob)
+
+    def _init_skip_modules(self):
+        """
+        If self.zipformer_downampling_factors = (1, 2, 4, 8, 4, 2), then at the input of layer
+        indexed 4 (in zero indexing), with has subsapling_factor=4, we combine the output of
+        layers 2 and 3; and at the input of layer indexed 5, which which has subsampling_factor=2,
+        we combine the outputs of layers 1 and 5.
+        """
+        skip_layers = []
+        skip_modules = []
+        z = self.zipformer_downsampling_factors
+        for i in range(len(z)):
+            if i <= 1 or z[i - 1] <= z[i]:
+                skip_layers.append(None)
+                skip_modules.append(SimpleCombinerIdentity())
+            else:
+                # TEMP
+                for j in range(i - 2, -1, -1):
+                    if z[j] <= z[i] or j == 0:
+                        # TEMP logging statement.
+                        logging.info(
+                            f"At encoder stack {i}, which has downsampling_factor={z[i]}, we will "
+                            f"combine the outputs of layers {j} and {i-1}, with downsampling_factors={z[j]} and {z[i-1]}."
+                        )
+                        skip_layers.append(j)
+                        skip_modules.append(
+                            SimpleCombiner(
+                                self.encoder_dims[j],
+                                self.encoder_dims[i - 1],
+                                min_weight=(0.0, 0.25),
+                            )
+                        )
+                        break
+        self.skip_layers = skip_layers
+        self.skip_modules = nn.ModuleList(skip_modules)
+
+    def get_feature_masks(self, x: torch.Tensor) -> List[float]:
+        # Note: The actual return type is Union[List[float], List[Tensor]],
+        # but to make torch.jit.script() work, we use List[float]
+        """
+        In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of
+        randomized feature masks, one per encoder.
+        On e.g. 15% of frames, these masks will zero out all encoder dims larger than
+        some supplied number, e.g. >256, so in effect on those frames we are using
+        a smaller encoder dim.
+
+        We generate the random masks at this level because we want the 2 masks to 'agree'
+        all the way up the encoder stack. This will mean that the 1st mask will have
+        mask values repeated self.zipformer_downsampling_factors times.
+
+        Args:
+           x: the embeddings (needed for the shape and dtype and device), of shape
+             (num_frames, batch_size, encoder_dims0)
+        """
+        num_encoders = len(self.encoder_dims)
+        if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing():
+            return [1.0] * num_encoders
+
+        (num_frames0, batch_size, _encoder_dims0) = x.shape
+
+        assert self.encoder_dims[0] == _encoder_dims0, (
+            self.encoder_dims,
+            _encoder_dims0,
+        )
+
+        max_downsampling_factor = max(self.zipformer_downsampling_factors)
+
+        num_frames_max = num_frames0 + max_downsampling_factor - 1
+
+        feature_mask_dropout_prob = 0.15
+
+        # frame_mask_max shape: (num_frames_max, batch_size, 1)
+        frame_mask_max = (
+            torch.rand(num_frames_max, batch_size, 1, device=x.device)
+            > feature_mask_dropout_prob
+        ).to(x.dtype)
+
+        feature_masks = []
+        for i in range(num_encoders):
+            ds = self.zipformer_downsampling_factors[i]
+            upsample_factor = max_downsampling_factor // ds
+
+            frame_mask = (
+                frame_mask_max.unsqueeze(1)
+                .expand(num_frames_max, upsample_factor, batch_size, 1)
+                .reshape(num_frames_max * upsample_factor, batch_size, 1)
+            )
+            num_frames = (num_frames0 + ds - 1) // ds
+            frame_mask = frame_mask[:num_frames]
+            feature_mask = torch.ones(
+                num_frames,
+                batch_size,
+                self.encoder_dims[i],
+                dtype=x.dtype,
+                device=x.device,
+            )
+            u = self.encoder_unmasked_dims[i]
+            feature_mask[:, :, u:] *= frame_mask
+            feature_masks.append(feature_mask)
+
+        return feature_masks
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1])
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        lengths = (x_lens - 7) >> 1
+        assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max())
+        mask = make_pad_mask(lengths, x.size(0))
+
+        outputs = []
+        feature_masks = self.get_feature_masks(x)
+
+        for i, (module, skip_module) in enumerate(
+            zip(self.encoders, self.skip_modules)
+        ):
+            ds = self.zipformer_downsampling_factors[i]
+            k = self.skip_layers[i]
+            if isinstance(k, int):
+                layer_skip_dropout_prob = self._get_layer_skip_dropout_prob()
+                if torch.jit.is_scripting() or torch.jit.is_tracing():
+                    x = skip_module(outputs[k], x)
+                elif (not self.training) or random.random() > layer_skip_dropout_prob:
+                    x = skip_module(outputs[k], x)
+            x = module(
+                x,
+                feature_mask=feature_masks[i],
+                src_key_padding_mask=None if mask is None else mask[..., ::ds],
+            )
+            outputs.append(x)
+
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        lengths = (lengths + 1) >> 1
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths
+
+
+class ZipformerEncoderLayer(nn.Module):
+    """
+    ZipformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        feedforward_dim: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        attention_dim: int,
+        nhead: int,
+        feedforward_dim: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        pos_dim: int = 4,
+    ) -> None:
+        super(ZipformerEncoderLayer, self).__init__()
+
+        self.d_model = d_model
+
+        # will be written to, see set_batch_count()
+        self.batch_count = 0
+
+        self.self_attn = RelPositionMultiheadAttention(
+            d_model,
+            attention_dim,
+            nhead,
+            pos_dim,
+            dropout=0.0,
+        )
+
+        self.pooling = PoolingModule(d_model)
+
+        self.feed_forward1 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.feed_forward2 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.feed_forward3 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.conv_module1 = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.conv_module2 = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        self.bypass_scale = nn.Parameter(torch.tensor(0.5))
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            d_model,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            max_abs=6.0,
+        )
+        self.whiten = Whiten(
+            num_groups=1, whitening_limit=5.0, prob=(0.025, 0.25), grad_scale=0.01
+        )
+
+    def get_bypass_scale(self):
+        if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing():
+            return self.bypass_scale
+        if random.random() < 0.1:
+            # ensure we get grads if self.bypass_scale becomes out of range
+            return self.bypass_scale
+        # hardcode warmup period for bypass scale
+        warmup_period = 20000.0
+        initial_clamp_min = 0.75
+        final_clamp_min = 0.25
+        if self.batch_count > warmup_period:
+            clamp_min = final_clamp_min
+        else:
+            clamp_min = initial_clamp_min - (self.batch_count / warmup_period) * (
+                initial_clamp_min - final_clamp_min
+            )
+        return self.bypass_scale.clamp(min=clamp_min, max=1.0)
+
+    def get_dynamic_dropout_rate(self):
+        # return dropout rate for the dynamic modules (self_attn, pooling, convolution); this
+        # starts at 0.2 and rapidly decreases to 0.  Its purpose is to keep the training stable
+        # at the beginning, by making the network focus on the feedforward modules.
+        if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing():
+            return 0.0
+        warmup_period = 2000.0
+        initial_dropout_rate = 0.2
+        final_dropout_rate = 0.0
+        if self.batch_count > warmup_period:
+            return final_dropout_rate
+        else:
+            return initial_dropout_rate - (
+                initial_dropout_rate * final_dropout_rate
+            ) * (self.batch_count / warmup_period)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+        batch_split: if not None, this layer will only be applied to
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        # macaron style feed forward module
+        src = src + self.feed_forward1(src)
+
+        # dropout rate for submodules that interact with time.
+        dynamic_dropout = self.get_dynamic_dropout_rate()
+
+        # pooling module
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            src = src + self.pooling(src, key_padding_mask=src_key_padding_mask)
+        elif random.random() >= dynamic_dropout:
+            src = src + self.pooling(src, key_padding_mask=src_key_padding_mask)
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            src_att, attn_weights = self.self_attn(
+                src,
+                pos_emb=pos_emb,
+                attn_mask=src_mask,
+                key_padding_mask=src_key_padding_mask,
+            )
+            src = src + src_att
+
+            src = src + self.conv_module1(
+                src, src_key_padding_mask=src_key_padding_mask
+            )
+
+            src = src + self.feed_forward2(src)
+
+            src = src + self.self_attn.forward2(src, attn_weights)
+
+            src = src + self.conv_module2(
+                src, src_key_padding_mask=src_key_padding_mask
+            )
+        else:
+            use_self_attn = random.random() >= dynamic_dropout
+            if use_self_attn:
+                src_att, attn_weights = self.self_attn(
+                    src,
+                    pos_emb=pos_emb,
+                    attn_mask=src_mask,
+                    key_padding_mask=src_key_padding_mask,
+                )
+                src = src + src_att
+
+            if random.random() >= dynamic_dropout:
+                src = src + self.conv_module1(
+                    src, src_key_padding_mask=src_key_padding_mask
+                )
+
+            src = src + self.feed_forward2(src)
+            if use_self_attn:
+                src = src + self.self_attn.forward2(src, attn_weights)
+
+            if random.random() >= dynamic_dropout:
+                src = src + self.conv_module2(
+                    src, src_key_padding_mask=src_key_padding_mask
+                )
+
+        src = src + self.feed_forward3(src)
+
+        src = self.norm_final(self.balancer(src))
+
+        delta = src - src_orig
+
+        src = src_orig + delta * self.get_bypass_scale()
+
+        return self.whiten(src)
+
+
+class ZipformerEncoder(nn.Module):
+    r"""ZipformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ZipformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8)
+        >>> zipformer_encoder = ZipformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = zipformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        dropout: float,
+        warmup_begin: float,
+        warmup_end: float,
+    ) -> None:
+        super().__init__()
+        # will be written to, see set_batch_count() Note: in inference time this
+        # may be zero but should be treated as large, we can check if
+        # self.training is true.
+        self.batch_count = 0
+        self.warmup_begin = warmup_begin
+        self.warmup_end = warmup_end
+        # module_seed is for when we need a random number that is unique to the module but
+        # shared across jobs.   It's used to randomly select how many layers to drop,
+        # so that we can keep this consistent across worker tasks (for efficiency).
+        self.module_seed = torch.randint(0, 1000, ()).item()
+
+        self.encoder_pos = RelPositionalEncoding(encoder_layer.d_model, dropout)
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert 0 <= warmup_begin <= warmup_end, (warmup_begin, warmup_end)
+
+        delta = (1.0 / num_layers) * (warmup_end - warmup_begin)
+        cur_begin = warmup_begin
+        for i in range(num_layers):
+            self.layers[i].warmup_begin = cur_begin
+            cur_begin += delta
+            self.layers[i].warmup_end = cur_begin
+
+    def get_layers_to_drop(self, rnd_seed: int):
+        ans = set()
+        if not self.training:
+            return ans
+
+        batch_count = self.batch_count
+        num_layers = len(self.layers)
+
+        def get_layerdrop_prob(layer: int) -> float:
+            layer_warmup_begin = self.layers[layer].warmup_begin
+            layer_warmup_end = self.layers[layer].warmup_end
+
+            initial_layerdrop_prob = 0.5
+            final_layerdrop_prob = 0.05
+
+            if batch_count == 0:
+                # As a special case, if batch_count == 0, return 0 (drop no
+                # layers).  This is rather ugly, I'm afraid; it is intended to
+                # enable our scan_pessimistic_batches_for_oom() code to work correctly
+                # so if we are going to get OOM it will happen early.
+                # also search for 'batch_count' with quotes in this file to see
+                # how we initialize the warmup count to a random number between
+                # 0 and 10.
+                return 0.0
+            elif batch_count < layer_warmup_begin:
+                return initial_layerdrop_prob
+            elif batch_count > layer_warmup_end:
+                return final_layerdrop_prob
+            else:
+                # linearly interpolate
+                t = (batch_count - layer_warmup_begin) / layer_warmup_end
+                assert 0.0 <= t < 1.001, t
+                return initial_layerdrop_prob + t * (
+                    final_layerdrop_prob - initial_layerdrop_prob
+                )
+
+        shared_rng = random.Random(batch_count + self.module_seed)
+        independent_rng = random.Random(rnd_seed)
+
+        layerdrop_probs = [get_layerdrop_prob(i) for i in range(num_layers)]
+        tot = sum(layerdrop_probs)
+        # Instead of drawing the samples independently, we first randomly decide
+        # how many layers to drop out, using the same random number generator between
+        # jobs so that all jobs drop out the same number (this is for speed).
+        # Then we use an approximate approach to drop out the individual layers
+        # with their specified probs while reaching this exact target.
+        num_to_drop = int(tot) + int(shared_rng.random() < (tot - int(tot)))
+
+        layers = list(range(num_layers))
+        independent_rng.shuffle(layers)
+
+        # go through the shuffled layers until we get the required number of samples.
+        if num_to_drop > 0:
+            for layer in itertools.cycle(layers):
+                if independent_rng.random() < layerdrop_probs[layer]:
+                    ans.add(layer)
+                if len(ans) == num_to_drop:
+                    break
+        if shared_rng.random() < 0.005 or __name__ == "__main__":
+            logging.info(
+                f"warmup_begin={self.warmup_begin:.1f}, warmup_end={self.warmup_end:.1f}, "
+                f"batch_count={batch_count:.1f}, num_to_drop={num_to_drop}, layers_to_drop={ans}"
+            )
+        return ans
+
+    def forward(
+        self,
+        src: Tensor,
+        # Note: The type of feature_mask should be Union[float, Tensor],
+        # but to make torch.jit.script() work, we use `float` here
+        feature_mask: float = 1.0,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        Returns: (x, x_no_combine), both of shape (S, N, E)
+        """
+        pos_emb = self.encoder_pos(src)
+        output = src
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            layers_to_drop = []
+        else:
+            rnd_seed = src.numel() + random.randint(0, 1000)
+            layers_to_drop = self.get_layers_to_drop(rnd_seed)
+
+        output = output * feature_mask
+
+        for i, mod in enumerate(self.layers):
+            if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+                if i in layers_to_drop:
+                    continue
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+            output = output * feature_mask
+
+        return output
+
+
+class DownsampledZipformerEncoder(nn.Module):
+    r"""
+    DownsampledZipformerEncoder is a zipformer encoder evaluated at a reduced frame rate,
+    after convolutional downsampling, and then upsampled again at the output, and combined
+    with the origin input, so that the output has the same shape as the input.
+    """
+
+    def __init__(
+        self, encoder: nn.Module, input_dim: int, output_dim: int, downsample: int
+    ):
+        super(DownsampledZipformerEncoder, self).__init__()
+        self.downsample_factor = downsample
+        self.downsample = AttentionDownsample(input_dim, output_dim, downsample)
+        self.encoder = encoder
+        self.upsample = SimpleUpsample(output_dim, downsample)
+        self.out_combiner = SimpleCombiner(
+            input_dim, output_dim, min_weight=(0.0, 0.25)
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        # Note: the type of feature_mask should be Unino[float, Tensor],
+        # but to make torch.jit.script() happ, we use float here
+        feature_mask: float = 1.0,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer.  feature_mask is expected to be already downsampled by
+               self.downsample_factor.
+            mask: the mask for the src sequence (optional).  CAUTION: we need to downsample
+                  this, if we are to support it.  Won't work correctly yet.
+            src_key_padding_mask: the mask for the src keys per batch (optional).  Should
+                  be downsampled already.
+
+        Shape:
+            src: (S, N, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        Returns: output of shape (S, N, F) where F is the number of output features
+            (output_dim to constructor)
+        """
+        src_orig = src
+        src = self.downsample(src)
+        ds = self.downsample_factor
+        if mask is not None:
+            mask = mask[::ds, ::ds]
+
+        src = self.encoder(
+            src,
+            feature_mask=feature_mask,
+            mask=mask,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src)
+
+
+class AttentionDownsample(torch.nn.Module):
+    """
+    Does downsampling with attention, by weighted sum, and a projection..
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, downsample: int):
+        """
+        Require out_channels > in_channels.
+        """
+        super(AttentionDownsample, self).__init__()
+        self.query = nn.Parameter(torch.randn(in_channels) * (in_channels**-0.5))
+
+        # fill in the extra dimensions with a projection of the input
+        if out_channels > in_channels:
+            self.extra_proj = nn.Linear(
+                in_channels * downsample, out_channels - in_channels, bias=False
+            )
+        else:
+            self.extra_proj = None
+        self.downsample = downsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, in_channels)
+        Returns a tensor of shape
+           ( (seq_len+downsample-1)//downsample, batch_size, out_channels)
+        """
+        (seq_len, batch_size, in_channels) = src.shape
+        ds = self.downsample
+        d_seq_len = (seq_len + ds - 1) // ds
+
+        # Pad to an exact multiple of self.downsample, could be 0 for onnx-export-compatibility
+        # right-pad src, repeating the last element.
+        pad = d_seq_len * ds - seq_len
+        src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2])
+        src = torch.cat((src, src_extra), dim=0)
+        assert src.shape[0] == d_seq_len * ds, (src.shape[0], d_seq_len, ds)
+
+        src = src.reshape(d_seq_len, ds, batch_size, in_channels)
+        scores = (src * self.query).sum(dim=-1, keepdim=True)
+
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04)
+
+        weights = scores.softmax(dim=1)
+
+        # ans1 is the first `in_channels` channels of the output
+        ans = (src * weights).sum(dim=1)
+        src = src.permute(0, 2, 1, 3).reshape(d_seq_len, batch_size, ds * in_channels)
+
+        if self.extra_proj is not None:
+            ans2 = self.extra_proj(src)
+            ans = torch.cat((ans, ans2), dim=2)
+        return ans
+
+
+class SimpleUpsample(torch.nn.Module):
+    """
+    A very simple form of upsampling that mostly just repeats the input, but
+    also adds a position-specific bias.
+    """
+
+    def __init__(self, num_channels: int, upsample: int):
+        super(SimpleUpsample, self).__init__()
+        self.bias = nn.Parameter(torch.randn(upsample, num_channels) * 0.01)
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, num_channels)
+        Returns a tensor of shape
+           ( (seq_len*upsample), batch_size, num_channels)
+        """
+        upsample = self.bias.shape[0]
+        (seq_len, batch_size, num_channels) = src.shape
+        src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels)
+        src = src + self.bias.unsqueeze(1)
+        src = src.reshape(seq_len * upsample, batch_size, num_channels)
+        return src
+
+
+class SimpleCombinerIdentity(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    def forward(self, src1: Tensor, src2: Tensor) -> Tensor:
+        return src1
+
+
+class SimpleCombiner(torch.nn.Module):
+    """
+    A very simple way of combining 2 vectors of 2 different dims, via a
+    learned weighted combination in the shared part of the dim.
+    Args:
+         dim1: the dimension of the first input, e.g. 256
+         dim2: the dimension of the second input, e.g. 384.
+    The output will have the same dimension as dim2.
+    """
+
+    def __init__(self, dim1: int, dim2: int, min_weight: Tuple[float] = (0.0, 0.0)):
+        super(SimpleCombiner, self).__init__()
+        assert dim2 >= dim1, (dim2, dim1)
+        self.weight1 = nn.Parameter(torch.zeros(()))
+        self.min_weight = min_weight
+
+    def forward(self, src1: Tensor, src2: Tensor) -> Tensor:
+        """
+        src1: (*, dim1)
+        src2: (*, dim2)
+
+        Returns: a tensor of shape (*, dim2)
+        """
+        assert src1.shape[:-1] == src2.shape[:-1], (src1.shape, src2.shape)
+
+        weight1 = self.weight1
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            if (
+                self.training
+                and random.random() < 0.25
+                and self.min_weight != (0.0, 0.0)
+            ):
+                weight1 = weight1.clamp(
+                    min=self.min_weight[0], max=1.0 - self.min_weight[1]
+                )
+
+        src1 = src1 * weight1
+        src2 = src2 * (1.0 - weight1)
+
+        src1_dim = src1.shape[-1]
+        src2_dim = src2.shape[-1]
+        if src1_dim != src2_dim:
+            if src1_dim < src2_dim:
+                src1 = torch.nn.functional.pad(src1, (0, src2_dim - src1_dim))
+            else:
+                src1 = src1[:src2_dim]
+
+        return src1 + src2
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct a PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        if is_jit_tracing():
+            # 10k frames correspond to ~100k ms, e.g., 100 seconds, i.e.,
+            # It assumes that the maximum input won't have more than
+            # 10k frames.
+            #
+            # TODO(fangjun): Use torch.jit.script() for this module
+            max_len = 10000
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(0) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(0), self.d_model)
+        pe_negative = torch.zeros(x.size(0), self.d_model)
+        position = torch.arange(0, x.size(0), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tensor:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (time, batch, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(0)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(0),
+        ]
+        return self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
+    we have to write up the differences.
+
+
+    Args:
+        embed_dim: total dimension of the model.
+        attention_dim: dimension in the attention module, may be less or more than embed_dim
+            but must be a multiple of num_heads.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        attention_dim: int,
+        num_heads: int,
+        pos_dim: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.attention_dim = attention_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = attention_dim // num_heads
+        self.pos_dim = pos_dim
+        assert self.head_dim % 2 == 0, self.head_dim
+        assert self.head_dim * num_heads == attention_dim, (
+            self.head_dim,
+            num_heads,
+            attention_dim,
+        )
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5, dividing it between the query and key.
+        in_proj_dim = (
+            2 * attention_dim  # query, key
+            + attention_dim // 2  # value
+            + pos_dim * num_heads  # positional encoding query
+        )
+
+        self.in_proj = ScaledLinear(
+            embed_dim, in_proj_dim, bias=True, initial_scale=self.head_dim**-0.25
+        )
+
+        # self.whiten_values is applied on the values in forward();
+        # it just copies the keys but prevents low-rank distribution by modifying grads.
+        self.whiten_values = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(
+            embed_dim, num_heads * pos_dim, bias=False, initial_scale=0.05
+        )
+
+        # the following are for diagnosics only, see --print-diagnostics option.
+        # they only copy their inputs.
+        self.copy_pos_query = Identity()
+        self.copy_query = Identity()
+
+        self.out_proj = ScaledLinear(
+            attention_dim // 2, embed_dim, bias=True, initial_scale=0.05
+        )
+
+        self.in_proj2 = nn.Linear(embed_dim, attention_dim // 2, bias=False)
+        self.out_proj2 = ScaledLinear(
+            attention_dim // 2, embed_dim, bias=True, initial_scale=0.05
+        )
+        # self.whiten_values2 is applied on the values in forward2()
+        self.whiten_values2 = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x: input to be projected to query, key, value
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Returns: (attn_output, attn_weights)
+
+             - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size,
+                E is the embedding dimension.
+              - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads
+                 and S is the sequence length.
+        """
+        x, weights = self.multi_head_attention_forward(
+            self.in_proj(x),
+            self.linear_pos(pos_emb),
+            self.attention_dim,
+            self.num_heads,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            attn_mask=attn_mask,
+        )
+        return x, weights
+
+    def multi_head_attention_forward(
+        self,
+        x_proj: Tensor,
+        pos: Tensor,
+        attention_dim: int,
+        num_heads: int,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x_proj: the projected input, to be split into query, key, value.
+            pos: head-specific biases arising from the positional embeddings.
+            attention_dim: dimension inside attention mechanism
+            num_heads: parallel attention heads.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is
+              the attention dimension.  Will be split into (query, key, value, pos).
+            - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence
+              length, N is the batch size, and A is the attention dim.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+              E is the embedding dimension.
+            - attn_weights: :math:`(N * H, S, S)` where N is the batch size,
+              H is the num-heads, S is the sequence length.
+        """
+
+        seq_len, bsz, _ = x_proj.size()
+
+        head_dim = attention_dim // num_heads
+        pos_dim = self.pos_dim  # positional-encoding dim per head
+        assert (
+            head_dim * num_heads == attention_dim
+        ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}"
+
+        # self-attention
+        q = x_proj[..., 0:attention_dim]
+        k = x_proj[..., attention_dim : 2 * attention_dim]
+        value_dim = attention_dim // 2
+        v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim]
+        # p is the position-encoding query, its dimension is num_heads*pos_dim..
+        p = x_proj[..., 2 * attention_dim + value_dim :]
+
+        k = self.whiten_keys(k)  # does nothing in the forward pass.
+        v = self.whiten_values(v)  # does nothing in the forward pass.
+        q = self.copy_query(q)  # for diagnostics only, does nothing.
+        p = self.copy_pos_query(p)  # for diagnostics only, does nothing.
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, seq_len, seq_len]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    seq_len,
+                    seq_len,
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.reshape(seq_len, bsz, num_heads, head_dim)
+        p = p.reshape(seq_len, bsz, num_heads, pos_dim)
+        k = k.reshape(seq_len, bsz, num_heads, head_dim)
+        v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == seq_len, "{} == {}".format(
+                key_padding_mask.size(1), seq_len
+            )
+
+        q = q.permute(1, 2, 0, 3)  # (batch, head, time1, head_dim)
+        p = p.permute(1, 2, 0, 3)  # (batch, head, time1, pos_dim)
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+
+        seq_len2 = 2 * seq_len - 1
+        pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1)
+        # pos shape now: (batch, head, pos_dim, seq_len2)
+
+        # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2)
+        #  [where seq_len2 represents relative position.]
+        pos_weights = torch.matmul(p, pos)
+        # the following .as_strided() expression converts the last axis of pos_weights from relative
+        # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+        # not, but let this code define which way round it is supposed to be.
+        if torch.jit.is_tracing():
+            (batch_size, num_heads, time1, n) = pos_weights.shape
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(seq_len)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+            pos_weights = pos_weights.reshape(-1, n)
+            pos_weights = torch.gather(pos_weights, dim=1, index=indexes)
+            pos_weights = pos_weights.reshape(batch_size, num_heads, time1, seq_len)
+        else:
+            pos_weights = pos_weights.as_strided(
+                (bsz, num_heads, seq_len, seq_len),
+                (
+                    pos_weights.stride(0),
+                    pos_weights.stride(1),
+                    pos_weights.stride(2) - pos_weights.stride(3),
+                    pos_weights.stride(3),
+                ),
+                storage_offset=pos_weights.stride(3) * (seq_len - 1),
+            )
+
+        # caution: they are really scores at this point.
+        attn_output_weights = torch.matmul(q, k) + pos_weights
+
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            if training and random.random() < 0.1:
+                # This is a harder way of limiting the attention scores to not be too large.
+                # It incurs a penalty if any of them has an absolute value greater than 50.0.
+                # this should be outside the normal range of the attention scores.  We use
+                # this mechanism instead of, say, a limit on entropy, because once the entropy
+                # gets very small gradients through the softmax can become very small, and
+                # some mechanisms like that become ineffective.
+                attn_output_weights = penalize_abs_values_gt(
+                    attn_output_weights, limit=25.0, penalty=1.0e-04
+                )
+
+        # attn_output_weights: (batch, head, time1, time2)
+        attn_output_weights = attn_output_weights.view(
+            bsz * num_heads, seq_len, seq_len
+        )
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights = attn_output_weights.masked_fill(
+                    attn_mask, float("-inf")
+                )
+            else:
+                attn_output_weights = attn_output_weights + attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, seq_len, seq_len
+            )
+
+        # Using this version of softmax, defined in scaling.py,
+        # should save a little of the memory used in backprop by, if
+        # we are in automatic mixed precision mode (amp) == autocast,
+        # only storing the half-precision output for backprop purposes.
+        attn_output_weights = softmax(attn_output_weights, dim=-1)
+
+        # If we are using chunk-wise attention mask and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`. At this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax). So we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, seq_len, seq_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, seq_len, seq_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2]
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, attention_dim // 2)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        return attn_output, attn_output_weights
+
+    def forward2(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """
+        Second forward function, where we re-use the attn_weights returned by the first forward function
+        but with different input.
+        Args:
+               x: input, of shape (seq_len, batch_size, embed_dim)
+           attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len)
+        Returns:
+               output of the same shape as x, i.e. (seq_len, batch_size, embed_dim)
+        """
+        num_heads = self.num_heads
+        (seq_len, bsz, embed_dim) = x.shape
+        head_dim = self.attention_dim // num_heads
+        # v: (tgt_len, bsz, embed_dim // 2)
+        v = self.in_proj2(x)
+        v = self.whiten_values2(v)  # does nothing in the forward pass.
+        v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        # now v: (bsz * num_heads, seq_len, head_dim // 2)
+        attn_output = torch.bmm(attn_weights, v)
+
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            if random.random() < 0.001 or __name__ == "__main__":
+                self._print_attn_stats(attn_weights, attn_output)
+
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, self.attention_dim // 2)
+        )
+        # returned value is of shape (seq_len, bsz, embed_dim), like x.
+        return self.out_proj2(attn_output)
+
+    def _print_attn_stats(self, attn_weights: Tensor, attn_output: Tensor):
+        # attn_weights: (batch_size * num_heads, seq_len, seq_len)
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        (n, seq_len, head_dim) = attn_output.shape
+        num_heads = self.num_heads
+        bsz = n // num_heads
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_output = attn_output.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .reshape(bsz, num_heads, seq_len)
+                    .mean(dim=(0, 2))
+                )
+                attn_output = attn_output.reshape(bsz, num_heads, seq_len, head_dim)
+                attn_output = attn_output.permute(1, 0, 2, 3).reshape(
+                    num_heads, bsz * seq_len, head_dim
+                )
+                attn_output_mean = attn_output.mean(dim=1, keepdim=True)
+                attn_output = attn_output - attn_output_mean
+                attn_covar = torch.matmul(attn_output.transpose(1, 2), attn_output) / (
+                    bsz * seq_len
+                )
+                # attn_covar: (num_heads, head_dim, head_dim)
+                # eigs, _ = torch.symeig(attn_covar)
+                # logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}")
+
+                attn_covar = _diag(attn_covar).mean(dim=1)  # (num_heads,)
+                embed_dim = self.in_proj2.weight.shape[1]
+                in_proj_covar = (
+                    self.in_proj2.weight.reshape(num_heads, head_dim, embed_dim) ** 2
+                ).mean(dim=(1, 2))
+                out_proj_covar = (
+                    self.out_proj2.weight.reshape(embed_dim, num_heads, head_dim) ** 2
+                ).mean(dim=(0, 2))
+                logging.info(
+                    f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}"
+                )
+
+
+class PoolingModule(nn.Module):
+    """
+    Averages the input over the time dimension and project with a square matrix.
+    """
+
+    def __init__(self, d_model: int):
+        super().__init__()
+        self.proj = ScaledLinear(d_model, d_model, initial_scale=0.1, bias=False)
+
+    def forward(self, x: Tensor, key_padding_mask: Optional[Tensor] = None):
+        """
+        Args:
+           x: a Tensor of shape (T, N, C)
+          key_padding_mask: a Tensor of bool, of shape (N, T), with True in masked
+            positions.
+        Returns:
+           a Tensor of shape (1, N, C)
+        """
+        if key_padding_mask is not None:
+            if torch.jit.is_tracing():
+                pooling_mask = (~key_padding_mask).to(x.dtype)
+            else:
+                pooling_mask = key_padding_mask.logical_not().to(x.dtype)  # (N, T)
+            pooling_mask = pooling_mask / pooling_mask.sum(dim=1, keepdim=True)
+            pooling_mask = pooling_mask.transpose(0, 1).contiguous().unsqueeze(-1)
+            # now pooling_mask: (T, N, 1)
+            x = (x * pooling_mask).sum(dim=0, keepdim=True)
+        else:
+            num_frames = x.shape[0]
+            pooling_mask = 1.0 / num_frames
+            x = (x * pooling_mask).sum(dim=0, keepdim=True)
+
+        x = self.proj(x)
+        return x
+
+
+class FeedforwardModule(nn.Module):
+    """Feedforward module in Zipformer model."""
+
+    def __init__(self, d_model: int, feedforward_dim: int, dropout: float):
+        super(FeedforwardModule, self).__init__()
+        self.in_proj = nn.Linear(d_model, feedforward_dim)
+        self.balancer = ActivationBalancer(
+            feedforward_dim, channel_dim=-1, max_abs=10.0, min_prob=0.25
+        )
+        self.activation = DoubleSwish()
+        self.dropout = nn.Dropout(dropout)
+        self.out_proj = ScaledLinear(feedforward_dim, d_model, initial_scale=0.01)
+
+    def forward(self, x: Tensor):
+        x = self.in_proj(x)
+        x = self.balancer(x)
+        x = self.activation(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Zipformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0, kernel_size
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            2 * channels,
+            channel_dim=1,
+            max_abs=10.0,
+            min_positive=0.05,
+            max_positive=1.0,
+        )
+
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channels,
+            channel_dim=1,
+            min_positive=0.05,
+            max_positive=1.0,
+            max_abs=20.0,
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.05,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+           src_key_padding_mask: the mask for the src keys per batch (optional):
+               (batch, #time), contains bool in masked positions.
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = (T-3)//2 - 2 == (T-7)//2
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        dropout: float = 0.1,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, (T-7)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer2_channels:
+            Number of channels in layer2
+          layer3_channels:
+            Number of channels in layer3
+        """
+        assert in_channels >= 7, in_channels
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=(0, 1),  # (time, freq)
+            ),
+            ActivationBalancer(layer1_channels, channel_dim=1),
+            DoubleSwish(),
+            nn.Conv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+                padding=0,
+            ),
+            ActivationBalancer(layer2_channels, channel_dim=1),
+            DoubleSwish(),
+            nn.Conv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=(1, 2),  # (time, freq)
+            ),
+            ActivationBalancer(layer3_channels, channel_dim=1),
+            DoubleSwish(),
+        )
+        out_height = (((in_channels - 1) // 2) - 1) // 2
+        self.out = ScaledLinear(out_height * layer3_channels, out_channels)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, (T-7)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, (T-7)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).reshape(b, t, c * f))
+        # Now x is of shape (N, (T-7)//2, odim)
+        x = self.dropout(x)
+        return x
+
+
+class AttentionCombine(nn.Module):
+    """
+    This module combines a list of Tensors, all with the same shape, to
+    produce a single output of that same shape which, in training time,
+    is a random combination of all the inputs; but which in test time
+    will be just the last input.
+
+    All but the last input will have a linear transform before we
+    randomly combine them; these linear transforms will be initialized
+    to the identity transform.
+
+    The idea is that the list of Tensors will be a list of outputs of multiple
+    zipformer layers.  This has a similar effect as iterated loss. (See:
+    DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER
+    NETWORKS).
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        num_inputs: int,
+        random_prob: float = 0.25,
+        single_prob: float = 0.333,
+    ) -> None:
+        """
+        Args:
+          num_channels:
+            the number of channels
+          num_inputs:
+            The number of tensor inputs, which equals the number of layers'
+            outputs that are fed into this module.  E.g. in an 18-layer neural
+            net if we output layers 16, 12, 18, num_inputs would be 3.
+          random_prob:
+            the probability with which we apply a nontrivial mask, in training
+            mode.
+         single_prob:
+            the probability with which we mask to allow just a single
+            module's output (in training)
+        """
+        super().__init__()
+
+        self.random_prob = random_prob
+        self.single_prob = single_prob
+        self.weight = torch.nn.Parameter(torch.zeros(num_channels, num_inputs))
+        self.bias = torch.nn.Parameter(torch.zeros(num_inputs))
+
+        assert 0 <= random_prob <= 1, random_prob
+        assert 0 <= single_prob <= 1, single_prob
+
+    def forward(self, inputs: List[Tensor]) -> Tensor:
+        """Forward function.
+        Args:
+          inputs:
+            A list of Tensor, e.g. from various layers of a transformer.
+            All must be the same shape, of (*, num_channels)
+        Returns:
+          A Tensor of shape (*, num_channels).  In test mode
+          this is just the final input.
+        """
+        num_inputs = self.weight.shape[1]
+        assert len(inputs) == num_inputs
+
+        # Shape of weights: (*, num_inputs)
+        num_channels = inputs[0].shape[-1]
+        num_frames = inputs[0].numel() // num_channels
+
+        ndim = inputs[0].ndim
+        # stacked_inputs: (num_frames, num_channels, num_inputs)
+        stacked_inputs = torch.stack(inputs, dim=ndim).reshape(
+            (num_frames, num_channels, num_inputs)
+        )
+
+        scores = (stacked_inputs * self.weight).sum(dim=(1,)) + self.bias
+
+        if random.random() < 0.002:
+            logging.info(f"Average scores are {scores.softmax(dim=1).mean(dim=0)}")
+
+        if self.training:
+            # random masking..
+            mask_start = torch.randint(
+                low=1,
+                high=int(num_inputs / self.random_prob),
+                size=(num_frames,),
+                device=scores.device,
+            ).unsqueeze(1)
+            # mask will have rows like: [ False, False, False, True, True, .. ]
+            arange = (
+                torch.arange(num_inputs, device=scores.device)
+                .unsqueeze(0)
+                .expand(num_frames, num_inputs)
+            )
+            mask = arange >= mask_start
+
+            apply_single_prob = torch.logical_and(
+                torch.rand(size=(num_frames, 1), device=scores.device)
+                < self.single_prob,
+                mask_start < num_inputs,
+            )
+            single_prob_mask = torch.logical_and(
+                apply_single_prob, arange < mask_start - 1
+            )
+
+            mask = torch.logical_or(mask, single_prob_mask)
+
+            scores = scores.masked_fill(mask, float("-inf"))
+
+        if self.training and random.random() < 0.1:
+            scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04)
+
+        weights = scores.softmax(dim=1)
+
+        # (num_frames, num_channels, num_inputs) * (num_frames, num_inputs, 1) -> (num_frames, num_channels, 1),
+        ans = torch.matmul(stacked_inputs, weights.unsqueeze(2))
+        # ans: (*, num_channels)
+        ans = ans.reshape(*tuple(inputs[0].shape[:-1]), num_channels)
+
+        if __name__ == "__main__":
+            # for testing only...
+            print("Weights = ", weights.reshape(num_frames, num_inputs))
+        return ans
+
+
+def _test_random_combine():
+    print("_test_random_combine()")
+    num_inputs = 3
+    num_channels = 50
+    m = AttentionCombine(
+        num_channels=num_channels,
+        num_inputs=num_inputs,
+        random_prob=0.5,
+        single_prob=0.0,
+    )
+
+    x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)]
+
+    y = m(x)
+    assert y.shape == x[0].shape
+    assert torch.allclose(y, x[0])  # .. since actually all ones.
+
+
+def _test_zipformer_main():
+    feature_dim = 50
+    batch_size = 5
+    seq_len = 20
+    feature_dim = 50
+    # Just make sure the forward pass runs.
+
+    c = Zipformer(
+        num_features=feature_dim,
+        encoder_dims=(64, 96),
+        encoder_unmasked_dims=(48, 64),
+        nhead=(4, 4),
+    )
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    assert ((seq_len - 7) // 2 + 1) // 2 == f[0].shape[1], (seq_len, f.shape[1])
+    f[0].sum().backward()
+    c.eval()
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+def _test_conv2d_subsampling():
+    num_features = 80
+    encoder_dims = 384
+    dropout = 0.1
+    encoder_embed = Conv2dSubsampling(num_features, encoder_dims, dropout=dropout)
+    for i in range(20, 40):
+        x = torch.rand(2, i, num_features)
+        y = encoder_embed(x)
+        assert (x.shape[1] - 7) // 2 == y.shape[1], (x.shape[1], y.shape[1])
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_random_combine()
+    _test_zipformer_main()
+    _test_conv2d_subsampling()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/ctc_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/ctc_decode.py
new file mode 100755
index 000000000..629bec058
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/ctc_decode.py
@@ -0,0 +1,812 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) ctc-decoding
+./pruned_transducer_stateless7_ctc/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) 1best
+./pruned_transducer_stateless7_ctc/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --decoding-method 1best
+
+(3) nbest
+./pruned_transducer_stateless7_ctc/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --decoding-method nbest
+
+(4) nbest-rescoring
+./pruned_transducer_stateless7_ctc/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+
+(5) whole-lattice-rescoring
+./pruned_transducer_stateless7_ctc/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.8,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(feature, feature_lens)
+    nnet_output = model.ctc_output(encoder_out)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decode.py
new file mode 100755
index 000000000..7641fa5af
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decode.py
@@ -0,0 +1,835 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_ctc/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/export.py
new file mode 100755
index 000000000..51e62d6a8
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/export.py
@@ -0,0 +1,322 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7_ctc/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless7_ctc/decode.py \
+        --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained.py
new file mode 100755
index 000000000..280b95984
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained.py
@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7_ctc/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py
new file mode 100755
index 000000000..d50d231d5
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py
@@ -0,0 +1,426 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+(1) ctc-decoding
+./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(3) nbest-rescoring
+./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.model_filename)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+    nnet_output = model.ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/model.py
new file mode 100644
index 000000000..a6e919e2f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/model.py
@@ -0,0 +1,198 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = nn.Linear(
+            encoder_dim,
+            vocab_size,
+        )
+        self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
+
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return a tuple containing simple loss, pruned loss, and ctc-output.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(encoder_out)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss, ctc_output)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained.py
new file mode 100755
index 000000000..78e0fa778
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained.py
@@ -0,0 +1,360 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_ctc/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7_ctc/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7_ctc/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7_ctc/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7_ctc/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7_ctc/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_ctc/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained_ctc.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained_ctc.py
new file mode 100755
index 000000000..904c1deae
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained_ctc.py
@@ -0,0 +1,444 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) ctc-decoding
+./pruned_transducer_stateless7_ctc/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./pruned_transducer_stateless7_ctc/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./bruned_transducer_stateless7_ctc/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./pruned_transducer_stateless7_ctc/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+    params.vocab_size = params.num_classes
+    params.blank_id = 0
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+    nnet_output = model.ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/test_model.py
new file mode 100755
index 000000000..e482d2040
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/test_model.py
@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless7_ctc/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    #  params.feedforward_dims = "1024,1024,1536,1536,1024"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    test_model_1()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py
new file mode 100755
index 000000000..d8fa08372
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py
@@ -0,0 +1,1276 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_ctc/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_ctc/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_ctc/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_ctc/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            # parameters for ctc loss
+            "beam_size": 10,
+            "use_double_scores": True,
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    token_ids = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(token_ids).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_output = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    # Compute ctc loss
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        supervision_segments, token_ids = encode_supervisions(
+            supervisions,
+            subsampling_factor=params.subsampling_factor,
+            token_ids=token_ids,
+        )
+
+    # Works with a BPE model
+    decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device)
+    dense_fsa_vec = k2.DenseFsaVec(
+        ctc_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction="sum",
+        use_double_scores=params.use_double_scores,
+    )
+    assert ctc_loss.requires_grad == is_training
+    loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/zipformer.py
new file mode 120000
index 000000000..79b076556
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/zipformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_decode.py
new file mode 100755
index 000000000..fa7144f0f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_decode.py
@@ -0,0 +1,803 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) ctc-decoding
+./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+(2) 1best
+./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --decoding-method 1best
+(3) nbest
+./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --decoding-method nbest
+(4) nbest-rescoring
+./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+(5) whole-lattice-rescoring
+./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --hlg-scale 0.8 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc_bs/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.8,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(feature, feature_lens)
+    nnet_output = model.ctc_output(encoder_out)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py
new file mode 100755
index 000000000..01ba7b711
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py
@@ -0,0 +1,857 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Yifan   Yang,)
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_ctc/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc_bs/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    # filter out blank frames using ctc outputs
+    ctc_output = model.ctc_output(encoder_out)
+    encoder_out = model.lconv(
+        x=encoder_out,
+        src_key_padding_mask=make_pad_mask(encoder_out_lens),
+    )
+    encoder_out, encoder_out_lens = model.frame_reducer(
+        x=encoder_out,
+        x_lens=encoder_out_lens,
+        ctc_output=ctc_output,
+        blank_id=0,
+    )
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/decode.py
new file mode 100755
index 000000000..e497787d3
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/decode.py
@@ -0,0 +1,835 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_ctc_bs/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc_bs/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export.py
new file mode 100755
index 000000000..9f35cf63e
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export.py
@@ -0,0 +1,321 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 13 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 13
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7_ctc_bs/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless7_ctc_bs/decode.py \
+        --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export_onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export_onnx.py
new file mode 100755
index 000000000..d3033b888
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export_onnx.py
@@ -0,0 +1,899 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                            Yifan   Yang)
+#           2023 NVIDIA Corporation (Author: Wen     Ding)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to ONNX format
+
+./pruned_transducer_stateless7_ctc_bs/export_onnx.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 13 \
+  --onnx 1
+
+It will generate the following files in the given `exp_dir`.
+Check `onnx_check.py` for how to use them.
+
+    - encoder.onnx
+    - decoder.onnx
+    - joiner.onnx
+    - joiner_encoder_proj.onnx
+    - joiner_decoder_proj.onnx
+    - lconv.onnx
+    - frame_reducer.onnx
+    - ctc_output.onnx
+
+(2) Export to ONNX format which can be used in Triton Server
+./pruned_transducer_stateless7_ctc_bs/export_onnx.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 13 \
+  --onnx-triton 1
+
+It will generate the following files in the given `exp_dir`.
+
+    - encoder.onnx
+    - decoder.onnx
+    - joiner.onnx
+    - joiner_encoder_proj.onnx
+    - joiner_decoder_proj.onnx
+    - lconv.onnx
+    - ctc_output.onnx
+
+Please see ./onnx_pretrained.py for usage of the generated files
+
+Check
+https://github.com/k2-fsa/sherpa-onnx
+for how to use the exported models outside of icefall.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from onnx_wrapper import TritonOnnxDecoder, TritonOnnxJoiner, TritonOnnxLconv
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc_bs/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--onnx",
+        type=str2bool,
+        default=False,
+        help="""If True, --jit is ignored and it exports the model
+        to onnx format.
+        It will generate the following files:
+
+            - encoder.onnx
+            - decoder.onnx
+            - joiner.onnx
+            - joiner_encoder_proj.onnx
+            - joiner_decoder_proj.onnx
+            - lconv.onnx
+            - frame_reducer.onnx
+            - ctc_output.onnx
+
+        Refer to ./onnx_check.py and ./onnx_pretrained.py for how to use them.
+        """,
+    )
+    parser.add_argument(
+        "--onnx-triton",
+        type=str2bool,
+        default=False,
+        help="""If True, and it exports the model
+        to onnx format which can be used in NVIDIA triton server. 
+        It will generate the following files:
+
+            - encoder.onnx
+            - decoder.onnx
+            - joiner.onnx
+            - joiner_encoder_proj.onnx
+            - joiner_decoder_proj.onnx
+            - lconv.onnx
+            - ctc_output.onnx
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_onnx(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T, C)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(15, 2000, 80, dtype=torch.float32)
+    x_lens = torch.tensor([2000] * 15, dtype=torch.int64)
+
+    #  encoder_model = torch.jit.script(encoder_model)
+    # It throws the following error for the above statement
+    #
+    # RuntimeError: Exporting the operator __is_ to ONNX opset version
+    # 11 is not supported. Please feel free to request support or
+    # submit a pull request on PyTorch GitHub.
+    #
+    # I cannot find which statement causes the above error.
+    # torch.onnx.export() will use torch.jit.trace() internally, which
+    # works well for the current reworked model
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_onnx(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, 1, C)
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    y = torch.zeros(15, decoder_model.context_size, dtype=torch.int64)
+    need_pad = False  # Always False, so we can use torch.jit.trace() here
+    # Note(fangjun): torch.jit.trace() is more efficient than torch.jit.script()
+    # in this case
+    torch.onnx.export(
+        decoder_model,
+        (y, need_pad),
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y", "need_pad"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_decoder_model_onnx_triton(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX-Triton format.
+    The exported model has one input:
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+    and has one output:
+        - decoder_out: a torch.float32 tensor of shape (N, 1, C)
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+
+    decoder_model = TritonOnnxDecoder(decoder_model)
+
+    torch.onnx.export(
+        decoder_model,
+        (y),
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+
+    The exported encoder_proj model has one input:
+
+        - encoder_out: a tensor of shape (N, encoder_out_dim)
+
+    and produces one output:
+
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+
+    The exported decoder_proj model has one input:
+
+        - decoder_out: a tensor of shape (N, decoder_out_dim)
+
+    and produces one output:
+
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+    """
+    encoder_proj_filename = str(joiner_filename).replace(".onnx", "_encoder_proj.onnx")
+    decoder_proj_filename = str(joiner_filename).replace(".onnx", "_decoder_proj.onnx")
+
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    joiner_dim = joiner_model.decoder_proj.weight.shape[0]
+
+    projected_encoder_out = torch.rand(1, 1, 1, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(1, 1, 1, joiner_dim, dtype=torch.float32)
+
+    project_input = False
+    # Note: It uses torch.jit.trace() internally
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out, project_input),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+            "project_input",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {joiner_filename}")
+
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.encoder_proj,
+        encoder_out,
+        encoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["encoder_out"],
+        output_names=["projected_encoder_out"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "projected_encoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_proj_filename}")
+
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.decoder_proj,
+        decoder_out,
+        decoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["decoder_out"],
+        output_names=["projected_decoder_out"],
+        dynamic_axes={
+            "decoder_out": {0: "N"},
+            "projected_decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_proj_filename}")
+
+
+def export_joiner_model_onnx_triton(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+    and produces one output:
+        - logit: a tensor of shape (N, vocab_size)
+    The exported encoder_proj model has one input:
+        - encoder_out: a tensor of shape (N, encoder_out_dim)
+    and produces one output:
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+    The exported decoder_proj model has one input:
+        - decoder_out: a tensor of shape (N, decoder_out_dim)
+    and produces one output:
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+    """
+    encoder_proj_filename = str(joiner_filename).replace(".onnx", "_encoder_proj.onnx")
+    decoder_proj_filename = str(joiner_filename).replace(".onnx", "_decoder_proj.onnx")
+
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    joiner_dim = joiner_model.decoder_proj.weight.shape[0]
+
+    projected_encoder_out = torch.rand(1, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(1, joiner_dim, dtype=torch.float32)
+
+    # Note: It uses torch.jit.trace() internally
+    joiner_model = TritonOnnxJoiner(joiner_model)
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+            "project_input",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {joiner_filename}")
+
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.encoder_proj,
+        encoder_out,
+        encoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["encoder_out"],
+        output_names=["projected_encoder_out"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "projected_encoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_proj_filename}")
+
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.decoder_proj,
+        decoder_out,
+        decoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["decoder_out"],
+        output_names=["projected_decoder_out"],
+        dynamic_axes={
+            "decoder_out": {0: "N"},
+            "projected_decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_proj_filename}")
+
+
+def export_lconv_onnx(
+    lconv: nn.Module,
+    lconv_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the lconv to ONNX format.
+
+    The exported lconv has two inputs:
+
+        - lconv_input: a tensor of shape (N, T, C)
+        - src_key_padding_mask: a tensor of shape (N, T)
+
+    and has one output:
+
+        - lconv_out: a tensor of shape (N, T, C)
+
+    Args:
+      lconv:
+        The lconv to be exported.
+      lconv_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    lconv_input = torch.zeros(15, 498, 384, dtype=torch.float32)
+    src_key_padding_mask = torch.zeros(15, 498, dtype=torch.bool)
+
+    torch.onnx.export(
+        lconv,
+        (lconv_input, src_key_padding_mask),
+        lconv_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["lconv_input", "src_key_padding_mask"],
+        output_names=["lconv_out"],
+        dynamic_axes={
+            "lconv_input": {0: "N", 1: "T"},
+            "src_key_padding_mask": {0: "N", 1: "T"},
+            "lconv_out": {0: "N", 1: "T"},
+        },
+    )
+    logging.info(f"Saved to {lconv_filename}")
+
+
+def export_lconv_onnx_triton(
+    lconv: nn.Module,
+    lconv_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the lconv to ONNX format.
+
+    The exported lconv has two inputs:
+
+        - lconv_input: a tensor of shape (N, T, C)
+        - lconv_input_lens: a tensor of shape (N, )
+
+    and has one output:
+
+        - lconv_out: a tensor of shape (N, T, C)
+
+    Args:
+      lconv:
+        The lconv to be exported.
+      lconv_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    lconv_input = torch.zeros(15, 498, 384, dtype=torch.float32)
+    lconv_input_lens = torch.tensor([498] * 15, dtype=torch.int64)
+
+    lconv = TritonOnnxLconv(lconv)
+
+    torch.onnx.export(
+        lconv,
+        (lconv_input, lconv_input_lens),
+        lconv_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["lconv_input", "lconv_input_lens"],
+        output_names=["lconv_out"],
+        dynamic_axes={
+            "lconv_input": {0: "N", 1: "T"},
+            "lconv_input_lens": {0: "N"},
+            "lconv_out": {0: "N", 1: "T"},
+        },
+    )
+    logging.info(f"Saved to {lconv_filename}")
+
+
+def export_frame_reducer_onnx(
+    frame_reducer: nn.Module,
+    frame_reducer_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the frame_reducer to ONNX format.
+
+    The exported frame_reducer has four inputs:
+
+        - x: a tensor of shape (N, T, C)
+        - x_lens: a tensor of shape (N, T)
+        - ctc_output: a tensor of shape (N, T, vocab_size)
+        - blank_id: an int, always 0
+
+    and has two outputs:
+
+        - x_fr: a tensor of shape (N, T, C)
+        - x_lens_fr: a tensor of shape (N, T)
+
+    Args:
+      frame_reducer:
+        The frame_reducer to be exported.
+      frame_reducer_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(15, 498, 384, dtype=torch.float32)
+    x_lens = torch.tensor([498] * 15, dtype=torch.int64)
+    ctc_output = torch.randn(15, 498, 500, dtype=torch.float32)
+
+    torch.onnx.export(
+        frame_reducer,
+        (x, x_lens, ctc_output),
+        frame_reducer_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens", "ctc_output"],
+        output_names=["out", "out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "ctc_output": {0: "N", 1: "T"},
+            "out": {0: "N", 1: "T"},
+            "out_lens": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {frame_reducer_filename}")
+
+
+def export_ctc_output_onnx(
+    ctc_output: nn.Module,
+    ctc_output_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the frame_reducer to ONNX format.
+
+    The exported frame_reducer has one inputs:
+
+        - encoder_out: a tensor of shape (N, T, C)
+
+    and has one output:
+
+        - ctc_output: a tensor of shape (N, T, vocab_size)
+
+    Args:
+      ctc_output:
+        The ctc_output to be exported.
+      ctc_output_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    encoder_out = torch.zeros(15, 498, 384, dtype=torch.float32)
+
+    torch.onnx.export(
+        ctc_output,
+        (encoder_out),
+        ctc_output_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["encoder_out"],
+        output_names=["ctc_output"],
+        dynamic_axes={
+            "encoder_out": {0: "N", 1: "T"},
+            "ctc_output": {0: "N", 1: "T"},
+        },
+    )
+    logging.info(f"Saved to {ctc_output_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+    opset_version = 13
+    logging.info("Exporting to onnx format")
+    encoder_filename = params.exp_dir / "encoder.onnx"
+    export_encoder_model_onnx(
+        model.encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+
+    decoder_filename = params.exp_dir / "decoder.onnx"
+    if params.onnx is True:
+        export_decoder_model_onnx(
+            model.decoder,
+            decoder_filename,
+            opset_version=opset_version,
+        )
+    elif params.onnx_triton is True:
+        export_decoder_model_onnx_triton(
+            model.decoder,
+            decoder_filename,
+            opset_version=opset_version,
+        )
+
+    joiner_filename = params.exp_dir / "joiner.onnx"
+    if params.onnx is True:
+        export_joiner_model_onnx(
+            model.joiner,
+            joiner_filename,
+            opset_version=opset_version,
+        )
+    elif params.onnx_triton is True:
+        export_joiner_model_onnx_triton(
+            model.joiner,
+            joiner_filename,
+            opset_version=opset_version,
+        )
+
+    lconv_filename = params.exp_dir / "lconv.onnx"
+    if params.onnx is True:
+        export_lconv_onnx(
+            model.lconv,
+            lconv_filename,
+            opset_version=opset_version,
+        )
+    elif params.onnx_triton is True:
+        export_lconv_onnx_triton(
+            model.lconv,
+            lconv_filename,
+            opset_version=opset_version,
+        )
+
+    if params.onnx is True:
+        frame_reducer_filename = params.exp_dir / "frame_reducer.onnx"
+        export_frame_reducer_onnx(
+            model.frame_reducer,
+            frame_reducer_filename,
+            opset_version=opset_version,
+        )
+
+    ctc_output_filename = params.exp_dir / "ctc_output.onnx"
+    export_ctc_output_onnx(
+        model.ctc_output,
+        ctc_output_filename,
+        opset_version=opset_version,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
new file mode 100644
index 000000000..c44cb1eaf
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
@@ -0,0 +1,166 @@
+#!/usr/bin/env python3
+#
+# Copyright      2022  Xiaomi Corp.        (authors: Yifan   Yang,
+#                                                    Zengwei Yao,
+#                                                    Wei     Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from icefall.utils import make_pad_mask
+
+
+class FrameReducer(nn.Module):
+    """The encoder output is first used to calculate
+    the CTC posterior probability; then for each output frame,
+    if its blank posterior is bigger than some thresholds,
+    it will be simply discarded from the encoder output.
+    """
+
+    def __init__(
+        self,
+    ):
+        super().__init__()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        ctc_output: torch.Tensor,
+        y_lens: Optional[torch.Tensor] = None,
+        blank_id: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            x:
+              The shared encoder output with shape [N, T, C].
+            x_lens:
+              A tensor of shape (batch_size,) containing the number of frames in
+              `x` before padding.
+            ctc_output:
+              The CTC output with shape [N, T, vocab_size].
+            y_lens:
+              A tensor of shape (batch_size,) containing the number of frames in
+              `y` before padding.
+            blank_id:
+              The blank id of ctc_output.
+        Returns:
+            out:
+              The frame reduced encoder output with shape [N, T', C].
+            out_lens:
+              A tensor of shape (batch_size,) containing the number of frames in
+              `out` before padding.
+        """
+        N, T, C = x.size()
+
+        padding_mask = make_pad_mask(x_lens)
+        non_blank_mask = (ctc_output[:, :, blank_id] < math.log(0.9)) * (~padding_mask)
+
+        if y_lens is not None:
+            # Limit the maximum number of reduced frames
+            limit_lens = T - y_lens
+            max_limit_len = limit_lens.max().int()
+            fake_limit_indexes = torch.topk(
+                ctc_output[:, :, blank_id], max_limit_len
+            ).indices
+            T_arange = (
+                torch.arange(max_limit_len)
+                .expand_as(
+                    fake_limit_indexes,
+                )
+                .to(device=x.device)
+            )
+            T_arange = torch.remainder(T_arange, limit_lens.unsqueeze(1))
+            limit_indexes = torch.gather(fake_limit_indexes, 1, T_arange)
+            limit_mask = torch.full_like(
+                non_blank_mask,
+                0,
+                device=x.device,
+            ).scatter_(1, limit_indexes, 1)
+
+            non_blank_mask = non_blank_mask | ~limit_mask
+
+        out_lens = non_blank_mask.sum(dim=1)
+        max_len = out_lens.max()
+        pad_lens_list = (
+            torch.full_like(
+                out_lens,
+                max_len.item(),
+                device=x.device,
+            )
+            - out_lens
+        )
+        max_pad_len = int(pad_lens_list.max())
+
+        out = F.pad(x, [0, 0, 0, max_pad_len])
+
+        valid_pad_mask = ~make_pad_mask(pad_lens_list)
+        total_valid_mask = torch.concat([non_blank_mask, valid_pad_mask], dim=1)
+
+        out = out[total_valid_mask].reshape(N, -1, C)
+
+        return out, out_lens
+
+
+if __name__ == "__main__":
+    import time
+
+    test_times = 10000
+    device = "cuda:0"
+    frame_reducer = FrameReducer()
+
+    # non zero case
+    x = torch.ones(15, 498, 384, dtype=torch.float32, device=device)
+    x_lens = torch.tensor([498] * 15, dtype=torch.int64, device=device)
+    y_lens = torch.tensor([150] * 15, dtype=torch.int64, device=device)
+    ctc_output = torch.log(
+        torch.randn(15, 498, 500, dtype=torch.float32, device=device),
+    )
+
+    avg_time = 0
+    for i in range(test_times):
+        torch.cuda.synchronize(device=x.device)
+        delta_time = time.time()
+        x_fr, x_lens_fr = frame_reducer(x, x_lens, ctc_output, y_lens)
+        torch.cuda.synchronize(device=x.device)
+        delta_time = time.time() - delta_time
+        avg_time += delta_time
+    print(x_fr.shape)
+    print(x_lens_fr)
+    print(avg_time / test_times)
+
+    # all zero case
+    x = torch.zeros(15, 498, 384, dtype=torch.float32, device=device)
+    x_lens = torch.tensor([498] * 15, dtype=torch.int64, device=device)
+    y_lens = torch.tensor([150] * 15, dtype=torch.int64, device=device)
+    ctc_output = torch.zeros(15, 498, 500, dtype=torch.float32, device=device)
+
+    avg_time = 0
+    for i in range(test_times):
+        torch.cuda.synchronize(device=x.device)
+        delta_time = time.time()
+        x_fr, x_lens_fr = frame_reducer(x, x_lens, ctc_output, y_lens)
+        torch.cuda.synchronize(device=x.device)
+        delta_time = time.time() - delta_time
+        avg_time += delta_time
+    print(x_fr.shape)
+    print(x_lens_fr)
+    print(avg_time / test_times)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/jit_pretrained.py
new file mode 100755
index 000000000..da2c6a39a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/jit_pretrained.py
@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py
new file mode 100755
index 000000000..653c25e06
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py
@@ -0,0 +1,426 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 13 \
+  --jit 1
+
+Usage of this script:
+
+(1) ctc-decoding
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(3) nbest-rescoring
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+  --model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.model_filename)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+    nnet_output = model.ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py
new file mode 100644
index 000000000..a902358ae
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py
@@ -0,0 +1,114 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from scaling import (
+    ActivationBalancer,
+    ScaledConv1d,
+)
+
+
+class LConv(nn.Module):
+    """A convolution module to prevent information loss."""
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int = 7,
+        bias: bool = True,
+    ):
+        """
+        Args:
+          channels:
+            Dimension of the input embedding, and of the lconv output.
+        """
+        super().__init__()
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        self.deriv_balancer1 = ActivationBalancer(
+            2 * channels,
+            channel_dim=1,
+            max_abs=10.0,
+            min_positive=0.05,
+            max_positive=1.0,
+        )
+
+        self.depthwise_conv = nn.Conv1d(
+            2 * channels,
+            2 * channels,
+            kernel_size=kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=2 * channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            2 * channels,
+            channel_dim=1,
+            min_positive=0.05,
+            max_positive=1.0,
+            max_abs=20.0,
+        )
+
+        self.pointwise_conv2 = ScaledConv1d(
+            2 * channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.05,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            x: A 3-D tensor of shape (N, T, C).
+        Returns:
+            Return a tensor of shape (N, T, C).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(0, 2, 1)  # (#batch, channels, time).
+
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channels, time)
+
+        return x.permute(0, 2, 1)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/model.py
new file mode 100644
index 000000000..0582b289f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/model.py
@@ -0,0 +1,226 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos, make_pad_mask
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        lconv: nn.Module,
+        frame_reducer: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+        self.lconv = lconv
+        self.frame_reducer = frame_reducer
+
+        self.simple_am_proj = nn.Linear(
+            encoder_dim,
+            vocab_size,
+        )
+        self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
+
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A floating point value which decides whether to do blank skip.
+        Returns:
+          Return a tuple containing simple loss, pruned loss, and ctc-output.
+        Note:
+          Regarding am_scale & lm_scale, it will make the loss-function one of
+          the form:
+            lm_scale * lm_probs + am_scale * am_probs +
+            (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(encoder_out)
+
+        # y_lens
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        # blank skip
+        blank_id = self.decoder.blank_id
+
+        if warmup >= 2.0:
+            # lconv
+            encoder_out = self.lconv(
+                x=encoder_out,
+                src_key_padding_mask=make_pad_mask(x_lens),
+            )
+
+            # frame reduce
+            encoder_out_fr, x_lens_fr = self.frame_reducer(
+                encoder_out,
+                x_lens,
+                ctc_output,
+                y_lens,
+                blank_id,
+            )
+        else:
+            encoder_out_fr = encoder_out
+            x_lens_fr = x_lens
+
+        # sos_y
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens_fr
+
+        am = self.simple_am_proj(encoder_out_fr)
+        lm = self.simple_lm_proj(decoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out_fr),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss, ctc_output)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py
new file mode 100755
index 000000000..494a34d97
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py
@@ -0,0 +1,469 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Yifan   Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc_bs/export_onnx.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 13
+
+Usage of this script:
+
+./pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/encoder.onnx \
+  --decoder-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/decoder.onnx \
+  --joiner-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/joiner_decoder_proj.onnx \
+  --lconv-filename ./pruned_transducer_stateless7_ctc_bs/exp/lconv.onnx \
+  --frame-reducer-filename ./pruned_transducer_stateless7_ctc_bs/exp/frame_reducer.onnx \
+  --ctc-output-filename ./pruned_transducer_stateless7_ctc_bs/exp/ctc_output.onnx \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import numpy as np
+import onnxruntime as ort
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence, pack_padded_sequence
+
+from icefall.utils import make_pad_mask
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-encoder-proj-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner encoder_proj onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-decoder-proj-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner decoder_proj onnx model. ",
+    )
+
+    parser.add_argument(
+        "--lconv-filename",
+        type=str,
+        required=True,
+        help="Path to the lconv onnx model. ",
+    )
+
+    parser.add_argument(
+        "--frame-reducer-filename",
+        type=str,
+        required=True,
+        help="Path to the frame reducer onnx model. ",
+    )
+
+    parser.add_argument(
+        "--ctc-output-filename",
+        type=str,
+        required=True,
+        help="Path to the ctc_output onnx model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: ort.InferenceSession,
+    joiner: ort.InferenceSession,
+    joiner_encoder_proj: ort.InferenceSession,
+    joiner_decoder_proj: ort.InferenceSession,
+    encoder_out: np.ndarray,
+    encoder_out_lens: np.ndarray,
+    context_size: int,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      decoder:
+        The decoder model.
+      joiner:
+        The joiner model.
+      joiner_encoder_proj:
+        The joiner encoder projection model.
+      joiner_decoder_proj:
+        The joiner decoder projection model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+      context_size:
+        The context size of the decoder model.
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    encoder_out = torch.from_numpy(encoder_out)
+    encoder_out_lens = torch.from_numpy(encoder_out_lens)
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    projected_encoder_out = joiner_encoder_proj.run(
+        [joiner_encoder_proj.get_outputs()[0].name],
+        {joiner_encoder_proj.get_inputs()[0].name: packed_encoder_out.data.numpy()},
+    )[0]
+
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input_nodes = decoder.get_inputs()
+    decoder_output_nodes = decoder.get_outputs()
+
+    joiner_input_nodes = joiner.get_inputs()
+    joiner_output_nodes = joiner.get_outputs()
+
+    decoder_input = torch.tensor(
+        hyps,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = decoder.run(
+        [decoder_output_nodes[0].name],
+        {
+            decoder_input_nodes[0].name: decoder_input.numpy(),
+        },
+    )[0].squeeze(1)
+    projected_decoder_out = joiner_decoder_proj.run(
+        [joiner_decoder_proj.get_outputs()[0].name],
+        {joiner_decoder_proj.get_inputs()[0].name: decoder_out},
+    )[0]
+
+    projected_decoder_out = torch.from_numpy(projected_decoder_out)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = projected_encoder_out[start:end]
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        projected_decoder_out = projected_decoder_out[:batch_size]
+
+        logits = joiner.run(
+            [joiner_output_nodes[0].name],
+            {
+                joiner_input_nodes[0].name: np.expand_dims(
+                    np.expand_dims(current_encoder_out, axis=1), axis=1
+                ),
+                joiner_input_nodes[1]
+                .name: projected_decoder_out.unsqueeze(1)
+                .unsqueeze(1)
+                .numpy(),
+            },
+        )[0]
+        logits = torch.from_numpy(logits).squeeze(1).squeeze(1)
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                dtype=torch.int64,
+            )
+            decoder_out = decoder.run(
+                [decoder_output_nodes[0].name],
+                {
+                    decoder_input_nodes[0].name: decoder_input.numpy(),
+                },
+            )[0].squeeze(1)
+            projected_decoder_out = joiner_decoder_proj.run(
+                [joiner_decoder_proj.get_outputs()[0].name],
+                {joiner_decoder_proj.get_inputs()[0].name: decoder_out},
+            )[0]
+            projected_decoder_out = torch.from_numpy(projected_decoder_out)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    session_opts = ort.SessionOptions()
+    session_opts.inter_op_num_threads = 1
+    session_opts.intra_op_num_threads = 1
+
+    encoder = ort.InferenceSession(
+        args.encoder_model_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    decoder = ort.InferenceSession(
+        args.decoder_model_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    joiner = ort.InferenceSession(
+        args.joiner_model_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    joiner_encoder_proj = ort.InferenceSession(
+        args.joiner_encoder_proj_model_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    joiner_decoder_proj = ort.InferenceSession(
+        args.joiner_decoder_proj_model_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    lconv = ort.InferenceSession(
+        args.lconv_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    frame_reducer = ort.InferenceSession(
+        args.frame_reducer_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    ctc_output = ort.InferenceSession(
+        args.ctc_output_filename,
+        sess_options=session_opts,
+        providers=["CPUExecutionProvider"],
+    )
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+
+    encoder_input_nodes = encoder.get_inputs()
+    encoder_out_nodes = encoder.get_outputs()
+    encoder_out, encoder_out_lens = encoder.run(
+        [encoder_out_nodes[0].name, encoder_out_nodes[1].name],
+        {
+            encoder_input_nodes[0].name: features.numpy(),
+            encoder_input_nodes[1].name: feature_lengths.numpy(),
+        },
+    )
+
+    ctc_output_input_nodes = ctc_output.get_inputs()
+    ctc_output_out_nodes = ctc_output.get_outputs()
+    ctc_out = ctc_output.run(
+        [ctc_output_out_nodes[0].name],
+        {
+            ctc_output_input_nodes[0].name: encoder_out,
+        },
+    )[0]
+
+    lconv_input_nodes = lconv.get_inputs()
+    lconv_out_nodes = lconv.get_outputs()
+    encoder_out = lconv.run(
+        [lconv_out_nodes[0].name],
+        {
+            lconv_input_nodes[0].name: encoder_out,
+            lconv_input_nodes[1]
+            .name: make_pad_mask(torch.from_numpy(encoder_out_lens))
+            .numpy(),
+        },
+    )[0]
+
+    frame_reducer_input_nodes = frame_reducer.get_inputs()
+    frame_reducer_out_nodes = frame_reducer.get_outputs()
+    encoder_out_fr, encoder_out_lens_fr = frame_reducer.run(
+        [frame_reducer_out_nodes[0].name, frame_reducer_out_nodes[1].name],
+        {
+            frame_reducer_input_nodes[0].name: encoder_out,
+            frame_reducer_input_nodes[1].name: encoder_out_lens,
+            frame_reducer_input_nodes[2].name: ctc_out,
+        },
+    )
+
+    hyps = greedy_search(
+        decoder=decoder,
+        joiner=joiner,
+        joiner_encoder_proj=joiner_encoder_proj,
+        joiner_decoder_proj=joiner_decoder_proj,
+        encoder_out=encoder_out_fr,
+        encoder_out_lens=encoder_out_lens_fr,
+        context_size=args.context_size,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py
new file mode 100755
index 000000000..247da0949
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py
@@ -0,0 +1,98 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from torch import nn
+from icefall.utils import make_pad_mask
+
+
+class TritonOnnxDecoder(nn.Module):
+    """
+    Triton wrapper for decoder model
+    """
+
+    def __init__(self, model):
+        """
+        Args:
+            model: decoder model
+        """
+        super().__init__()
+
+        self.model = model
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        need_pad = False
+        return self.model(y, need_pad)
+
+
+class TritonOnnxJoiner(nn.Module):
+    def __init__(
+        self,
+        model,
+    ):
+        super().__init__()
+
+        self.model = model
+        self.encoder_proj = model.encoder_proj
+        self.decoder_proj = model.decoder_proj
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, C).
+        Returns:
+          Return a tensor of shape (N, T, C).
+        """
+        project_input = False
+        return self.model(encoder_out, decoder_out, project_input)
+
+
+class TritonOnnxLconv(nn.Module):
+    def __init__(
+        self,
+        model,
+    ):
+        super().__init__()
+
+        self.model = model
+
+    def forward(
+        self,
+        lconv_input: torch.Tensor,
+        lconv_input_lens: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          lconv_input: Its shape is (N, T, C).
+          lconv_input_lens: Its shape is (N, ).
+        Returns:
+          Return a tensor of shape (N, T, C).
+        """
+        mask = make_pad_mask(lconv_input_lens)
+
+        return self.model(x=lconv_input, src_key_padding_mask=mask)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/pretrained.py
new file mode 100755
index 000000000..5d240cf30
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/pretrained.py
@@ -0,0 +1,359 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 13
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7_ctc_bs/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_ctc_bs/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py
new file mode 100755
index 000000000..914107526
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py
@@ -0,0 +1,440 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) ctc-decoding
+./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./bruned_transducer_stateless7_ctc/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py \
+  --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+    params.vocab_size = params.num_classes
+    params.blank_id = 0
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+    nnet_output = model.ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/test_model.py
new file mode 100755
index 000000000..7f0893985
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/test_model.py
@@ -0,0 +1,55 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless7_ctc_bs/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    test_model_1()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py
new file mode 100755
index 000000000..25a1aa674
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py
@@ -0,0 +1,1265 @@
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./pruned_transducer_stateless7_ctc_bs/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+  --full-libri 1 \
+  --max-duration 300
+# For mix precision training:
+./pruned_transducer_stateless7_ctc_bs/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+  --full-libri 1 \
+  --max-duration 750
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from frame_reducer import FrameReducer
+from joiner import Joiner
+from lconv import LConv
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_ctc_bs/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram, 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.5,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+    Explanation of options saved in `params`:
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+        - best_train_epoch: It is the epoch that has the best training loss.
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+        - subsampling_factor:  The subsampling factor for the model.
+        - encoder_dim: Hidden dim for multi-head attention model.
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            # parameters for ctc loss
+            "beam_size": 10,
+            "use_double_scores": True,
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_lconv(params: AttributeDict) -> nn.Module:
+    lconv = LConv(
+        channels=int(params.encoder_dims.split(",")[-1]),
+    )
+    return lconv
+
+
+def get_frame_reducer(params: AttributeDict) -> nn.Module:
+    frame_reducer = FrameReducer()
+    return frame_reducer
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+    lconv = get_lconv(params)
+    frame_reducer = get_frame_reducer(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        lconv=lconv,
+        frame_reducer=frame_reducer,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+    warmup = batch_idx_train / warm_step
+
+    texts = batch["supervisions"]["text"]
+    token_ids = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(token_ids).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_output = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    # Compute ctc loss
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        supervision_segments, token_ids = encode_supervisions(
+            supervisions,
+            subsampling_factor=params.subsampling_factor,
+            token_ids=token_ids,
+        )
+
+    # Works with a BPE model
+    decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device)
+    dense_fsa_vec = k2.DenseFsaVec(
+        ctc_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction="sum",
+        use_double_scores=params.use_double_scores,
+    )
+    assert ctc_loss.requires_grad == is_training
+    loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/zipformer.py
new file mode 120000
index 000000000..79b076556
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/zipformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md
new file mode 100644
index 000000000..0f3c63e75
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md
@@ -0,0 +1,10 @@
+This recipe implements Streaming Zipformer-Transducer model.
+
+See https://k2-fsa.github.io/icefall/recipes/Streaming-ASR/librispeech/zipformer_transducer.html for detailed tutorials.
+
+[./emformer.py](./emformer.py) and [./train.py](./train.py)
+are basically the same as
+[./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py).
+The only purpose of [./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py)
+is for exporting to [sherpa-ncnn](https://github.com/k2-fsa/sherpa-ncnn).
+
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode.py
new file mode 100755
index 000000000..02029c108
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode.py
@@ -0,0 +1,991 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_rescore,
+    modified_beam_search_lm_rescore_LODR,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=2,
+        help="""The order of the ngram lm.
+        """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="ID of the backoff symbol in the ngram LM",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network language model.
+      ngram_lm:
+        A ngram language model
+      ngram_lm_scale:
+        The scale for the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += 30
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, 30),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_rescore":
+        lm_scale_list = [0.01 * i for i in range(10, 50)]
+        ans_dict = modified_beam_search_lm_rescore(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        lm_scale_list = [0.02 * i for i in range(2, 30)]
+        ans_dict = modified_beam_search_lm_rescore_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            LODR_lm=ngram_lm,
+            sp=sp,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    elif params.decoding_method in (
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+    ):
+        ans = dict()
+        assert ans_dict is not None
+        for key, hyps in ans_dict.items():
+            hyps = [sp.decode(hyp).split() for hyp in hyps]
+            ans[f"beam_size_{params.beam_size}_{key}"] = hyps
+        return ans
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      ngram_lm:
+        A n-gram LM to be used for LODR.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_shallow_fusion:
+        params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load the neural network LM if required
+    if params.use_shallow_fusion or params.decoding_method in (
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    ):
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+    else:
+        LM = None
+
+    # only load N-gram LM when needed
+    if params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        try:
+            import kenlm
+        except ImportError:
+            print("Please install kenlm first. You can use")
+            print(" pip install https://github.com/kpu/kenlm/archive/master.zip")
+            print("to install it")
+            import sys
+
+            sys.exit(-1)
+        ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa")
+        logging.info(f"lm filename: {ngram_file_name}")
+        ngram_lm = kenlm.Model(ngram_file_name)
+        ngram_lm_scale = None  # use a list to search
+
+    elif params.decoding_method == "modified_beam_search_LODR":
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"Loading token level lm: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+    import time
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start = time.time()
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+        logging.info(f"Elasped time for {test_set}: {time.time() - start}")
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
new file mode 100644
index 000000000..2c4b144fc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
@@ -0,0 +1,151 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class DecodeStream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        initial_states: List[torch.Tensor],
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+    ) -> None:
+        """
+        Args:
+          initial_states:
+            Initial decode states of the model, e.g. the return value of
+            `get_init_state` in conformer.py
+          decoding_graph:
+            Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+            Used only when decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+        """
+        if params.decoding_method == "fast_beam_search":
+            assert decoding_graph is not None
+            assert device == decoding_graph.device
+
+        self.params = params
+        self.cut_id = cut_id
+        self.LOG_EPS = math.log(1e-10)
+
+        self.states = initial_states
+
+        # It contains a 2-D tensors representing the feature frames.
+        self.features: torch.Tensor = None
+
+        self.num_frames: int = 0
+        # how many frames have been processed. (before subsampling).
+        # we only modify this value in `func:get_feature_frames`.
+        self.num_processed_frames: int = 0
+
+        self._done: bool = False
+
+        # The transcript of current utterance.
+        self.ground_truth: str = ""
+
+        # The decoding result (partial or final) of current utterance.
+        self.hyp: List = []
+
+        # how many frames have been processed, after subsampling (i.e. a
+        # cumulative sum of the second return value of
+        # encoder.streaming_forward
+        self.done_frames: int = 0
+
+        # It has two steps of feature subsampling in zipformer: out_lens=((x_lens-7)//2+1)//2
+        # 1) feature embedding: out_lens=(x_lens-7)//2
+        # 2) output subsampling: out_lens=(out_lens+1)//2
+        self.pad_length = 7
+
+        if params.decoding_method == "greedy_search":
+            self.hyp = [-1] * (params.context_size - 1) + [params.blank_id]
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[-1] * (params.context_size - 1) + [params.blank_id],
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    @property
+    def done(self) -> bool:
+        """Return True if all the features are processed."""
+        return self._done
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    def set_features(
+        self,
+        features: torch.Tensor,
+        tail_pad_len: int = 0,
+    ) -> None:
+        """Set features tensor of current utterance."""
+        assert features.dim() == 2, features.dim()
+        self.features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, self.pad_length + tail_pad_len),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+        self.num_frames = self.features.size(0)
+
+    def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]:
+        """Consume chunk_size frames of features"""
+        chunk_length = chunk_size + self.pad_length
+
+        ret_length = min(self.num_frames - self.num_processed_frames, chunk_length)
+
+        ret_features = self.features[
+            self.num_processed_frames : self.num_processed_frames + ret_length  # noqa
+        ]
+
+        self.num_processed_frames += chunk_size
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_features, ret_length
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.params.decoding_method == "greedy_search":
+            return self.hyp[self.params.context_size :]  # noqa
+        elif self.params.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.params.context_size :]  # noqa
+        else:
+            assert self.params.decoding_method == "fast_beam_search"
+            return self.hyp
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
new file mode 100755
index 000000000..9a6d2155b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
@@ -0,0 +1,1255 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer_for_ncnn_export_only import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+        is_pnnx=True,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
new file mode 100755
index 000000000..a7d06a5dd
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
@@ -0,0 +1,372 @@
+#!/usr/bin/env python3
+
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for more details about how to use this file.
+
+We use
+https://huggingface.co/pfluo/k2fsa-zipformer-chinese-english-mixed
+to demonstrate the usage of this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/pfluo/k2fsa-zipformer-chinese-english-mixed
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char_bpe/L.pt"
+git lfs pull --include "data/lang_char_bpe/L_disambig.pt"
+git lfs pull --include "data/lang_char_bpe/Linv.pt"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export to ncnn
+
+./pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py \
+  --lang-dir $repo/data/lang_char_bpe \
+  --exp-dir $repo/exp \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims "1024,1024,1536,1536,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+cd $repo/exp
+
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+You can find converted models at
+https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-bilingual-zh-en-2023-02-13
+
+See ./streaming-ncnn-decode.py
+and
+https://github.com/k2-fsa/sherpa-ncnn
+for usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+from scaling_converter import convert_scaled_to_non_scaled
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_char/tokens.txt",
+        help="The tokens.txt file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
+
+    decode_chunk_len = encoder_model.decode_chunk_size * 2
+    pad_length = 7
+    T = decode_chunk_len + pad_length  # 32 + 7 = 39
+
+    logging.info(f"decode_chunk_len: {decode_chunk_len}")
+    logging.info(f"T: {T}")
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.get_init_state()
+
+    traced_model = torch.jit.trace(encoder_model, (x, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_pnnx=True)
+
+    encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
new file mode 100755
index 000000000..8f2178b1d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
@@ -0,0 +1,370 @@
+#!/usr/bin/env python3
+
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for more details about how to use this file.
+
+We use
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+to demonstrate the usage of this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export to ncnn
+
+./pruned_transducer_stateless7_streaming/export-for-ncnn.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --exp-dir $repo/exp \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims "1024,1024,2048,2048,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+cd $repo/exp
+
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+You can find converted models at
+https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-en-2023-02-13
+
+See ./streaming-ncnn-decode.py
+and
+https://github.com/k2-fsa/sherpa-ncnn
+for usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+from scaling_converter import convert_scaled_to_non_scaled
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
+
+    decode_chunk_len = encoder_model.decode_chunk_size * 2
+    pad_length = 7
+    T = decode_chunk_len + pad_length  # 32 + 7 = 39
+
+    logging.info(f"decode_chunk_len: {decode_chunk_len}")
+    logging.info(f"T: {T}")
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.get_init_state()
+
+    traced_model = torch.jit.trace(encoder_model, (x, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_pnnx=True)
+
+    encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
new file mode 100755
index 000000000..2de56837e
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
@@ -0,0 +1,684 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/pfluo/k2fsa-zipformer-chinese-english-mixed
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/pfluo/k2fsa-zipformer-chinese-english-mixed
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char_bpe/L.pt"
+git lfs pull --include "data/lang_char_bpe/Linv.pt"
+git lfs pull --include "data/lang_char_bpe/L_disambig.pt"
+git lfs pull --include "exp/pretrained.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7_streaming/export-onnx-zh.py \
+  --tokens $repo/data/lang_char_bpe/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --decode-chunk-len 32 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims "1024,1024,1536,1536,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files in $repo/exp
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py for how to use the exported models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_char/tokens.txt",
+        help="The tokens.txt file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(self, x: Tensor, states: List[Tensor]) -> Tuple[Tensor, List[Tensor]]:
+        """Please see the help information of Zipformer.streaming_forward"""
+        N = x.size(0)
+        T = x.size(1)
+        x_lens = torch.tensor([T] * N, device=x.device)
+
+        output, _, new_states = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=states,
+        )
+
+        output = self.encoder_proj(output)
+        # Now output is of shape (N, T, joiner_dim)
+
+        return output, new_states
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """
+    Onnx model inputs:
+      - 0: src
+      - many state tensors (the exact number depending on the actual model)
+
+    Onnx model outputs:
+      - 0: output, its shape is (N, T, joiner_dim)
+      - many state tensors (the exact number depending on the actual model)
+
+    Args:
+      encoder_model:
+        The model to be exported
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+
+    encoder_model.encoder.__class__.forward = (
+        encoder_model.encoder.__class__.streaming_forward
+    )
+
+    decode_chunk_len = encoder_model.encoder.decode_chunk_size * 2
+    pad_length = 7
+    T = decode_chunk_len + pad_length
+    logging.info(f"decode_chunk_len: {decode_chunk_len}")
+    logging.info(f"pad_length: {pad_length}")
+    logging.info(f"T: {T}")
+
+    x = torch.rand(1, T, 80, dtype=torch.float32)
+
+    init_state = encoder_model.encoder.get_init_state()
+
+    num_encoders = encoder_model.encoder.num_encoders
+    logging.info(f"num_encoders: {num_encoders}")
+    logging.info(f"len(init_state): {len(init_state)}")
+
+    inputs = {}
+    input_names = ["x"]
+
+    outputs = {}
+    output_names = ["encoder_out"]
+
+    def build_inputs_outputs(tensors, name, N):
+        for i, s in enumerate(tensors):
+            logging.info(f"{name}_{i}.shape: {s.shape}")
+            inputs[f"{name}_{i}"] = {N: "N"}
+            outputs[f"new_{name}_{i}"] = {N: "N"}
+            input_names.append(f"{name}_{i}")
+            output_names.append(f"new_{name}_{i}")
+
+    num_encoder_layers = ",".join(map(str, encoder_model.encoder.num_encoder_layers))
+    encoder_dims = ",".join(map(str, encoder_model.encoder.encoder_dims))
+    attention_dims = ",".join(map(str, encoder_model.encoder.attention_dims))
+    cnn_module_kernels = ",".join(map(str, encoder_model.encoder.cnn_module_kernels))
+    ds = encoder_model.encoder.zipformer_downsampling_factors
+    left_context_len = encoder_model.encoder.left_context_len
+    left_context_len = [left_context_len // k for k in ds]
+    left_context_len = ",".join(map(str, left_context_len))
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "decode_chunk_len": str(decode_chunk_len),  # 32
+        "T": str(T),  # 39
+        "num_encoder_layers": num_encoder_layers,
+        "encoder_dims": encoder_dims,
+        "attention_dims": attention_dims,
+        "cnn_module_kernels": cnn_module_kernels,
+        "left_context_len": left_context_len,
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    # (num_encoder_layers, 1)
+    cached_len = init_state[num_encoders * 0 : num_encoders * 1]
+
+    # (num_encoder_layers, 1, encoder_dim)
+    cached_avg = init_state[num_encoders * 1 : num_encoders * 2]
+
+    # (num_encoder_layers, left_context_len, 1, attention_dim)
+    cached_key = init_state[num_encoders * 2 : num_encoders * 3]
+
+    # (num_encoder_layers, left_context_len, 1, attention_dim//2)
+    cached_val = init_state[num_encoders * 3 : num_encoders * 4]
+
+    # (num_encoder_layers, left_context_len, 1, attention_dim//2)
+    cached_val2 = init_state[num_encoders * 4 : num_encoders * 5]
+
+    # (num_encoder_layers, 1, encoder_dim, cnn_module_kernel-1)
+    cached_conv1 = init_state[num_encoders * 5 : num_encoders * 6]
+
+    # (num_encoder_layers, 1, encoder_dim, cnn_module_kernel-1)
+    cached_conv2 = init_state[num_encoders * 6 : num_encoders * 7]
+
+    build_inputs_outputs(cached_len, "cached_len", 1)
+    build_inputs_outputs(cached_avg, "cached_avg", 1)
+    build_inputs_outputs(cached_key, "cached_key", 2)
+    build_inputs_outputs(cached_val, "cached_val", 2)
+    build_inputs_outputs(cached_val2, "cached_val2", 2)
+    build_inputs_outputs(cached_conv1, "cached_conv1", 1)
+    build_inputs_outputs(cached_conv2, "cached_conv2", 1)
+
+    logging.info(inputs)
+    logging.info(outputs)
+    logging.info(input_names)
+    logging.info(output_names)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, init_state),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes={
+            "x": {0: "N"},
+            "encoder_out": {0: "N"},
+            **inputs,
+            **outputs,
+        },
+    )
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+    if params.use_averaged_model:
+        suffix += "-with-averaged-model"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
new file mode 100755
index 000000000..d71080760
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
@@ -0,0 +1,672 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7_streaming/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files in $repo/exp
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py for how to use the exported models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(self, x: Tensor, states: List[Tensor]) -> Tuple[Tensor, List[Tensor]]:
+        """Please see the help information of Zipformer.streaming_forward"""
+        N = x.size(0)
+        T = x.size(1)
+        x_lens = torch.tensor([T] * N, device=x.device)
+
+        output, _, new_states = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=states,
+        )
+
+        output = self.encoder_proj(output)
+        # Now output is of shape (N, T, joiner_dim)
+
+        return output, new_states
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """
+    Onnx model inputs:
+      - 0: src
+      - many state tensors (the exact number depending on the actual model)
+
+    Onnx model outputs:
+      - 0: output, its shape is (N, T, joiner_dim)
+      - many state tensors (the exact number depending on the actual model)
+
+    Args:
+      encoder_model:
+        The model to be exported
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+
+    encoder_model.encoder.__class__.forward = (
+        encoder_model.encoder.__class__.streaming_forward
+    )
+
+    decode_chunk_len = encoder_model.encoder.decode_chunk_size * 2
+    pad_length = 7
+    T = decode_chunk_len + pad_length
+    logging.info(f"decode_chunk_len: {decode_chunk_len}")
+    logging.info(f"pad_length: {pad_length}")
+    logging.info(f"T: {T}")
+
+    x = torch.rand(1, T, 80, dtype=torch.float32)
+
+    init_state = encoder_model.encoder.get_init_state()
+
+    num_encoders = encoder_model.encoder.num_encoders
+    logging.info(f"num_encoders: {num_encoders}")
+    logging.info(f"len(init_state): {len(init_state)}")
+
+    inputs = {}
+    input_names = ["x"]
+
+    outputs = {}
+    output_names = ["encoder_out"]
+
+    def build_inputs_outputs(tensors, name, N):
+        for i, s in enumerate(tensors):
+            logging.info(f"{name}_{i}.shape: {s.shape}")
+            inputs[f"{name}_{i}"] = {N: "N"}
+            outputs[f"new_{name}_{i}"] = {N: "N"}
+            input_names.append(f"{name}_{i}")
+            output_names.append(f"new_{name}_{i}")
+
+    num_encoder_layers = ",".join(map(str, encoder_model.encoder.num_encoder_layers))
+    encoder_dims = ",".join(map(str, encoder_model.encoder.encoder_dims))
+    attention_dims = ",".join(map(str, encoder_model.encoder.attention_dims))
+    cnn_module_kernels = ",".join(map(str, encoder_model.encoder.cnn_module_kernels))
+    ds = encoder_model.encoder.zipformer_downsampling_factors
+    left_context_len = encoder_model.encoder.left_context_len
+    left_context_len = [left_context_len // k for k in ds]
+    left_context_len = ",".join(map(str, left_context_len))
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "decode_chunk_len": str(decode_chunk_len),  # 32
+        "T": str(T),  # 39
+        "num_encoder_layers": num_encoder_layers,
+        "encoder_dims": encoder_dims,
+        "attention_dims": attention_dims,
+        "cnn_module_kernels": cnn_module_kernels,
+        "left_context_len": left_context_len,
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    # (num_encoder_layers, 1)
+    cached_len = init_state[num_encoders * 0 : num_encoders * 1]
+
+    # (num_encoder_layers, 1, encoder_dim)
+    cached_avg = init_state[num_encoders * 1 : num_encoders * 2]
+
+    # (num_encoder_layers, left_context_len, 1, attention_dim)
+    cached_key = init_state[num_encoders * 2 : num_encoders * 3]
+
+    # (num_encoder_layers, left_context_len, 1, attention_dim//2)
+    cached_val = init_state[num_encoders * 3 : num_encoders * 4]
+
+    # (num_encoder_layers, left_context_len, 1, attention_dim//2)
+    cached_val2 = init_state[num_encoders * 4 : num_encoders * 5]
+
+    # (num_encoder_layers, 1, encoder_dim, cnn_module_kernel-1)
+    cached_conv1 = init_state[num_encoders * 5 : num_encoders * 6]
+
+    # (num_encoder_layers, 1, encoder_dim, cnn_module_kernel-1)
+    cached_conv2 = init_state[num_encoders * 6 : num_encoders * 7]
+
+    build_inputs_outputs(cached_len, "cached_len", 1)
+    build_inputs_outputs(cached_avg, "cached_avg", 1)
+    build_inputs_outputs(cached_key, "cached_key", 2)
+    build_inputs_outputs(cached_val, "cached_val", 2)
+    build_inputs_outputs(cached_val2, "cached_val2", 2)
+    build_inputs_outputs(cached_conv1, "cached_conv1", 1)
+    build_inputs_outputs(cached_conv2, "cached_conv2", 1)
+
+    logging.info(inputs)
+    logging.info(outputs)
+    logging.info(input_names)
+    logging.info(output_names)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, init_state),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes={
+            "x": {0: "N"},
+            "encoder_out": {0: "N"},
+            **inputs,
+            **outputs,
+        },
+    )
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+    if params.use_averaged_model:
+        suffix += "-with-averaged-model"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py
new file mode 100755
index 000000000..59a7eb589
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export.py
@@ -0,0 +1,884 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7_streaming/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless7_streaming/decode.py \
+        --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp
+
+(3) Export to ONNX format with pretrained.pt
+
+cd ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+ln -s pretrained.pt epoch-999.pt
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --use-averaged-model False \
+  --epoch 999 \
+  --avg 1 \
+  --fp16 \
+  --onnx 1
+
+It will generate the following files in the given `exp_dir`.
+Check `onnx_check.py` for how to use them.
+
+    - encoder.onnx
+    - decoder.onnx
+    - joiner.onnx
+    - joiner_encoder_proj.onnx
+    - joiner_decoder_proj.onnx
+
+Check
+https://github.com/k2-fsa/sherpa-onnx
+for how to use the exported models outside of icefall.
+
+(4) Export to ONNX format for triton server
+
+cd ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp
+ln -s pretrained.pt epoch-999.pt
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --use-averaged-model False \
+  --epoch 999 \
+  --avg 1 \
+  --fp16 \
+  --onnx-triton 1 \
+  --onnx 1
+
+It will generate the following files in the given `exp_dir`.
+Check `onnx_check.py` for how to use them.
+
+    - encoder.onnx
+    - decoder.onnx
+    - joiner.onnx
+
+Check
+https://github.com/k2-fsa/sherpa/tree/master/triton
+for how to use the exported models outside of icefall.
+
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import onnxruntime
+import torch
+import torch.nn as nn
+from onnx_model_wrapper import OnnxStreamingEncoder, TritonOnnxDecoder, TritonOnnxJoiner
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--onnx",
+        type=str2bool,
+        default=False,
+        help="""If True, --jit is ignored and it exports the model
+        to onnx format. It will generate the following files:
+
+            - encoder.onnx
+            - decoder.onnx
+            - joiner.onnx
+            - joiner_encoder_proj.onnx
+            - joiner_decoder_proj.onnx
+
+        Refer to ./onnx_check.py and ./onnx_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--onnx-triton",
+        type=str2bool,
+        default=False,
+        help="""If True, --onnx would export model into the following files:
+
+            - encoder.onnx
+            - decoder.onnx
+            - joiner.onnx
+        These files would be used for https://github.com/k2-fsa/sherpa/tree/master/triton.
+        """,
+    )
+
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="whether to export fp16 onnx model, default false",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def test_acc(xlist, blist, rtol=1e-3, atol=1e-5, tolerate_small_mismatch=True):
+    for a, b in zip(xlist, blist):
+        try:
+            torch.testing.assert_allclose(a, b, rtol=rtol, atol=atol)
+        except AssertionError as error:
+            if tolerate_small_mismatch:
+                print("small mismatch detected", error)
+            else:
+                return False
+    return True
+
+
+def export_encoder_model_onnx(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T, C)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    batch_size = 17
+    seq_len = 101
+    torch.manual_seed(0)
+    x = torch.rand(batch_size, seq_len, 80, dtype=torch.float32)
+    x_lens = torch.tensor([seq_len - i for i in range(batch_size)], dtype=torch.int64)
+
+    #  encoder_model = torch.jit.script(encoder_model)
+    # It throws the following error for the above statement
+    #
+    # RuntimeError: Exporting the operator __is_ to ONNX opset version
+    # 11 is not supported. Please feel free to request support or
+    # submit a pull request on PyTorch GitHub.
+    #
+    # I cannot find which statement causes the above error.
+    # torch.onnx.export() will use torch.jit.trace() internally, which
+    # works well for the current reworked model
+    initial_states = [encoder_model.get_init_state() for _ in range(batch_size)]
+    states = stack_states(initial_states)
+
+    left_context_len = encoder_model.decode_chunk_size * encoder_model.num_left_chunks
+    encoder_attention_dim = encoder_model.encoders[0].attention_dim
+
+    len_cache = torch.cat(states[: encoder_model.num_encoders]).transpose(0, 1)  # B,15
+    avg_cache = torch.cat(
+        states[encoder_model.num_encoders : 2 * encoder_model.num_encoders]
+    ).transpose(
+        0, 1
+    )  # [B,15,384]
+    cnn_cache = torch.cat(states[5 * encoder_model.num_encoders :]).transpose(
+        0, 1
+    )  # [B,2*15,384,cnn_kernel-1]
+    pad_tensors = [
+        torch.nn.functional.pad(
+            tensor,
+            (
+                0,
+                encoder_attention_dim - tensor.shape[-1],
+                0,
+                0,
+                0,
+                left_context_len - tensor.shape[1],
+                0,
+                0,
+            ),
+        )
+        for tensor in states[
+            2 * encoder_model.num_encoders : 5 * encoder_model.num_encoders
+        ]
+    ]
+    attn_cache = torch.cat(pad_tensors).transpose(0, 2)  # [B,64,15*3,192]
+
+    encoder_model_wrapper = OnnxStreamingEncoder(encoder_model)
+
+    torch.onnx.export(
+        encoder_model_wrapper,
+        (x, x_lens, len_cache, avg_cache, attn_cache, cnn_cache),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "x",
+            "x_lens",
+            "len_cache",
+            "avg_cache",
+            "attn_cache",
+            "cnn_cache",
+        ],
+        output_names=[
+            "encoder_out",
+            "encoder_out_lens",
+            "new_len_cache",
+            "new_avg_cache",
+            "new_attn_cache",
+            "new_cnn_cache",
+        ],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+            "len_cache": {0: "N"},
+            "avg_cache": {0: "N"},
+            "attn_cache": {0: "N"},
+            "cnn_cache": {0: "N"},
+            "new_len_cache": {0: "N"},
+            "new_avg_cache": {0: "N"},
+            "new_attn_cache": {0: "N"},
+            "new_cnn_cache": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_filename}")
+
+    # Test onnx encoder with torch native encoder
+    encoder_model.eval()
+    (
+        encoder_out_torch,
+        encoder_out_lens_torch,
+        new_states_torch,
+    ) = encoder_model.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=states,
+    )
+    ort_session = onnxruntime.InferenceSession(
+        str(encoder_filename), providers=["CPUExecutionProvider"]
+    )
+    ort_inputs = {
+        "x": x.numpy(),
+        "x_lens": x_lens.numpy(),
+        "len_cache": len_cache.numpy(),
+        "avg_cache": avg_cache.numpy(),
+        "attn_cache": attn_cache.numpy(),
+        "cnn_cache": cnn_cache.numpy(),
+    }
+    ort_outs = ort_session.run(None, ort_inputs)
+
+    assert test_acc(
+        [encoder_out_torch.numpy(), encoder_out_lens_torch.numpy()], ort_outs[:2]
+    )
+    logging.info(f"{encoder_filename} acc test succeeded.")
+
+
+def export_decoder_model_onnx(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, 1, C)
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = False  # Always False, so we can use torch.jit.trace() here
+    # Note(fangjun): torch.jit.trace() is more efficient than torch.jit.script()
+    # in this case
+    torch.onnx.export(
+        decoder_model,
+        (y, need_pad),
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y", "need_pad"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_decoder_model_onnx_triton(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, 1, C)
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+
+    decoder_model = TritonOnnxDecoder(decoder_model)
+
+    torch.onnx.export(
+        decoder_model,
+        (y,),
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+
+    The exported encoder_proj model has one input:
+
+        - encoder_out: a tensor of shape (N, encoder_out_dim)
+
+    and produces one output:
+
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+
+    The exported decoder_proj model has one input:
+
+        - decoder_out: a tensor of shape (N, decoder_out_dim)
+
+    and produces one output:
+
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+    """
+    encoder_proj_filename = str(joiner_filename).replace(".onnx", "_encoder_proj.onnx")
+    decoder_proj_filename = str(joiner_filename).replace(".onnx", "_decoder_proj.onnx")
+
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    joiner_dim = joiner_model.decoder_proj.weight.shape[0]
+
+    projected_encoder_out = torch.rand(1, 1, 1, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(1, 1, 1, joiner_dim, dtype=torch.float32)
+
+    project_input = False
+    # Note: It uses torch.jit.trace() internally
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out, project_input),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+            "project_input",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {joiner_filename}")
+
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.encoder_proj,
+        encoder_out,
+        encoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["encoder_out"],
+        output_names=["projected_encoder_out"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "projected_encoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_proj_filename}")
+
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.decoder_proj,
+        decoder_out,
+        decoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["decoder_out"],
+        output_names=["projected_decoder_out"],
+        dynamic_axes={
+            "decoder_out": {0: "N"},
+            "projected_decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_proj_filename}")
+
+
+def export_joiner_model_onnx_triton(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported model has two inputs:
+        - encoder_out: a tensor of shape (N, encoder_out_dim)
+        - decoder_out: a tensor of shape (N, decoder_out_dim)
+    and has one output:
+        - joiner_out: a tensor of shape (N, vocab_size)
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    joiner_model = TritonOnnxJoiner(joiner_model)
+    # Note: It uses torch.jit.trace() internally
+    torch.onnx.export(
+        joiner_model,
+        (encoder_out, decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["encoder_out", "decoder_out"],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.onnx:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        opset_version = 13
+        logging.info("Exporting to onnx format")
+        encoder_filename = params.exp_dir / "encoder.onnx"
+        export_encoder_model_onnx(
+            model.encoder,
+            encoder_filename,
+            opset_version=opset_version,
+        )
+        if not params.onnx_triton:
+            decoder_filename = params.exp_dir / "decoder.onnx"
+            export_decoder_model_onnx(
+                model.decoder,
+                decoder_filename,
+                opset_version=opset_version,
+            )
+
+            joiner_filename = params.exp_dir / "joiner.onnx"
+            export_joiner_model_onnx(
+                model.joiner,
+                joiner_filename,
+                opset_version=opset_version,
+            )
+        else:
+            decoder_filename = params.exp_dir / "decoder.onnx"
+            export_decoder_model_onnx_triton(
+                model.decoder,
+                decoder_filename,
+                opset_version=opset_version,
+            )
+
+            joiner_filename = params.exp_dir / "joiner.onnx"
+            export_joiner_model_onnx_triton(
+                model.joiner,
+                joiner_filename,
+                opset_version=opset_version,
+            )
+
+        if params.fp16:
+            try:
+                import onnxmltools
+                from onnxmltools.utils.float16_converter import convert_float_to_float16
+            except ImportError:
+                print("Please install onnxmltools!")
+                import sys
+
+                sys.exit(1)
+
+            def export_onnx_fp16(onnx_fp32_path, onnx_fp16_path):
+                onnx_fp32_model = onnxmltools.utils.load_model(onnx_fp32_path)
+                onnx_fp16_model = convert_float_to_float16(onnx_fp32_model)
+                onnxmltools.utils.save_model(onnx_fp16_model, onnx_fp16_path)
+
+            encoder_fp16_filename = params.exp_dir / "encoder_fp16.onnx"
+            export_onnx_fp16(encoder_filename, encoder_fp16_filename)
+
+            decoder_fp16_filename = params.exp_dir / "decoder_fp16.onnx"
+            export_onnx_fp16(decoder_filename, decoder_fp16_filename)
+
+            joiner_fp16_filename = params.exp_dir / "joiner_fp16.onnx"
+            export_onnx_fp16(joiner_filename, joiner_fp16_filename)
+
+            if not params.onnx_triton:
+                encoder_proj_filename = str(joiner_filename).replace(
+                    ".onnx", "_encoder_proj.onnx"
+                )
+                encoder_proj_fp16_filename = (
+                    params.exp_dir / "joiner_encoder_proj_fp16.onnx"
+                )
+                export_onnx_fp16(encoder_proj_filename, encoder_proj_fp16_filename)
+
+                decoder_proj_filename = str(joiner_filename).replace(
+                    ".onnx", "_decoder_proj.onnx"
+                )
+                decoder_proj_fp16_filename = (
+                    params.exp_dir / "joiner_decoder_proj_fp16.onnx"
+                )
+                export_onnx_fp16(decoder_proj_filename, decoder_proj_fp16_filename)
+
+    elif params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        model.encoder.__class__.non_streaming_forward = model.encoder.__class__.forward
+        model.encoder.__class__.non_streaming_forward = torch.jit.export(
+            model.encoder.__class__.non_streaming_forward
+        )
+        model.encoder.__class__.forward = model.encoder.__class__.streaming_forward
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
new file mode 100755
index 000000000..c8301b2da
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
@@ -0,0 +1,278 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7_streaming/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7_streaming/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+    model.encoder.decode_chunk_size = args.decode_chunk_len // 2
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder.non_streaming_forward(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
new file mode 100755
index 000000000..4af742316
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
@@ -0,0 +1,313 @@
+#!/usr/bin/env python3
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --decode-chunk-len 32
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+    params: AttributeDict,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    decode_chunk_len = params.decode_chunk_len  # before subsampling
+    pad_length = 7
+    s = f"decode_chunk_len: {decode_chunk_len}"
+    logging.info(s)
+    assert encoder_model.decode_chunk_size == decode_chunk_len // 2, (
+        encoder_model.decode_chunk_size,
+        decode_chunk_len,
+    )
+
+    T = decode_chunk_len + pad_length
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    x_lens = torch.full((1,), T, dtype=torch.int32)
+    states = encoder_model.get_init_state(device=x.device)
+
+    encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename, params)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
new file mode 100755
index 000000000..f2ac1914d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
@@ -0,0 +1,295 @@
+#!/usr/bin/env python3
+# flake8: noqa
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models exported by `torch.jit.trace()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 10 \
+  --use-averaged-model=True \
+  --decode-chunk-len 32
+
+Usage of this script:
+
+./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless7_streaming/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./pruned_transducer_stateless7_streaming/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./pruned_transducer_stateless7_streaming/exp/joiner_jit_trace.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --decode-chunk-len 32 \
+  /path/to/foo.wav \
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Optional
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    assert encoder_out.ndim == 2
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32).unsqueeze(0)
+        # decoder_input.shape (1,, 1 context_size)
+        decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1)
+    else:
+        assert decoder_out.ndim == 2
+        assert hyp is not None, hyp
+
+    T = encoder_out.size(0)
+    for i in range(T):
+        cur_encoder_out = encoder_out[i : i + 1]
+        joiner_out = joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32).unsqueeze(0)
+            decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1)
+
+    return hyp, decoder_out
+
+
+def create_streaming_feature_extractor(sample_rate) -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor(args.sample_rate)
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=args.sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    logging.info("Decoding started")
+    chunk_length = args.decode_chunk_len
+    assert encoder.decode_chunk_size == chunk_length // 2, (
+        encoder.decode_chunk_size,
+        chunk_length,
+    )
+
+    # we subsample features with ((x_len - 7) // 2 + 1) // 2
+    pad_length = 7
+    T = chunk_length + pad_length
+
+    logging.info(f"chunk_length: {chunk_length}")
+
+    states = encoder.get_init_state(device)
+
+    tail_padding = torch.zeros(int(0.3 * args.sample_rate), dtype=torch.float32)
+
+    wave_samples = torch.cat([wave_samples, tail_padding])
+
+    chunk = int(0.25 * args.sample_rate)  # 0.2 second
+    num_processed_frames = 0
+
+    hyp = None
+    decoder_out = None
+
+    start = 0
+    while start < wave_samples.numel():
+        logging.info(f"{start}/{wave_samples.numel()}")
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+        online_fbank.accept_waveform(
+            sampling_rate=args.sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= T:
+            frames = []
+            for i in range(T):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            frames = torch.cat(frames, dim=0).unsqueeze(0)
+            x_lens = torch.tensor([T], dtype=torch.int32)
+            encoder_out, out_lens, states = encoder(
+                x=frames,
+                x_lens=x_lens,
+                states=states,
+            )
+            num_processed_frames += chunk_length
+
+            hyp, decoder_out = greedy_search(
+                decoder, joiner, encoder_out.squeeze(0), decoder_out, hyp
+            )
+
+    context_size = 2
+    logging.info(args.sound_file)
+    logging.info(sp.decode(hyp[context_size:]))
+
+    logging.info("Decoding Done")
+
+
+torch.set_num_threads(4)
+torch.set_num_interop_threads(1)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/model.py
new file mode 120000
index 000000000..e17d4f734
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
new file mode 100644
index 000000000..442a0a8af
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+#
+# Copyright      2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+import ncnn
+import numpy as np
+
+
+layer_list = []
+
+
+def RegisterCustomLayers(net):
+    RegisterPoolingModuleNoProj(net)
+    RegisterTensorAsStrided(net)
+    RegisterSimpleUpsample(net)
+    RegisterStack(net)
+
+
+def RegisterPoolingModuleNoProj(net):
+    net.register_custom_layer(
+        "PoolingModuleNoProj",
+        PoolingModuleNoProjCreator,
+        PoolingModuleNoProjDeleter,
+    )
+
+
+def PoolingModuleNoProjCreator():
+    return PoolingModuleNoProj()
+
+
+def PoolingModuleNoProjDeleter(l):
+    for i, layer in enumerate(layer_list):
+        if layer == l:
+            del layer_list[i]
+            break
+
+
+def TensorAsStridedCreator():
+    return TensorAsStrided()
+
+
+def TensorAsStridedDeleter(l):
+    for i, layer in enumerate(layer_list):
+        if layer == l:
+            del layer_list[i]
+            break
+
+
+def RegisterTensorAsStrided(net):
+    net.register_custom_layer(
+        "TensorAsStrided",
+        TensorAsStridedCreator,
+        TensorAsStridedDeleter,
+    )
+
+
+def SimpleUpsampleCreator():
+    return SimpleUpsample()
+
+
+def SimpleUpsampleDeleter(l):
+    for i, layer in enumerate(layer_list):
+        if layer == l:
+            del layer_list[i]
+            break
+
+
+def RegisterSimpleUpsample(net):
+    net.register_custom_layer(
+        "SimpleUpsample",
+        SimpleUpsampleCreator,
+        SimpleUpsampleDeleter,
+    )
+
+
+def StackCreator():
+    return Stack()
+
+
+def StackDeleter(l):
+    for i, layer in enumerate(layer_list):
+        if layer == l:
+            del layer_list[i]
+            break
+
+
+def RegisterStack(net):
+    net.register_custom_layer(
+        "Stack",
+        StackCreator,
+        StackDeleter,
+    )
+
+
+class PoolingModuleNoProj(ncnn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.one_blob_only = False
+        self.support_inplace = False
+        layer_list.append(self)
+
+    def forward(self, bottom_blobs, top_blobs, opt):
+        x = bottom_blobs[0]
+        cached_len = bottom_blobs[1]
+        cached_avg = bottom_blobs[2]
+
+        # x.dims = 2, x.w = C, x.h = T, e.g., C=384, T=16
+        # cached_len.dims = 1, cached_len.w = 1
+        # cached_avg.dims = 2, cached_avg.w = C, cached_len.h = 1, e.g., C=384
+
+        x = x.numpy()  # x is of shape (T, C), e.g., (16, 384)
+        x = x.cumsum(axis=0)
+
+        cached_len = cached_len.numpy()
+        cached_avg = cached_avg.numpy()
+
+        x = x + cached_len * cached_avg[0]
+        scale = np.arange(1, x.shape[0] + 1, dtype=np.float32).reshape(-1, 1)
+        x = x / (scale + cached_len)
+
+        out_cached_len = cached_len + x.shape[0]
+        out_cached_avg = x[-1:]
+
+        top_blobs[0].clone_from(ncnn.Mat(x), opt.blob_allocator)
+        top_blobs[1].clone_from(ncnn.Mat(out_cached_len), opt.blob_allocator)
+        top_blobs[2].clone_from(ncnn.Mat(out_cached_avg), opt.blob_allocator)
+
+        #  print(top_blobs[0].numpy().shape)
+        #  print(top_blobs[1].numpy().shape)
+        #  print(top_blobs[2].numpy().shape)
+        return 0
+
+
+class TensorAsStrided(ncnn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.one_blob_only = True
+        self.support_inplace = False
+
+        layer_list.append(self)
+
+    def load_param(self, pd):
+        sizes = pd.get(0, ncnn.Mat())
+        strides = pd.get(1, ncnn.Mat())
+        storage_offset = pd.get(2, 0)
+
+        assert sizes.dims == 1, sizes.dims
+        assert strides.dims == 1, strides.dims
+
+        assert sizes.w == strides.w, (sizes.w, strides.w)
+
+        self.sizes = sizes.numpy("i").tolist()
+        self.strides = strides.numpy("i").tolist()
+        self.storage_offset = storage_offset
+
+        return 0
+
+    def forward(self, bottom_blob, top_blob, opt):
+        if bottom_blob.dims != 3:
+            raise ValueError(
+                f"Only 3-D tensors are supported. Given {bottom_blob.dims}"
+            )
+        in_c = bottom_blob.c
+        in_h = bottom_blob.h
+        in_w = bottom_blob.w
+
+        out_c = self.sizes[0]
+        out_h = self.sizes[1]
+        out_w = self.sizes[2]
+
+        assert in_c == out_c, (in_c, out_c)
+        assert self.strides[0] == in_h * in_w, (
+            self.strides[0],
+            in_h,
+            in_w,
+            in_h * in_w,
+        )
+
+        bottom_blob = bottom_blob.numpy()
+        out = np.empty((out_c, out_h, out_w), dtype=np.float32)
+
+        for c in range(out_c):
+            p = bottom_blob[c].reshape(-1)[self.storage_offset :]
+            for h in range(out_h):
+                q = p[h * self.strides[1] :]
+                if True:
+                    for w in range(out_w):
+                        out[c][h][w] = q[w * self.strides[2]]
+                else:
+                    out[c][h] = q[: (out_w * self.strides[2]) : self.strides[2]]
+
+        top_blob.clone_from(ncnn.Mat(out), opt.blob_allocator)
+
+        return 0
+
+
+class SimpleUpsample(ncnn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.one_blob_only = True
+        self.support_inplace = False
+
+        layer_list.append(self)
+
+    def load_param(self, pd):
+        upsample = pd.get(0, 0)
+        num_channels = pd.get(1, 0)
+        bias_data_size = pd.get(2, 0)
+
+        assert upsample * num_channels == bias_data_size, (
+            upsample,
+            num_channels,
+            bias_data_size,
+            upsample * num_channels,
+        )
+
+        self.upsample = upsample
+        self.num_channels = num_channels
+        self.bias_data_size = bias_data_size
+
+        return 0
+
+    def load_model(self, md):
+        bias = md.load(self.num_channels, self.upsample, 0)
+        assert bias.w == self.num_channels, (bias.w, self.num_channels)
+        assert bias.h == self.upsample, (bias.h, self.upsample)
+
+        self.bias = bias.numpy()  # its shape is (upsample, num_channels)
+
+        return 0
+
+    def forward(self, bottom_blob, top_blob, opt):
+        assert bottom_blob.dims == 2, bottom_blob.dims
+        assert bottom_blob.w == self.num_channels, (bottom_blob.w, self.num_channels)
+
+        bottom_blob = bottom_blob.numpy()
+
+        out = np.expand_dims(bottom_blob, axis=1) + self.bias
+        out = out.reshape(-1, self.num_channels)
+
+        top_blob.clone_from(ncnn.Mat(out), opt.blob_allocator)
+
+        return 0
+
+
+class Stack(ncnn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.one_blob_only = False
+        self.support_inplace = False
+
+        layer_list.append(self)
+
+    def load_param(self, pd):
+        axis = pd.get(0, 0)
+
+        self.axis = axis
+
+        return 0
+
+    def forward(self, bottom_blobs, top_blobs, opt):
+        bottom_blobs = [b.numpy() for b in bottom_blobs]
+        out = np.stack(bottom_blobs, axis=self.axis)
+
+        top_blobs[0].clone_from(ncnn.Mat(out), opt.blob_allocator)
+
+        return 0
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
new file mode 100755
index 000000000..d7a4b9551
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
@@ -0,0 +1,260 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script checks that exported ONNX models produce the same output
+with the given torchscript model for the same input.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model via torch.jit.trace()
+
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp
+
+  - encoder_jit_trace.pt
+  - decoder_jit_trace.pt
+  - joiner_jit_trace.pt
+
+3. Export the model to ONNX
+
+./pruned_transducer_stateless7_streaming/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+4. Run this file
+
+./pruned_transducer_stateless7_streaming/onnx_check.py \
+  --jit-encoder-filename $repo/exp/encoder_jit_trace.pt \
+  --jit-decoder-filename $repo/exp/decoder_jit_trace.pt \
+  --jit-joiner-filename $repo/exp/joiner_jit_trace.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+"""
+
+import argparse
+import logging
+
+import torch
+from onnx_pretrained import OnnxModel
+from zipformer import stack_states
+
+from icefall import is_module_available
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript encoder model",
+    )
+
+    parser.add_argument(
+        "--jit-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript decoder model",
+    )
+
+    parser.add_argument(
+        "--jit-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript joiner model",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the ONNX encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the ONNX decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the ONNX joiner model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    torch_encoder_model: torch.jit.ScriptModule,
+    torch_encoder_proj_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    N = torch.randint(1, 100, size=(1,)).item()
+    T = onnx_model.segment
+    C = 80
+    x_lens = torch.tensor([T] * N)
+    torch_states = [torch_encoder_model.get_init_state() for _ in range(N)]
+    torch_states = stack_states(torch_states)
+
+    onnx_model.init_encoder_states(N)
+
+    for i in range(5):
+        logging.info(f"test_encoder: iter {i}")
+        x = torch.rand(N, T, C)
+        torch_encoder_out, _, torch_states = torch_encoder_model(
+            x, x_lens, torch_states
+        )
+        torch_encoder_out = torch_encoder_proj_model(torch_encoder_out)
+
+        onnx_encoder_out = onnx_model.run_encoder(x)
+
+        assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-4), (
+            (torch_encoder_out - onnx_encoder_out).abs().max()
+        )
+
+
+def test_decoder(
+    torch_decoder_model: torch.jit.ScriptModule,
+    torch_decoder_proj_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    context_size = onnx_model.context_size
+    vocab_size = onnx_model.vocab_size
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_decoder: iter {i}, N={N}")
+        x = torch.randint(
+            low=1,
+            high=vocab_size,
+            size=(N, context_size),
+            dtype=torch.int64,
+        )
+        torch_decoder_out = torch_decoder_model(x, need_pad=torch.tensor([False]))
+        torch_decoder_out = torch_decoder_proj_model(torch_decoder_out)
+        torch_decoder_out = torch_decoder_out.squeeze(1)
+
+        onnx_decoder_out = onnx_model.run_decoder(x)
+        assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
+            (torch_decoder_out - onnx_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    torch_joiner_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    encoder_dim = torch_joiner_model.encoder_proj.weight.shape[1]
+    decoder_dim = torch_joiner_model.decoder_proj.weight.shape[1]
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_joiner: iter {i}, N={N}")
+        encoder_out = torch.rand(N, encoder_dim)
+        decoder_out = torch.rand(N, decoder_dim)
+
+        projected_encoder_out = torch_joiner_model.encoder_proj(encoder_out)
+        projected_decoder_out = torch_joiner_model.decoder_proj(decoder_out)
+
+        torch_joiner_out = torch_joiner_model(encoder_out, decoder_out)
+        onnx_joiner_out = onnx_model.run_joiner(
+            projected_encoder_out, projected_decoder_out
+        )
+
+        assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
+            (torch_joiner_out - onnx_joiner_out).abs().max()
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_encoder_model = torch.jit.load(args.jit_encoder_filename)
+    torch_decoder_model = torch.jit.load(args.jit_decoder_filename)
+    torch_joiner_model = torch.jit.load(args.jit_joiner_filename)
+
+    onnx_model = OnnxModel(
+        encoder_model_filename=args.onnx_encoder_filename,
+        decoder_model_filename=args.onnx_decoder_filename,
+        joiner_model_filename=args.onnx_joiner_filename,
+    )
+
+    logging.info("Test encoder")
+    # When exporting the model to onnx, we have already put the encoder_proj
+    # inside the encoder.
+    test_encoder(torch_encoder_model, torch_joiner_model.encoder_proj, onnx_model)
+
+    logging.info("Test decoder")
+    # When exporting the model to onnx, we have already put the decoder_proj
+    # inside the decoder.
+    test_decoder(torch_decoder_model, torch_joiner_model.decoder_proj, onnx_model)
+
+    logging.info("Test joiner")
+    test_joiner(torch_joiner_model, onnx_model)
+
+    logging.info("Finished checking ONNX models")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# See https://github.com/pytorch/pytorch/issues/38342
+# and https://github.com/pytorch/pytorch/issues/33354
+#
+# If we don't do this, the delay increases whenever there is
+# a new request that changes the actual batch size.
+# If you use `py-spy dump --pid <server-pid> --native`, you will
+# see a lot of time is spent in re-compiling the torch script model.
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    torch.manual_seed(20230207)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
new file mode 100644
index 000000000..71a418742
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
@@ -0,0 +1,231 @@
+from typing import Optional, Tuple
+
+import torch
+
+
+class OnnxStreamingEncoder(torch.nn.Module):
+    """This class warps the streaming Zipformer to reduce the number of
+    state tensors for onnx.
+    https://github.com/k2-fsa/icefall/pull/831
+    """
+
+    def __init__(self, encoder):
+        """
+        Args:
+            encoder: An instance of Zipformer Class
+        """
+        super().__init__()
+        self.model = encoder
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        len_cache: torch.tensor,
+        avg_cache: torch.tensor,
+        attn_cache: torch.tensor,
+        cnn_cache: torch.tensor,
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+    ]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          len_cache:
+            The cached numbers of past frames.
+          avg_cache:
+            The cached average tensors.
+          attn_cache:
+            The cached key tensors of the first attention modules.
+            The cached value tensors of the first attention modules.
+            The cached value tensors of the second attention modules.
+          cnn_cache:
+            The cached left contexts of the first convolution modules.
+            The cached left contexts of the second convolution modules.
+
+        Returns:
+          Return a tuple containing 2 tensors:
+
+        """
+        num_encoder_layers = []
+        encoder_attention_dims = []
+        states = []
+        for i, encoder in enumerate(self.model.encoders):
+            num_encoder_layers.append(encoder.num_layers)
+            encoder_attention_dims.append(encoder.attention_dim)
+
+        len_cache = len_cache.transpose(0, 1)  # sum(num_encoder_layers)==15, [15, B]
+        offset = 0
+        for num_layer in num_encoder_layers:
+            states.append(len_cache[offset : offset + num_layer])
+            offset += num_layer
+
+        avg_cache = avg_cache.transpose(0, 1)  # [15, B, 384]
+        offset = 0
+        for num_layer in num_encoder_layers:
+            states.append(avg_cache[offset : offset + num_layer])
+            offset += num_layer
+
+        attn_cache = attn_cache.transpose(0, 2)  # [15*3, 64, B, 192]
+        left_context_len = attn_cache.shape[1]
+        offset = 0
+        for i, num_layer in enumerate(num_encoder_layers):
+            ds = self.model.zipformer_downsampling_factors[i]
+            states.append(
+                attn_cache[offset : offset + num_layer, : left_context_len // ds]
+            )
+            offset += num_layer
+        for i, num_layer in enumerate(num_encoder_layers):
+            encoder_attention_dim = encoder_attention_dims[i]
+            ds = self.model.zipformer_downsampling_factors[i]
+            states.append(
+                attn_cache[
+                    offset : offset + num_layer,
+                    : left_context_len // ds,
+                    :,
+                    : encoder_attention_dim // 2,
+                ]
+            )
+            offset += num_layer
+        for i, num_layer in enumerate(num_encoder_layers):
+            ds = self.model.zipformer_downsampling_factors[i]
+            states.append(
+                attn_cache[
+                    offset : offset + num_layer,
+                    : left_context_len // ds,
+                    :,
+                    : encoder_attention_dim // 2,
+                ]
+            )
+            offset += num_layer
+
+        cnn_cache = cnn_cache.transpose(0, 1)  # [30, B, 384, cnn_kernel-1]
+        offset = 0
+        for num_layer in num_encoder_layers:
+            states.append(cnn_cache[offset : offset + num_layer])
+            offset += num_layer
+        for num_layer in num_encoder_layers:
+            states.append(cnn_cache[offset : offset + num_layer])
+            offset += num_layer
+
+        encoder_out, encoder_out_lens, new_states = self.model.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=states,
+        )
+
+        new_len_cache = torch.cat(states[: self.model.num_encoders]).transpose(
+            0, 1
+        )  # [B,15]
+        new_avg_cache = torch.cat(
+            states[self.model.num_encoders : 2 * self.model.num_encoders]
+        ).transpose(
+            0, 1
+        )  # [B,15,384]
+        new_cnn_cache = torch.cat(states[5 * self.model.num_encoders :]).transpose(
+            0, 1
+        )  # [B,2*15,384,cnn_kernel-1]
+        assert len(set(encoder_attention_dims)) == 1
+        pad_tensors = [
+            torch.nn.functional.pad(
+                tensor,
+                (
+                    0,
+                    encoder_attention_dims[0] - tensor.shape[-1],
+                    0,
+                    0,
+                    0,
+                    left_context_len - tensor.shape[1],
+                    0,
+                    0,
+                ),
+            )
+            for tensor in states[
+                2 * self.model.num_encoders : 5 * self.model.num_encoders
+            ]
+        ]
+        new_attn_cache = torch.cat(pad_tensors).transpose(0, 2)  # [B,64,15*3,192]
+
+        return (
+            encoder_out,
+            encoder_out_lens,
+            new_len_cache,
+            new_avg_cache,
+            new_attn_cache,
+            new_cnn_cache,
+        )
+
+
+class TritonOnnxDecoder(torch.nn.Module):
+    """This class warps the Decoder in decoder.py
+    to remove the scalar input "need_pad".
+    Triton currently doesn't support scalar input.
+    https://github.com/triton-inference-server/server/issues/2333
+    """
+
+    def __init__(
+        self,
+        decoder: torch.nn.Module,
+    ):
+        """
+        Args:
+          decoder: A instance of Decoder
+        """
+        super().__init__()
+        self.model = decoder
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        # False to not pad the input. Should be False during inference.
+        need_pad = False
+        return self.model(y, need_pad)
+
+
+class TritonOnnxJoiner(torch.nn.Module):
+    """This class warps the Joiner in joiner.py
+    to remove the scalar input "project_input".
+    Triton currently doesn't support scalar input.
+    https://github.com/triton-inference-server/server/issues/2333
+    "project_input" is set to True.
+    Triton solutions only need export joiner to a single joiner.onnx.
+    """
+
+    def __init__(
+        self,
+        joiner: torch.nn.Module,
+    ):
+        super().__init__()
+        self.model = joiner
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        # Apply input projections encoder_proj and decoder_proj.
+        project_input = True
+        return self.model(encoder_out, decoder_out, project_input)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
new file mode 100755
index 000000000..04861ea37
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
@@ -0,0 +1,515 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This script loads ONNX models exported by ./export-onnx.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless7_streaming/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files in $repo/exp
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file with the exported ONNX models
+
+./pruned_transducer_stateless7_streaming/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_encoder_states()
+
+    def init_encoder_states(self, batch_size: int = 1):
+        encoder_meta = self.encoder.get_modelmeta().custom_metadata_map
+
+        model_type = encoder_meta["model_type"]
+        assert model_type == "zipformer", model_type
+
+        decode_chunk_len = int(encoder_meta["decode_chunk_len"])
+        T = int(encoder_meta["T"])
+
+        num_encoder_layers = encoder_meta["num_encoder_layers"]
+        encoder_dims = encoder_meta["encoder_dims"]
+        attention_dims = encoder_meta["attention_dims"]
+        cnn_module_kernels = encoder_meta["cnn_module_kernels"]
+        left_context_len = encoder_meta["left_context_len"]
+
+        def to_int_list(s):
+            return list(map(int, s.split(",")))
+
+        num_encoder_layers = to_int_list(num_encoder_layers)
+        encoder_dims = to_int_list(encoder_dims)
+        attention_dims = to_int_list(attention_dims)
+        cnn_module_kernels = to_int_list(cnn_module_kernels)
+        left_context_len = to_int_list(left_context_len)
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+        logging.info(f"encoder_dims: {encoder_dims}")
+        logging.info(f"attention_dims: {attention_dims}")
+        logging.info(f"cnn_module_kernels: {cnn_module_kernels}")
+        logging.info(f"left_context_len: {left_context_len}")
+
+        num_encoders = len(num_encoder_layers)
+
+        cached_len = []
+        cached_avg = []
+        cached_key = []
+        cached_val = []
+        cached_val2 = []
+        cached_conv1 = []
+        cached_conv2 = []
+
+        N = batch_size
+
+        for i in range(num_encoders):
+            cached_len.append(torch.zeros(num_encoder_layers[i], N, dtype=torch.int64))
+            cached_avg.append(torch.zeros(num_encoder_layers[i], N, encoder_dims[i]))
+            cached_key.append(
+                torch.zeros(
+                    num_encoder_layers[i], left_context_len[i], N, attention_dims[i]
+                )
+            )
+            cached_val.append(
+                torch.zeros(
+                    num_encoder_layers[i],
+                    left_context_len[i],
+                    N,
+                    attention_dims[i] // 2,
+                )
+            )
+            cached_val2.append(
+                torch.zeros(
+                    num_encoder_layers[i],
+                    left_context_len[i],
+                    N,
+                    attention_dims[i] // 2,
+                )
+            )
+            cached_conv1.append(
+                torch.zeros(
+                    num_encoder_layers[i], N, encoder_dims[i], cnn_module_kernels[i] - 1
+                )
+            )
+            cached_conv2.append(
+                torch.zeros(
+                    num_encoder_layers[i], N, encoder_dims[i], cnn_module_kernels[i] - 1
+                )
+            )
+
+        self.cached_len = cached_len
+        self.cached_avg = cached_avg
+        self.cached_key = cached_key
+        self.cached_val = cached_val
+        self.cached_val2 = cached_val2
+        self.cached_conv1 = cached_conv1
+        self.cached_conv2 = cached_conv2
+
+        self.num_encoders = num_encoders
+
+        self.segment = T
+        self.offset = decode_chunk_len
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def _build_encoder_input_output(
+        self,
+        x: torch.Tensor,
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        encoder_input = {"x": x.numpy()}
+        encoder_output = ["encoder_out"]
+
+        def build_states_input(states: List[torch.Tensor], name: str):
+            for i, s in enumerate(states):
+                if isinstance(s, torch.Tensor):
+                    encoder_input[f"{name}_{i}"] = s.numpy()
+                else:
+                    encoder_input[f"{name}_{i}"] = s
+
+                encoder_output.append(f"new_{name}_{i}")
+
+        build_states_input(self.cached_len, "cached_len")
+        build_states_input(self.cached_avg, "cached_avg")
+        build_states_input(self.cached_key, "cached_key")
+        build_states_input(self.cached_val, "cached_val")
+        build_states_input(self.cached_val2, "cached_val2")
+        build_states_input(self.cached_conv1, "cached_conv1")
+        build_states_input(self.cached_conv2, "cached_conv2")
+
+        return encoder_input, encoder_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        num_encoders = self.num_encoders
+
+        self.cached_len = states[num_encoders * 0 : num_encoders * 1]
+        self.cached_avg = states[num_encoders * 1 : num_encoders * 2]
+        self.cached_key = states[num_encoders * 2 : num_encoders * 3]
+        self.cached_val = states[num_encoders * 3 : num_encoders * 4]
+        self.cached_val2 = states[num_encoders * 4 : num_encoders * 5]
+        self.cached_conv1 = states[num_encoders * 5 : num_encoders * 6]
+        self.cached_conv2 = states[num_encoders * 6 : num_encoders * 7]
+
+    def run_encoder(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+        Returns:
+          Return a 3-D tensor of shape (N, T', joiner_dim) where
+          T' is usually equal to ((T-7)//2+1)//2
+        """
+        encoder_input, encoder_output_names = self._build_encoder_input_output(x)
+        out = self.encoder.run(encoder_output_names, encoder_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    context_size: int,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+) -> List[int]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (1, T, joiner_dim)
+      context_size:
+        The context size of the decoder model.
+      decoder_out:
+        Optional. Decoder output of the previous chunk.
+      hyp:
+        Decoding results for previous chunks.
+    Returns:
+      Return the decoded results so far.
+    """
+
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert hyp is not None, hyp
+
+    encoder_out = encoder_out.squeeze(0)
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t : t + 1]
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+            decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    context_size = model.context_size
+    hyp = None
+    decoder_out = None
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            encoder_out = model.run_encoder(frames)
+            hyp, decoder_out = greedy_search(
+                model,
+                encoder_out,
+                context_size,
+                decoder_out,
+                hyp,
+            )
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
new file mode 100755
index 000000000..bc42e8d05
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
@@ -0,0 +1,360 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7_streaming/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7_streaming/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_streaming/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
new file mode 100755
index 000000000..883fdcbdd
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
@@ -0,0 +1,446 @@
+#!/usr/bin/env python3
+#
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py \
+  --tokens ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/tokens.txt \
+  --encoder-param-filename ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/encoder_jit_trace-pnnx.ncnn.param \
+  --encoder-bin-filename ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/encoder_jit_trace-pnnx.ncnn.bin \
+  --decoder-param-filename ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/decoder_jit_trace-pnnx.ncnn.param \
+  --decoder-bin-filename ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/decoder_jit_trace-pnnx.ncnn.bin \
+  --joiner-param-filename ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/joiner_jit_trace-pnnx.ncnn.param \
+  --joiner-bin-filename ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/joiner_jit_trace-pnnx.ncnn.bin \
+  ./sherpa-ncnn-streaming-zipformer-en-2023-02-13/test_wavs/1089-134686-0001.wav
+
+You can find pretrained models at
+- English: https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-en-2023-02-13
+- Bilingual (Chinese + English): https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-bilingual-zh-en-2023-02-13
+"""
+
+import argparse
+import logging
+from typing import List, Optional, Tuple
+
+import k2
+import ncnn
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+from ncnn_custom_layer import RegisterCustomLayers
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument(
+        "--encoder-param-filename",
+        type=str,
+        help="Path to encoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--encoder-bin-filename",
+        type=str,
+        help="Path to encoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--decoder-param-filename",
+        type=str,
+        help="Path to decoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--decoder-bin-filename",
+        type=str,
+        help="Path to decoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--joiner-param-filename",
+        type=str,
+        help="Path to joiner.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--joiner-bin-filename",
+        type=str,
+        help="Path to joiner.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="Path to foo.wav",
+    )
+
+    return parser.parse_args()
+
+
+def to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+class Model:
+    def __init__(self, args):
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+
+        # Please change the parameters according to your model
+
+        # 20M
+        # self.num_encoder_layers = to_int_tuple("2,2,2,2,2")
+        # self.encoder_dims = to_int_tuple("256,256,256,256,256")  # also known as d_model
+        # self.attention_dims = to_int_tuple("192,192,192,192,192")
+        # self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
+        # self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
+
+        # 9.6M
+        # self.num_encoder_layers = to_int_tuple("2,3,2,2,3")
+        # self.encoder_dims = to_int_tuple("160,160,160,160,160")  # also known as d_model
+        # self.attention_dims = to_int_tuple("96,96,96,96,96")
+        # self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
+        # self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
+
+        # 5.5M or 6M
+
+        # self.num_encoder_layers = to_int_tuple("2,2,2,2,2")
+        # self.encoder_dims = to_int_tuple("128,128,128,128,128")  # also known as d_model
+        # self.attention_dims = to_int_tuple("96,96,96,96,96")
+        # self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
+        # self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
+
+        self.num_encoder_layers = to_int_tuple("2,4,3,2,4")
+        self.encoder_dims = to_int_tuple("384,384,384,384,384")  # also known as d_model
+        self.attention_dims = to_int_tuple("192,192,192,192,192")
+        self.zipformer_downsampling_factors = to_int_tuple("1,2,4,8,2")
+        self.cnn_module_kernels = to_int_tuple("31,31,31,31,31")
+
+        self.decode_chunk_size = 32 // 2
+        num_left_chunks = 4
+        self.left_context_length = self.decode_chunk_size * num_left_chunks  # 64
+
+        self.chunk_length = self.decode_chunk_size * 2
+        pad_length = 7
+        self.T = self.chunk_length + pad_length
+
+    def get_init_states(self) -> List[torch.Tensor]:
+        cached_len_list = []
+        cached_avg_list = []
+        cached_key_list = []
+        cached_val_list = []
+        cached_val2_list = []
+        cached_conv1_list = []
+        cached_conv2_list = []
+
+        for i in range(len(self.num_encoder_layers)):
+            num_layers = self.num_encoder_layers[i]
+            ds = self.zipformer_downsampling_factors[i]
+            attention_dim = self.attention_dims[i]
+            left_context_length = self.left_context_length // ds
+            encoder_dim = self.encoder_dims[i]
+            cnn_module_kernel = self.cnn_module_kernels[i]
+
+            cached_len_list.append(torch.zeros(num_layers))
+            cached_avg_list.append(torch.zeros(num_layers, encoder_dim))
+            cached_key_list.append(
+                torch.zeros(num_layers, left_context_length, attention_dim)
+            )
+            cached_val_list.append(
+                torch.zeros(num_layers, left_context_length, attention_dim // 2)
+            )
+            cached_val2_list.append(
+                torch.zeros(num_layers, left_context_length, attention_dim // 2)
+            )
+            cached_conv1_list.append(
+                torch.zeros(num_layers, encoder_dim, cnn_module_kernel - 1)
+            )
+            cached_conv2_list.append(
+                torch.zeros(num_layers, encoder_dim, cnn_module_kernel - 1)
+            )
+
+        states = (
+            cached_len_list
+            + cached_avg_list
+            + cached_key_list
+            + cached_val_list
+            + cached_val2_list
+            + cached_conv1_list
+            + cached_conv2_list
+        )
+
+        return states
+
+    def init_encoder(self, args):
+        encoder_net = ncnn.Net()
+        encoder_net.opt.use_packing_layout = False
+        encoder_net.opt.use_fp16_storage = False
+        encoder_net.opt.num_threads = 4
+
+        encoder_param = args.encoder_param_filename
+        encoder_model = args.encoder_bin_filename
+
+        RegisterCustomLayers(encoder_net)
+
+        encoder_net.load_param(encoder_param)
+        encoder_net.load_model(encoder_model)
+
+        self.encoder_net = encoder_net
+
+    def init_decoder(self, args):
+        decoder_param = args.decoder_param_filename
+        decoder_model = args.decoder_bin_filename
+
+        decoder_net = ncnn.Net()
+        decoder_net.opt.num_threads = 4
+
+        decoder_net.load_param(decoder_param)
+        decoder_net.load_model(decoder_model)
+
+        self.decoder_net = decoder_net
+
+    def init_joiner(self, args):
+        joiner_param = args.joiner_param_filename
+        joiner_model = args.joiner_bin_filename
+        joiner_net = ncnn.Net()
+        joiner_net.opt.num_threads = 4
+
+        joiner_net.load_param(joiner_param)
+        joiner_net.load_model(joiner_model)
+
+        self.joiner_net = joiner_net
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          x:
+            A tensor of shape (T, C)
+          states:
+            A list of tensors. len(states) == self.num_layers * 4
+        Returns:
+          Return a tuple containing:
+           - encoder_out, a tensor of shape (T, encoder_dim).
+           - next_states, a list of tensors containing the next states
+        """
+        with self.encoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(x.numpy()).clone())
+
+            for i in range(len(states)):
+                name = f"in{i+1}"
+                ex.input(name, ncnn.Mat(states[i].squeeze().numpy()).clone())
+
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            encoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+
+            out_states: List[torch.Tensor] = []
+            for i in range(len(states)):
+                name = f"out{i+1}"
+                ret, ncnn_out_state = ex.extract(name)
+                assert ret == 0, ret
+                ncnn_out_state = torch.from_numpy(ncnn_out_state.numpy())
+
+                if i < len(self.num_encoder_layers):
+                    # for cached_len, we need to discard the last dim
+                    ncnn_out_state = ncnn_out_state.squeeze(1)
+
+                out_states.append(ncnn_out_state)
+
+            return encoder_out, out_states
+
+    def run_decoder(self, decoder_input):
+        assert decoder_input.dtype == torch.int32
+
+        with self.decoder_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(decoder_input.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            decoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return decoder_out
+
+    def run_joiner(self, encoder_out, decoder_out):
+        with self.joiner_net.create_extractor() as ex:
+            ex.input("in0", ncnn.Mat(encoder_out.numpy()).clone())
+            ex.input("in1", ncnn.Mat(decoder_out.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            joiner_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return joiner_out
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: Model,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32)  # (1, context_size)
+        decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    else:
+        assert decoder_out.ndim == 1
+        assert hyp is not None, hyp
+
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t]
+
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32)
+            decoder_out = model.run_decoder(decoder_input).squeeze(0)
+
+    return hyp, decoder_out
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+
+    wave_samples = torch.cat([wave_samples, tail_padding])
+
+    states = model.get_init_states()
+    logging.info(f"number of states: {len(states)}")
+
+    hyp = None
+    decoder_out = None
+
+    num_processed_frames = 0
+    segment = model.T
+    offset = model.chunk_length
+
+    chunk = int(1 * sample_rate)  # 0.2 second
+
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            encoder_out, states = model.run_encoder(frames, states)
+            hyp, decoder_out = greedy_search(model, encoder_out, decoder_out, hyp)
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    context_size = 2
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(sound_file)
+    logging.info(text)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
new file mode 120000
index 000000000..3a5f89833
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
new file mode 100755
index 000000000..a0f54b6e1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
@@ -0,0 +1,609 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py
new file mode 100755
index 000000000..de12d7af1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py
@@ -0,0 +1,187 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless7_streaming/test_model.py
+"""
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    # Test jit script
+    convert_scaled_to_non_scaled(model, inplace=True)
+    # We won't use the forward() method of the model in C++, so just ignore
+    # it here.
+    # Otherwise, one of its arguments is a ragged tensor and is not
+    # torch scriptabe.
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    print("Using torch.jit.script")
+    model = torch.jit.script(model)
+
+
+def test_model_small():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,2,2,2,2"
+    params.feedforward_dims = "256,256,512,512,256"
+    params.nhead = "4,4,4,4,4"
+    params.encoder_dims = "128,128,128,128,128"
+    params.attention_dims = "96,96,96,96,96"
+    params.encoder_unmasked_dims = "96,96,96,96,96"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 320
+    params.joiner_dim = 320
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    import pdb
+
+    pdb.set_trace()
+
+    # Test jit script
+    convert_scaled_to_non_scaled(model, inplace=True)
+    # We won't use the forward() method of the model in C++, so just ignore
+    # it here.
+    # Otherwise, one of its arguments is a ragged tensor and is not
+    # torch scriptabe.
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    print("Using torch.jit.script")
+    model = torch.jit.script(model)
+
+
+def test_model_jit_trace():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    params.num_left_chunks = 4
+    params.short_chunk_size = 50
+    params.decode_chunk_len = 32
+    model = get_transducer_model(params)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    # Test encoder
+    def _test_encoder():
+        encoder = model.encoder
+        assert encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+            encoder.decode_chunk_size,
+            params.decode_chunk_len,
+        )
+        T = params.decode_chunk_len + 7
+
+        x = torch.zeros(1, T, 80, dtype=torch.float32)
+        x_lens = torch.full((1,), T, dtype=torch.int32)
+        states = encoder.get_init_state(device=x.device)
+        encoder.__class__.forward = encoder.__class__.streaming_forward
+        traced_encoder = torch.jit.trace(encoder, (x, x_lens, states))
+
+        states1 = encoder.get_init_state(device=x.device)
+        states2 = traced_encoder.get_init_state(device=x.device)
+        for i in range(5):
+            x = torch.randn(1, T, 80, dtype=torch.float32)
+            x_lens = torch.full((1,), T, dtype=torch.int32)
+            y1, _, states1 = encoder.streaming_forward(x, x_lens, states1)
+            y2, _, states2 = traced_encoder(x, x_lens, states2)
+            assert torch.allclose(y1, y2, atol=1e-6), (i, (y1 - y2).abs().mean())
+
+    # Test decoder
+    def _test_decoder():
+        decoder = model.decoder
+        y = torch.zeros(10, decoder.context_size, dtype=torch.int64)
+        need_pad = torch.tensor([False])
+
+        traced_decoder = torch.jit.trace(decoder, (y, need_pad))
+        d1 = decoder(y, need_pad)
+        d2 = traced_decoder(y, need_pad)
+        assert torch.equal(d1, d2), (d1 - d2).abs().mean()
+
+    # Test joiner
+    def _test_joiner():
+        joiner = model.joiner
+        encoder_out_dim = joiner.encoder_proj.weight.shape[1]
+        decoder_out_dim = joiner.decoder_proj.weight.shape[1]
+        encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+        decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+        traced_joiner = torch.jit.trace(joiner, (encoder_out, decoder_out))
+        j1 = joiner(encoder_out, decoder_out)
+        j2 = traced_joiner(encoder_out, decoder_out)
+        assert torch.equal(j1, j2), (j1 - j2).abs().mean()
+
+    _test_encoder()
+    _test_decoder()
+    _test_joiner()
+
+
+def main():
+    test_model_small()
+    test_model_jit_trace()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py
new file mode 100755
index 000000000..2d915ff87
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -0,0 +1,1259 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.mini_libri:
+        train_cuts = librispeech.train_clean_5_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+        if params.full_libri:
+            train_cuts += librispeech.train_clean_360_cuts()
+            train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    # train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if params.mini_libri:
+        valid_cuts = librispeech.dev_clean_2_cuts()
+    else:
+        valid_cuts = librispeech.dev_clean_cuts()
+        valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         sp=sp,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
new file mode 100644
index 000000000..c7e45564f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
@@ -0,0 +1,2891 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import itertools
+import logging
+import math
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    Identity,
+    MaxEig,
+    ScaledConv1d,
+    Whiten,
+    _diag,
+    penalize_abs_values_gt,
+    random_clamp,
+    softmax,
+)
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask, subsequent_chunk_mask
+
+
+def stack_states(state_list: List[List[Tensor]]) -> List[Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+        ``states[i]`` is a list of 7 * num_encoders elements of i-th utterance.
+        ``states[i][0:num_encoders]`` is the cached numbers of past frames.
+        ``states[i][num_encoders:2*num_encoders]`` is the cached average tensors.
+        ``states[i][2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+        ``states[i][3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+        ``states[i][4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+        ``states[i][5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+        ``states[i][6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+    assert len(state_list[0]) % 7 == 0, len(state_list[0])
+    num_encoders = len(state_list[0]) // 7
+
+    cached_len = []
+    cached_avg = []
+    cached_key = []
+    cached_val = []
+    cached_val2 = []
+    cached_conv1 = []
+    cached_conv2 = []
+
+    # For cached_len
+    len_list = [state_list[n][0:num_encoders] for n in range(batch_size)]
+    for i in range(num_encoders):
+        # len_avg: (num_layers, batch_size)
+        len_avg = torch.cat([len_list[n][i] for n in range(batch_size)], dim=1)
+        cached_len.append(len_avg)
+
+    # For cached_avg
+    avg_list = [
+        state_list[n][num_encoders : 2 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # avg: (num_layers, batch_size, D)
+        avg = torch.cat([avg_list[n][i] for n in range(batch_size)], dim=1)
+        cached_avg.append(avg)
+
+    # For cached_key
+    key_list = [
+        state_list[n][2 * num_encoders : 3 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # key: (num_layers, left_context_size, batch_size, D)
+        key = torch.cat([key_list[n][i] for n in range(batch_size)], dim=2)
+        cached_key.append(key)
+
+    # For cached_val
+    val_list = [
+        state_list[n][3 * num_encoders : 4 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # val: (num_layers, left_context_size, batch_size, D)
+        val = torch.cat([val_list[n][i] for n in range(batch_size)], dim=2)
+        cached_val.append(val)
+
+    # For cached_val2
+    val2_list = [
+        state_list[n][4 * num_encoders : 5 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # val2: (num_layers, left_context_size, batch_size, D)
+        val2 = torch.cat([val2_list[n][i] for n in range(batch_size)], dim=2)
+        cached_val2.append(val2)
+
+    # For cached_conv1
+    conv1_list = [
+        state_list[n][5 * num_encoders : 6 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # conv1: (num_layers, batch_size, D, kernel-1)
+        conv1 = torch.cat([conv1_list[n][i] for n in range(batch_size)], dim=1)
+        cached_conv1.append(conv1)
+
+    # For cached_conv2
+    conv2_list = [
+        state_list[n][6 * num_encoders : 7 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # conv2: (num_layers, batch_size, D, kernel-1)
+        conv2 = torch.cat([conv2_list[n][i] for n in range(batch_size)], dim=1)
+        cached_conv2.append(conv2)
+
+    states = (
+        cached_len
+        + cached_avg
+        + cached_key
+        + cached_val
+        + cached_val2
+        + cached_conv1
+        + cached_conv2
+    )
+    return states
+
+
+def unstack_states(states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+      states:
+        A list of 7 * num_encoders elements:
+        ``states[0:num_encoders]`` is the cached numbers of past frames.
+        ``states[num_encoders:2*num_encoders]`` is the cached average tensors.
+        ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+        ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+        ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+        ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+        ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+
+    Returns:
+      A list of states.
+      ``states[i]`` is a list of 7 * num_encoders elements of i-th utterance.
+    """
+    assert len(states) % 7 == 0, len(states)
+    num_encoders = len(states) // 7
+    (
+        cached_len,
+        cached_avg,
+        cached_key,
+        cached_val,
+        cached_val2,
+        cached_conv1,
+        cached_conv2,
+    ) = (states[i * num_encoders : (i + 1) * num_encoders] for i in range(7))
+
+    batch_size = cached_len[0].shape[1]
+
+    len_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_len[i]: (num_layers, batch_size)
+        len_avg = cached_len[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            len_list[n].append(len_avg[n])
+
+    avg_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_avg[i]: (num_layers, batch_size, D)
+        avg = cached_avg[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            avg_list[n].append(avg[n])
+
+    key_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_key[i]: (num_layers, left_context, batch_size, D)
+        key = cached_key[i].chunk(chunks=batch_size, dim=2)
+        for n in range(batch_size):
+            key_list[n].append(key[n])
+
+    val_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_val[i]: (num_layers, left_context, batch_size, D)
+        val = cached_val[i].chunk(chunks=batch_size, dim=2)
+        for n in range(batch_size):
+            val_list[n].append(val[n])
+
+    val2_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_val2[i]: (num_layers, left_context, batch_size, D)
+        val2 = cached_val2[i].chunk(chunks=batch_size, dim=2)
+        for n in range(batch_size):
+            val2_list[n].append(val2[n])
+
+    conv1_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_conv1[i]: (num_layers, batch_size, D, kernel-1)
+        conv1 = cached_conv1[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            conv1_list[n].append(conv1[n])
+
+    conv2_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_conv2[i]: (num_layers, batch_size, D, kernel-1)
+        conv2 = cached_conv2[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            conv2_list[n].append(conv2[n])
+
+    state_list = [
+        (
+            len_list[i]
+            + avg_list[i]
+            + key_list[i]
+            + val_list[i]
+            + val2_list[i]
+            + conv1_list[i]
+            + conv2_list[i]
+        )
+        for i in range(batch_size)
+    ]
+    return state_list
+
+
+class Zipformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        d_model: (int,int): embedding dimension of 2 encoder stacks
+        attention_dim: (int,int): attention dimension of 2 encoder stacks
+        nhead (int, int): number of heads
+        dim_feedforward (int, int): feedforward dimension in 2 encoder stacks
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        cnn_module_kernels (int): Kernel size of convolution module
+        warmup_batches (float): number of batches to warm up over
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_downsampling_factor: int = 2,
+        encoder_dims: Tuple[int] = (384, 384),
+        attention_dim: Tuple[int] = (256, 256),
+        encoder_unmasked_dims: Tuple[int] = (256, 256),
+        zipformer_downsampling_factors: Tuple[int] = (2, 4),
+        nhead: Tuple[int] = (8, 8),
+        feedforward_dim: Tuple[int] = (1536, 2048),
+        num_encoder_layers: Tuple[int] = (12, 12),
+        dropout: float = 0.1,
+        cnn_module_kernels: Tuple[int] = (31, 31),
+        pos_dim: int = 4,
+        num_left_chunks: int = 4,
+        short_chunk_threshold: float = 0.75,
+        short_chunk_size: int = 50,
+        decode_chunk_size: int = 16,
+        warmup_batches: float = 4000.0,
+    ) -> None:
+        super(Zipformer, self).__init__()
+
+        self.num_features = num_features
+        assert 0 < encoder_dims[0] <= encoder_dims[1]
+        self.encoder_dims = encoder_dims
+        self.encoder_unmasked_dims = encoder_unmasked_dims
+        self.zipformer_downsampling_factors = zipformer_downsampling_factors
+        self.output_downsampling_factor = output_downsampling_factor
+
+        self.num_left_chunks = num_left_chunks
+        self.short_chunk_threshold = short_chunk_threshold
+        self.short_chunk_size = short_chunk_size
+
+        # Used in decoding
+        self.decode_chunk_size = decode_chunk_size
+
+        self.left_context_len = self.decode_chunk_size * self.num_left_chunks
+
+        # will be written to, see set_batch_count()
+        self.batch_count = 0
+        self.warmup_end = warmup_batches
+
+        for u, d in zip(encoder_unmasked_dims, encoder_dims):
+            assert u <= d, (u, d)
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, (T - 7)//2, encoder_dims).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> (T - 7)//2
+        #   (2) embedding: num_features -> encoder_dims
+        self.encoder_embed = Conv2dSubsampling(
+            num_features, encoder_dims[0], dropout=dropout
+        )
+
+        # each one will be ZipformerEncoder or DownsampledZipformerEncoder
+        encoders = []
+
+        self.num_encoder_layers = num_encoder_layers
+        self.num_encoders = len(encoder_dims)
+        self.attention_dims = attention_dim
+        self.cnn_module_kernels = cnn_module_kernels
+        for i in range(self.num_encoders):
+            encoder_layer = ZipformerEncoderLayer(
+                encoder_dims[i],
+                attention_dim[i],
+                nhead[i],
+                feedforward_dim[i],
+                dropout,
+                cnn_module_kernels[i],
+                pos_dim,
+            )
+
+            # For the segment of the warmup period, we let the Conv2dSubsampling
+            # layer learn something.  Then we start to warm up the other encoders.
+            encoder = ZipformerEncoder(
+                encoder_layer,
+                num_encoder_layers[i],
+                dropout,
+                warmup_begin=warmup_batches * (i + 1) / (self.num_encoders + 1),
+                warmup_end=warmup_batches * (i + 2) / (self.num_encoders + 1),
+            )
+
+            if zipformer_downsampling_factors[i] != 1:
+                encoder = DownsampledZipformerEncoder(
+                    encoder,
+                    input_dim=encoder_dims[i - 1] if i > 0 else encoder_dims[0],
+                    output_dim=encoder_dims[i],
+                    downsample=zipformer_downsampling_factors[i],
+                )
+            encoders.append(encoder)
+        self.encoders = nn.ModuleList(encoders)
+
+        # initializes self.skip_layers and self.skip_modules
+        self._init_skip_modules()
+
+        self.downsample_output = AttentionDownsample(
+            encoder_dims[-1], encoder_dims[-1], downsample=output_downsampling_factor
+        )
+
+    def _get_layer_skip_dropout_prob(self):
+        if not self.training:
+            return 0.0
+        batch_count = self.batch_count
+        min_dropout_prob = 0.025
+
+        if batch_count > self.warmup_end:
+            return min_dropout_prob
+        else:
+            return 0.5 - (batch_count / self.warmup_end) * (0.5 - min_dropout_prob)
+
+    def _init_skip_modules(self):
+        """
+        If self.zipformer_downsampling_factors = (1, 2, 4, 8, 4, 2), then at the input of layer
+        indexed 4 (in zero indexing), which has subsampling_factor=4, we combine the output of
+        layers 2 and 3; and at the input of layer indexed 5, which has subsampling_factor=2,
+        we combine the outputs of layers 1 and 4.
+        """
+        skip_layers = []
+        skip_modules = []
+        z = self.zipformer_downsampling_factors
+        for i in range(len(z)):
+            if i <= 1 or z[i - 1] <= z[i]:
+                skip_layers.append(None)
+                skip_modules.append(SimpleCombinerIdentity())
+            else:
+                # TEMP
+                for j in range(i - 2, -1, -1):
+                    if z[j] <= z[i] or j == 0:
+                        # TEMP logging statement.
+                        logging.info(
+                            f"At encoder stack {i}, which has downsampling_factor={z[i]}, we will "
+                            f"combine the outputs of layers {j} and {i-1}, with downsampling_factors={z[j]} and {z[i-1]}."
+                        )
+                        skip_layers.append(j)
+                        skip_modules.append(
+                            SimpleCombiner(
+                                self.encoder_dims[j],
+                                self.encoder_dims[i - 1],
+                                min_weight=(0.0, 0.25),
+                            )
+                        )
+                        break
+        self.skip_layers = skip_layers
+        self.skip_modules = nn.ModuleList(skip_modules)
+
+    def get_feature_masks(self, x: torch.Tensor) -> List[float]:
+        # Note: The actual return type is Union[List[float], List[Tensor]],
+        # but to make torch.jit.script() work, we use List[float]
+        """
+        In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of
+        randomized feature masks, one per encoder.
+        On e.g. 15% of frames, these masks will zero out all encoder dims larger than
+        some supplied number, e.g. >256, so in effect on those frames we are using
+        a smaller encoder dim.
+
+        We generate the random masks at this level because we want the 2 masks to 'agree'
+        all the way up the encoder stack. This will mean that the 1st mask will have
+        mask values repeated self.zipformer_downsampling_factors times.
+
+        Args:
+           x: the embeddings (needed for the shape and dtype and device), of shape
+             (num_frames, batch_size, encoder_dims0)
+        """
+        num_encoders = len(self.encoder_dims)
+        if torch.jit.is_scripting() or not self.training:
+            return [1.0] * num_encoders
+
+        (num_frames0, batch_size, _encoder_dims0) = x.shape
+
+        assert self.encoder_dims[0] == _encoder_dims0, (
+            self.encoder_dims,
+            _encoder_dims0,
+        )
+
+        max_downsampling_factor = max(self.zipformer_downsampling_factors)
+
+        num_frames_max = num_frames0 + max_downsampling_factor - 1
+
+        feature_mask_dropout_prob = 0.15
+
+        # frame_mask_max shape: (num_frames_max, batch_size, 1)
+        frame_mask_max = (
+            torch.rand(num_frames_max, batch_size, 1, device=x.device)
+            > feature_mask_dropout_prob
+        ).to(x.dtype)
+
+        feature_masks = []
+        for i in range(num_encoders):
+            ds = self.zipformer_downsampling_factors[i]
+            upsample_factor = max_downsampling_factor // ds
+
+            frame_mask = (
+                frame_mask_max.unsqueeze(1)
+                .expand(num_frames_max, upsample_factor, batch_size, 1)
+                .reshape(num_frames_max * upsample_factor, batch_size, 1)
+            )
+            num_frames = (num_frames0 + ds - 1) // ds
+            frame_mask = frame_mask[:num_frames]
+            feature_mask = torch.ones(
+                num_frames,
+                batch_size,
+                self.encoder_dims[i],
+                dtype=x.dtype,
+                device=x.device,
+            )
+            u = self.encoder_unmasked_dims[i]
+            feature_mask[:, :, u:] *= frame_mask
+            feature_masks.append(feature_mask)
+
+        return feature_masks
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          chunk_size:
+            The chunk size used in evaluation mode.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1])
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        lengths = (x_lens - 7) >> 1
+        assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max())
+        mask = make_pad_mask(lengths)
+
+        outputs = []
+        feature_masks = self.get_feature_masks(x)
+
+        if self.training:
+            # Training mode
+            max_ds = max(self.zipformer_downsampling_factors)
+            # Generate dynamic chunk-wise attention mask during training
+            max_len = x.size(0) // max_ds
+            short_chunk_size = self.short_chunk_size // max_ds
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * self.short_chunk_threshold):
+                # Full attention
+                chunk_size = x.size(0)
+            else:
+                # Chunk-wise attention
+                chunk_size = chunk_size % short_chunk_size + 1
+                chunk_size *= max_ds
+        else:
+            chunk_size = self.decode_chunk_size
+            # Evaluation mode
+            for ds in self.zipformer_downsampling_factors:
+                assert chunk_size % ds == 0, (chunk_size, ds)
+
+        attn_mask = ~subsequent_chunk_mask(
+            size=x.size(0),
+            chunk_size=chunk_size,
+            num_left_chunks=self.num_left_chunks,
+            device=x.device,
+        )
+
+        for i, (module, skip_module) in enumerate(
+            zip(self.encoders, self.skip_modules)
+        ):
+            ds = self.zipformer_downsampling_factors[i]
+            k = self.skip_layers[i]
+            if isinstance(k, int):
+                layer_skip_dropout_prob = self._get_layer_skip_dropout_prob()
+                if torch.jit.is_scripting():
+                    x = skip_module(outputs[k], x)
+                elif (not self.training) or random.random() > layer_skip_dropout_prob:
+                    x = skip_module(outputs[k], x)
+            x = module(
+                x,
+                feature_mask=feature_masks[i],
+                src_key_padding_mask=None if mask is None else mask[..., ::ds],
+                attn_mask=attn_mask[::ds, ::ds],
+            )
+            outputs.append(x)
+
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        lengths = (lengths + 1) >> 1
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: List[Tensor],
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+            seq_len is the input chunk length.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states:
+            A list of 7 * num_encoders elements:
+            ``states[0:num_encoders]`` is the cached numbers of past frames.
+            ``states[num_encoders:2*num_encoders]`` is the cached average tensors.
+            ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+            ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+            ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+            ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+            ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1])
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+            - updated states.
+        """
+        assert len(states) == 7 * self.num_encoders, (len(states), self.num_encoders)
+
+        cached_len = states[: self.num_encoders]
+        cached_avg = states[self.num_encoders : 2 * self.num_encoders]
+        cached_key = states[2 * self.num_encoders : 3 * self.num_encoders]
+        cached_val = states[3 * self.num_encoders : 4 * self.num_encoders]
+        cached_val2 = states[4 * self.num_encoders : 5 * self.num_encoders]
+        cached_conv1 = states[5 * self.num_encoders : 6 * self.num_encoders]
+        cached_conv2 = states[6 * self.num_encoders : 7 * self.num_encoders]
+
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        lengths = (x_lens - 7) >> 1
+        assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max())
+
+        outputs = []
+        new_cached_len = []
+        new_cached_avg = []
+        new_cached_key = []
+        new_cached_val = []
+        new_cached_val2 = []
+        new_cached_conv1 = []
+        new_cached_conv2 = []
+
+        for i, (module, skip_module) in enumerate(
+            zip(self.encoders, self.skip_modules)
+        ):
+            k = self.skip_layers[i]
+            if isinstance(k, int):
+                x = skip_module(outputs[k], x)
+            x, len_avg, avg, key, val, val2, conv1, conv2 = module.streaming_forward(
+                x,
+                cached_len=cached_len[i],
+                cached_avg=cached_avg[i],
+                cached_key=cached_key[i],
+                cached_val=cached_val[i],
+                cached_val2=cached_val2[i],
+                cached_conv1=cached_conv1[i],
+                cached_conv2=cached_conv2[i],
+            )
+            outputs.append(x)
+            # Update caches
+            new_cached_len.append(len_avg)
+            new_cached_avg.append(avg)
+            new_cached_key.append(key)
+            new_cached_val.append(val)
+            new_cached_val2.append(val2)
+            new_cached_conv1.append(conv1)
+            new_cached_conv2.append(conv2)
+
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        lengths = (lengths + 1) >> 1
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = (
+            new_cached_len
+            + new_cached_avg
+            + new_cached_key
+            + new_cached_val
+            + new_cached_val2
+            + new_cached_conv1
+            + new_cached_conv2
+        )
+        return x, lengths, new_states
+
+    @torch.jit.export
+    def get_init_state(
+        self,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[Tensor]:
+        """Get initial states.
+        A list of 7 * num_encoders elements:
+        ``states[0:num_encoders]`` is the cached numbers of past frames.
+        ``states[num_encoders:2*num_encoders]`` is the cached average tensors.
+        ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+        ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+        ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+        ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+        ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+        """
+        cached_len = []
+        cached_avg = []
+        cached_key = []
+        cached_val = []
+        cached_val2 = []
+        cached_conv1 = []
+        cached_conv2 = []
+
+        left_context_len = self.decode_chunk_size * self.num_left_chunks
+
+        for i, encoder in enumerate(self.encoders):
+            num_layers = encoder.num_layers
+            ds = self.zipformer_downsampling_factors[i]
+
+            len_avg = torch.zeros(num_layers, 1, dtype=torch.int64, device=device)
+            cached_len.append(len_avg)
+
+            avg = torch.zeros(num_layers, 1, encoder.d_model, device=device)
+            cached_avg.append(avg)
+
+            key = torch.zeros(
+                num_layers,
+                left_context_len // ds,
+                1,
+                encoder.attention_dim,
+                device=device,
+            )
+            cached_key.append(key)
+
+            val = torch.zeros(
+                num_layers,
+                left_context_len // ds,
+                1,
+                encoder.attention_dim // 2,
+                device=device,
+            )
+            cached_val.append(val)
+
+            val2 = torch.zeros(
+                num_layers,
+                left_context_len // ds,
+                1,
+                encoder.attention_dim // 2,
+                device=device,
+            )
+            cached_val2.append(val2)
+
+            conv1 = torch.zeros(
+                num_layers,
+                1,
+                encoder.d_model,
+                encoder.cnn_module_kernel - 1,
+                device=device,
+            )
+            cached_conv1.append(conv1)
+
+            conv2 = torch.zeros(
+                num_layers,
+                1,
+                encoder.d_model,
+                encoder.cnn_module_kernel - 1,
+                device=device,
+            )
+            cached_conv2.append(conv2)
+
+        states = (
+            cached_len
+            + cached_avg
+            + cached_key
+            + cached_val
+            + cached_val2
+            + cached_conv1
+            + cached_conv2
+        )
+        return states
+
+
+class ZipformerEncoderLayer(nn.Module):
+    """
+    ZipformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        feedforward_dim: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        attention_dim: int,
+        nhead: int,
+        feedforward_dim: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        pos_dim: int = 4,
+    ) -> None:
+        super(ZipformerEncoderLayer, self).__init__()
+
+        self.d_model = d_model
+        self.attention_dim = attention_dim
+        self.cnn_module_kernel = cnn_module_kernel
+
+        # will be written to, see set_batch_count()
+        self.batch_count = 0
+
+        self.self_attn = RelPositionMultiheadAttention(
+            d_model,
+            attention_dim,
+            nhead,
+            pos_dim,
+            dropout=0.0,
+        )
+
+        self.pooling = PoolingModule(d_model)
+
+        self.feed_forward1 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.feed_forward2 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.feed_forward3 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.conv_module1 = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.conv_module2 = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        self.bypass_scale = nn.Parameter(torch.tensor(0.5))
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            d_model,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            max_abs=6.0,
+        )
+        self.whiten = Whiten(
+            num_groups=1, whitening_limit=5.0, prob=(0.025, 0.25), grad_scale=0.01
+        )
+
+    def get_bypass_scale(self):
+        if torch.jit.is_scripting() or not self.training:
+            return self.bypass_scale
+        if random.random() < 0.1:
+            # ensure we get grads if self.bypass_scale becomes out of range
+            return self.bypass_scale
+        # hardcode warmup period for bypass scale
+        warmup_period = 20000.0
+        initial_clamp_min = 0.75
+        final_clamp_min = 0.25
+        if self.batch_count > warmup_period:
+            clamp_min = final_clamp_min
+        else:
+            clamp_min = initial_clamp_min - (self.batch_count / warmup_period) * (
+                initial_clamp_min - final_clamp_min
+            )
+        return self.bypass_scale.clamp(min=clamp_min, max=1.0)
+
+    def get_dynamic_dropout_rate(self):
+        # return dropout rate for the dynamic modules (self_attn, pooling, convolution); this
+        # starts at 0.2 and rapidly decreases to 0.  Its purpose is to keep the training stable
+        # at the beginning, by making the network focus on the feedforward modules.
+        if torch.jit.is_scripting() or not self.training:
+            return 0.0
+        warmup_period = 2000.0
+        initial_dropout_rate = 0.2
+        final_dropout_rate = 0.0
+        if self.batch_count > warmup_period:
+            return final_dropout_rate
+        else:
+            return initial_dropout_rate - (
+                initial_dropout_rate - final_dropout_rate
+            ) * (self.batch_count / warmup_period)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+        batch_split: if not None, this layer will only be applied to
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        # macaron style feed forward module
+        src = src + self.feed_forward1(src)
+
+        # dropout rate for submodules that interact with time.
+        dynamic_dropout = self.get_dynamic_dropout_rate()
+
+        # pooling module
+        if torch.jit.is_scripting():
+            src = src + self.pooling(src, src_key_padding_mask=src_key_padding_mask)
+        elif random.random() >= dynamic_dropout:
+            src = src + self.pooling(src, src_key_padding_mask=src_key_padding_mask)
+
+        if torch.jit.is_scripting():
+            src_att, attn_weights = self.self_attn(
+                src,
+                pos_emb=pos_emb,
+                attn_mask=attn_mask,
+                key_padding_mask=src_key_padding_mask,
+            )
+            src = src + src_att
+
+            src = src + self.conv_module1(
+                src, src_key_padding_mask=src_key_padding_mask
+            )
+
+            src = src + self.feed_forward2(src)
+
+            src = src + self.self_attn.forward2(src, attn_weights)
+
+            src = src + self.conv_module2(
+                src, src_key_padding_mask=src_key_padding_mask
+            )
+        else:
+            use_self_attn = random.random() >= dynamic_dropout
+            if use_self_attn:
+                src_att, attn_weights = self.self_attn(
+                    src,
+                    pos_emb=pos_emb,
+                    attn_mask=attn_mask,
+                    key_padding_mask=src_key_padding_mask,
+                )
+                src = src + src_att
+
+            if random.random() >= dynamic_dropout:
+                src = src + self.conv_module1(
+                    src, src_key_padding_mask=src_key_padding_mask
+                )
+
+            src = src + self.feed_forward2(src)
+
+            if use_self_attn:
+                src = src + self.self_attn.forward2(src, attn_weights)
+
+            if random.random() >= dynamic_dropout:
+                src = src + self.conv_module2(
+                    src, src_key_padding_mask=src_key_padding_mask
+                )
+
+        src = src + self.feed_forward3(src)
+
+        src = self.norm_final(self.balancer(src))
+
+        delta = src - src_orig
+
+        src = src_orig + delta * self.get_bypass_scale()
+
+        return self.whiten(src)
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            cached_len: processed number of past frames.
+            cached_avg: cached average of past frames.
+            cached_key: cached key tensor of left context for the first attention module.
+            cached_val: cached value tensor of left context for the first attention module.
+            cached_val2: cached value tensor of left context for the second attention module.
+            cached_conv1: cached left context for the first convolution module.
+            cached_conv2: cached left context for the second convolution module.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, left_context_len+2*S-1, E)
+            cached_len: (N,)
+              N is the batch size.
+            cached_avg: (N, C).
+              N is the batch size, C is the feature dimension.
+            cached_key: (left_context_len, N, K).
+              N is the batch size, K is the key dimension.
+            cached_val: (left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_val2: (left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_conv1: (N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+            cached_conv2: (N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+        """
+        src_orig = src
+
+        # macaron style feed forward module
+        src = src + self.feed_forward1(src)
+
+        src_pool, cached_len, cached_avg = self.pooling.streaming_forward(
+            src,
+            cached_len=cached_len,
+            cached_avg=cached_avg,
+        )
+        src = src + src_pool
+
+        (
+            src_attn,
+            attn_weights,
+            cached_key,
+            cached_val,
+        ) = self.self_attn.streaming_forward(
+            src,
+            pos_emb=pos_emb,
+            cached_key=cached_key,
+            cached_val=cached_val,
+        )
+        src = src + src_attn
+
+        src_conv, cached_conv1 = self.conv_module1.streaming_forward(
+            src,
+            cache=cached_conv1,
+        )
+        src = src + src_conv
+
+        src = src + self.feed_forward2(src)
+
+        src_attn, cached_val2 = self.self_attn.streaming_forward2(
+            src,
+            attn_weights,
+            cached_val=cached_val2,
+        )
+        src = src + src_attn
+
+        src_conv, cached_conv2 = self.conv_module2.streaming_forward(
+            src,
+            cache=cached_conv2,
+        )
+        src = src + src_conv
+
+        src = src + self.feed_forward3(src)
+
+        src = self.norm_final(self.balancer(src))
+
+        delta = src - src_orig
+
+        src = src_orig + delta * self.bypass_scale
+
+        return (
+            src,
+            cached_len,
+            cached_avg,
+            cached_key,
+            cached_val,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        )
+
+
+class ZipformerEncoder(nn.Module):
+    r"""ZipformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ZipformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8)
+        >>> zipformer_encoder = ZipformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = zipformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        dropout: float,
+        warmup_begin: float,
+        warmup_end: float,
+    ) -> None:
+        super().__init__()
+        # will be written to, see set_batch_count() Note: in inference time this
+        # may be zero but should be treated as large, we can check if
+        # self.training is true.
+        self.batch_count = 0
+        self.warmup_begin = warmup_begin
+        self.warmup_end = warmup_end
+        # module_seed is for when we need a random number that is unique to the module but
+        # shared across jobs.   It's used to randomly select how many layers to drop,
+        # so that we can keep this consistent across worker tasks (for efficiency).
+        self.module_seed = torch.randint(0, 1000, ()).item()
+
+        self.encoder_pos = RelPositionalEncoding(encoder_layer.d_model, dropout)
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        self.d_model = encoder_layer.d_model
+        self.attention_dim = encoder_layer.attention_dim
+        self.cnn_module_kernel = encoder_layer.cnn_module_kernel
+
+        assert 0 <= warmup_begin <= warmup_end, (warmup_begin, warmup_end)
+
+        delta = (1.0 / num_layers) * (warmup_end - warmup_begin)
+        cur_begin = warmup_begin
+        for i in range(num_layers):
+            self.layers[i].warmup_begin = cur_begin
+            cur_begin += delta
+            self.layers[i].warmup_end = cur_begin
+
+    def get_layers_to_drop(self, rnd_seed: int):
+        ans = set()
+        if not self.training:
+            return ans
+
+        batch_count = self.batch_count
+        num_layers = len(self.layers)
+
+        def get_layerdrop_prob(layer: int) -> float:
+            layer_warmup_begin = self.layers[layer].warmup_begin
+            layer_warmup_end = self.layers[layer].warmup_end
+
+            initial_layerdrop_prob = 0.5
+            final_layerdrop_prob = 0.05
+
+            if batch_count == 0:
+                # As a special case, if batch_count == 0, return 0 (drop no
+                # layers).  This is rather ugly, I'm afraid; it is intended to
+                # enable our scan_pessimistic_batches_for_oom() code to work correctly
+                # so if we are going to get OOM it will happen early.
+                # also search for 'batch_count' with quotes in this file to see
+                # how we initialize the warmup count to a random number between
+                # 0 and 10.
+                return 0.0
+            elif batch_count < layer_warmup_begin:
+                return initial_layerdrop_prob
+            elif batch_count > layer_warmup_end:
+                return final_layerdrop_prob
+            else:
+                # linearly interpolate
+                t = (batch_count - layer_warmup_begin) / layer_warmup_end
+                assert 0.0 <= t < 1.001, t
+                return initial_layerdrop_prob + t * (
+                    final_layerdrop_prob - initial_layerdrop_prob
+                )
+
+        shared_rng = random.Random(batch_count + self.module_seed)
+        independent_rng = random.Random(rnd_seed)
+
+        layerdrop_probs = [get_layerdrop_prob(i) for i in range(num_layers)]
+        tot = sum(layerdrop_probs)
+        # Instead of drawing the samples independently, we first randomly decide
+        # how many layers to drop out, using the same random number generator between
+        # jobs so that all jobs drop out the same number (this is for speed).
+        # Then we use an approximate approach to drop out the individual layers
+        # with their specified probs while reaching this exact target.
+        num_to_drop = int(tot) + int(shared_rng.random() < (tot - int(tot)))
+
+        layers = list(range(num_layers))
+        independent_rng.shuffle(layers)
+
+        # go through the shuffled layers until we get the required number of samples.
+        if num_to_drop > 0:
+            for layer in itertools.cycle(layers):
+                if independent_rng.random() < layerdrop_probs[layer]:
+                    ans.add(layer)
+                if len(ans) == num_to_drop:
+                    break
+        if shared_rng.random() < 0.005 or __name__ == "__main__":
+            logging.info(
+                f"warmup_begin={self.warmup_begin:.1f}, warmup_end={self.warmup_end:.1f}, "
+                f"batch_count={batch_count:.1f}, num_to_drop={num_to_drop}, layers_to_drop={ans}"
+            )
+        return ans
+
+    def forward(
+        self,
+        src: Tensor,
+        # Note: The type of feature_mask should be Union[float, Tensor],
+        # but to make torch.jit.script() work, we use `float` here
+        feature_mask: float = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        Returns: (x, x_no_combine), both of shape (S, N, E)
+        """
+        pos_emb = self.encoder_pos(src)
+        output = src
+
+        if torch.jit.is_scripting():
+            layers_to_drop = []
+        else:
+            rnd_seed = src.numel() + random.randint(0, 1000)
+            layers_to_drop = self.get_layers_to_drop(rnd_seed)
+
+        output = output * feature_mask
+
+        for i, mod in enumerate(self.layers):
+            if not torch.jit.is_scripting():
+                if i in layers_to_drop:
+                    continue
+            output = mod(
+                output,
+                pos_emb,
+                attn_mask=attn_mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+            output = output * feature_mask
+
+        return output
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        src: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            cached_len: number of past frames.
+            cached_avg: cached average of past frames.
+            cached_key: cached key tensor for first attention module.
+            cached_val: cached value tensor for first attention module.
+            cached_val2: cached value tensor for second attention module.
+            cached_conv1: cached left contexts for the first convolution module.
+            cached_conv2: cached left contexts for the second convolution module.
+
+        Shape:
+            src: (S, N, E).
+            cached_len: (num_layers,)
+            cached_avg: (num_layers, N, C).
+              N is the batch size, C is the feature dimension.
+            cached_key: (num_layers, left_context_len, N, K).
+              N is the batch size, K is the key dimension.
+            cached_val: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_val2: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_conv1: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+            cached_conv2: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+
+        Returns: A tuple of 8 tensors:
+            - output tensor
+            - updated cached number of past frames.
+            - updated cached average of past frames.
+            - updated cached key tensor of of the first attention module.
+            - updated cached value tensor of of the first attention module.
+            - updated cached value tensor of of the second attention module.
+            - updated cached left contexts of the first convolution module.
+            - updated cached left contexts of the second convolution module.
+        """
+        assert cached_len.size(0) == self.num_layers, (
+            cached_len.size(0),
+            self.num_layers,
+        )
+        assert cached_avg.size(0) == self.num_layers, (
+            cached_avg.size(0),
+            self.num_layers,
+        )
+        assert cached_key.size(0) == self.num_layers, (
+            cached_key.size(0),
+            self.num_layers,
+        )
+        assert cached_val.size(0) == self.num_layers, (
+            cached_val.size(0),
+            self.num_layers,
+        )
+        assert cached_val2.size(0) == self.num_layers, (
+            cached_val2.size(0),
+            self.num_layers,
+        )
+        assert cached_conv1.size(0) == self.num_layers, (
+            cached_conv1.size(0),
+            self.num_layers,
+        )
+        assert cached_conv2.size(0) == self.num_layers, (
+            cached_conv2.size(0),
+            self.num_layers,
+        )
+
+        left_context_len = cached_key.shape[1]
+        pos_emb = self.encoder_pos(src, left_context_len)
+        output = src
+
+        new_cached_len = []
+        new_cached_avg = []
+        new_cached_key = []
+        new_cached_val = []
+        new_cached_val2 = []
+        new_cached_conv1 = []
+        new_cached_conv2 = []
+        for i, mod in enumerate(self.layers):
+            output, len_avg, avg, key, val, val2, conv1, conv2 = mod.streaming_forward(
+                output,
+                pos_emb,
+                cached_len=cached_len[i],
+                cached_avg=cached_avg[i],
+                cached_key=cached_key[i],
+                cached_val=cached_val[i],
+                cached_val2=cached_val2[i],
+                cached_conv1=cached_conv1[i],
+                cached_conv2=cached_conv2[i],
+            )
+            # Update caches
+            new_cached_len.append(len_avg)
+            new_cached_avg.append(avg)
+            new_cached_key.append(key)
+            new_cached_val.append(val)
+            new_cached_val2.append(val2)
+            new_cached_conv1.append(conv1)
+            new_cached_conv2.append(conv2)
+
+        return (
+            output,
+            torch.stack(new_cached_len, dim=0),
+            torch.stack(new_cached_avg, dim=0),
+            torch.stack(new_cached_key, dim=0),
+            torch.stack(new_cached_val, dim=0),
+            torch.stack(new_cached_val2, dim=0),
+            torch.stack(new_cached_conv1, dim=0),
+            torch.stack(new_cached_conv2, dim=0),
+        )
+
+
+class DownsampledZipformerEncoder(nn.Module):
+    r"""
+    DownsampledZipformerEncoder is a zipformer encoder evaluated at a reduced frame rate,
+    after convolutional downsampling, and then upsampled again at the output, and combined
+    with the origin input, so that the output has the same shape as the input.
+    """
+
+    def __init__(
+        self, encoder: nn.Module, input_dim: int, output_dim: int, downsample: int
+    ):
+        super(DownsampledZipformerEncoder, self).__init__()
+        self.downsample_factor = downsample
+        self.downsample = AttentionDownsample(input_dim, output_dim, downsample)
+        self.encoder = encoder
+        self.num_layers = encoder.num_layers
+        self.d_model = encoder.d_model
+        self.attention_dim = encoder.attention_dim
+        self.cnn_module_kernel = encoder.cnn_module_kernel
+        self.upsample = SimpleUpsample(output_dim, downsample)
+        self.out_combiner = SimpleCombiner(
+            input_dim, output_dim, min_weight=(0.0, 0.25)
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        # Note: the type of feature_mask should be Unino[float, Tensor],
+        # but to make torch.jit.script() happ, we use float here
+        feature_mask: float = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer.  feature_mask is expected to be already downsampled by
+               self.downsample_factor.
+            attn_mask: attention mask (optional). Should be downsampled already.
+            src_key_padding_mask: the mask for the src keys per batch (optional).  Should be downsampled already.
+
+        Shape:
+            src: (S, N, E).
+            attn_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        Returns: output of shape (S, N, F) where F is the number of output features
+            (output_dim to constructor)
+        """
+        src_orig = src
+        src = self.downsample(src)
+
+        src = self.encoder(
+            src,
+            feature_mask=feature_mask,
+            attn_mask=attn_mask,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src)
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required).
+            cached_avg: cached average value of past frames.
+            cached_len: length of past frames.
+            cached_key: cached key tensor for the first attention module.
+            cached_val: cached value tensor for the first attention module.
+            cached_val2: cached value tensor for the second attention module.
+            cached_conv1: cached left context for the first convolution module.
+            cached_conv2: cached left context for the second convolution module.
+
+        Shape:
+            src: (S, N, E).
+            cached_len: (N,)
+              N is the batch size.
+            cached_avg: (num_layers, N, C).
+              N is the batch size, C is the feature dimension.
+            cached_key: (num_layers, left_context_len, N, K).
+              N is the batch size, K is the key dimension.
+            cached_val: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_val2: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_conv1: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+            cached_conv2: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+        Returns: output of shape (S, N, F) where F is the number of output features
+            (output_dim to constructor)
+        """
+        src_orig = src
+        src = self.downsample(src)
+
+        (
+            src,
+            cached_len,
+            cached_avg,
+            cached_key,
+            cached_val,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ) = self.encoder.streaming_forward(
+            src,
+            cached_len=cached_len,
+            cached_avg=cached_avg,
+            cached_key=cached_key,
+            cached_val=cached_val,
+            cached_val2=cached_val2,
+            cached_conv1=cached_conv1,
+            cached_conv2=cached_conv2,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return (
+            self.out_combiner(src_orig, src),
+            cached_len,
+            cached_avg,
+            cached_key,
+            cached_val,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        )
+
+
+class AttentionDownsample(torch.nn.Module):
+    """
+    Does downsampling with attention, by weighted sum, and a projection..
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, downsample: int):
+        super(AttentionDownsample, self).__init__()
+        self.query = nn.Parameter(torch.randn(in_channels) * (in_channels**-0.5))
+
+        # fill in the extra dimensions with a projection of the input
+        if out_channels > in_channels:
+            self.extra_proj = nn.Linear(
+                in_channels * downsample, out_channels - in_channels, bias=False
+            )
+        else:
+            self.extra_proj = None
+        self.downsample = downsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, 1, in_channels)
+        Returns a tensor of shape
+           ( (seq_len+downsample-1)//downsample, batch_size, out_channels)
+        """
+        (seq_len, batch_size, in_channels) = src.shape
+        ds = self.downsample
+        d_seq_len = (seq_len + ds - 1) // ds
+
+        # Pad to an exact multiple of self.downsample
+        if seq_len != d_seq_len * ds:
+            # right-pad src, repeating the last element.
+            pad = d_seq_len * ds - seq_len
+            src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2])
+            src = torch.cat((src, src_extra), dim=0)
+            assert src.shape[0] == d_seq_len * ds, (src.shape[0], d_seq_len, ds)
+
+        src = src.reshape(d_seq_len, ds, batch_size, in_channels)
+        scores = (src * self.query).sum(dim=-1, keepdim=True)
+
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04)
+
+        weights = scores.softmax(dim=1)
+
+        # ans1 is the first `in_channels` channels of the output
+        ans = (src * weights).sum(dim=1)
+        src = src.permute(0, 2, 1, 3).reshape(d_seq_len, batch_size, ds * in_channels)
+
+        if self.extra_proj is not None:
+            ans2 = self.extra_proj(src)
+            ans = torch.cat((ans, ans2), dim=2)
+        return ans
+
+
+class SimpleUpsample(torch.nn.Module):
+    """
+    A very simple form of upsampling that mostly just repeats the input, but
+    also adds a position-specific bias.
+    """
+
+    def __init__(self, num_channels: int, upsample: int):
+        super(SimpleUpsample, self).__init__()
+        self.bias = nn.Parameter(torch.randn(upsample, num_channels) * 0.01)
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, num_channels)
+        Returns a tensor of shape
+           ( (seq_len*upsample), batch_size, num_channels)
+        """
+        upsample = self.bias.shape[0]
+        (seq_len, batch_size, num_channels) = src.shape
+        src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels)
+        src = src + self.bias.unsqueeze(1)
+        src = src.reshape(seq_len * upsample, batch_size, num_channels)
+        return src
+
+
+class SimpleCombinerIdentity(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    def forward(self, src1: Tensor, src2: Tensor) -> Tensor:
+        return src1
+
+
+class SimpleCombiner(torch.nn.Module):
+    """
+    A very simple way of combining 2 vectors of 2 different dims, via a
+    learned weighted combination in the shared part of the dim.
+    Args:
+         dim1: the dimension of the first input, e.g. 256
+         dim2: the dimension of the second input, e.g. 384.
+    The output will have the same dimension as dim2.
+    """
+
+    def __init__(self, dim1: int, dim2: int, min_weight: Tuple[float] = (0.0, 0.0)):
+        super(SimpleCombiner, self).__init__()
+        assert dim2 >= dim1, (dim2, dim1)
+        self.weight1 = nn.Parameter(torch.zeros(()))
+        self.min_weight = min_weight
+
+    def forward(self, src1: Tensor, src2: Tensor) -> Tensor:
+        """
+        src1: (*, dim1)
+        src2: (*, dim2)
+
+        Returns: a tensor of shape (*, dim2)
+        """
+        assert src1.shape[:-1] == src2.shape[:-1], (src1.shape, src2.shape)
+
+        weight1 = self.weight1
+        if not torch.jit.is_scripting():
+            if (
+                self.training
+                and random.random() < 0.25
+                and self.min_weight != (0.0, 0.0)
+            ):
+                weight1 = weight1.clamp(
+                    min=self.min_weight[0], max=1.0 - self.min_weight[1]
+                )
+
+        src1 = src1 * weight1
+        src2 = src2 * (1.0 - weight1)
+
+        src1_dim = src1.shape[-1]
+        src2_dim = src2.shape[-1]
+        if src1_dim != src2_dim:
+            if src1_dim < src2_dim:
+                src1 = torch.nn.functional.pad(src1, (0, src2_dim - src1_dim))
+            else:
+                src1 = src1[:src2_dim]
+
+        return src1 + src2
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        dropout_rate: float,
+        max_len: int = 5000,
+    ) -> None:
+        """Construct a PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(max_len))
+
+    def extend_pe(self, x: Tensor, left_context_len: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_left = x.size(0) + left_context_len
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_left * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use positive relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_left, self.d_model)
+        pe_negative = torch.zeros(x_size_left, self.d_model)
+        position = torch.arange(0, x_size_left, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor, left_context_len: int = 0) -> Tensor:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (time, batch, `*`).
+            left_context_len: (int): Length of cached left context.
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, left_context_len + 2*time-1, `*`).
+
+        """
+        self.extend_pe(x, left_context_len)
+        x_size_left = x.size(0) + left_context_len
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_left
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(0),
+        ]
+        return self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
+    we have to write up the differences.
+
+
+    Args:
+        embed_dim: total dimension of the model.
+        attention_dim: dimension in the attention module, may be less or more than embed_dim
+            but must be a multiple of num_heads.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        attention_dim: int,
+        num_heads: int,
+        pos_dim: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.attention_dim = attention_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = attention_dim // num_heads
+        self.pos_dim = pos_dim
+        assert self.head_dim % 2 == 0, self.head_dim
+        assert self.head_dim * num_heads == attention_dim, (
+            self.head_dim,
+            num_heads,
+            attention_dim,
+        )
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5, dividing it between the query and key.
+        in_proj_dim = (
+            2 * attention_dim  # query, key
+            + attention_dim // 2  # value
+            + pos_dim * num_heads  # positional encoding query
+        )
+
+        self.in_proj = ScaledLinear(
+            embed_dim, in_proj_dim, bias=True, initial_scale=self.head_dim**-0.25
+        )
+
+        # self.whiten_values is applied on the values in forward();
+        # it just copies the keys but prevents low-rank distribution by modifying grads.
+        self.whiten_values = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(
+            embed_dim, num_heads * pos_dim, bias=False, initial_scale=0.05
+        )
+
+        # the following are for diagnosics only, see --print-diagnostics option.
+        # they only copy their inputs.
+        self.copy_pos_query = Identity()
+        self.copy_query = Identity()
+
+        self.out_proj = ScaledLinear(
+            attention_dim // 2, embed_dim, bias=True, initial_scale=0.05
+        )
+
+        self.in_proj2 = nn.Linear(embed_dim, attention_dim // 2, bias=False)
+        self.out_proj2 = ScaledLinear(
+            attention_dim // 2, embed_dim, bias=True, initial_scale=0.05
+        )
+        # self.whiten_values2 is applied on the values in forward2()
+        self.whiten_values2 = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x: input to be projected to query, key, value
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Returns: (attn_output, attn_weights)
+
+             - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size,
+                E is the embedding dimension.
+              - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads
+                 and S is the sequence length.
+        """
+        x, weights = self.multi_head_attention_forward(
+            self.in_proj(x),
+            self.linear_pos(pos_emb),
+            self.attention_dim,
+            self.num_heads,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            attn_mask=attn_mask,
+        )
+        return x, weights
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
+        r"""
+        Args:
+            x: input to be projected to query, key, value
+            pos_emb: Positional embedding tensor
+
+        Shape:
+            - Inputs:
+            - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - cached_key: :math:`(left_context_len, N, K)`, where N is the batch size, K is the key dimension.
+            - cached_val: :math:`(left_context_len, N, V)`, where N is the batch size, V is the value dimension.
+
+            - Returns: (attn_output, attn_weights, cached_key, cached_val)
+
+              - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size,
+                E is the embedding dimension.
+              - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads
+                 and S is the sequence length.
+              - cached_key: :math:`(left_context_len, N, K)`, updated cached attention key tensor of
+                left context
+              - cached_val: :math:`(left_context_len, N, K)`, updated cached attention value tensor of
+        """
+        (
+            x,
+            weights,
+            cached_key,
+            cached_val,
+        ) = self.streaming_multi_head_attention_forward(
+            self.in_proj(x),
+            self.linear_pos(pos_emb),
+            self.attention_dim,
+            self.num_heads,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            cached_key=cached_key,
+            cached_val=cached_val,
+        )
+        return x, weights, cached_key, cached_val
+
+    def multi_head_attention_forward(
+        self,
+        x_proj: Tensor,
+        pos: Tensor,
+        attention_dim: int,
+        num_heads: int,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x_proj: the projected input, to be split into query, key, value.
+            pos: head-specific biases arising from the positional embeddings.
+            attention_dim: dimension inside attention mechanism
+            num_heads: parallel attention heads.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is
+              the attention dimension.  Will be split into (query, key, value, pos).
+            - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence
+              length, N is the batch size, and A is the attention dim.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+              E is the embedding dimension.
+            - attn_weights: :math:`(N * H, S, S)` where N is the batch size,
+              H is the num-heads, S is the sequence length.
+        """
+
+        seq_len, bsz, _ = x_proj.size()
+
+        head_dim = attention_dim // num_heads
+        pos_dim = self.pos_dim  # positional-encoding dim per head
+        assert (
+            head_dim * num_heads == attention_dim
+        ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}"
+
+        # self-attention
+        q = x_proj[..., 0:attention_dim]
+        k = x_proj[..., attention_dim : 2 * attention_dim]
+        value_dim = attention_dim // 2
+        v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim]
+        # p is the position-encoding query, its dimension is num_heads*pos_dim..
+        p = x_proj[..., 2 * attention_dim + value_dim :]
+
+        k = self.whiten_keys(k)  # does nothing in the forward pass.
+        v = self.whiten_values(v)  # does nothing in the forward pass.
+        q = self.copy_query(q)  # for diagnostics only, does nothing.
+        p = self.copy_pos_query(p)  # for diagnostics only, does nothing.
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, seq_len, seq_len]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    seq_len,
+                    seq_len,
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.reshape(seq_len, bsz, num_heads, head_dim)
+        p = p.reshape(seq_len, bsz, num_heads, pos_dim)
+        k = k.reshape(seq_len, bsz, num_heads, head_dim)
+        v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == seq_len, "{} == {}".format(
+                key_padding_mask.size(1), seq_len
+            )
+
+        q = q.permute(1, 2, 0, 3)  # (batch, head, time1, head_dim)
+        p = p.permute(1, 2, 0, 3)  # (batch, head, time1, pos_dim)
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+
+        seq_len2 = 2 * seq_len - 1
+        pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1)
+        # pos shape now: (batch, head, pos_dim, seq_len2)
+
+        # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2)
+        #  [where seq_len2 represents relative position.]
+        pos_weights = torch.matmul(p, pos)
+        # the following .as_strided() expression converts the last axis of pos_weights from relative
+        # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+        # not, but let this code define which way round it is supposed to be.
+        if torch.jit.is_tracing():
+            (batch_size, num_heads, time1, n) = pos_weights.shape
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(seq_len)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+            pos_weights = pos_weights.reshape(-1, n)
+            pos_weights = torch.gather(pos_weights, dim=1, index=indexes)
+            pos_weights = pos_weights.reshape(batch_size, num_heads, time1, seq_len)
+        else:
+            pos_weights = pos_weights.as_strided(
+                (bsz, num_heads, seq_len, seq_len),
+                (
+                    pos_weights.stride(0),
+                    pos_weights.stride(1),
+                    pos_weights.stride(2) - pos_weights.stride(3),
+                    pos_weights.stride(3),
+                ),
+                storage_offset=pos_weights.stride(3) * (seq_len - 1),
+            )
+
+        # caution: they are really scores at this point.
+        attn_output_weights = torch.matmul(q, k) + pos_weights
+
+        if not torch.jit.is_scripting():
+            if training and random.random() < 0.1:
+                # This is a harder way of limiting the attention scores to not be too large.
+                # It incurs a penalty if any of them has an absolute value greater than 50.0.
+                # this should be outside the normal range of the attention scores.  We use
+                # this mechanism instead of, say, a limit on entropy, because once the entropy
+                # gets very small gradients through the softmax can become very small, and
+                # some mechanisms like that become ineffective.
+                attn_output_weights = penalize_abs_values_gt(
+                    attn_output_weights, limit=25.0, penalty=1.0e-04
+                )
+
+        # attn_output_weights: (batch, head, time1, time2)
+        attn_output_weights = attn_output_weights.view(
+            bsz * num_heads, seq_len, seq_len
+        )
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights = attn_output_weights.masked_fill(
+                    attn_mask, float("-inf")
+                )
+            else:
+                attn_output_weights = attn_output_weights + attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, seq_len, seq_len
+            )
+
+        # Using this version of softmax, defined in scaling.py,
+        # should save a little of the memory used in backprop by, if
+        # we are in automatic mixed precision mode (amp) == autocast,
+        # only storing the half-precision output for backprop purposes.
+        attn_output_weights = softmax(attn_output_weights, dim=-1)
+
+        # If we are using chunk-wise attention mask and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`. At this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax). So we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, seq_len, seq_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, seq_len, seq_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2]
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, attention_dim // 2)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        return attn_output, attn_output_weights
+
+    def streaming_multi_head_attention_forward(
+        self,
+        x_proj: Tensor,
+        pos: Tensor,
+        attention_dim: int,
+        num_heads: int,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
+        r"""
+        Args:
+            x_proj: the projected input, to be split into query, key, value.
+            pos: head-specific biases arising from the positional embeddings.
+            attention_dim: dimension inside attention mechanism
+            num_heads: parallel attention heads.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            cached_key: cached attention key tensor of left context.
+            cached_val: cached attention value tensor of left context.
+
+        Shape:
+            Inputs:
+            - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is
+              the attention dimension.  Will be split into (query, key, value, pos).
+            - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence
+              length, N is the batch size, and A is the attention dim.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+              E is the embedding dimension.
+            - attn_weights: :math:`(N * H, S, S)` where N is the batch size,
+              H is the num-heads, S is the sequence length.
+            - cached_key: :math:`(left_context_len, N, K)`, updated cached attention key tensor of left context.
+            - cached_val: :math:`(left_context_len, N, K)`, updated cached attention value tensor of left context.
+        """
+
+        seq_len, bsz, _ = x_proj.size()
+
+        head_dim = attention_dim // num_heads
+        pos_dim = self.pos_dim  # positional-encoding dim per head
+        assert (
+            head_dim * num_heads == attention_dim
+        ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}"
+
+        # self-attention
+        q = x_proj[..., 0:attention_dim]
+        k = x_proj[..., attention_dim : 2 * attention_dim]
+        value_dim = attention_dim // 2
+        v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim]
+        # p is the position-encoding query, its dimension is num_heads*pos_dim..
+        p = x_proj[..., 2 * attention_dim + value_dim :]
+
+        left_context_len = cached_key.shape[0]
+        assert left_context_len > 0, left_context_len
+        assert cached_key.shape[0] == cached_val.shape[0], (
+            cached_key.shape,
+            cached_val.shape,
+        )
+        # Pad cached left contexts
+        k = torch.cat([cached_key, k], dim=0)
+        v = torch.cat([cached_val, v], dim=0)
+        # Update cached left contexts
+        cached_key = k[-left_context_len:, ...]
+        cached_val = v[-left_context_len:, ...]
+
+        # The length of key and value
+        kv_len = k.shape[0]
+
+        q = q.reshape(seq_len, bsz, num_heads, head_dim)
+        p = p.reshape(seq_len, bsz, num_heads, pos_dim)
+        k = k.reshape(kv_len, bsz, num_heads, head_dim)
+        v = v.reshape(kv_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        q = q.permute(1, 2, 0, 3)  # (batch, head, time1, head_dim)
+        p = p.permute(1, 2, 0, 3)  # (batch, head, time1, pos_dim)
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+
+        seq_len2 = 2 * seq_len - 1 + left_context_len
+        pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1)
+        # pos shape now: (batch, head, pos_dim, seq_len2)
+
+        # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2)
+        #  [where seq_len2 represents relative position.]
+        pos_weights = torch.matmul(p, pos)
+        # the following .as_strided() expression converts the last axis of pos_weights from relative
+        # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+        # not, but let this code define which way round it is supposed to be.
+        if torch.jit.is_tracing():
+            (batch_size, num_heads, time1, n) = pos_weights.shape
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(kv_len)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+            pos_weights = pos_weights.reshape(-1, n)
+            pos_weights = torch.gather(pos_weights, dim=1, index=indexes)
+            pos_weights = pos_weights.reshape(batch_size, num_heads, time1, kv_len)
+        else:
+            pos_weights = pos_weights.as_strided(
+                (bsz, num_heads, seq_len, kv_len),
+                (
+                    pos_weights.stride(0),
+                    pos_weights.stride(1),
+                    pos_weights.stride(2) - pos_weights.stride(3),
+                    pos_weights.stride(3),
+                ),
+                storage_offset=pos_weights.stride(3) * (seq_len - 1),
+            )
+
+        # caution: they are really scores at this point.
+        attn_output_weights = torch.matmul(q, k) + pos_weights
+
+        # attn_output_weights: (batch, head, time1, time2)
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, seq_len, kv_len)
+
+        # Using this version of softmax, defined in scaling.py,
+        # should save a little of the memory used in backprop by, if
+        # we are in automatic mixed precision mode (amp) == autocast,
+        # only storing the half-precision output for backprop purposes.
+        attn_output_weights = softmax(attn_output_weights, dim=-1)
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2]
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, attention_dim // 2)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        return attn_output, attn_output_weights, cached_key, cached_val
+
+    def forward2(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """
+        Second forward function, where we re-use the attn_weights returned by the first forward function
+        but with different input.
+        Args:
+               x: input, of shape (seq_len, batch_size, embed_dim)
+           attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len)
+        Returns:
+               output of the same shape as x, i.e. (seq_len, batch_size, embed_dim)
+        """
+        num_heads = self.num_heads
+        (seq_len, bsz, embed_dim) = x.shape
+        head_dim = self.attention_dim // num_heads
+        # v: (tgt_len, bsz, embed_dim // 2)
+        v = self.in_proj2(x)
+        v = self.whiten_values2(v)  # does nothing in the forward pass.
+        v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        # now v: (bsz * num_heads, seq_len, head_dim // 2)
+        attn_output = torch.bmm(attn_weights, v)
+
+        if not torch.jit.is_scripting():
+            if random.random() < 0.001 or __name__ == "__main__":
+                self._print_attn_stats(attn_weights, attn_output)
+
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, self.attention_dim // 2)
+        )
+        # returned value is of shape (seq_len, bsz, embed_dim), like x.
+        return self.out_proj2(attn_output)
+
+    def streaming_forward2(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_val: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Second forward function, where we re-use the attn_weights returned by the first forward function
+        but with different input.
+        Args:
+            x: input, of shape (seq_len, batch_size, embed_dim)
+            attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len)
+            cached_val: cached attention value tensor of left context.
+        Returns:
+            - output of the same shape as x, i.e. (seq_len, batch_size, embed_dim)
+            - updated cached attention value tensor of left context.
+        """
+        num_heads = self.num_heads
+        (seq_len, bsz, embed_dim) = x.shape
+        head_dim = self.attention_dim // num_heads
+        # v: (tgt_len, bsz, embed_dim // 2)
+        v = self.in_proj2(x)
+
+        left_context_len = cached_val.shape[0]
+        assert left_context_len > 0, left_context_len
+        v = torch.cat([cached_val, v], dim=0)
+        cached_val = v[-left_context_len:]
+
+        seq_len2 = left_context_len + seq_len
+        v = v.reshape(seq_len2, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        # now v: (bsz * num_heads, seq_len, head_dim // 2)
+        attn_output = torch.bmm(attn_weights, v)
+
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, self.attention_dim // 2)
+        )
+        # returned value is of shape (seq_len, bsz, embed_dim), like x.
+        return self.out_proj2(attn_output), cached_val
+
+    def _print_attn_stats(self, attn_weights: Tensor, attn_output: Tensor):
+        # attn_weights: (batch_size * num_heads, seq_len, seq_len)
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        (n, seq_len, head_dim) = attn_output.shape
+        num_heads = self.num_heads
+        bsz = n // num_heads
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_output = attn_output.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .reshape(bsz, num_heads, seq_len)
+                    .mean(dim=(0, 2))
+                )
+                attn_output = attn_output.reshape(bsz, num_heads, seq_len, head_dim)
+                attn_output = attn_output.permute(1, 0, 2, 3).reshape(
+                    num_heads, bsz * seq_len, head_dim
+                )
+                attn_output_mean = attn_output.mean(dim=1, keepdim=True)
+                attn_output = attn_output - attn_output_mean
+                attn_covar = torch.matmul(attn_output.transpose(1, 2), attn_output) / (
+                    bsz * seq_len
+                )
+                # attn_covar: (num_heads, head_dim, head_dim)
+                # eigs, _ = torch.symeig(attn_covar)
+                # logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}")
+
+                attn_covar = _diag(attn_covar).mean(dim=1)  # (num_heads,)
+                embed_dim = self.in_proj2.weight.shape[1]
+                in_proj_covar = (
+                    self.in_proj2.weight.reshape(num_heads, head_dim, embed_dim) ** 2
+                ).mean(dim=(1, 2))
+                out_proj_covar = (
+                    self.out_proj2.weight.reshape(embed_dim, num_heads, head_dim) ** 2
+                ).mean(dim=(0, 2))
+                logging.info(
+                    f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}"
+                )
+
+
+class PoolingModule(nn.Module):
+    """
+    Averages the input over the time dimension and project with a square matrix.
+    """
+
+    def __init__(self, d_model: int):
+        super().__init__()
+        self.proj = ScaledLinear(d_model, d_model, initial_scale=0.1, bias=False)
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Args:
+           x: a Tensor of shape (T, N, C)
+           src_key_padding_mask: a Tensor of bool, of shape (N, T), with True in masked
+               positions.
+
+        Returns:
+           - output, a Tensor of shape (T, N, C).
+        """
+        if src_key_padding_mask is not None:
+            # False in padding positions
+            padding_mask = src_key_padding_mask.logical_not().to(x.dtype)  # (N, T)
+            # Cumulated numbers of frames from start
+            cum_mask = padding_mask.cumsum(dim=1)  # (N, T)
+            x = x.cumsum(dim=0)  # (T, N, C)
+            pooling_mask = padding_mask / cum_mask
+            pooling_mask = pooling_mask.transpose(0, 1).contiguous().unsqueeze(-1)
+            # now pooling_mask: (T, N, 1)
+            x = x * pooling_mask  # (T, N, C)
+        else:
+            num_frames = x.shape[0]
+            cum_mask = torch.arange(1, num_frames + 1).unsqueeze(1)  # (T, 1)
+            x = x.cumsum(dim=0)  # (T, N, C)
+            pooling_mask = (1.0 / cum_mask).unsqueeze(2)
+            # now pooling_mask: (T, N, 1)
+            x = x * pooling_mask
+
+        x = self.proj(x)
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor]:
+        """
+        Args:
+           x: a Tensor of shape (T, N, C)
+           cached_len: a Tensor of int, of shape (N,), containing the number of
+               past frames in batch.
+           cached_avg: a Tensor of shape (N, C), the average over all past frames
+               in batch.
+
+        Returns:
+           A tuple of 2 tensors:
+           - output, a Tensor of shape (T, N, C).
+           - updated cached_avg, a Tensor of shape (N, C).
+        """
+        x = x.cumsum(dim=0)  # (T, N, C)
+        x = x + (cached_avg * cached_len.unsqueeze(1)).unsqueeze(0)
+        # Cumulated numbers of frames from start
+        cum_mask = torch.arange(1, x.size(0) + 1, device=x.device)
+        cum_mask = cum_mask.unsqueeze(1) + cached_len.unsqueeze(0)  # (T, N)
+        pooling_mask = (1.0 / cum_mask).unsqueeze(2)
+        # now pooling_mask: (T, N, 1)
+        x = x * pooling_mask  # (T, N, C)
+
+        cached_len = cached_len + x.size(0)
+        cached_avg = x[-1]
+
+        x = self.proj(x)
+        return x, cached_len, cached_avg
+
+
+class FeedforwardModule(nn.Module):
+    """Feedforward module in Zipformer model."""
+
+    def __init__(self, d_model: int, feedforward_dim: int, dropout: float):
+        super(FeedforwardModule, self).__init__()
+        self.in_proj = nn.Linear(d_model, feedforward_dim)
+        self.balancer = ActivationBalancer(
+            feedforward_dim, channel_dim=-1, max_abs=10.0, min_prob=0.25
+        )
+        self.activation = DoubleSwish()
+        self.dropout = nn.Dropout(dropout)
+        self.out_proj = ScaledLinear(feedforward_dim, d_model, initial_scale=0.01)
+
+    def forward(self, x: Tensor):
+        x = self.in_proj(x)
+        x = self.balancer(x)
+        x = self.activation(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Zipformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0, kernel_size
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            2 * channels,
+            channel_dim=1,
+            max_abs=10.0,
+            min_positive=0.05,
+            max_positive=1.0,
+        )
+
+        # Will pad cached left context
+        self.lorder = kernel_size - 1
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channels,
+            channel_dim=1,
+            min_positive=0.05,
+            max_positive=1.0,
+            max_abs=20.0,
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.05,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional):
+               (batch, #time), contains bool in masked positions.
+
+        Returns:
+            - Output tensor (#time, batch, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        # 1D Depthwise Conv
+        # Make depthwise_conv causal by
+        # manualy padding self.lorder zeros to the left
+        x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cache: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch:
+               (batch, #time), contains bool in masked positions.
+            cache: Cached left context for depthwise_conv, with shape of
+               (batch, channels, #kernel_size-1). Only used in real streaming decoding.
+
+        Returns:
+            A tuple of 2 tensors:
+            - Output tensor (#time, batch, channels).
+            - New cached left context, with shape of (batch, channels, #kernel_size-1).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        assert cache.shape == (x.size(0), x.size(1), self.lorder), (
+            cache.shape,
+            (x.size(0), x.size(1), self.lorder),
+        )
+        x = torch.cat([cache, x], dim=2)
+        # Update cache
+        cache = x[:, :, -self.lorder :]
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1), cache
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = (T-3)//2 - 2 == (T-7)//2
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        dropout: float = 0.1,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, (T-7)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer2_channels:
+            Number of channels in layer2
+          layer3_channels:
+            Number of channels in layer3
+        """
+        assert in_channels >= 7, in_channels
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=(0, 1),  # (time, freq)
+            ),
+            ActivationBalancer(layer1_channels, channel_dim=1),
+            DoubleSwish(),
+            nn.Conv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+                padding=0,
+            ),
+            ActivationBalancer(layer2_channels, channel_dim=1),
+            DoubleSwish(),
+            nn.Conv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=(1, 2),  # (time, freq)
+            ),
+            ActivationBalancer(layer3_channels, channel_dim=1),
+            DoubleSwish(),
+        )
+        out_height = (((in_channels - 1) // 2) - 1) // 2
+        self.out = ScaledLinear(out_height * layer3_channels, out_channels)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, (T-7)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, (T-7)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).reshape(b, t, c * f))
+        # Now x is of shape (N, (T-7)//2, odim)
+        x = self.dropout(x)
+        return x
+
+
+def _test_zipformer_main():
+    feature_dim = 50
+    batch_size = 5
+    seq_len = 47
+    feature_dim = 50
+    # Just make sure the forward pass runs.
+
+    c = Zipformer(
+        num_features=feature_dim,
+        encoder_dims=(64, 96),
+        encoder_unmasked_dims=(48, 64),
+        nhead=(4, 4),
+        decode_chunk_size=4,
+    )
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    assert ((seq_len - 7) // 2 + 1) // 2 == f[0].shape[1], (seq_len, f.shape[1])
+    f[0].sum().backward()
+    c.eval()
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+def _test_conv2d_subsampling():
+    num_features = 80
+    encoder_dims = 384
+    dropout = 0.1
+    encoder_embed = Conv2dSubsampling(num_features, encoder_dims, dropout=dropout)
+    for i in range(20, 40):
+        x = torch.rand(2, i, num_features)
+        y = encoder_embed(x)
+        assert (x.shape[1] - 7) // 2 == y.shape[1], (x.shape[1], y.shape[1])
+
+
+def _test_pooling_module():
+    N, S, C = 2, 12, 32
+    chunk_len = 4
+    m = PoolingModule(d_model=C)
+
+    # test chunk-wise forward with padding_mask
+    x = torch.randn(S, N, C)
+    y = m(x)
+    cached_len = torch.zeros(N, dtype=torch.int32)
+    cached_avg = torch.zeros(N, C)
+    for i in range(S // chunk_len):
+        start = i * chunk_len
+        end = start + chunk_len
+        x_chunk = x[start:end]
+        y_chunk, cached_len, cached_avg = m.streaming_forward(
+            x_chunk,
+            cached_len=cached_len,
+            cached_avg=cached_avg,
+        )
+        assert torch.allclose(y_chunk, y[start:end]), (y_chunk, y[start:end])
+
+
+def _test_state_stack_unstack():
+    m = Zipformer(
+        num_features=80,
+        encoder_dims=(64, 96),
+        encoder_unmasked_dims=(48, 64),
+        nhead=(4, 4),
+        zipformer_downsampling_factors=(4, 8),
+        num_left_chunks=2,
+        decode_chunk_size=8,
+    )
+    s1 = m.get_init_state()
+    s2 = m.get_init_state()
+    states = stack_states([s1, s2])
+    new_s1, new_s2 = unstack_states(states)
+    for i in range(m.num_encoders * 7):
+        for x, y in zip(s1[i], new_s1[i]):
+            assert torch.equal(x, y)
+        for x, y in zip(s2[i], new_s2[i]):
+            assert torch.equal(x, y)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_zipformer_main()
+    _test_conv2d_subsampling()
+    _test_pooling_module()
+    _test_state_stack_unstack()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
new file mode 100644
index 000000000..5284ed627
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
@@ -0,0 +1,3144 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey,)
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import itertools
+import logging
+import math
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (  # not as in other dirs.. just scales down initial parameter values.
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    Identity,
+    MaxEig,
+    ScaledConv1d,
+    ScaledLinear,
+    Whiten,
+    _diag,
+    penalize_abs_values_gt,
+    random_clamp,
+    softmax,
+)
+from torch import Tensor, nn
+from zipformer import PoolingModule
+
+from icefall.utils import make_pad_mask, subsequent_chunk_mask
+
+
+def stack_states(state_list: List[List[Tensor]]) -> List[Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+        ``states[i]`` is a list of 7 * num_encoders elements of i-th utterance.
+        ``states[i][0:num_encoders]`` is the cached numbers of past frames.
+        ``states[i][num_encoders:2*num_encoders]`` is the cached average tensors.
+        ``states[i][2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+        ``states[i][3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+        ``states[i][4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+        ``states[i][5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+        ``states[i][6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+
+    Returns:
+      A new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+    assert len(state_list[0]) % 7 == 0, len(state_list[0])
+    num_encoders = len(state_list[0]) // 7
+
+    cached_len = []
+    cached_avg = []
+    cached_key = []
+    cached_val = []
+    cached_val2 = []
+    cached_conv1 = []
+    cached_conv2 = []
+
+    # For cached_len
+    len_list = [state_list[n][0:num_encoders] for n in range(batch_size)]
+    for i in range(num_encoders):
+        # len_avg: (num_layers, batch_size)
+        len_avg = torch.cat([len_list[n][i] for n in range(batch_size)], dim=1)
+        cached_len.append(len_avg)
+
+    # For cached_avg
+    avg_list = [
+        state_list[n][num_encoders : 2 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # avg: (num_layers, batch_size, D)
+        avg = torch.cat([avg_list[n][i] for n in range(batch_size)], dim=1)
+        cached_avg.append(avg)
+
+    # For cached_key
+    key_list = [
+        state_list[n][2 * num_encoders : 3 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # key: (num_layers, left_context_size, batch_size, D)
+        key = torch.cat([key_list[n][i] for n in range(batch_size)], dim=2)
+        cached_key.append(key)
+
+    # For cached_val
+    val_list = [
+        state_list[n][3 * num_encoders : 4 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # val: (num_layers, left_context_size, batch_size, D)
+        val = torch.cat([val_list[n][i] for n in range(batch_size)], dim=2)
+        cached_val.append(val)
+
+    # For cached_val2
+    val2_list = [
+        state_list[n][4 * num_encoders : 5 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # val2: (num_layers, left_context_size, batch_size, D)
+        val2 = torch.cat([val2_list[n][i] for n in range(batch_size)], dim=2)
+        cached_val2.append(val2)
+
+    # For cached_conv1
+    conv1_list = [
+        state_list[n][5 * num_encoders : 6 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # conv1: (num_layers, batch_size, D, kernel-1)
+        conv1 = torch.cat([conv1_list[n][i] for n in range(batch_size)], dim=1)
+        cached_conv1.append(conv1)
+
+    # For cached_conv2
+    conv2_list = [
+        state_list[n][6 * num_encoders : 7 * num_encoders] for n in range(batch_size)
+    ]
+    for i in range(num_encoders):
+        # conv2: (num_layers, batch_size, D, kernel-1)
+        conv2 = torch.cat([conv2_list[n][i] for n in range(batch_size)], dim=1)
+        cached_conv2.append(conv2)
+
+    states = (
+        cached_len
+        + cached_avg
+        + cached_key
+        + cached_val
+        + cached_val2
+        + cached_conv1
+        + cached_conv2
+    )
+    return states
+
+
+def unstack_states(states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+      states:
+        A list of 7 * num_encoders elements:
+        ``states[0:num_encoders]`` is the cached numbers of past frames.
+        ``states[num_encoders:2*num_encoders]`` is the cached average tensors.
+        ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+        ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+        ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+        ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+        ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+
+    Returns:
+      A list of states.
+      ``states[i]`` is a list of 7 * num_encoders elements of i-th utterance.
+    """
+    assert len(states) % 7 == 0, len(states)
+    num_encoders = len(states) // 7
+    (
+        cached_len,
+        cached_avg,
+        cached_key,
+        cached_val,
+        cached_val2,
+        cached_conv1,
+        cached_conv2,
+    ) = (states[i * num_encoders : (i + 1) * num_encoders] for i in range(7))
+
+    batch_size = cached_len[0].shape[1]
+
+    len_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_len[i]: (num_layers, batch_size)
+        len_avg = cached_len[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            len_list[n].append(len_avg[n])
+
+    avg_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_avg[i]: (num_layers, batch_size, D)
+        avg = cached_avg[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            avg_list[n].append(avg[n])
+
+    key_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_key[i]: (num_layers, left_context, batch_size, D)
+        key = cached_key[i].chunk(chunks=batch_size, dim=2)
+        for n in range(batch_size):
+            key_list[n].append(key[n])
+
+    val_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_val[i]: (num_layers, left_context, batch_size, D)
+        val = cached_val[i].chunk(chunks=batch_size, dim=2)
+        for n in range(batch_size):
+            val_list[n].append(val[n])
+
+    val2_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_val2[i]: (num_layers, left_context, batch_size, D)
+        val2 = cached_val2[i].chunk(chunks=batch_size, dim=2)
+        for n in range(batch_size):
+            val2_list[n].append(val2[n])
+
+    conv1_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_conv1[i]: (num_layers, batch_size, D, kernel-1)
+        conv1 = cached_conv1[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            conv1_list[n].append(conv1[n])
+
+    conv2_list = [[] for _ in range(batch_size)]
+    for i in range(num_encoders):
+        # cached_conv2[i]: (num_layers, batch_size, D, kernel-1)
+        conv2 = cached_conv2[i].chunk(chunks=batch_size, dim=1)
+        for n in range(batch_size):
+            conv2_list[n].append(conv2[n])
+
+    state_list = [
+        (
+            len_list[i]
+            + avg_list[i]
+            + key_list[i]
+            + val_list[i]
+            + val2_list[i]
+            + conv1_list[i]
+            + conv2_list[i]
+        )
+        for i in range(batch_size)
+    ]
+    return state_list
+
+
+class Zipformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        d_model: (int,int): embedding dimension of 2 encoder stacks
+        attention_dim: (int,int): attention dimension of 2 encoder stacks
+        nhead (int, int): number of heads
+        dim_feedforward (int, int): feedforward dimension in 2 encoder stacks
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        cnn_module_kernels (int): Kernel size of convolution module
+        warmup_batches (float): number of batches to warm up over
+        is_pnnx (bool): True if we are going to convert this model via pnnx.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_downsampling_factor: int = 2,
+        encoder_dims: Tuple[int] = (384, 384),
+        attention_dim: Tuple[int] = (256, 256),
+        encoder_unmasked_dims: Tuple[int] = (256, 256),
+        zipformer_downsampling_factors: Tuple[int] = (2, 4),
+        nhead: Tuple[int] = (8, 8),
+        feedforward_dim: Tuple[int] = (1536, 2048),
+        num_encoder_layers: Tuple[int] = (12, 12),
+        dropout: float = 0.1,
+        cnn_module_kernels: Tuple[int] = (31, 31),
+        pos_dim: int = 4,
+        num_left_chunks: int = 4,
+        short_chunk_threshold: float = 0.75,
+        short_chunk_size: int = 50,
+        decode_chunk_size: int = 16,
+        warmup_batches: float = 4000.0,
+        is_pnnx: bool = False,
+    ) -> None:
+        super(Zipformer, self).__init__()
+        self.is_pnnx = is_pnnx
+
+        self.num_features = num_features
+        assert 0 < encoder_dims[0] <= encoder_dims[1]
+        self.encoder_dims = encoder_dims
+        self.encoder_unmasked_dims = encoder_unmasked_dims
+        self.zipformer_downsampling_factors = zipformer_downsampling_factors
+        self.output_downsampling_factor = output_downsampling_factor
+
+        self.num_left_chunks = num_left_chunks
+        self.short_chunk_threshold = short_chunk_threshold
+        self.short_chunk_size = short_chunk_size
+
+        # Used in decoding
+        self.decode_chunk_size = decode_chunk_size
+
+        self.left_context_len = self.decode_chunk_size * self.num_left_chunks
+
+        # will be written to, see set_batch_count()
+        self.batch_count = 0
+        self.warmup_end = warmup_batches
+
+        for u, d in zip(encoder_unmasked_dims, encoder_dims):
+            assert u <= d, (u, d)
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, (T - 7)//2, encoder_dims).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> (T - 7)//2
+        #   (2) embedding: num_features -> encoder_dims
+        self.encoder_embed = Conv2dSubsampling(
+            num_features, encoder_dims[0], dropout=dropout, is_pnnx=is_pnnx
+        )
+
+        # each one will be ZipformerEncoder or DownsampledZipformerEncoder
+        encoders = []
+
+        self.num_encoders = len(encoder_dims)
+        for i in range(self.num_encoders):
+            ds = zipformer_downsampling_factors[i]
+            encoder_layer = ZipformerEncoderLayer(
+                encoder_dims[i],
+                attention_dim[i],
+                nhead[i],
+                feedforward_dim[i],
+                dropout,
+                cnn_module_kernels[i],
+                pos_dim,
+                is_pnnx=self.is_pnnx,
+                left_context_len=self.left_context_len // ds,
+                x_size=self.decode_chunk_size // ds,
+            )
+
+            # For the segment of the warmup period, we let the Conv2dSubsampling
+            # layer learn something.  Then we start to warm up the other encoders.
+            encoder = ZipformerEncoder(
+                encoder_layer,
+                num_encoder_layers[i],
+                dropout,
+                warmup_begin=warmup_batches * (i + 1) / (self.num_encoders + 1),
+                warmup_end=warmup_batches * (i + 2) / (self.num_encoders + 1),
+                is_pnnx=is_pnnx,
+                left_context_len=self.left_context_len // ds,
+                x_size=self.decode_chunk_size // ds,
+            )
+
+            if zipformer_downsampling_factors[i] != 1:
+                in_x_size = self.decode_chunk_size
+                encoder = DownsampledZipformerEncoder(
+                    encoder,
+                    input_dim=encoder_dims[i - 1] if i > 0 else encoder_dims[0],
+                    output_dim=encoder_dims[i],
+                    downsample=zipformer_downsampling_factors[i],
+                    is_pnnx=is_pnnx,
+                    left_context_len=self.left_context_len // ds,
+                    in_x_size=in_x_size,
+                )
+            encoders.append(encoder)
+        self.encoders = nn.ModuleList(encoders)
+
+        # initializes self.skip_layers and self.skip_modules
+        self._init_skip_modules()
+
+        self.downsample_output = AttentionDownsample(
+            encoder_dims[-1],
+            encoder_dims[-1],
+            downsample=output_downsampling_factor,
+            is_pnnx=is_pnnx,
+            in_x_size=self.decode_chunk_size,
+        )
+
+    def _get_layer_skip_dropout_prob(self):
+        if not self.training:
+            return 0.0
+        batch_count = self.batch_count
+        min_dropout_prob = 0.025
+
+        if batch_count > self.warmup_end:
+            return min_dropout_prob
+        else:
+            return 0.5 - (batch_count / self.warmup_end) * (0.5 - min_dropout_prob)
+
+    def _init_skip_modules(self):
+        """
+        If self.zipformer_downsampling_factors = (1, 2, 4, 8, 4, 2), then at the input of layer
+        indexed 4 (in zero indexing), which has subsampling_factor=4, we combine the output of
+        layers 2 and 3; and at the input of layer indexed 5, which has subsampling_factor=2,
+        we combine the outputs of layers 1 and 4.
+        """
+        skip_layers = []
+        skip_modules = []
+        z = self.zipformer_downsampling_factors
+        for i in range(len(z)):
+            if i <= 1 or z[i - 1] <= z[i]:
+                skip_layers.append(None)
+                skip_modules.append(SimpleCombinerIdentity())
+            else:
+                # TEMP
+                for j in range(i - 2, -1, -1):
+                    if z[j] <= z[i] or j == 0:
+                        # TEMP logging statement.
+                        logging.info(
+                            f"At encoder stack {i}, which has downsampling_factor={z[i]}, we will "
+                            f"combine the outputs of layers {j} and {i-1}, with downsampling_factors={z[j]} and {z[i-1]}."
+                        )
+                        skip_layers.append(j)
+                        skip_modules.append(
+                            SimpleCombiner(
+                                self.encoder_dims[j],
+                                self.encoder_dims[i - 1],
+                                min_weight=(0.0, 0.25),
+                            )
+                        )
+                        break
+        self.skip_layers = skip_layers
+        self.skip_modules = nn.ModuleList(skip_modules)
+
+    def get_feature_masks(self, x: torch.Tensor) -> List[float]:
+        # Note: The actual return type is Union[List[float], List[Tensor]],
+        # but to make torch.jit.script() work, we use List[float]
+        """
+        In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of
+        randomized feature masks, one per encoder.
+        On e.g. 15% of frames, these masks will zero out all encoder dims larger than
+        some supplied number, e.g. >256, so in effect on those frames we are using
+        a smaller encoder dim.
+
+        We generate the random masks at this level because we want the 2 masks to 'agree'
+        all the way up the encoder stack. This will mean that the 1st mask will have
+        mask values repeated self.zipformer_downsampling_factors times.
+
+        Args:
+           x: the embeddings (needed for the shape and dtype and device), of shape
+             (num_frames, batch_size, encoder_dims0)
+        """
+        num_encoders = len(self.encoder_dims)
+        if torch.jit.is_scripting() or not self.training:
+            return [1.0] * num_encoders
+
+        (num_frames0, batch_size, _encoder_dims0) = x.shape
+
+        assert self.encoder_dims[0] == _encoder_dims0, (
+            self.encoder_dims,
+            _encoder_dims0,
+        )
+
+        max_downsampling_factor = max(self.zipformer_downsampling_factors)
+
+        num_frames_max = num_frames0 + max_downsampling_factor - 1
+
+        feature_mask_dropout_prob = 0.15
+
+        # frame_mask_max shape: (num_frames_max, batch_size, 1)
+        frame_mask_max = (
+            torch.rand(num_frames_max, batch_size, 1, device=x.device)
+            > feature_mask_dropout_prob
+        ).to(x.dtype)
+
+        feature_masks = []
+        for i in range(num_encoders):
+            ds = self.zipformer_downsampling_factors[i]
+            upsample_factor = max_downsampling_factor // ds
+
+            frame_mask = (
+                frame_mask_max.unsqueeze(1)
+                .expand(num_frames_max, upsample_factor, batch_size, 1)
+                .reshape(num_frames_max * upsample_factor, batch_size, 1)
+            )
+            num_frames = (num_frames0 + ds - 1) // ds
+            frame_mask = frame_mask[:num_frames]
+            feature_mask = torch.ones(
+                num_frames,
+                batch_size,
+                self.encoder_dims[i],
+                dtype=x.dtype,
+                device=x.device,
+            )
+            u = self.encoder_unmasked_dims[i]
+            feature_mask[:, :, u:] *= frame_mask
+            feature_masks.append(feature_mask)
+
+        return feature_masks
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          chunk_size:
+            The chunk size used in evaluation mode.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1])
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        lengths = (x_lens - 7) >> 1
+        assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max())
+        mask = make_pad_mask(lengths)
+
+        outputs = []
+        feature_masks = self.get_feature_masks(x)
+
+        if self.training:
+            # Training mode
+            max_ds = max(self.zipformer_downsampling_factors)
+            # Generate dynamic chunk-wise attention mask during training
+            max_len = x.size(0) // max_ds
+            short_chunk_size = self.short_chunk_size // max_ds
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * self.short_chunk_threshold):
+                # Full attention
+                chunk_size = x.size(0)
+            else:
+                # Chunk-wise attention
+                chunk_size = chunk_size % short_chunk_size + 1
+                chunk_size *= max_ds
+        else:
+            chunk_size = self.decode_chunk_size
+            # Evaluation mode
+            for ds in self.zipformer_downsampling_factors:
+                assert chunk_size % ds == 0, (chunk_size, ds)
+
+        attn_mask = ~subsequent_chunk_mask(
+            size=x.size(0),
+            chunk_size=chunk_size,
+            num_left_chunks=self.num_left_chunks,
+            device=x.device,
+        )
+
+        for i, (module, skip_module) in enumerate(
+            zip(self.encoders, self.skip_modules)
+        ):
+            ds = self.zipformer_downsampling_factors[i]
+            k = self.skip_layers[i]
+            if isinstance(k, int):
+                layer_skip_dropout_prob = self._get_layer_skip_dropout_prob()
+                if torch.jit.is_scripting():
+                    x = skip_module(outputs[k], x)
+                elif (not self.training) or random.random() > layer_skip_dropout_prob:
+                    x = skip_module(outputs[k], x)
+            x = module(
+                x,
+                feature_mask=feature_masks[i],
+                src_key_padding_mask=None if mask is None else mask[..., ::ds],
+                attn_mask=attn_mask[::ds, ::ds],
+            )
+            outputs.append(x)
+
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        lengths = (lengths + 1) >> 1
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        states: List[Tensor],
+    ) -> Tuple[Tensor, List[Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+            seq_len is the input chunk length.
+          states:
+            A list of 7 * num_encoders elements:
+            ``states[0:num_encoders]`` is the cached numbers of past frames.
+            ``states[num_encoders:2*num_encoders]`` is the cached average tensors.
+            ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+            ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+            ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+            ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+            ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1])
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+            - updated states.
+        """
+        assert len(states) == 7 * self.num_encoders, (len(states), self.num_encoders)
+
+        cached_len = states[: self.num_encoders]
+        cached_avg = states[self.num_encoders : 2 * self.num_encoders]
+        cached_key = states[2 * self.num_encoders : 3 * self.num_encoders]
+        cached_val = states[3 * self.num_encoders : 4 * self.num_encoders]
+        cached_val2 = states[4 * self.num_encoders : 5 * self.num_encoders]
+        cached_conv1 = states[5 * self.num_encoders : 6 * self.num_encoders]
+        cached_conv2 = states[6 * self.num_encoders : 7 * self.num_encoders]
+
+        x = self.encoder_embed(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        outputs = []
+        new_cached_len = []
+        new_cached_avg = []
+        new_cached_key = []
+        new_cached_val = []
+        new_cached_val2 = []
+        new_cached_conv1 = []
+        new_cached_conv2 = []
+
+        for i, (module, skip_module) in enumerate(
+            zip(self.encoders, self.skip_modules)
+        ):
+            k = self.skip_layers[i]
+            if isinstance(k, int):
+                x = skip_module(outputs[k], x)
+            x, len_avg, avg, key, val, val2, conv1, conv2 = module.streaming_forward(
+                x,
+                cached_len=cached_len[i],
+                cached_avg=cached_avg[i],
+                cached_key=cached_key[i],
+                cached_val=cached_val[i],
+                cached_val2=cached_val2[i],
+                cached_conv1=cached_conv1[i],
+                cached_conv2=cached_conv2[i],
+            )
+
+            outputs.append(x)
+            # Update caches
+            new_cached_len.append(len_avg)
+            new_cached_avg.append(avg)
+            new_cached_key.append(key)
+            new_cached_val.append(val)
+            new_cached_val2.append(val2)
+            new_cached_conv1.append(conv1)
+            new_cached_conv2.append(conv2)
+
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = (
+            new_cached_len
+            + new_cached_avg
+            + new_cached_key
+            + new_cached_val
+            + new_cached_val2
+            + new_cached_conv1
+            + new_cached_conv2
+        )
+        return x, new_states
+
+    @torch.jit.export
+    def get_init_state(
+        self,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[Tensor]:
+        """Get initial states.
+        A list of 7 * num_encoders elements:
+        ``states[0:num_encoders]`` is the cached numbers of past frames.
+        ``states[num_encoders:2*num_encoders]`` is the cached average tensors.
+        ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules.
+        ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules.
+        ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules.
+        ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules.
+        ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules.
+        """
+        cached_len = []
+        cached_avg = []
+        cached_key = []
+        cached_val = []
+        cached_val2 = []
+        cached_conv1 = []
+        cached_conv2 = []
+
+        for i, encoder in enumerate(self.encoders):
+            num_layers = encoder.num_layers
+            ds = self.zipformer_downsampling_factors[i]
+
+            len_avg = torch.zeros(num_layers, 1, device=device)
+            cached_len.append(len_avg)
+
+            avg = torch.zeros(num_layers, 1, encoder.d_model, device=device)
+            cached_avg.append(avg)
+
+            key = torch.zeros(
+                num_layers,
+                self.left_context_len // ds,
+                1,
+                encoder.attention_dim,
+                device=device,
+            )
+            cached_key.append(key)
+
+            val = torch.zeros(
+                num_layers,
+                self.left_context_len // ds,
+                1,
+                encoder.attention_dim // 2,
+                device=device,
+            )
+            cached_val.append(val)
+
+            val2 = torch.zeros(
+                num_layers,
+                self.left_context_len // ds,
+                1,
+                encoder.attention_dim // 2,
+                device=device,
+            )
+            cached_val2.append(val2)
+
+            conv1 = torch.zeros(
+                num_layers,
+                1,
+                encoder.d_model,
+                encoder.cnn_module_kernel - 1,
+                device=device,
+            )
+            cached_conv1.append(conv1)
+
+            conv2 = torch.zeros(
+                num_layers,
+                1,
+                encoder.d_model,
+                encoder.cnn_module_kernel - 1,
+                device=device,
+            )
+            cached_conv2.append(conv2)
+
+        states = (
+            cached_len
+            + cached_avg
+            + cached_key
+            + cached_val
+            + cached_val2
+            + cached_conv1
+            + cached_conv2
+        )
+        return states
+
+
+class ZipformerEncoderLayer(nn.Module):
+    """
+    ZipformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        feedforward_dim: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        attention_dim: int,
+        nhead: int,
+        feedforward_dim: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        pos_dim: int = 4,
+        is_pnnx: bool = False,
+        left_context_len: int = 0,
+        x_size: int = 0,
+    ) -> None:
+        super(ZipformerEncoderLayer, self).__init__()
+
+        self.d_model = d_model
+        self.attention_dim = attention_dim
+        self.cnn_module_kernel = cnn_module_kernel
+
+        # will be written to, see set_batch_count()
+        self.batch_count = 0
+
+        self.self_attn = RelPositionMultiheadAttention(
+            d_model,
+            attention_dim,
+            nhead,
+            pos_dim,
+            dropout=0.0,
+            is_pnnx=is_pnnx,
+            left_context_len=left_context_len,
+            x_size=x_size,
+        )
+
+        self.pooling = PoolingModule(d_model)
+
+        self.feed_forward1 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.feed_forward2 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.feed_forward3 = FeedforwardModule(d_model, feedforward_dim, dropout)
+
+        self.conv_module1 = ConvolutionModule(
+            d_model, cnn_module_kernel, is_pnnx=is_pnnx, x_size=x_size
+        )
+
+        self.conv_module2 = ConvolutionModule(
+            d_model, cnn_module_kernel, is_pnnx=is_pnnx, x_size=x_size
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        self.bypass_scale = nn.Parameter(torch.tensor(0.5))
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            d_model,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            max_abs=6.0,
+        )
+        self.whiten = Whiten(
+            num_groups=1, whitening_limit=5.0, prob=(0.025, 0.25), grad_scale=0.01
+        )
+
+    def get_bypass_scale(self):
+        if torch.jit.is_scripting() or not self.training:
+            return self.bypass_scale
+        if random.random() < 0.1:
+            # ensure we get grads if self.bypass_scale becomes out of range
+            return self.bypass_scale
+        # hardcode warmup period for bypass scale
+        warmup_period = 20000.0
+        initial_clamp_min = 0.75
+        final_clamp_min = 0.25
+        if self.batch_count > warmup_period:
+            clamp_min = final_clamp_min
+        else:
+            clamp_min = initial_clamp_min - (self.batch_count / warmup_period) * (
+                initial_clamp_min - final_clamp_min
+            )
+        return self.bypass_scale.clamp(min=clamp_min, max=1.0)
+
+    def get_dynamic_dropout_rate(self):
+        # return dropout rate for the dynamic modules (self_attn, pooling, convolution); this
+        # starts at 0.2 and rapidly decreases to 0.  Its purpose is to keep the training stable
+        # at the beginning, by making the network focus on the feedforward modules.
+        if torch.jit.is_scripting() or not self.training:
+            return 0.0
+        warmup_period = 2000.0
+        initial_dropout_rate = 0.2
+        final_dropout_rate = 0.0
+        if self.batch_count > warmup_period:
+            return final_dropout_rate
+        else:
+            return initial_dropout_rate - (
+                initial_dropout_rate * final_dropout_rate
+            ) * (self.batch_count / warmup_period)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+        batch_split: if not None, this layer will only be applied to
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        # macaron style feed forward module
+        src = src + self.feed_forward1(src)
+
+        # dropout rate for submodules that interact with time.
+        dynamic_dropout = self.get_dynamic_dropout_rate()
+
+        # pooling module
+        if torch.jit.is_scripting():
+            src = src + self.pooling(src, src_key_padding_mask=src_key_padding_mask)
+        elif random.random() >= dynamic_dropout:
+            src = src + self.pooling(src, src_key_padding_mask=src_key_padding_mask)
+
+        if torch.jit.is_scripting():
+            src_att, attn_weights = self.self_attn(
+                src,
+                pos_emb=pos_emb,
+                attn_mask=attn_mask,
+                key_padding_mask=src_key_padding_mask,
+            )
+            src = src + src_att
+
+            src = src + self.conv_module1(
+                src, src_key_padding_mask=src_key_padding_mask
+            )
+
+            src = src + self.feed_forward2(src)
+
+            src = src + self.self_attn.forward2(src, attn_weights)
+
+            src = src + self.conv_module2(
+                src, src_key_padding_mask=src_key_padding_mask
+            )
+        else:
+            use_self_attn = random.random() >= dynamic_dropout
+            if use_self_attn:
+                src_att, attn_weights = self.self_attn(
+                    src,
+                    pos_emb=pos_emb,
+                    attn_mask=attn_mask,
+                    key_padding_mask=src_key_padding_mask,
+                )
+                src = src + src_att
+
+            if random.random() >= dynamic_dropout:
+                src = src + self.conv_module1(
+                    src, src_key_padding_mask=src_key_padding_mask
+                )
+
+            src = src + self.feed_forward2(src)
+
+            if use_self_attn:
+                src = src + self.self_attn.forward2(src, attn_weights)
+
+            if random.random() >= dynamic_dropout:
+                src = src + self.conv_module2(
+                    src, src_key_padding_mask=src_key_padding_mask
+                )
+
+        src = src + self.feed_forward3(src)
+
+        src = self.norm_final(self.balancer(src))
+
+        delta = src - src_orig
+
+        src = src_orig + delta * self.get_bypass_scale()
+
+        return self.whiten(src)
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            cached_len: processed number of past frames.
+            cached_avg: cached average of past frames.
+            cached_key: cached key tensor of left context for the first attention module.
+            cached_val: cached value tensor of left context for the first attention module.
+            cached_val2: cached value tensor of left context for the second attention module.
+            cached_conv1: cached left context for the first convolution module.
+            cached_conv2: cached left context for the second convolution module.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, left_context_len+2*S-1, E)
+            cached_len: (N,)
+              N is the batch size.
+            cached_avg: (N, C).
+              N is the batch size, C is the feature dimension.
+            cached_key: (left_context_len, N, K).
+              N is the batch size, K is the key dimension.
+            cached_val: (left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_val2: (left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_conv1: (N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+            cached_conv2: (N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+        """
+        src_orig = src
+
+        # macaron style feed forward module
+        src = src + self.feed_forward1(src)
+
+        src_pool, cached_len, cached_avg = self.pooling.streaming_forward(
+            src,
+            cached_len=cached_len,
+            cached_avg=cached_avg,
+        )
+        src = src + src_pool
+
+        (
+            src_attn,
+            attn_weights,
+            cached_key,
+            cached_val,
+        ) = self.self_attn.streaming_forward(
+            src,
+            pos_emb=pos_emb,
+            cached_key=cached_key,
+            cached_val=cached_val,
+        )
+
+        src = src + src_attn
+
+        src_conv, cached_conv1 = self.conv_module1.streaming_forward(
+            src,
+            cache=cached_conv1,
+        )
+
+        src = src + src_conv
+
+        src = src + self.feed_forward2(src)
+
+        src_attn, cached_val2 = self.self_attn.streaming_forward2(
+            src,
+            attn_weights,
+            cached_val=cached_val2,
+        )
+        src = src + src_attn
+
+        src_conv, cached_conv2 = self.conv_module2.streaming_forward(
+            src,
+            cache=cached_conv2,
+        )
+        src = src + src_conv
+
+        src = src + self.feed_forward3(src)
+
+        src = self.norm_final(self.balancer(src))
+
+        delta = src - src_orig
+
+        src = src_orig + delta * self.bypass_scale
+
+        return (
+            src,
+            cached_len,
+            cached_avg,
+            cached_key,
+            cached_val,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        )
+
+
+class ZipformerStateSelect(nn.Module):
+    """ncnn does not support selecting along batch index.
+    This class provides a workaround for it. We
+    need to change pnnx accordingly.
+    """
+
+    def __init__(self, i: int):
+        super().__init__()
+        self.i = i
+
+    def forward(self, x: torch.Tensor):
+        return x[self.i]
+
+
+class ZipformerEncoder(nn.Module):
+    r"""ZipformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ZipformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8)
+        >>> zipformer_encoder = ZipformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = zipformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        dropout: float,
+        warmup_begin: float,
+        warmup_end: float,
+        is_pnnx: bool = False,
+        x_size: int = 0,
+        left_context_len: int = 0,
+    ) -> None:
+        super().__init__()
+        # will be written to, see set_batch_count() Note: in inference time this
+        # may be zero but should be treated as large, we can check if
+        # self.training is true.
+        self.batch_count = 0
+        self.warmup_begin = warmup_begin
+        self.warmup_end = warmup_end
+        # module_seed is for when we need a random number that is unique to the module but
+        # shared across jobs.   It's used to randomly select how many layers to drop,
+        # so that we can keep this consistent across worker tasks (for efficiency).
+        self.module_seed = torch.randint(0, 1000, ()).item()
+        self.left_context_len = left_context_len
+
+        self.encoder_pos = RelPositionalEncoding(
+            encoder_layer.d_model,
+            dropout,
+            is_pnnx=is_pnnx,
+            x_size=x_size,
+            left_context_len=left_context_len,
+        )
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        state_select_list = []
+        for i in range(num_layers):
+            state_select_list.append(ZipformerStateSelect(i))
+        self.state_select_list = nn.ModuleList(state_select_list)
+
+        self.d_model = encoder_layer.d_model
+        self.attention_dim = encoder_layer.attention_dim
+        self.cnn_module_kernel = encoder_layer.cnn_module_kernel
+
+        assert 0 <= warmup_begin <= warmup_end, (warmup_begin, warmup_end)
+
+        delta = (1.0 / num_layers) * (warmup_end - warmup_begin)
+        cur_begin = warmup_begin
+        for i in range(num_layers):
+            self.layers[i].warmup_begin = cur_begin
+            cur_begin += delta
+            self.layers[i].warmup_end = cur_begin
+
+    def get_layers_to_drop(self, rnd_seed: int):
+        ans = set()
+        if not self.training:
+            return ans
+
+        batch_count = self.batch_count
+        num_layers = len(self.layers)
+
+        def get_layerdrop_prob(layer: int) -> float:
+            layer_warmup_begin = self.layers[layer].warmup_begin
+            layer_warmup_end = self.layers[layer].warmup_end
+
+            initial_layerdrop_prob = 0.5
+            final_layerdrop_prob = 0.05
+
+            if batch_count == 0:
+                # As a special case, if batch_count == 0, return 0 (drop no
+                # layers).  This is rather ugly, I'm afraid; it is intended to
+                # enable our scan_pessimistic_batches_for_oom() code to work correctly
+                # so if we are going to get OOM it will happen early.
+                # also search for 'batch_count' with quotes in this file to see
+                # how we initialize the warmup count to a random number between
+                # 0 and 10.
+                return 0.0
+            elif batch_count < layer_warmup_begin:
+                return initial_layerdrop_prob
+            elif batch_count > layer_warmup_end:
+                return final_layerdrop_prob
+            else:
+                # linearly interpolate
+                t = (batch_count - layer_warmup_begin) / layer_warmup_end
+                assert 0.0 <= t < 1.001, t
+                return initial_layerdrop_prob + t * (
+                    final_layerdrop_prob - initial_layerdrop_prob
+                )
+
+        shared_rng = random.Random(batch_count + self.module_seed)
+        independent_rng = random.Random(rnd_seed)
+
+        layerdrop_probs = [get_layerdrop_prob(i) for i in range(num_layers)]
+        tot = sum(layerdrop_probs)
+        # Instead of drawing the samples independently, we first randomly decide
+        # how many layers to drop out, using the same random number generator between
+        # jobs so that all jobs drop out the same number (this is for speed).
+        # Then we use an approximate approach to drop out the individual layers
+        # with their specified probs while reaching this exact target.
+        num_to_drop = int(tot) + int(shared_rng.random() < (tot - int(tot)))
+
+        layers = list(range(num_layers))
+        independent_rng.shuffle(layers)
+
+        # go through the shuffled layers until we get the required number of samples.
+        if num_to_drop > 0:
+            for layer in itertools.cycle(layers):
+                if independent_rng.random() < layerdrop_probs[layer]:
+                    ans.add(layer)
+                if len(ans) == num_to_drop:
+                    break
+        if shared_rng.random() < 0.005 or __name__ == "__main__":
+            logging.info(
+                f"warmup_begin={self.warmup_begin:.1f}, warmup_end={self.warmup_end:.1f}, "
+                f"batch_count={batch_count:.1f}, num_to_drop={num_to_drop}, layers_to_drop={ans}"
+            )
+        return ans
+
+    def forward(
+        self,
+        src: Tensor,
+        # Note: The type of feature_mask should be Union[float, Tensor],
+        # but to make torch.jit.script() work, we use `float` here
+        feature_mask: float = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        Returns: (x, x_no_combine), both of shape (S, N, E)
+        """
+        pos_emb = self.encoder_pos(src)
+        output = src
+
+        if torch.jit.is_scripting():
+            layers_to_drop = []
+        else:
+            rnd_seed = src.numel() + random.randint(0, 1000)
+            layers_to_drop = self.get_layers_to_drop(rnd_seed)
+
+        output = output * feature_mask
+
+        for i, mod in enumerate(self.layers):
+            if not torch.jit.is_scripting():
+                if i in layers_to_drop:
+                    continue
+            output = mod(
+                output,
+                pos_emb,
+                attn_mask=attn_mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+            output = output * feature_mask
+
+        return output
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        src: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            cached_len: number of past frames.
+            cached_avg: cached average of past frames.
+            cached_key: cached key tensor for first attention module.
+            cached_val: cached value tensor for first attention module.
+            cached_val2: cached value tensor for second attention module.
+            cached_conv1: cached left contexts for the first convolution module.
+            cached_conv2: cached left contexts for the second convolution module.
+
+        Shape:
+            src: (S, N, E).
+            cached_len: (num_layers,)
+            cached_avg: (num_layers, N, C).
+              N is the batch size, C is the feature dimension.
+            cached_key: (num_layers, left_context_len, N, K).
+              N is the batch size, K is the key dimension.
+            cached_val: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_val2: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_conv1: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+            cached_conv2: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+
+        Returns: A tuple of 8 tensors:
+            - output tensor
+            - updated cached number of past frames.
+            - updated cached average of past frames.
+            - updated cached key tensor of of the first attention module.
+            - updated cached value tensor of of the first attention module.
+            - updated cached value tensor of of the second attention module.
+            - updated cached left contexts of the first convolution module.
+            - updated cached left contexts of the second convolution module.
+        """
+        assert cached_len.size(0) == self.num_layers, (
+            cached_len.size(0),
+            self.num_layers,
+        )
+        assert cached_avg.size(0) == self.num_layers, (
+            cached_avg.size(0),
+            self.num_layers,
+        )
+        assert cached_key.size(0) == self.num_layers, (
+            cached_key.size(0),
+            self.num_layers,
+        )
+        assert cached_val.size(0) == self.num_layers, (
+            cached_val.size(0),
+            self.num_layers,
+        )
+        assert cached_val2.size(0) == self.num_layers, (
+            cached_val2.size(0),
+            self.num_layers,
+        )
+        assert cached_conv1.size(0) == self.num_layers, (
+            cached_conv1.size(0),
+            self.num_layers,
+        )
+        assert cached_conv2.size(0) == self.num_layers, (
+            cached_conv2.size(0),
+            self.num_layers,
+        )
+
+        assert self.left_context_len == cached_key.shape[1], (
+            self.left_context_len,
+            cached_key.shape[1],
+        )
+
+        left_context_len = self.left_context_len
+        pos_emb = self.encoder_pos(src, left_context_len)
+
+        output = src
+
+        new_cached_len = []
+        new_cached_avg = []
+        new_cached_key = []
+        new_cached_val = []
+        new_cached_val2 = []
+        new_cached_conv1 = []
+        new_cached_conv2 = []
+        for i, (mod, state_select) in enumerate(
+            zip(self.layers, self.state_select_list)
+        ):
+            output, len_avg, avg, key, val, val2, conv1, conv2 = mod.streaming_forward(
+                output,
+                pos_emb,
+                cached_len=state_select(cached_len),
+                cached_avg=state_select(cached_avg),
+                cached_key=state_select(cached_key),
+                cached_val=state_select(cached_val),
+                cached_val2=state_select(cached_val2),
+                cached_conv1=state_select(cached_conv1),
+                cached_conv2=state_select(cached_conv2),
+            )
+            # Update caches
+            new_cached_len.append(len_avg)
+            new_cached_avg.append(avg)
+            new_cached_key.append(key)
+            new_cached_val.append(val)
+            new_cached_val2.append(val2)
+            new_cached_conv1.append(conv1)
+            new_cached_conv2.append(conv2)
+
+        return (
+            output,
+            torch.stack(new_cached_len, dim=0),
+            torch.stack(new_cached_avg, dim=0),
+            torch.stack(new_cached_key, dim=0),
+            torch.stack(new_cached_val, dim=0),
+            torch.stack(new_cached_val2, dim=0),
+            torch.stack(new_cached_conv1, dim=0),
+            torch.stack(new_cached_conv2, dim=0),
+        )
+
+
+class DownsampledZipformerEncoder(nn.Module):
+    r"""
+    DownsampledZipformerEncoder is a zipformer encoder evaluated at a reduced frame rate,
+    after convolutional downsampling, and then upsampled again at the output, and combined
+    with the origin input, so that the output has the same shape as the input.
+    """
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        input_dim: int,
+        output_dim: int,
+        downsample: int,
+        is_pnnx: bool = False,
+        left_context_len: int = 0,
+        in_x_size: int = 0,
+    ):
+        super(DownsampledZipformerEncoder, self).__init__()
+        self.downsample_factor = downsample
+        self.downsample = AttentionDownsample(
+            input_dim, output_dim, downsample, is_pnnx=is_pnnx, in_x_size=in_x_size
+        )
+        self.encoder = encoder
+        self.num_layers = encoder.num_layers
+        self.d_model = encoder.d_model
+        self.attention_dim = encoder.attention_dim
+        self.cnn_module_kernel = encoder.cnn_module_kernel
+        self.upsample = SimpleUpsample(output_dim, downsample)
+        self.out_combiner = SimpleCombiner(
+            input_dim, output_dim, min_weight=(0.0, 0.25)
+        )
+        self.in_x_size = in_x_size
+
+    def forward(
+        self,
+        src: Tensor,
+        # Note: the type of feature_mask should be Unino[float, Tensor],
+        # but to make torch.jit.script() happ, we use float here
+        feature_mask: float = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer.  feature_mask is expected to be already downsampled by
+               self.downsample_factor.
+            attn_mask: attention mask (optional). Should be downsampled already.
+            src_key_padding_mask: the mask for the src keys per batch (optional).  Should be downsampled already.
+
+        Shape:
+            src: (S, N, E).
+            attn_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        Returns: output of shape (S, N, F) where F is the number of output features
+            (output_dim to constructor)
+        """
+        src_orig = src
+        src = self.downsample(src)
+
+        src = self.encoder(
+            src,
+            feature_mask=feature_mask,
+            attn_mask=attn_mask,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src)
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        cached_len: Tensor,
+        cached_avg: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required).
+            cached_avg: cached average value of past frames.
+            cached_len: length of past frames.
+            cached_key: cached key tensor for the first attention module.
+            cached_val: cached value tensor for the first attention module.
+            cached_val2: cached value tensor for the second attention module.
+            cached_conv1: cached left context for the first convolution module.
+            cached_conv2: cached left context for the second convolution module.
+
+        Shape:
+            src: (S, N, E).
+            cached_len: (N,)
+              N is the batch size.
+            cached_avg: (num_layers, N, C).
+              N is the batch size, C is the feature dimension.
+            cached_key: (num_layers, left_context_len, N, K).
+              N is the batch size, K is the key dimension.
+            cached_val: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_val2: (num_layers, left_context_len, N, V).
+              N is the batch size, V is the key dimension.
+            cached_conv1: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+            cached_conv2: (num_layers, N, C, kernel_size-1).
+              N is the batch size, C is the convolution channels.
+        Returns: output of shape (S, N, F) where F is the number of output features
+            (output_dim to constructor)
+        """
+        assert src.shape[0] == self.in_x_size, (src.shape[0], self.in_x_size)
+
+        src_orig = src
+
+        src = self.downsample(src)
+
+        (
+            src,
+            cached_len,
+            cached_avg,
+            cached_key,
+            cached_val,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ) = self.encoder.streaming_forward(
+            src,
+            cached_len=cached_len,
+            cached_avg=cached_avg,
+            cached_key=cached_key,
+            cached_val=cached_val,
+            cached_val2=cached_val2,
+            cached_conv1=cached_conv1,
+            cached_conv2=cached_conv2,
+        )
+
+        src = self.upsample(src)
+
+        if src.shape[0] != self.in_x_size:
+            # remove any extra frames that are not a multiple of downsample_factor
+            src = src[: self.in_x_size]
+
+        return (
+            self.out_combiner(src_orig, src),
+            cached_len,
+            cached_avg,
+            cached_key,
+            cached_val,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        )
+
+
+class AttentionDownsampleUnsqueeze(torch.nn.Module):
+    """We apply this operation only in PyTorch
+    and discards in ncnn.
+    """
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x.unsqueeze(1)
+
+
+class AttentionDownsample(torch.nn.Module):
+    """
+    Does downsampling with attention, by weighted sum, and a projection..
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        downsample: int,
+        is_pnnx: bool = False,
+        in_x_size: int = 0,
+    ):
+        super(AttentionDownsample, self).__init__()
+
+        self.query = nn.Parameter(torch.randn(in_channels) * (in_channels**-0.5))
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.is_pnnx = is_pnnx
+        self.in_x_size = in_x_size
+
+        self.unsqueeze = AttentionDownsampleUnsqueeze()
+
+        # fill in the extra dimensions with a projection of the input
+        if out_channels > in_channels:
+            self.extra_proj = nn.Linear(
+                in_channels * downsample, out_channels - in_channels, bias=False
+            )
+        else:
+            self.extra_proj = None
+        self.downsample = downsample
+
+        self.d_seq_len = (in_x_size + downsample - 1) // downsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, 1, in_channels)
+        Returns a tensor of shape
+           ( (seq_len+downsample-1)//downsample, batch_size, out_channels)
+        """
+        assert src.shape[0] == self.in_x_size, (
+            src.shape[0],
+            self.in_x_size,
+            src.shape,
+            type(src),
+        )
+        assert src.shape[2] == self.in_channels, (src.shape[2], self.in_channels)
+        if not self.is_pnnx:
+            (seq_len, batch_size, in_channels) = src.shape
+        else:
+            seq_len = self.in_x_size
+            batch_size = 1
+            in_channels = self.in_channels
+
+        ds = self.downsample
+        d_seq_len = self.d_seq_len
+
+        # Pad to an exact multiple of self.downsample
+        if seq_len != d_seq_len * ds:
+            assert self.is_pnnx is False, "TODO(fangjun): Handle it!"
+            # right-pad src, repeating the last element.
+            pad = d_seq_len * ds - seq_len
+            src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2])
+            src = torch.cat((src, src_extra), dim=0)
+            assert src.shape[0] == d_seq_len * ds, (src.shape[0], d_seq_len, ds)
+
+        if not self.is_pnnx:
+            src = src.reshape(d_seq_len, ds, batch_size, in_channels)
+            scores = (src * self.query).sum(dim=-1, keepdim=True)
+
+            if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+                scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04)
+
+            weights = scores.softmax(dim=1)
+
+            # ans1 is the first `in_channels` channels of the output
+            ans = (src * weights).sum(dim=1)
+            src = src.permute(0, 2, 1, 3).reshape(
+                d_seq_len, batch_size, ds * in_channels
+            )
+
+            if self.extra_proj is not None:
+                ans2 = self.extra_proj(src)
+                ans = torch.cat((ans, ans2), dim=2)
+        else:
+            src = src.reshape(d_seq_len, ds, in_channels)
+            scores = (src * self.query).sum(dim=-1, keepdim=True)
+
+            if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+                scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04)
+
+            weights = scores.softmax(dim=1)
+
+            # ans1 is the first `in_channels` channels of the output
+            ans = (src * weights).sum(dim=1)
+
+            assert (
+                self.extra_proj is None
+            ), "The code for it being not None is not tested"
+            #  ans = ans.unsqueeze(1)
+            ans = self.unsqueeze(ans)
+            # Note: In ncnn, we ignore self.unsqueeze
+            # so ans in ncnn is still a 2-D tensor, e.g., (8, 384)
+
+        return ans
+
+
+class SimpleUpsample(torch.nn.Module):
+    """
+    A very simple form of upsampling that mostly just repeats the input, but
+    also adds a position-specific bias.
+    """
+
+    def __init__(self, num_channels: int, upsample: int):
+        super(SimpleUpsample, self).__init__()
+        self.bias = nn.Parameter(torch.randn(upsample, num_channels) * 0.01)
+        self.upsample = upsample
+        self.num_channels = num_channels
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, num_channels)
+        Returns a tensor of shape
+           ( (seq_len*upsample), batch_size, num_channels)
+        """
+        upsample = self.bias.shape[0]
+        (seq_len, batch_size, num_channels) = src.shape
+        src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels)
+        src = src + self.bias.unsqueeze(1)
+        src = src.reshape(seq_len * upsample, batch_size, num_channels)
+        return src
+
+
+class SimpleCombinerIdentity(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    def forward(self, src1: Tensor, src2: Tensor) -> Tensor:
+        return src1
+
+
+class SimpleCombiner(torch.nn.Module):
+    """
+    A very simple way of combining 2 vectors of 2 different dims, via a
+    learned weighted combination in the shared part of the dim.
+    Args:
+         dim1: the dimension of the first input, e.g. 256
+         dim2: the dimension of the second input, e.g. 384.
+    The output will have the same dimension as dim2.
+    """
+
+    def __init__(self, dim1: int, dim2: int, min_weight: Tuple[float] = (0.0, 0.0)):
+        super(SimpleCombiner, self).__init__()
+        assert dim2 >= dim1, (dim2, dim1)
+        self.weight1 = nn.Parameter(torch.zeros(()))
+        self.min_weight = min_weight
+        self.dim1 = dim1
+        self.dim2 = dim2
+
+    def forward(self, src1: Tensor, src2: Tensor) -> Tensor:
+        """
+        src1: (*, dim1)
+        src2: (*, dim2)
+
+        Returns: a tensor of shape (*, dim2)
+        """
+        assert src1.shape[:-1] == src2.shape[:-1], (src1.shape, src2.shape)
+
+        weight1 = self.weight1
+        if not torch.jit.is_scripting():
+            if (
+                self.training
+                and random.random() < 0.25
+                and self.min_weight != (0.0, 0.0)
+            ):
+                weight1 = weight1.clamp(
+                    min=self.min_weight[0], max=1.0 - self.min_weight[1]
+                )
+
+        src1 = src1 * weight1
+        src2 = src2 * (1.0 - weight1)
+
+        assert src1.shape[-1] == self.dim1, (src1.shape[-1], self.dim1)
+        assert src2.shape[-1] == self.dim2, (src2.shape[-1], self.dim2)
+
+        src1_dim = self.dim1
+        src2_dim = self.dim2
+
+        if src1_dim != src2_dim:
+            if src1_dim < src2_dim:
+                src1 = torch.nn.functional.pad(src1, (0, src2_dim - src1_dim))
+            else:
+                src1 = src1[:src2_dim]
+
+        return src1 + src2
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        dropout_rate: float,
+        max_len: int = 5000,
+        is_pnnx: bool = False,
+        x_size: int = 0,
+        left_context_len: int = 0,
+    ) -> None:
+        """Construct a PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(dropout_rate)
+        self.is_pnnx = is_pnnx
+        self.x_size = x_size
+        self.left_context_len = left_context_len
+        self.pe = None
+        if is_pnnx:
+            x_size_left = x_size + left_context_len
+            self.extend_pe(torch.tensor(0.0).expand(x_size_left))
+            self.pe = self.pe[:, :-left_context_len]
+            assert self.pe.size(1) == x_size + left_context_len - 1 + x_size, (
+                self.pe.size(1),
+                x_size,
+                left_context_len,
+                x_size,
+                self.pe.shape,
+            )
+        else:
+            self.extend_pe(torch.tensor(0.0).expand(max_len))
+
+    def extend_pe(self, x: Tensor, left_context_len: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_left = x.size(0) + left_context_len
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_left * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use positive relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_left, self.d_model)
+        pe_negative = torch.zeros(x_size_left, self.d_model)
+        position = torch.arange(0, x_size_left, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor, left_context_len: int = 0) -> Tensor:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (time, batch, `*`).
+            left_context_len: (int): Length of cached left context.
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, left_context_len + 2*time-1, `*`).
+
+        """
+        if self.is_pnnx:
+            assert self.x_size == x.size(0), (self.x_size, x.size(0))
+            assert self.left_context_len == left_context_len, (
+                self.left_context_len,
+                left_context_len,
+            )
+            return self.pe
+
+        self.extend_pe(x, left_context_len)
+        x_size_left = x.size(0) + left_context_len
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_left
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(0),
+        ]
+        return self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttentionPermute(nn.Module):
+    """ncnn does not support permuatation relating to the batch axis 0.
+    This is a workaround for exporting to ncnn via PNNX.
+    """
+
+    def __init__(self, kind: int):
+        super().__init__()
+        self.kind = kind
+        assert self.kind in (2, 3), self.kind
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.kind == 2:
+            return x.permute(1, 0, 2)
+        elif self.kind == 3:
+            return x.permute(1, 2, 0)
+        else:
+            assert False, f"Unsupported kind {self.kind}"
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
+    we have to write up the differences.
+
+
+    Args:
+        embed_dim: total dimension of the model.
+        attention_dim: dimension in the attention module, may be less or more than embed_dim
+            but must be a multiple of num_heads.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        attention_dim: int,
+        num_heads: int,
+        pos_dim: int,
+        dropout: float = 0.0,
+        is_pnnx: bool = False,
+        left_context_len: int = 0,
+        x_size: int = 0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.attention_dim = attention_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = attention_dim // num_heads
+        self.pos_dim = pos_dim
+        assert self.head_dim % 2 == 0, self.head_dim
+        assert self.head_dim * num_heads == attention_dim, (
+            self.head_dim,
+            num_heads,
+            attention_dim,
+        )
+
+        self.is_pnnx = is_pnnx
+
+        self.my_permute_pqv = RelPositionMultiheadAttentionPermute(kind=2)
+        self.my_permute_k_pos = RelPositionMultiheadAttentionPermute(kind=3)
+        self.left_context_len = left_context_len
+        self.x_size = x_size
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5, dividing it between the query and key.
+        in_proj_dim = (
+            2 * attention_dim
+            + attention_dim // 2  # query (attention_dim,), key (attention_dim,)
+            + pos_dim * num_heads  # value (attention_dim // 2,)
+        )  # positional encoding query (pos_dim * num_heads, )
+
+        self.in_proj = ScaledLinear(
+            embed_dim, in_proj_dim, bias=True, initial_scale=self.head_dim**-0.25
+        )
+
+        # self.whiten_values is applied on the values in forward();
+        # it just copies the keys but prevents low-rank distribution by modifying grads.
+        self.whiten_values = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(
+            embed_dim, num_heads * pos_dim, bias=False, initial_scale=0.05
+        )
+
+        # the following are for diagnosics only, see --print-diagnostics option.
+        # they only copy their inputs.
+        self.copy_pos_query = Identity()
+        self.copy_query = Identity()
+
+        self.out_proj = ScaledLinear(
+            attention_dim // 2, embed_dim, bias=True, initial_scale=0.05
+        )
+
+        self.in_proj2 = nn.Linear(embed_dim, attention_dim // 2, bias=False)
+        self.out_proj2 = ScaledLinear(
+            attention_dim // 2, embed_dim, bias=True, initial_scale=0.05
+        )
+        # self.whiten_values2 is applied on the values in forward2()
+        self.whiten_values2 = Whiten(
+            num_groups=num_heads,
+            whitening_limit=2.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x: input to be projected to query, key, value
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Returns: (attn_output, attn_weights)
+
+             - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size,
+                E is the embedding dimension.
+              - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads
+                 and S is the sequence length.
+        """
+        x, weights = self.multi_head_attention_forward(
+            self.in_proj(x),
+            self.linear_pos(pos_emb),
+            self.attention_dim,
+            self.num_heads,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            attn_mask=attn_mask,
+        )
+        return x, weights
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
+        r"""
+        Args:
+            x: input to be projected to query, key, value
+            pos_emb: Positional embedding tensor
+
+        Shape:
+            - Inputs:
+            - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - cached_key: :math:`(left_context_len, N, K)`, where N is the batch size, K is the key dimension.
+            - cached_val: :math:`(left_context_len, N, V)`, where N is the batch size, V is the value dimension.
+
+            - Returns: (attn_output, attn_weights, cached_key, cached_val)
+
+              - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size,
+                E is the embedding dimension.
+              - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads
+                 and S is the sequence length.
+              - cached_key: :math:`(left_context_len, N, K)`, updated cached attention key tensor of
+                left context
+              - cached_val: :math:`(left_context_len, N, K)`, updated cached attention value tensor of
+        """
+        (
+            x,
+            weights,
+            cached_key,
+            cached_val,
+        ) = self.streaming_multi_head_attention_forward(
+            self.in_proj(x),
+            self.linear_pos(pos_emb),
+            self.attention_dim,
+            self.num_heads,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            cached_key=cached_key,
+            cached_val=cached_val,
+        )
+        return x, weights, cached_key, cached_val
+
+    def multi_head_attention_forward(
+        self,
+        x_proj: Tensor,
+        pos: Tensor,
+        attention_dim: int,
+        num_heads: int,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x_proj: the projected input, to be split into query, key, value.
+            pos: head-specific biases arising from the positional embeddings.
+            attention_dim: dimension inside attention mechanism
+            num_heads: parallel attention heads.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is
+              the attention dimension.  Will be split into (query, key, value, pos).
+            - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence
+              length, N is the batch size, and A is the attention dim.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+              E is the embedding dimension.
+            - attn_weights: :math:`(N * H, S, S)` where N is the batch size,
+              H is the num-heads, S is the sequence length.
+        """
+
+        seq_len, bsz, _ = x_proj.size()
+
+        head_dim = attention_dim // num_heads
+        pos_dim = self.pos_dim  # positional-encoding dim per head
+        assert (
+            head_dim * num_heads == attention_dim
+        ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}"
+
+        # self-attention
+        q = x_proj[..., 0:attention_dim]
+        k = x_proj[..., attention_dim : 2 * attention_dim]
+        value_dim = attention_dim // 2
+        v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim]
+        # p is the position-encoding query, its dimension is num_heads*pos_dim..
+        p = x_proj[..., 2 * attention_dim + value_dim :]
+
+        k = self.whiten_keys(k)  # does nothing in the forward pass.
+        v = self.whiten_values(v)  # does nothing in the forward pass.
+        q = self.copy_query(q)  # for diagnostics only, does nothing.
+        p = self.copy_pos_query(p)  # for diagnostics only, does nothing.
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, seq_len, seq_len]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    seq_len,
+                    seq_len,
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.reshape(seq_len, bsz, num_heads, head_dim)
+        p = p.reshape(seq_len, bsz, num_heads, pos_dim)
+        k = k.reshape(seq_len, bsz, num_heads, head_dim)
+        v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == seq_len, "{} == {}".format(
+                key_padding_mask.size(1), seq_len
+            )
+
+        q = q.permute(1, 2, 0, 3)  # (batch, head, time1, head_dim)
+        p = p.permute(1, 2, 0, 3)  # (batch, head, time1, pos_dim)
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+
+        seq_len2 = 2 * seq_len - 1
+        pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1)
+        # pos shape now: (batch, head, pos_dim, seq_len2)
+
+        # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2)
+        #  [where seq_len2 represents relative position.]
+        pos_weights = torch.matmul(p, pos)
+        # the following .as_strided() expression converts the last axis of pos_weights from relative
+        # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+        # not, but let this code define which way round it is supposed to be.
+        pos_weights = pos_weights.as_strided(
+            (bsz, num_heads, seq_len, seq_len),
+            (
+                pos_weights.stride(0),
+                pos_weights.stride(1),
+                pos_weights.stride(2) - pos_weights.stride(3),
+                pos_weights.stride(3),
+            ),
+            storage_offset=pos_weights.stride(3) * (seq_len - 1),
+        )
+
+        # caution: they are really scores at this point.
+        attn_output_weights = torch.matmul(q, k) + pos_weights
+
+        if not torch.jit.is_scripting():
+            if training and random.random() < 0.1:
+                # This is a harder way of limiting the attention scores to not be too large.
+                # It incurs a penalty if any of them has an absolute value greater than 50.0.
+                # this should be outside the normal range of the attention scores.  We use
+                # this mechanism instead of, say, a limit on entropy, because once the entropy
+                # gets very small gradients through the softmax can become very small, and
+                # some mechanisms like that become ineffective.
+                attn_output_weights = penalize_abs_values_gt(
+                    attn_output_weights, limit=25.0, penalty=1.0e-04
+                )
+
+        # attn_output_weights: (batch, head, time1, time2)
+        attn_output_weights = attn_output_weights.view(
+            bsz * num_heads, seq_len, seq_len
+        )
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights = attn_output_weights.masked_fill(
+                    attn_mask, float("-inf")
+                )
+            else:
+                attn_output_weights = attn_output_weights + attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, seq_len, seq_len
+            )
+
+        # Using this version of softmax, defined in scaling.py,
+        # should save a little of the memory used in backprop by, if
+        # we are in automatic mixed precision mode (amp) == autocast,
+        # only storing the half-precision output for backprop purposes.
+        attn_output_weights = softmax(attn_output_weights, dim=-1)
+
+        # If we are using chunk-wise attention mask and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`. At this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax). So we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, seq_len, seq_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, seq_len, seq_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2]
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, attention_dim // 2)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        return attn_output, attn_output_weights
+
+    def streaming_multi_head_attention_forward(
+        self,
+        x_proj: Tensor,
+        pos: Tensor,
+        attention_dim: int,
+        num_heads: int,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        cached_key: Tensor,
+        cached_val: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
+        r"""
+        Args:
+            x_proj: the projected input, to be split into query, key, value.
+            pos: head-specific biases arising from the positional embeddings.
+            attention_dim: dimension inside attention mechanism
+            num_heads: parallel attention heads.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            cached_key: cached attention key tensor of left context.
+            cached_val: cached attention value tensor of left context.
+
+        Shape:
+            Inputs:
+            - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is
+              the attention dimension.  Will be split into (query, key, value, pos).
+            - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence
+              length, N is the batch size, and A is the attention dim.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+              E is the embedding dimension.
+            - attn_weights: :math:`(N * H, S, S)` where N is the batch size,
+              H is the num-heads, S is the sequence length.
+            - cached_key: :math:`(left_context_len, N, K)`, updated cached attention key tensor of left context.
+            - cached_val: :math:`(left_context_len, N, K)`, updated cached attention value tensor of left context.
+        """
+        if not self.is_pnnx:
+            seq_len, bsz, _ = x_proj.size()
+            assert seq_len == self.x_size, (seq_len, self.x_size)
+        else:
+            seq_len = self.x_size
+            bsz = 1
+
+        head_dim = attention_dim // num_heads
+        pos_dim = self.pos_dim  # positional-encoding dim per head
+        assert (
+            head_dim * num_heads == attention_dim
+        ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}"
+
+        # self-attention
+        q = x_proj[:, :, 0:attention_dim]  # (x_size, N, attention_dim)
+        #  return q, q, q, q
+        k = x_proj[:, :, attention_dim : 2 * attention_dim]
+        # k is (x_size, N, attention_dim)
+        value_dim = attention_dim // 2
+        v = x_proj[:, :, 2 * attention_dim : 2 * attention_dim + value_dim]
+        # v is (x_size, 0, attention_dim//2)
+
+        # p is the position-encoding query, its dimension is num_heads*pos_dim..
+        p = x_proj[:, :, 2 * attention_dim + value_dim :]
+        # p is (x_size, N, pos_dim * num_heads)
+
+        if not self.is_pnnx:
+            left_context_len = cached_key.shape[0]
+        else:
+            assert cached_key.shape[0] == self.left_context_len, (
+                cached_key.shape,
+                self.left_context_len,
+            )
+            left_context_len = self.left_context_len
+
+        assert left_context_len > 0, left_context_len
+        assert cached_key.shape[0] == cached_val.shape[0], (
+            cached_key.shape,
+            cached_val.shape,
+        )
+        # Note: We need to fix the Concat in ncnn
+        # cached_key is (1, 64, 192) in ncnn
+        # k is (16, 192) in ncnn
+        # Pad cached left contexts
+        k = torch.cat([cached_key, k], dim=0)
+        # (left_context_len + x_size, N, attention_dim)
+
+        v = torch.cat([cached_val, v], dim=0)
+        # v: (left_context_len + x_size, N, attention_dim//2)
+        # Update cached left contexts
+        if not self.is_pnnx:
+            cached_key = k[-left_context_len:, ...]
+            cached_val = v[-left_context_len:, ...]
+        else:
+            cached_key = k[self.x_size :]
+            cached_val = v[self.x_size :]
+            assert cached_key.shape[0] == left_context_len, (
+                cached_key.shape,
+                left_context_len,
+            )
+            assert cached_val.shape[0] == left_context_len, (
+                cached_val.shape,
+                left_context_len,
+            )
+
+        if not self.is_pnnx:
+            # The length of key and value
+            kv_len = k.shape[0]
+        else:
+            kv_len = left_context_len + self.x_size
+            assert kv_len == k.shape[0], (kv_len, k.shape)
+
+        if not self.is_pnnx:
+            q = q.reshape(seq_len, bsz, num_heads, head_dim)
+            p = p.reshape(seq_len, bsz, num_heads, pos_dim)
+            k = k.reshape(kv_len, bsz, num_heads, head_dim)
+
+            v = v.reshape(kv_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+            # v is (bsz * num_heads, kv_len, head_dim//2)
+
+            q = q.permute(1, 2, 0, 3)  # (batch, head, time1, head_dim)
+            p = p.permute(1, 2, 0, 3)  # (batch, head, time1, pos_dim)
+            k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+
+            seq_len2 = 2 * seq_len - 1 + left_context_len
+            pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1)
+            # pos shape now: (batch, head, pos_dim, seq_len2)
+        else:
+            q = q.reshape(seq_len, num_heads, head_dim)
+            p = p.reshape(seq_len, num_heads, pos_dim)
+            k = k.reshape(kv_len, num_heads, head_dim)
+            #  v = v.reshape(kv_len, num_heads, head_dim // 2).permute(1, 0, 2)
+            v = v.reshape(kv_len, num_heads, head_dim // 2)
+            v = self.my_permute_pqv(v)
+            # v is (num_heads, kv_len, head_dim//2) e.g., (8, 80, 12)
+
+            #  q = q.permute(1, 0, 2)  # (head, time1, head_dim)
+            #  p = p.permute(1, 0, 2)  # (head, time1, pos_dim)
+            #  k = k.permute(1, 2, 0)  # (head, d_k, time2)
+
+            q = self.my_permute_pqv(q)  # (head, time1, head_dim), e.g., (8, 16, 24)
+            p = self.my_permute_pqv(p)  # (head, time1, pos_dim), e.g., (8, 16, 4)
+            k = self.my_permute_k_pos(k)  # (head, d_k, time2) e.g., (8, 24, 80)
+
+            seq_len2 = 2 * seq_len - 1 + left_context_len
+            #  pos = pos.reshape(seq_len2, num_heads, pos_dim).permute(1, 2, 0)
+            # pos shape now: (head, pos_dim, seq_len2)
+
+            pos = pos.reshape(seq_len2, num_heads, pos_dim)
+            pos = self.my_permute_k_pos(
+                pos
+            )  # (head, pos_dim, seq_len2), e.g, (8, 4, 95)
+
+        # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2) ,e.g., (1, 8, 16, 95)
+        #  [where seq_len2 represents relative position.]
+        pos_weights = torch.matmul(p, pos)
+
+        # the following .as_strided() expression converts the last axis of pos_weights from relative
+        # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+        # not, but let this code define which way round it is supposed to be.
+
+        if not self.is_pnnx:
+            pos_weights = pos_weights.as_strided(
+                (bsz, num_heads, seq_len, kv_len),
+                (
+                    pos_weights.stride(0),
+                    pos_weights.stride(1),
+                    pos_weights.stride(2) - pos_weights.stride(3),
+                    pos_weights.stride(3),
+                ),
+                storage_offset=pos_weights.stride(3) * (seq_len - 1),
+            )
+        else:
+            pos_weights = pos_weights.as_strided(
+                (num_heads, seq_len, kv_len),
+                (
+                    pos_weights.stride(0),
+                    pos_weights.stride(1) - pos_weights.stride(2),
+                    pos_weights.stride(2),
+                ),
+                storage_offset=pos_weights.stride(2) * (seq_len - 1),
+            )
+
+        # caution: they are really scores at this point.
+        attn_output_weights = torch.matmul(q, k) + pos_weights
+
+        # attn_output_weights: (batch, head, time1, time2)
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, seq_len, kv_len)
+
+        # Using this version of softmax, defined in scaling.py,
+        # should save a little of the memory used in backprop by, if
+        # we are in automatic mixed precision mode (amp) == autocast,
+        # only storing the half-precision output for backprop purposes.
+        attn_output_weights = softmax(attn_output_weights, dim=-1)
+
+        attn_output = torch.bmm(attn_output_weights, v)
+
+        assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2]
+        # (8, 16, 12)
+
+        if not self.is_pnnx:
+            attn_output = (
+                attn_output.transpose(0, 1)
+                .contiguous()
+                .view(seq_len, bsz, attention_dim // 2)
+            )
+            attn_output = nn.functional.linear(
+                attn_output, out_proj_weight, out_proj_bias
+            )
+        else:
+            attn_output = self.my_permute_pqv(attn_output)  # (1, 0, 2)
+            attn_output = attn_output.reshape(seq_len, bsz, attention_dim // 2)
+            # We have changed InnerProduct in ncnn to treat
+            # (seq_len, bsz, attention_dim//2) as
+            # (seq_len, attention_dim//2)
+
+            attn_output = nn.functional.linear(
+                attn_output, out_proj_weight, out_proj_bias
+            )
+
+        return attn_output, attn_output_weights, cached_key, cached_val
+
+    def forward2(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """
+        Second forward function, where we re-use the attn_weights returned by the first forward function
+        but with different input.
+        Args:
+               x: input, of shape (seq_len, batch_size, embed_dim)
+           attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len)
+        Returns:
+               output of the same shape as x, i.e. (seq_len, batch_size, embed_dim)
+        """
+        num_heads = self.num_heads
+        (seq_len, bsz, embed_dim) = x.shape
+        head_dim = self.attention_dim // num_heads
+        # v: (tgt_len, bsz, embed_dim // 2)
+        v = self.in_proj2(x)
+        v = self.whiten_values2(v)  # does nothing in the forward pass.
+        v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1)
+
+        # now v: (bsz * num_heads, seq_len, head_dim // 2)
+        attn_output = torch.bmm(attn_weights, v)
+
+        if not torch.jit.is_scripting():
+            if random.random() < 0.001 or __name__ == "__main__":
+                self._print_attn_stats(attn_weights, attn_output)
+
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, bsz, self.attention_dim // 2)
+        )
+        # returned value is of shape (seq_len, bsz, embed_dim), like x.
+        return self.out_proj2(attn_output)
+
+    def streaming_forward2(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_val: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Second forward function, where we re-use the attn_weights returned by the first forward function
+        but with different input.
+        Args:
+            x: input, of shape (seq_len, batch_size, embed_dim)
+            attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len)
+            cached_val: cached attention value tensor of left context.
+        Returns:
+            - output of the same shape as x, i.e. (seq_len, batch_size, embed_dim)
+            - updated cached attention value tensor of left context.
+        """
+        num_heads = self.num_heads
+
+        assert x.shape[0] == self.x_size, (x.shape[0], self.x_size)
+        assert x.shape[2] == self.embed_dim, (x.shape[2], self.embed_dim)
+
+        if not self.is_pnnx:
+            (seq_len, bsz, embed_dim) = x.shape
+        else:
+            seq_len = self.x_size
+            bsz = 1
+            embed_dim = self.embed_dim
+
+        head_dim = self.attention_dim // num_heads
+        # v: (tgt_len, bsz, embed_dim // 2)
+        v = self.in_proj2(x)
+
+        assert cached_val.shape[0] == self.left_context_len, (
+            cached_val.shape[0],
+            self.left_context_len,
+        )
+
+        left_context_len = self.left_context_len
+        assert left_context_len > 0, left_context_len
+        v = torch.cat([cached_val, v], dim=0)
+        cached_val = v[-left_context_len:]
+
+        seq_len2 = left_context_len + seq_len
+        if not self.is_pnnx:
+            v = v.reshape(seq_len2, bsz * num_heads, head_dim // 2).transpose(0, 1)
+        else:
+            v = v.reshape(seq_len2, bsz * num_heads, head_dim // 2)
+            #  v = v.permute(1, 0, 2)
+            v = self.my_permute_pqv(v)
+
+        # now v: (bsz * num_heads, seq_len, head_dim // 2)
+        attn_output = torch.bmm(attn_weights, v)
+
+        if not self.is_pnnx:
+            # attn_output: (bsz * num_heads, seq_len, head_dim)
+            attn_output = (
+                attn_output.transpose(0, 1)
+                .contiguous()
+                .view(seq_len, bsz, self.attention_dim // 2)
+            )
+        else:
+            attn_output = self.my_permute_pqv(attn_output)  # (1, 0, 2)
+            attn_output = attn_output.reshape(seq_len, bsz, self.attention_dim // 2)
+            # We have changed InnerProduct in ncnn to ignore bsz
+            # when invoking self.out_proj2(attn_output)
+
+        # returned value is of shape (seq_len, bsz, embed_dim), like x.
+        return self.out_proj2(attn_output), cached_val
+
+    def _print_attn_stats(self, attn_weights: Tensor, attn_output: Tensor):
+        # attn_weights: (batch_size * num_heads, seq_len, seq_len)
+        # attn_output: (bsz * num_heads, seq_len, head_dim)
+        (n, seq_len, head_dim) = attn_output.shape
+        num_heads = self.num_heads
+        bsz = n // num_heads
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_output = attn_output.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .reshape(bsz, num_heads, seq_len)
+                    .mean(dim=(0, 2))
+                )
+                attn_output = attn_output.reshape(bsz, num_heads, seq_len, head_dim)
+                attn_output = attn_output.permute(1, 0, 2, 3).reshape(
+                    num_heads, bsz * seq_len, head_dim
+                )
+                attn_output_mean = attn_output.mean(dim=1, keepdim=True)
+                attn_output = attn_output - attn_output_mean
+                attn_covar = torch.matmul(attn_output.transpose(1, 2), attn_output) / (
+                    bsz * seq_len
+                )
+                # attn_covar: (num_heads, head_dim, head_dim)
+                # eigs, _ = torch.symeig(attn_covar)
+                # logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}")
+
+                attn_covar = _diag(attn_covar).mean(dim=1)  # (num_heads,)
+                embed_dim = self.in_proj2.weight.shape[1]
+                in_proj_covar = (
+                    self.in_proj2.weight.reshape(num_heads, head_dim, embed_dim) ** 2
+                ).mean(dim=(1, 2))
+                out_proj_covar = (
+                    self.out_proj2.weight.reshape(embed_dim, num_heads, head_dim) ** 2
+                ).mean(dim=(0, 2))
+                logging.info(
+                    f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}"
+                )
+
+
+class FeedforwardModule(nn.Module):
+    """Feedforward module in Zipformer model."""
+
+    def __init__(self, d_model: int, feedforward_dim: int, dropout: float):
+        super(FeedforwardModule, self).__init__()
+        self.in_proj = nn.Linear(d_model, feedforward_dim)
+        self.balancer = ActivationBalancer(
+            feedforward_dim, channel_dim=-1, max_abs=10.0, min_prob=0.25
+        )
+        self.activation = DoubleSwish()
+        self.dropout = nn.Dropout(dropout)
+        self.out_proj = ScaledLinear(feedforward_dim, d_model, initial_scale=0.01)
+
+    def forward(self, x: Tensor):
+        x = self.in_proj(x)
+        x = self.balancer(x)
+        x = self.activation(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Zipformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        is_pnnx: bool = False,
+        x_size: int = 0,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0, kernel_size
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            2 * channels,
+            channel_dim=1,
+            max_abs=10.0,
+            min_positive=0.05,
+            max_positive=1.0,
+        )
+
+        # Will pad cached left context
+        self.lorder = kernel_size - 1
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channels,
+            channel_dim=1,
+            min_positive=0.05,
+            max_positive=1.0,
+            max_abs=20.0,
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.05,
+        )
+
+        self.is_pnnx = is_pnnx
+        self.x_size = x_size
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional):
+               (batch, #time), contains bool in masked positions.
+
+        Returns:
+            - Output tensor (#time, batch, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        # 1D Depthwise Conv
+        # Make depthwise_conv causal by
+        # manualy padding self.lorder zeros to the left
+        x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cache: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch:
+               (batch, #time), contains bool in masked positions.
+            cache: Cached left context for depthwise_conv, with shape of
+               (batch, channels, #kernel_size-1). Only used in real streaming decoding.
+
+        Returns:
+            A tuple of 2 tensors:
+            - Output tensor (#time, batch, channels).
+            - New cached left context, with shape of (batch, channels, #kernel_size-1).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        assert cache.shape == (x.size(0), x.size(1), self.lorder), (
+            cache.shape,
+            (x.size(0), x.size(1), self.lorder),
+        )
+        x = torch.cat([cache, x], dim=2)
+
+        cache = x[:, :, self.x_size :]
+
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1), cache
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = (T-3)//2 - 2 == (T-7)//2
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        dropout: float = 0.1,
+        is_pnnx: bool = False,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, (T-7)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer2_channels:
+            Number of channels in layer2
+          layer3_channels:
+            Number of channels in layer3
+          is_pnnx:
+            True if we are converting the model to PNNX format.
+            False otherwise.
+        """
+        assert in_channels >= 7, in_channels
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=(0, 1),  # (time, freq)
+            ),
+            # After this layer (N, 1, T, C) -> (N, layer1_channels, T-2, C)
+            ActivationBalancer(layer1_channels, channel_dim=1),
+            DoubleSwish(),
+            nn.Conv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+                padding=0,
+            ),
+            # After this layer (N, layer1_channels, T-2, C) -> (N, layer2_channels, ((T-2) - 3)//2+1, (C-3)//2+1)
+            # i.e., (N, layer2_channels, (T-5)//2+1, (C-3)//2+1)
+            # i.e., (N, layer2_channels, (T-3)//2, (C-1)//2)
+            ActivationBalancer(layer2_channels, channel_dim=1),
+            DoubleSwish(),
+            nn.Conv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=(1, 2),  # (time, freq)
+            ),
+            # After this layer, (N, layer2_channels, (T-3)//2, (C-1)//2)
+            # ->
+            # (N, layer3_channels, (T-3)//2-2, ((C-1)//2 - 3)//2 + 1)
+            # (N, layer3_channels, (T-7)//2, (C-3)//4)
+            ActivationBalancer(layer3_channels, channel_dim=1),
+            DoubleSwish(),
+        )
+        out_height = (((in_channels - 1) // 2) - 1) // 2
+        self.out = ScaledLinear(out_height * layer3_channels, out_channels)
+        self.dropout = nn.Dropout(dropout)
+
+        # ncnn supports only batch size == 1
+        self.is_pnnx = is_pnnx
+        self.conv_out_dim = self.out.weight.shape[1]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, (T-7)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+
+        if torch.jit.is_tracing() and self.is_pnnx:
+            x = x.permute(0, 2, 1, 3).reshape(1, -1, self.conv_out_dim)
+            x = self.out(x)
+        else:
+            # Now x is of shape (N, odim, (T-7)//2, ((idim-1)//2 - 1)//2)
+            b, c, t, f = x.size()
+            x = self.out(x.transpose(1, 2).reshape(b, t, c * f))
+        # Now x is of shape (N, (T-7)//2, odim)
+        x = self.dropout(x)
+        return x
+
+
+def _test_zipformer_main():
+    feature_dim = 50
+    batch_size = 5
+    seq_len = 47
+    feature_dim = 50
+    # Just make sure the forward pass runs.
+
+    c = Zipformer(
+        num_features=feature_dim,
+        encoder_dims=(64, 96),
+        encoder_unmasked_dims=(48, 64),
+        nhead=(4, 4),
+        decode_chunk_size=4,
+    )
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    assert ((seq_len - 7) // 2 + 1) // 2 == f[0].shape[1], (seq_len, f.shape[1])
+    f[0].sum().backward()
+    c.eval()
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+def _test_conv2d_subsampling():
+    num_features = 80
+    encoder_dims = 384
+    dropout = 0.1
+    encoder_embed = Conv2dSubsampling(num_features, encoder_dims, dropout=dropout)
+    for i in range(20, 40):
+        x = torch.rand(2, i, num_features)
+        y = encoder_embed(x)
+        assert (x.shape[1] - 7) // 2 == y.shape[1], (x.shape[1], y.shape[1])
+
+
+def _test_pooling_module():
+    N, S, C = 2, 12, 32
+    chunk_len = 4
+    m = PoolingModule(d_model=C)
+
+    # test chunk-wise forward with padding_mask
+    x = torch.randn(S, N, C)
+    y = m(x)
+    cached_len = torch.zeros(N, dtype=torch.int32)
+    cached_avg = torch.zeros(N, C)
+    for i in range(S // chunk_len):
+        start = i * chunk_len
+        end = start + chunk_len
+        x_chunk = x[start:end]
+        y_chunk, cached_len, cached_avg = m.streaming_forward(
+            x_chunk,
+            cached_len=cached_len,
+            cached_avg=cached_avg,
+        )
+        assert torch.allclose(y_chunk, y[start:end]), (y_chunk, y[start:end])
+
+
+def _test_state_stack_unstack():
+    m = Zipformer(
+        num_features=80,
+        encoder_dims=(64, 96),
+        encoder_unmasked_dims=(48, 64),
+        nhead=(4, 4),
+        zipformer_downsampling_factors=(4, 8),
+        num_left_chunks=2,
+        decode_chunk_size=8,
+    )
+    s1 = m.get_init_state()
+    s2 = m.get_init_state()
+    states = stack_states([s1, s2])
+    new_s1, new_s2 = unstack_states(states)
+    for i in range(m.num_encoders * 7):
+        for x, y in zip(s1[i], new_s1[i]):
+            assert torch.equal(x, y)
+        for x, y in zip(s2[i], new_s2[i]):
+            assert torch.equal(x, y)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_zipformer_main()
+    _test_conv2d_subsampling()
+    _test_pooling_module()
+    _test_state_stack_unstack()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/asr_datamodule.py
new file mode 120000
index 000000000..a3a1584d1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless8/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode.py
new file mode 100755
index 000000000..35158ced4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode.py
@@ -0,0 +1,818 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from librispeech import LibriSpeech
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--right-padding",
+        type=int,
+        default=64,
+        help="Padding frames at the end of features",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += params.right_padding
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.right_padding),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        params.suffix += f"-right-padding-{params.right_padding}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode_gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode_gigaspeech.py
new file mode 100755
index 000000000..a4f52ad7f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode_gigaspeech.py
@@ -0,0 +1,837 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+
+# from asr_datamodule import LibriSpeechAsrDataModule
+from gigaspeech_asrmodule import GigaSpeechAsrDataModule
+from gigaspeech_scoring import asr_text_post_processing
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-padding",
+        type=int,
+        default=64,
+        help="Padding frames at the end of features",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += params.right_padding
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.right_padding),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = post_processing(results)
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    """
+    This scripts test a libri model with libri BPE
+    on Gigaspeech.
+    """
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / (params.decoding_method + "_gigaspeech")
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        params.suffix += f"-right-padding-{params.right_padding}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode_stream.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode_stream.py
new file mode 120000
index 000000000..2b4596e0b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decode_stream.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/decode_stream.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decoder.py
new file mode 120000
index 000000000..4e79a10e0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/do_not_use_it_directly.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/do_not_use_it_directly.py
new file mode 120000
index 000000000..beeffaa03
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/do_not_use_it_directly.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/encoder_interface.py
new file mode 120000
index 000000000..24f414dd1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn-zh.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn-zh.py
new file mode 120000
index 000000000..c0e71accf
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn-zh.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py
new file mode 100755
index 000000000..8f2178b1d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py
@@ -0,0 +1,370 @@
+#!/usr/bin/env python3
+
+"""
+Please see
+https://k2-fsa.github.io/icefall/model-export/export-ncnn.html
+for more details about how to use this file.
+
+We use
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+to demonstrate the usage of this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export to ncnn
+
+./pruned_transducer_stateless7_streaming/export-for-ncnn.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --exp-dir $repo/exp \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --decode-chunk-len 32 \
+  --num-encoder-layers "2,4,3,2,4" \
+  --feedforward-dims "1024,1024,2048,2048,1024" \
+  --nhead "8,8,8,8,8" \
+  --encoder-dims "384,384,384,384,384" \
+  --attention-dims "192,192,192,192,192" \
+  --encoder-unmasked-dims "256,256,256,256,256" \
+  --zipformer-downsampling-factors "1,2,4,8,2" \
+  --cnn-module-kernels "31,31,31,31,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+cd $repo/exp
+
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+You can find converted models at
+https://huggingface.co/csukuangfj/sherpa-ncnn-streaming-zipformer-en-2023-02-13
+
+See ./streaming-ncnn-decode.py
+and
+https://github.com/k2-fsa/sherpa-ncnn
+for usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model
+from scaling_converter import convert_scaled_to_non_scaled
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: torch.nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    encoder_model.__class__.forward = encoder_model.__class__.streaming_forward
+
+    decode_chunk_len = encoder_model.decode_chunk_size * 2
+    pad_length = 7
+    T = decode_chunk_len + pad_length  # 32 + 7 = 39
+
+    logging.info(f"decode_chunk_len: {decode_chunk_len}")
+    logging.info(f"T: {T}")
+
+    x = torch.zeros(1, T, 80, dtype=torch.float32)
+    states = encoder_model.get_init_state()
+
+    traced_model = torch.jit.trace(encoder_model, (x, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: torch.nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: torch.nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-ncnn")
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_pnnx=True)
+
+    encoder_num_param = sum([p.numel() for p in model.encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in model.decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in model.joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    logging.info("Using torch.jit.trace()")
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-onnx.py
new file mode 120000
index 000000000..137fa8cec
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-onnx.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/export-onnx.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export.py
new file mode 120000
index 000000000..6a009311c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/export.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech.py
new file mode 120000
index 000000000..6c6b08d3f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless8/gigaspeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech_asrmodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech_asrmodule.py
new file mode 120000
index 000000000..54f18a4f0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech_asrmodule.py
@@ -0,0 +1 @@
+../../../gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech_scoring.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech_scoring.py
new file mode 120000
index 000000000..fdfa6ce4b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/gigaspeech_scoring.py
@@ -0,0 +1 @@
+../../../gigaspeech/ASR/pruned_transducer_stateless2/gigaspeech_scoring.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_pretrained.py
new file mode 120000
index 000000000..c427e7709
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_trace_export.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_trace_export.py
new file mode 120000
index 000000000..44ecf1780
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_trace_export.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/jit_trace_export.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_trace_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_trace_pretrained.py
new file mode 120000
index 000000000..762d38b73
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/jit_trace_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/joiner.py
new file mode 120000
index 000000000..2a9c1ca5f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/librispeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/librispeech.py
new file mode 120000
index 000000000..7c22bc4b7
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/librispeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless8/librispeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/model.py
new file mode 120000
index 000000000..17ced2998
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_check.py
new file mode 120000
index 000000000..8ed81ba1c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_check.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/onnx_check.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_model_wrapper.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_model_wrapper.py
new file mode 120000
index 000000000..c780015d1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_model_wrapper.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_pretrained.py
new file mode 120000
index 000000000..da0236c2d
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/onnx_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/optim.py
new file mode 120000
index 000000000..6c5f3fc3e
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/pretrained.py
new file mode 120000
index 000000000..4c519b771
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/pretrained.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/scaling.py
new file mode 120000
index 000000000..420bc4149
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/scaling_converter.py
new file mode 120000
index 000000000..b6cc7dc13
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming-ncnn-decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming-ncnn-decode.py
new file mode 120000
index 000000000..d137d28ad
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming-ncnn-decode.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_beam_search.py
new file mode 120000
index 000000000..dee9005d0
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_decode.py
new file mode 100755
index 000000000..2904f086c
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_decode.py
@@ -0,0 +1,610 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding_method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from librispeech import LibriSpeech
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming_multi/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/test_model.py
new file mode 120000
index 000000000..1a9ba93e6
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/test_model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/test_model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
new file mode 100755
index 000000000..565dc7a16
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
@@ -0,0 +1,1365 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --full-libri 1 \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decoder import Decoder
+from gigaspeech import GigaSpeech
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from librispeech import LibriSpeech
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--full-libri",
+        type=str2bool,
+        default=True,
+        help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--giga-prob",
+        type=float,
+        default=0.5,
+        help="The probability to select a batch from the GigaSpeech dataset",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_libri(c: Cut) -> bool:
+    """Return True if this cut is from the LibriSpeech dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of GigaSpeech to dict(origin='giga')
+      See ../local/preprocess_gigaspeech.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    giga_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    libri_tot_loss = MetricsTracker()
+    giga_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.giga_prob, params.giga_prob]
+    iter_libri = iter(train_dl)
+    iter_giga = iter(giga_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_libri if idx == 0 else iter_giga
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            name = "libri" if idx == 0 else "giga"
+            logging.info(f"{name} reaches end of dataloader")
+            break
+
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        libri = is_libri(batch["supervisions"]["cut"][0])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            if libri:
+                libri_tot_loss = (
+                    libri_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "libri"  # for logging only
+            else:
+                giga_tot_loss = (
+                    giga_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "giga"
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], "
+                f"libri_tot_loss[{libri_tot_loss}], "
+                f"giga_tot_loss[{giga_tot_loss}], "
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                libri_tot_loss.write_summary(
+                    tb_writer, "train/libri_tot_", params.batch_idx_train
+                )
+                giga_tot_loss.write_summary(
+                    tb_writer, "train/giga_tot_", params.batch_idx_train
+                )
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(
+    cuts: CutSet, sp: spm.SentencePieceProcessor
+) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            return False
+
+        return True
+
+    cuts = cuts.filter(remove_short_and_long_utt)
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    train_cuts = filter_short_and_long_utterances(train_cuts, sp)
+
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+    # XL 10k hours
+    # L  2.5k hours
+    # M  1k hours
+    # S  250 hours
+    # XS 10 hours
+    # DEV 12 hours
+    # Test 40 hours
+    if params.full_libri:
+        logging.info("Using the XL subset of GigaSpeech (10k hours)")
+        train_giga_cuts = gigaspeech.train_XL_cuts()
+    else:
+        logging.info("Using the S subset of GigaSpeech (250 hours)")
+        train_giga_cuts = gigaspeech.train_S_cuts()
+
+    train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts, sp)
+    train_giga_cuts = train_giga_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    giga_train_dl = asr_datamodule.train_dataloaders(
+        train_giga_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         sp=sp,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            giga_train_dl=giga_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert 0 <= args.giga_prob < 1, args.giga_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/zipformer.py
new file mode 120000
index 000000000..be9e75bfa
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/zipformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless8/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless8/asr_datamodule.py
new file mode 120000
index 000000000..3ba9ada4f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless8/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless8/decode.py
new file mode 100755
index 000000000..e07777c9f
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/decode.py
@@ -0,0 +1,797 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless8/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless8/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from librispeech import LibriSpeech
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless8/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless8/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless8/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py
new file mode 100755
index 000000000..3fef231a1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py
@@ -0,0 +1,604 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless8/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless7",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # <blk> is defined in local/train_bpe_model.py
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/export.py b/egs/librispeech/ASR/pruned_transducer_stateless8/export.py
new file mode 100755
index 000000000..d9697680b
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/export.py
@@ -0,0 +1,333 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless8/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless8/decode.py \
+        --exp-dir ./pruned_transducer_stateless8/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless8/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device),
+                strict=False,
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless8/gigaspeech.py
new file mode 120000
index 000000000..5242c652a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/gigaspeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/gigaspeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/jit_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless8/jit_pretrained.py
new file mode 100755
index 000000000..129497d5a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/jit_pretrained.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless8/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless8/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/librispeech.py b/egs/librispeech/ASR/pruned_transducer_stateless8/librispeech.py
new file mode 120000
index 000000000..b76723bf5
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/librispeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/librispeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/model.py b/egs/librispeech/ASR/pruned_transducer_stateless8/model.py
new file mode 100644
index 000000000..39a360796
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/model.py
@@ -0,0 +1,220 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import penalize_abs_values_gt
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        decoder_giga: Optional[nn.Module] = None,
+        joiner_giga: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_giga = decoder_giga
+        self.joiner_giga = joiner_giga
+
+        self.simple_am_proj = nn.Linear(
+            encoder_dim,
+            vocab_size,
+        )
+        self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
+
+        if decoder_giga is not None:
+            self.simple_am_proj_giga = nn.Linear(encoder_dim, vocab_size)
+            self.simple_lm_proj_giga = nn.Linear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        libri: bool = True,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          libri:
+            True to use the decoder and joiner for the LibriSpeech dataset.
+            False to use the decoder and joiner for the GigaSpeech dataset.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        if libri:
+            decoder = self.decoder
+            simple_lm_proj = self.simple_lm_proj
+            simple_am_proj = self.simple_am_proj
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_giga
+            simple_lm_proj = self.simple_lm_proj_giga
+            simple_am_proj = self.simple_am_proj_giga
+            joiner = self.joiner_giga
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = simple_lm_proj(decoder_out)
+        am = simple_am_proj(encoder_out)
+
+        # if self.training and random.random() < 0.25:
+        #    lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04)
+        # if self.training and random.random() < 0.25:
+        #    am = penalize_abs_values_gt(am, 30.0, 1.0e-04)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=joiner.encoder_proj(encoder_out),
+            lm=joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless8/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless8/pretrained.py
new file mode 100755
index 000000000..64b38c9d5
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/pretrained.py
@@ -0,0 +1,360 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless8/export.py \
+  --exp-dir ./pruned_transducer_stateless8/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless8/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless8/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless8/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless8/exp/pretrained.pt is generated by
+./pruned_transducer_stateless8/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless8/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/scaling_converter.py b/egs/librispeech/ASR/pruned_transducer_stateless8/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/train.py b/egs/librispeech/ASR/pruned_transducer_stateless8/train.py
new file mode 100755
index 000000000..3f271c5b4
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/train.py
@@ -0,0 +1,1371 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+cd egs/librispeech/ASR/
+./prepare.sh
+./prepare_giga_speech.sh
+
+./pruned_transducer_stateless8/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless8/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless8/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless8/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decoder import Decoder
+from gigaspeech import GigaSpeech
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from librispeech import LibriSpeech
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--full-libri",
+        type=str2bool,
+        default=True,
+        help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless8/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--giga-prob",
+        type=float,
+        default=0.5,
+        help="The probability to select a batch from the GigaSpeech dataset",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(
+    params: AttributeDict,
+    enable_giga: bool = True,
+) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    if enable_giga:
+        logging.info("Use giga")
+        decoder_giga = get_decoder_model(params)
+        joiner_giga = get_joiner_model(params)
+    else:
+        logging.info("Disable giga")
+        decoder_giga = None
+        joiner_giga = None
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        decoder_giga=decoder_giga,
+        joiner_giga=joiner_giga,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_libri(c: Cut) -> bool:
+    """Return True if this cut is from the LibriSpeech dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of GigaSpeech to dict(origin='giga')
+      See ../local/preprocess_gigaspeech.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    libri = is_libri(supervisions["cut"][0])
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            libri=libri,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    giga_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      giga_train_dl:
+        Dataloader for the GigaSpeech training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      rng:
+        For selecting which dataset to use.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    libri_tot_loss = MetricsTracker()
+    giga_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.giga_prob, params.giga_prob]
+
+    iter_libri = iter(train_dl)
+    iter_giga = iter(giga_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_libri if idx == 0 else iter_giga
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            name = "libri" if idx == 0 else "giga"
+            logging.info(f"{name} reaches end of dataloader")
+            break
+
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        libri = is_libri(batch["supervisions"]["cut"][0])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            if libri:
+                libri_tot_loss = (
+                    libri_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "libri"  # for logging only
+            else:
+                giga_tot_loss = (
+                    giga_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "giga"
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], "
+                f"libri_tot_loss[{libri_tot_loss}], "
+                f"giga_tot_loss[{giga_tot_loss}], "
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                libri_tot_loss.write_summary(
+                    tb_writer, "train/libri_tot_", params.batch_idx_train
+                )
+                giga_tot_loss.write_summary(
+                    tb_writer, "train/giga_tot_", params.batch_idx_train
+                )
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(
+    cuts: CutSet, sp: spm.SentencePieceProcessor
+) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    cuts = cuts.filter(remove_short_and_long_utt)
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=True)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    train_cuts = filter_short_and_long_utterances(train_cuts, sp)
+
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+    # XL 10k hours
+    # L  2.5k hours
+    # M  1k hours
+    # S  250 hours
+    # XS 10 hours
+    # DEV 12 hours
+    # Test 40 hours
+    if params.full_libri:
+        logging.info("Using the XL subset of GigaSpeech (10k hours)")
+        train_giga_cuts = gigaspeech.train_XL_cuts()
+    else:
+        logging.info("Using the S subset of GigaSpeech (250 hours)")
+        train_giga_cuts = gigaspeech.train_S_cuts()
+
+    train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts, sp)
+    train_giga_cuts = train_giga_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    giga_train_dl = asr_datamodule.train_dataloaders(
+        train_giga_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    if False and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            giga_train_dl=giga_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert 0 <= args.giga_prob < 1, args.giga_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless8/zipformer.py
new file mode 120000
index 000000000..79b076556
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/zipformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/shared b/egs/librispeech/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/librispeech/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/README.md b/egs/librispeech/ASR/streaming_conformer_ctc/README.md
new file mode 100644
index 000000000..53f383c99
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/README.md
@@ -0,0 +1,90 @@
+## Train and Decode
+Commands of data preparation/train/decode steps are almost the same with
+../conformer_ctc experiment except some options.
+
+Please read the code and understand following new added options before running this experiment:
+
+  For data preparation:
+
+    Nothing new.
+
+  For streaming_conformer_ctc/train.py:
+
+    --dynamic-chunk-training
+    --short-chunk-proportion
+
+  For streaming_conformer_ctc/streaming_decode.py:
+
+    --chunk-size
+    --tailing-num-frames
+    --simulate-streaming
+
+## Performence and a trained model.
+
+The latest results with this streaming code is shown in following table:
+
+chunk size | wer on test-clean | wer on test-other
+-- | -- | --
+full | 3.53 | 8.52
+40(1.96s) | 3.78 | 9.38
+32(1.28s) | 3.82 | 9.44
+24(0.96s) | 3.95 | 9.76
+16(0.64s) | 4.06 | 9.98
+8(0.32s) | 4.30 | 10.55
+4(0.16s) | 5.88 | 12.01
+
+
+A trained model is also provided.
+By run
+```
+git clone https://huggingface.co/GuoLiyong/streaming_conformer
+
+# You may want to manually check md5sum values of downloaded files
+# 8e633bc1de37f5ae57a2694ceee32a93  trained_streaming_conformer.pt
+# 4c0aeefe26c784ec64873cc9b95420f1  L.pt
+# d1f91d81005fb8ce4d65953a4a984ee7  Linv.pt
+# e1c1902feb7b9fc69cd8d26e663c2608  bpe.model
+# 8617e67159b0ff9118baa54f04db24cc  tokens.txt
+# 72b075ab5e851005cd854e666c82c3bb  words.txt
+```
+
+If there is any different md5sum values, please run
+```
+cd streaming_models
+git lfs pull
+```
+And check md5sum values again.
+
+Finally, following files will be downloaded:
+<pre>
+streaming_models/
+|-- lang_bpe
+|   |-- L.pt
+|   |-- Linv.pt
+|   |-- bpe.model
+|   |-- tokens.txt
+|   `-- words.txt
+`-- trained_streaming_conformer.pt
+</pre>
+
+
+And run commands you will get the same results of previous table:
+```
+trained_models=/path/to/downloaded/streaming_models/
+for chunk_size in 4 8 16 24 36 40 -1; do
+    ./streaming_conformer_ctc/streaming_decode.py \
+      --chunk-size=${chunk_size} \
+      --trained-dir=${trained_models}
+done
+```
+Results of following command is indentical to previous one,
+but model consumes features chunk_by_chunk, i.e. a streaming way.
+```
+trained_models=/path/to/downloaded/streaming_models/
+for chunk_size in 4 8 16 24 36 40 -1; do
+    ./streaming_conformer_ctc/streaming_decode.py \
+      --simulate-streaming=True \
+      --chunk-size=${chunk_size} \
+      --trained-dir=${trained_models}
+done
+```
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/asr_datamodule.py b/egs/librispeech/ASR/streaming_conformer_ctc/asr_datamodule.py
new file mode 120000
index 000000000..a73848de9
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/asr_datamodule.py
@@ -0,0 +1 @@
+../conformer_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py b/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py
new file mode 100644
index 000000000..0b982f4bf
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py
@@ -0,0 +1,1306 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+# from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/mask.py#L42
+def subsequent_chunk_mask(
+    size: int,
+    chunk_size: int,
+    num_left_chunks: int = -1,
+    device: torch.device = torch.device("cpu"),
+) -> torch.Tensor:
+    """Create mask for subsequent steps (size, size) with chunk size,
+       this is for streaming encoder
+    Args:
+        size (int): size of mask
+        chunk_size (int): size of chunk
+        num_left_chunks (int): number of left chunks
+            <0: use full chunk
+            >=0: use num_left_chunks
+        device (torch.device): "cpu" or "cuda" or torch.Tensor.device
+    Returns:
+        torch.Tensor: mask
+    Examples:
+        >>> subsequent_chunk_mask(4, 2)
+        [[1, 1, 0, 0],
+         [1, 1, 0, 0],
+         [1, 1, 1, 1],
+         [1, 1, 1, 1]]
+    """
+    ret = torch.zeros(size, size, device=device, dtype=torch.bool)
+    for i in range(size):
+        if num_left_chunks < 0:
+            start = 0
+        else:
+            start = max((i // chunk_size - num_left_chunks) * chunk_size, 0)
+        ending = min((i // chunk_size + 1) * chunk_size, size)
+        ret[i, start:ending] = True
+    return ret
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+        causal: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+            causal,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self,
+        x: Tensor,
+        supervisions: Optional[Supervisions] = None,
+        dynamic_chunk_training: bool = False,
+        short_chunk_proportion: float = 0.5,
+        chunk_size: int = -1,
+        simulate_streaming: bool = False,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+          dynamic_chunk_training:
+            For training only.
+            IF True, train with dynamic right context for some batches
+            sampled with a distribution
+            if False, train with full right context all the time.
+          short_chunk_proportion:
+            For training only.
+            Proportion of samples that will be trained with dynamic chunk.
+          chunk_size:
+            For eval only.
+            right context when evaluating test utts.
+            -1 means all right context.
+          simulate_streaming=False,
+            For eval only.
+            If true, the feature will be feeded into the model chunk by chunk.
+            If false, the whole utts if feeded into the model together i.e. the
+            model only foward once.
+
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        if self.encoder.training:
+            return self.train_run_encoder(
+                x, supervisions, dynamic_chunk_training, short_chunk_proportion
+            )
+        else:
+            return self.eval_run_encoder(
+                x, supervisions, chunk_size, simulate_streaming
+            )
+
+    def train_run_encoder(
+        self,
+        x: Tensor,
+        supervisions: Optional[Supervisions] = None,
+        dynamic_chunk_training: bool = False,
+        short_chunk_threshold: float = 0.5,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+          dynamic_chunk_training:
+            IF True, train with dynamic right context for some batches
+            sampled with a distribution
+            if False, train with full right context all the time.
+          short_chunk_proportion:
+            Proportion of samples that will be trained with dynamic chunk.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        src_key_padding_mask = encoder_padding_mask(x.size(0), supervisions)
+        if src_key_padding_mask is not None:
+            src_key_padding_mask = src_key_padding_mask.to(x.device)
+
+        if dynamic_chunk_training:
+            max_len = x.size(0)
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * short_chunk_threshold):
+                chunk_size = max_len
+            else:
+                chunk_size = chunk_size % 25 + 1
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0), chunk_size=chunk_size, device=x.device
+            )
+            x = self.encoder(
+                x, pos_emb, mask=mask, src_key_padding_mask=src_key_padding_mask
+            )  # (T, B, F)
+        else:
+            x = self.encoder(
+                x, pos_emb, src_key_padding_mask=src_key_padding_mask
+            )  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, src_key_padding_mask
+
+    def eval_run_encoder(
+        self,
+        feature: Tensor,
+        supervisions: Optional[Supervisions] = None,
+        chunk_size: int = -1,
+        simulate_streaming=False,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          feature:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        # feature.shape: N T C
+        num_frames = feature.size(1)
+
+        # As temporarily in icefall only subsampling_rate == 4 is supported,
+        # following parameters are hard-coded here.
+        # Change it accordingly if other subsamling_rate are supported.
+        embed_left_context = 7
+        embed_conv_right_context = 3
+        subsampling_rate = 4
+        stride = chunk_size * subsampling_rate
+        decoding_window = embed_conv_right_context + stride
+
+        # This is also only compatible to sumsampling_rate == 4
+        length_after_subsampling = ((feature.size(1) - 1) // 2 - 1) // 2
+        src_key_padding_mask = encoder_padding_mask(
+            length_after_subsampling, supervisions
+        )
+        if src_key_padding_mask is not None:
+            src_key_padding_mask = src_key_padding_mask.to(feature.device)
+
+        if chunk_size < 0:
+            # non-streaming decoding
+            x = self.encoder_embed(feature)
+            x, pos_emb = self.encoder_pos(x)
+            x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+            x = self.encoder(
+                x, pos_emb, src_key_padding_mask=src_key_padding_mask
+            )  # (T, B, F)
+        else:
+            if simulate_streaming:
+                # simulate chunk_by_chunk streaming decoding
+                # Results of this branch should be identical to following
+                # "else" branch.
+                # But this branch is a little slower
+                # as the feature is feeded chunk by chunk
+
+                # store the result  of chunk_by_chunk decoding
+                encoder_output = []
+
+                # caches
+                pos_emb_positive = []
+                pos_emb_negative = []
+                pos_emb_central = None
+                encoder_cache = [None for i in range(len(self.encoder.layers))]
+                conv_cache = [None for i in range(len(self.encoder.layers))]
+
+                # start chunk_by_chunk decoding
+                offset = 0
+                for cur in range(0, num_frames - embed_left_context + 1, stride):
+                    end = min(cur + decoding_window, num_frames)
+                    cur_feature = feature[:, cur:end, :]
+                    cur_feature = self.encoder_embed(cur_feature)
+                    cur_embed, cur_pos_emb = self.encoder_pos(cur_feature, offset)
+                    cur_embed = cur_embed.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+
+                    cur_T = cur_feature.size(1)
+                    if cur == 0:
+                        # for first chunk extract the central pos embedding
+                        pos_emb_central = cur_pos_emb[0, (chunk_size - 1), :].view(
+                            1, 1, -1
+                        )
+                        cur_T -= 1
+                    pos_emb_positive.append(cur_pos_emb[0, :cur_T].flip(0))
+                    pos_emb_negative.append(cur_pos_emb[0, -cur_T:])
+                    assert pos_emb_positive[-1].size(0) == cur_T
+
+                    pos_emb_pos = torch.cat(pos_emb_positive, dim=0).unsqueeze(0)
+                    pos_emb_neg = torch.cat(pos_emb_negative, dim=0).unsqueeze(0)
+                    cur_pos_emb = torch.cat(
+                        [pos_emb_pos.flip(1), pos_emb_central, pos_emb_neg],
+                        dim=1,
+                    )
+
+                    x = self.encoder.chunk_forward(
+                        cur_embed,
+                        cur_pos_emb,
+                        src_key_padding_mask=src_key_padding_mask[
+                            :, : offset + cur_embed.size(0)
+                        ],
+                        encoder_cache=encoder_cache,
+                        conv_cache=conv_cache,
+                        offset=offset,
+                    )  # (T, B, F)
+                    encoder_output.append(x)
+                    offset += cur_embed.size(0)
+
+                x = torch.cat(encoder_output, dim=0)
+            else:
+                # NOT simulate chunk_by_chunk decoding
+                # Results of this branch should be identical to previous
+                # simulate chunk_by_chunk decoding branch.
+                # But this branch is faster.
+                x = self.encoder_embed(feature)
+                x, pos_emb = self.encoder_pos(x)
+                x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+                mask = ~subsequent_chunk_mask(
+                    size=x.size(0), chunk_size=chunk_size, device=x.device
+                )
+                x = self.encoder(
+                    x,
+                    pos_emb,
+                    mask=mask,
+                    src_key_padding_mask=src_key_padding_mask,
+                )  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, src_key_padding_mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        causal: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel, causal=causal)
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(self.conv_module(src))
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        encoder_cache: Optional[Tensor] = None,
+        conv_cache: Optional[Tensor] = None,
+        offset=0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        if encoder_cache is None:
+            # src: [chunk_size, N, F] e.g. [8, 41, 512]
+            key = src
+            val = key
+            encoder_cache = key
+        else:
+            key = torch.cat([encoder_cache, src], dim=0)
+            val = key
+            encoder_cache = key
+        src_att = self.self_attn(
+            src,
+            key,
+            val,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+            offset=offset,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src  # [chunk_size, N, F] e.g. [8, 41, 512]
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        if conv_cache is not None:
+            src = torch.cat([conv_cache, src], dim=0)
+        conv_cache = src
+
+        src = self.conv_module(src)
+        src = src[-residual.size(0) :, :, :]  # noqa: E203
+
+        src = residual + self.dropout(src)
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src, encoder_cache, conv_cache
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        encoder_cache=None,
+        conv_cache=None,
+        offset=0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            output, e_cache, c_cache = mod.chunk_forward(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                encoder_cache=encoder_cache[layer_index],
+                conv_cache=conv_cache[layer_index],
+                offset=offset,
+            )
+            encoder_cache[layer_index] = e_cache
+            conv_cache[layer_index] = c_cache
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor, offset: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_1 = offset + x.size(1)
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_1 * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_1, self.d_model)
+        pe_negative = torch.zeros(x_size_1, self.d_model)
+        position = torch.arange(0, x_size_1, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor, offset: int = 0) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x, offset)
+        x = x * self.xscale
+        x_size_1 = offset + x.size(1)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_1
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x_size_1,
+        ]
+        x_T = x.size(1)
+        if offset > 0:
+            pos_emb = torch.cat([pos_emb[:, :x_T], pos_emb[:, -x_T:]], dim=1)
+        else:
+            pos_emb = torch.cat(
+                [
+                    pos_emb[:, : (x_T - 1)],
+                    self.pe[0, self.pe.size(1) // 2].view(1, 1, self.pe.size(-1)),
+                    pos_emb[:, -(x_T - 1) :],  # noqa: E203
+                ],
+                dim=1,
+            )
+
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        offset=0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            offset=offset,
+        )
+
+    def rel_shift(self, x: Tensor, offset=0) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 == time1 + offset, since it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        time2 = time1 + offset
+        assert n == 2 * time2 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+
+        return x.as_strided(
+            (batch_size, num_heads, time1, time2),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        offset=0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+
+        # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1)
+        p = p.permute(0, 2, 3, 1)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(q_with_bias_v, p)  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd, offset=offset)  # [B, head, time1, time2]
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        causal: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        # from https://github.com/wenet-e2e/wenet/blob/main/wenet/transformer/convolution.py#L41
+        if causal:
+            self.lorder = kernel_size - 1
+            padding = 0  # manualy padding self.lorder zeros to the left later
+        else:
+            assert (kernel_size - 1) % 2 == 0
+            self.lorder = 0
+            padding = (kernel_size - 1) // 2
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if self.lorder > 0:
+            # manualy padding self.lorder zeros to the left
+            # make depthwise_conv causal
+            x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+        x = self.depthwise_conv(x)
+        x = self.activation(self.norm(x))
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/label_smoothing.py b/egs/librispeech/ASR/streaming_conformer_ctc/label_smoothing.py
new file mode 120000
index 000000000..08734abd7
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/label_smoothing.py
@@ -0,0 +1 @@
+../conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/streaming_decode.py b/egs/librispeech/ASR/streaming_conformer_ctc/streaming_decode.py
new file mode 100755
index 000000000..a26d0b789
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/streaming_decode.py
@@ -0,0 +1,507 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+# from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/common.py#L166
+def remove_duplicates_and_blank(hyp: List[int]) -> List[int]:
+    new_hyp: List[int] = []
+    cur = 0
+    while cur < len(hyp):
+        if hyp[cur] != 0:
+            new_hyp.append(hyp[cur])
+        prev = cur
+        while cur < len(hyp) and hyp[cur] == hyp[prev]:
+            cur += 1
+    return new_hyp
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=int,
+        default=8,
+        help="Frames of right context"
+        "-1 for whole right context, i.e. non-streaming decoding",
+    )
+
+    parser.add_argument(
+        "--tailing-num-frames",
+        type=int,
+        default=20,
+        help="tailing dummy frames padded to the right, only used during decoding",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="simulate chunk by chunk decoding",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="ctc-greedy-search",
+        help="Streaming Decoding method",
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        conformer_ctc/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="streaming_conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--trained-dir",
+        type=Path,
+        default=None,
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--avg-models",
+        type=str,
+        default=None,
+        help="Manually select models to average, seperated by comma;"
+        "e.g. 60,62,63,72",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("conformer_ctc/exp"),
+            "lang_dir": Path("data/lang_bpe"),
+            # parameters for conformer
+            "causal": True,
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    chunk_size: int = -1,
+    simulate_streaming=False,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+      model:
+        The neural model.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    feature = batch["inputs"]
+    device = torch.device("cuda")
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    # Extra dummy tailing frames my reduce deletion error
+    # example WITHOUT padding:
+    # CHAPTER SEVEN ON THE RACES OF MAN
+    # example WITH padding:
+    # CHAPTER SEVEN ON THE RACES OF (MAN->*)
+    tailing_frames = (
+        torch.tensor([-23.0259])
+        .expand([feature.size(0), params.tailing_num_frames, 80])
+        .to(feature.device)
+    )
+    feature = torch.cat([feature, tailing_frames], dim=1)
+    supervisions["num_frames"] += params.tailing_num_frames
+
+    nnet_output, memory, memory_key_padding_mask = model(
+        feature,
+        supervisions,
+        chunk_size=chunk_size,
+        simulate_streaming=simulate_streaming,
+    )
+
+    assert params.method == "ctc-greedy-search"
+    key = "ctc-greedy-search"
+    batch_size = nnet_output.size(0)
+    maxlen = nnet_output.size(1)
+    topk_prob, topk_index = nnet_output.topk(1, dim=2)  # (B, maxlen, 1)
+    topk_index = topk_index.view(batch_size, maxlen)  # (B, maxlen)
+    topk_index = topk_index.masked_fill_(memory_key_padding_mask, 0)  # (B, maxlen)
+    token_ids = [token_id.tolist() for token_id in topk_index]
+    token_ids = [remove_duplicates_and_blank(token_id) for token_id in token_ids]
+    hyps = bpe_model.decode(token_ids)
+    hyps = [s.split() for s in hyps]
+    return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    chunk_size: int = -1,
+    simulate_streaming=False,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      chunk_size:
+        right context to simulate streaming decoding
+        -1 for whole right context, i.e. non-stream decoding
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            chunk_size=chunk_size,
+            simulate_streaming=simulate_streaming,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    if params.avg_models is not None:
+        avg_models = params.avg_models.replace(",", "_")
+        result_file_prefix = f"epoch-avg-{avg_models}-chunksize \
+        -{params.chunk_size}-tailing-num-frames-{params.tailing_num_frames}-"
+    else:
+        result_file_prefix = f"epoch-{params.epoch}-avg-{params.avg}-chunksize \
+        -{params.chunk_size}-tailing-num-frames-{params.tailing_num_frames}-"
+    for key, results in results_dict.items():
+        recog_path = (
+            params.exp_dir / f"{result_file_prefix}recogs-{test_set_name}-{key}.txt"
+        )
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.exp_dir / f"{result_file_prefix}-errs-{test_set_name}-{key}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    if params.trained_dir is not None:
+        params.lang_dir = Path(params.trained_dir) / "lang_bpe"
+        # used naming result files
+        params.epoch = "trained_model"
+        params.avg = 1
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+        causal=params.causal,
+    )
+
+    if params.trained_dir is not None:
+        model_name = f"{params.trained_dir}/trained_streaming_conformer.pt"
+        load_checkpoint(model_name, model)
+    elif params.avg == 1 and params.avg_models is not None:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        filenames = []
+        if params.avg_models is not None:
+            model_ids = params.avg_models.split(",")
+            for i in model_ids:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        else:
+            start = params.epoch - params.avg + 1
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+    # CAUTION: `test_sets` is for displaying only.
+    # If you want to skip test-clean, you have to skip
+    # it inside the for loop. That is, use
+    #
+    #   if test_set == 'test-clean': continue
+    #
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+    test_sets = ["test-clean", "test-other"]
+    for test_set, test_dl in zip(test_sets, librispeech.test_dataloaders()):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            chunk_size=params.chunk_size,
+            simulate_streaming=params.simulate_streaming,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/subsampling.py b/egs/librispeech/ASR/streaming_conformer_ctc/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/train.py b/egs/librispeech/ASR/streaming_conformer_ctc/train.py
new file mode 100755
index 000000000..14d7274c2
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/train.py
@@ -0,0 +1,771 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=78,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=True,
+        help="Whether to use dynamic right context during training.",
+    )
+
+    parser.add_argument(
+        "--short-chunk-proportion",
+        type=float,
+        default=0.7,
+        help="Proportion of samples trained with short right context",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Whether to do batch normalization for the
+                              input features.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - lr_factor: The lr_factor for Noam optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            "num_decoder_layers": 6,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "lr_factor": 5.0,
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(
+            feature,
+            supervisions,
+            dynamic_chunk_training=params.dynamic_chunk_training,
+            short_chunk_proportion=params.short_chunk_proportion,
+        )
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    token_ids = graph_compiler.texts_to_ids(texts)
+
+    decoding_graph = graph_compiler.compile(token_ids)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/transformer.py b/egs/librispeech/ASR/streaming_conformer_ctc/transformer.py
new file mode 100644
index 000000000..0c87fdf1b
--- /dev/null
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/transformer.py
@@ -0,0 +1,946 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from label_smoothing import LabelSmoothingLoss
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        if use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        supervision: Optional[Supervisions] = None,
+        dynamic_chunk_training: bool = False,
+        short_chunk_proportion: float = 0.5,
+        chunk_size: int = -1,
+        simulate_streaming=False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        encoder_memory, memory_key_padding_mask = self.run_encoder(
+            x,
+            supervision,
+            dynamic_chunk_training=dynamic_chunk_training,
+            short_chunk_proportion=short_chunk_proportion,
+            chunk_size=chunk_size,
+            simulate_streaming=simulate_streaming,
+        )
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        chunk_size: int = -1,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+          chunk_size: right chunk_size to simulate streaming decoding
+            -1 for whole right context
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        # streaming decoding(chunk_size >= 0) is only verified with Conformer
+        assert chunk_size == -1
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(
+            x, src_key_padding_mask=mask, chunk_size=chunk_size
+        )  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/README.md b/egs/librispeech/ASR/tdnn_lstm_ctc/README.md
new file mode 100644
index 000000000..b1e01a218
--- /dev/null
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/README.md
@@ -0,0 +1,4 @@
+
+Please visit
+<https://k2-fsa.github.io/icefall/recipes/Non-streaming-ASR/librispeech/tdnn_lstm_ctc.html>
+for how to run this recipe.
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/__init__.py b/egs/librispeech/ASR/tdnn_lstm_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py
new file mode 100644
index 000000000..c500eb3e5
--- /dev/null
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py
@@ -0,0 +1,475 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--full-libri",
+            type=str2bool,
+            default=True,
+            help="""Used only when --mini-libri is False.When enabled,
+            use 960h LibriSpeech. Otherwise, use 100h subset.""",
+        )
+        group.add_argument(
+            "--mini-libri",
+            type=str2bool,
+            default=False,
+            help="True for mini librispeech",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_clean_5_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get train-clean-5 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-5.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-100.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_360_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-360 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-360.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_other_500_cuts(self) -> CutSet:
+        logging.info("About to get train-other-500 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-other-500.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_all_shuf_cuts(self) -> CutSet:
+        logging.info(
+            "About to get the shuffled train-clean-100, \
+            train-clean-360 and train-other-500 cuts"
+        )
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_2_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get dev-clean-2 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean-2.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/decode.py b/egs/librispeech/ASR/tdnn_lstm_ctc/decode.py
new file mode 100755
index 000000000..92529e06c
--- /dev/null
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/decode.py
@@ -0,0 +1,501 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from model import TdnnLstm
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=19,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="whole-lattice-rescoring",
+        help="""Decoding method.
+        Supported values are:
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp/"),
+            "lang_dir": Path("data/lang_phone"),
+            "lm_dir": Path("data/lm"),
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    feature = feature.permute(0, 2, 1)  # now feature is (N, C, T)
+
+    nnet_output = model(feature)
+    # nnet_output is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-{params.num_paths}"
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in ["nbest-rescoring", "whole-lattice-rescoring"]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+
+    ans = dict()
+    for lm_scale_str, best_path in best_path_dict.items():
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            lexicon=lexicon,
+            G=G,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu"))
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in ["nbest-rescoring", "whole-lattice-rescoring"]:
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location="cpu")
+            G = k2.Fsa.from_dict(d).to(device)
+
+        if params.method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            lexicon=lexicon,
+            G=G,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/model.py b/egs/librispeech/ASR/tdnn_lstm_ctc/model.py
new file mode 100644
index 000000000..1731e1ebe
--- /dev/null
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/model.py
@@ -0,0 +1,100 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class TdnnLstm(nn.Module):
+    def __init__(
+        self, num_features: int, num_classes: int, subsampling_factor: int = 3
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            It reduces the number of output frames by this factor.
+        """
+        super().__init__()
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=500,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+            nn.Conv1d(
+                in_channels=500,
+                out_channels=500,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+            nn.Conv1d(
+                in_channels=500,
+                out_channels=500,
+                kernel_size=3,
+                stride=self.subsampling_factor,  # stride: subsampling_factor!
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+        )
+        self.lstms = nn.ModuleList(
+            [nn.LSTM(input_size=500, hidden_size=500, num_layers=1) for _ in range(5)]
+        )
+        self.lstm_bnorms = nn.ModuleList(
+            [nn.BatchNorm1d(num_features=500, affine=False) for _ in range(5)]
+        )
+        self.dropout = nn.Dropout(0.2)
+        self.linear = nn.Linear(in_features=500, out_features=self.num_classes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Its shape is [N, C, T]
+
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x = self.tdnn(x)
+        x = x.permute(2, 0, 1)  # (N, C, T) -> (T, N, C) -> how LSTM expects it
+        for lstm, bnorm in zip(self.lstms, self.lstm_bnorms):
+            x_new, _ = lstm(x)
+            x_new = bnorm(x_new.permute(1, 2, 0)).permute(
+                2, 0, 1
+            )  # (T, N, C) -> (N, C, T) -> (T, N, C)
+            x_new = self.dropout(x_new)
+            x = x_new + x  # skip connections
+        x = x.transpose(
+            1, 0
+        )  # (T, N, C) -> (N, T, C) -> linear expects "features" in the last dim
+        x = self.linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/pretrained.py b/egs/librispeech/ASR/tdnn_lstm_ctc/pretrained.py
new file mode 100755
index 000000000..fde724866
--- /dev/null
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/pretrained.py
@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from model import TdnnLstm
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding, rescore_with_whole_lattice
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.8,
+        help="""
+        Used only when method is whole-lattice-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "num_classes": 72,
+            "sample_rate": 16000,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+    if not hasattr(HLG, "lm_scores"):
+        # For whole-lattice-rescoring and attention-decoder
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method == "whole-lattice-rescoring":
+        logging.info(f"Loading G from {params.G}")
+        G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+        # Add epsilon self-loops to G as we will compose
+        # it with the whole lattice later
+        G = G.to(device)
+        G = k2.add_epsilon_self_loops(G)
+        G = k2.arc_sort(G)
+        G.lm_scores = G.scores.clone()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    features = features.permute(0, 2, 1)  # now features is (N, C, T)
+
+    with torch.no_grad():
+        nnet_output = model(features)
+        # nnet_output is (N, T, C)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "1best":
+        logging.info("Use HLG decoding")
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+    elif params.method == "whole-lattice-rescoring":
+        logging.info("Use HLG decoding + LM rescoring")
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=[params.ngram_lm_scale],
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/train.py b/egs/librispeech/ASR/tdnn_lstm_ctc/train.py
new file mode 100755
index 000000000..90245ed46
--- /dev/null
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/train.py
@@ -0,0 +1,642 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./tdnn_lstm_ctc/train.py \
+     --world-size 4 \
+     --full-libri 1 \
+     --max-duration 300 \
+     --num-epochs 20
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import LibriSpeechAsrDataModule
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import TdnnLstm
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn_lstm_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - exp_dir: It specifies the directory where all training related
+                   files, e.g., checkpoints, log, etc, are saved
+
+        - lang_dir: It contains language related input files such as
+                    "lexicon.txt"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp"),
+            "lang_dir": Path("data/lang_phone"),
+            "lr": 1e-4,
+            "feature_dim": 80,
+            "weight_decay": 5e-4,
+            "subsampling_factor": 3,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of TdnnLstm in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    feature = feature.permute(0, 2, 1)  # now feature is (N, C, T)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(feature)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervisions = batch["supervisions"]
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+    decoding_graph = graph_compiler.compile(texts)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(2) - supervisions["num_frames"]) / feature.size(2)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.AdamW(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    scheduler = StepLR(optimizer, step_size=8, gamma=0.1)
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if epoch > params.start_epoch:
+            logging.info(f"epoch {epoch}, lr: {scheduler.get_last_lr()[0]}")
+
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/lr",
+                scheduler.get_last_lr()[0],
+                params.batch_idx_train,
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        scheduler.step()
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/README.md b/egs/librispeech/ASR/tiny_transducer_ctc/README.md
new file mode 100644
index 000000000..78dbc12c9
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/README.md
@@ -0,0 +1,184 @@
+## Introduction
+
+This recipe is intended for streaming ASR on very low cost devices, with model parameters in the range of 1-2M. It uses a small convolutional net as the encoder. It is trained with combined transducer and CTC losses, and supports both phone and BPE lexicons. For phone lexicon, you can do transducer decoding using a method with LG, but the results were bad. 
+
+The encoder consists of 2 subsampling layers followed by a stack of Conv1d-batchnorm-activation-causal_squeeze_excite blocks, with optional skip connections. To reduce latency (at the cost of slightly higher WER), half of the blocks use causal convolution.
+
+A few remarks & observations:
+
+1. Phone lexicon works better than BPE for CTC decoding (with HLG) but worse for transducer decoding. 
+
+2. SpecAugment is not helpful for very small models as they tend to underfit rather than overfit. For the large model, a less aggressive SpecAugment (see asr_datamodule.py) improved the result a little.
+
+3. Squeeze-and-excitation worked like a charm! It reduces WER quite a bit with marginal increase of parameters and MAC ops. To make it causal I changed the global average pooling layer to a moving average filter, so only historical context is used.
+
+## Pretrained models
+
+You can find pretrained models, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/wangtiance/tiny_transducer_ctc/tree/main>
+
+## Results on full libri
+
+I tried 3 different sizes of the encoder. The parameters are around 1M, 2M and 4M, respectively. For CTC decoding, whole-lattice-rescoring frequently causes OOM error so the result is not shown.
+
+### Small encoder
+
+The small encoder uses 10 layers of 1D convolution block with 256 channels, without skip connections. The encoder, decoder and joiner dim is 256. Algorithmic latency is 280ms. Multiply-add ops for the encoder is 22.0Mops. It is more applicable for ASR products with limited vocabulary (like a fixed set of phrases or short sentences). 
+
+#### CTC decoding with phone lexicon
+Total parameters: 1073392
+
+Parameters for CTC decoding: 865816
+
+|                 | test-clean | test-other | comment              |
+|-----------------|------------|------------|----------------------|
+| 1best           | 9.68       | 24.9       | --epoch 30 --avg 2   |
+| nbest-rescoring | 8.2        | 22.7       | --epoch 30 --avg 2   |
+
+The training commands are:
+```bash
+
+./tiny_transducer_ctc/train.py \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --max-duration 600 \
+  --exp-dir tiny_transducer_ctc/exp_small_phone \
+  --ctc-loss-scale 0.7 \
+  --enable-spec-aug 0 \
+  --lang-dir lang_phone \
+  --encoder-dim 256 \
+  --decoder-dim 256 \
+  --joiner-dim 256 \
+  --conv-layers 10 \
+  --channels 256 \
+  --skip-add 0 \
+```
+
+#### Transducer decoding with BPE 500 lexicon
+Total parameters: 1623264
+
+Parameters for transducer decoding: 1237764
+
+|                    | test-clean | test-other | comment              |
+|--------------------|------------|------------|----------------------|
+| greedy_search      | 14.47      | 32.03      | --epoch 30 --avg 1   |
+| fast_beam_search   | 13.38      | 29.61      | --epoch 30 --avg 1   |
+|modified_beam_search| 13.02      | 29.32      | --epoch 30 --avg 1   |
+
+The training commands are:
+```bash
+
+./tiny_transducer_ctc/train.py \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --max-duration 600 \
+  --exp-dir tiny_transducer_ctc/exp_small_bpe \
+  --ctc-loss-scale 0.2 \
+  --enable-spec-aug 0 \
+  --lang-dir lang_bpe_500 \
+  --encoder-dim 256 \
+  --decoder-dim 256 \
+  --joiner-dim 256 \
+  --conv-layers 10 \
+  --channels 256 \
+  --skip-add 0 \
+```
+
+### Middle encoder
+
+The middle encoder uses 18 layers of 1D convolution block with 300 channels, with skip connections. The encoder, decoder and joiner dim is 256. Algorithmic latency is 440ms. Multiply-add ops for the encoder is 50.1Mops. Note that the nbest-rescoring result is better than the tdnn_lstm_ctc recipe with whole-lattice-rescoring.
+
+#### CTC decoding with phone lexicon
+Total parameters: 2186242
+
+Parameters for CTC decoding: 1978666
+
+|                 | test-clean | test-other | comment              |
+|-----------------|------------|------------|----------------------|
+| 1best           | 7.48       | 18.94      | --epoch 30 --avg 1   |
+| nbest-rescoring | 6.31       | 16.89      | --epoch 30 --avg 1   |
+
+The training commands are:
+```bash
+
+./tiny_transducer_ctc/train.py \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --max-duration 600 \
+  --exp-dir tiny_transducer_ctc/exp_middle_phone \
+  --ctc-loss-scale 0.7 \
+  --enable-spec-aug 0 \
+  --lang-dir lang_phone \
+  --encoder-dim 256 \
+  --decoder-dim 256 \
+  --joiner-dim 256 \
+  --conv-layers 18 \
+  --channels 300 \
+  --skip-add 1 \
+```
+
+#### Transducer decoding with BPE 500 lexicon
+Total parameters: 2735794
+
+Parameters for transducer decoding: 2350294
+
+|                    | test-clean | test-other | comment              |
+|--------------------|------------|------------|----------------------|
+| greedy_search      | 10.26      | 25.13      | --epoch 30 --avg 2   |
+| fast_beam_search   | 9.69       | 23.58      | --epoch 30 --avg 2   |
+|modified_beam_search| 9.43       | 23.53      | --epoch 30 --avg 2   |
+
+The training commands are:
+```bash
+
+./tiny_transducer_ctc/train.py \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --max-duration 600 \
+  --exp-dir tiny_transducer_ctc/exp_middle_bpe \
+  --ctc-loss-scale 0.2 \
+  --enable-spec-aug 0 \
+  --lang-dir lang_bpe_500 \
+  --encoder-dim 256 \
+  --decoder-dim 256 \
+  --joiner-dim 256 \
+  --conv-layers 18 \
+  --channels 300 \
+  --skip-add 1 \
+```
+
+### Large encoder
+
+The large encoder uses 18 layers of 1D convolution block with 400 channels, with skip connections. The encoder, decoder and joiner dim is 400. Algorithmic latency is 440ms. Multiply-add ops for the encoder is 88.8Mops. It is interesting to see how much the gap is if we simply scale down more complicated models like Zipformer or emformer.
+
+
+#### Transducer decoding with BPE 500 lexicon
+Total parameters: 4821330
+
+Parameters for transducer decoding: 4219830
+
+|                    | test-clean | test-other | comment              |
+|--------------------|------------|------------|----------------------|
+| greedy_search      | 8.29       | 21.11      | --epoch 30 --avg 1   |
+| fast_beam_search   | 7.91       | 20.1       | --epoch 30 --avg 1   |
+|modified_beam_search| 7.74       | 19.89      | --epoch 30 --avg 1   |
+
+
+The training commands are:
+```bash
+
+./tiny_transducer_ctc/train.py \
+  --num-epochs 30 \
+  --full-libri 1 \
+  --max-duration 600 \
+  --exp-dir tiny_transducer_ctc/exp_large_bpe \
+  --ctc-loss-scale 0.2 \
+  --enable-spec-aug 1 \
+  --lang-dir lang_bpe_500 \
+  --encoder-dim 400 \
+  --decoder-dim 400 \
+  --joiner-dim 400 \
+  --conv-layers 18 \
+  --channels 400 \
+  --skip-add 1 \
+```
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/asr_datamodule.py b/egs/librispeech/ASR/tiny_transducer_ctc/asr_datamodule.py
new file mode 100644
index 000000000..3acd22ae4
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/asr_datamodule.py
@@ -0,0 +1,454 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--full-libri",
+            type=str2bool,
+            default=True,
+            help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=False,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=0,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_feature_masks=2,
+                    features_mask_size=5,
+                    num_frame_masks=10,
+                    frames_mask_size=5,
+                    p=0.5,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-100.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_360_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-360 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-360.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_other_500_cuts(self) -> CutSet:
+        logging.info("About to get train-other-500 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-other-500.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_all_shuf_cuts(self) -> CutSet:
+        logging.info(
+            "About to get the shuffled train-clean-100, \
+            train-clean-360 and train-other-500 cuts"
+        )
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/beam_search.py b/egs/librispeech/ASR/tiny_transducer_ctc/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/ctc_decode.py b/egs/librispeech/ASR/tiny_transducer_ctc/ctc_decode.py
new file mode 100644
index 000000000..cda03b56e
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/ctc_decode.py
@@ -0,0 +1,771 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import math
+import pprint
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="tiny_transducer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_phone",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.7,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "context_size": 2,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, _ = model.encoder(feature, feature_lens)
+    nnet_output = model.ctc_output(encoder_out)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    # lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    # lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    # lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+    lm_scale_list = [0.6, 0.7, 0.8, 0.9]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"{wer}-{test_set_name}-{key}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-hlg-scale-{params.hlg_scale}"
+
+    if params.use_averaged_model:
+        params.suffix += "-uam"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(pprint.pformat(params, indent=2))
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        assert "lang_bpe" in str(
+            params.lang_dir
+        ), "ctc-decoding only supports BPE lexicons."
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    enc_param = sum([p.numel() for p in model.encoder.parameters()])
+    ctc_param = sum([p.numel() for p in model.ctc_output.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+    logging.info(f"Parameters for CTC decoding: {enc_param + ctc_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/decode.py b/egs/librispeech/ASR/tiny_transducer_ctc/decode.py
new file mode 100644
index 000000000..6c2bf9ea1
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/decode.py
@@ -0,0 +1,717 @@
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+import pprint
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="tiny_transducer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="fast_beam_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.1,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG or 
+        fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest, 
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"{wer}-{test_set_name}-{key}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    if "lang_phone" in str(params.lang_dir):
+        assert params.decoding_method in (
+            "fast_beam_search_LG",
+            "fast_beam_search_nbest_LG",
+        ), "For phone lexicon, use a decoding method with LG."
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-uam"
+
+    setup_logger(f"{params.res_dir}/log-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    if "lang_bpe" in str(params.lang_dir):
+        sp = spm.SentencePieceProcessor()
+        sp.load(str(params.lang_dir / "bpe.model"))
+
+        # <blk> and <unk> are defined in local/train_bpe_model.py
+        params.blank_id = sp.piece_to_id("<blk>")
+        params.unk_id = sp.piece_to_id("<unk>")
+        params.vocab_size = sp.get_piece_size()
+    else:
+        params.blank_id = lexicon.token_table.get("<eps>")
+        params.unk_id = lexicon.token_table.get("SPN")
+        params.vocab_size = max(lexicon.tokens) + 1
+        sp = None
+
+    logging.info(pprint.pformat(params, indent=2))
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    enc_param = sum([p.numel() for p in model.encoder.parameters()])
+    dec_param = sum([p.numel() for p in model.decoder.parameters()])
+    join_param = sum([p.numel() for p in model.joiner.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+    logging.info(
+        f"Parameters for transducer decoding: {enc_param + dec_param + join_param}"
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/decoder.py b/egs/librispeech/ASR/tiny_transducer_ctc/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/encoder.py b/egs/librispeech/ASR/tiny_transducer_ctc/encoder.py
new file mode 100644
index 000000000..afdd00293
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/encoder.py
@@ -0,0 +1,378 @@
+#!/usr/bin/env python3
+# Copyright (c)  2022 Spacetouch Inc. (author: Tiance Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn.functional as F
+from encoder_interface import EncoderInterface
+from scaling import ActivationBalancer, DoubleSwish
+from torch import Tensor, nn
+
+
+class Conv1dNet(EncoderInterface):
+    """
+    1D Convolution network with causal squeeze and excitation
+    module and optional skip connections.
+
+    Latency: 80ms + (conv_layers+1) // 2 * 40ms, assuming 10ms stride.
+
+    Args:
+        output_dim (int): Number of output channels of the last layer.
+        input_dim (int): Number of input features
+        conv_layers (int): Number of convolution layers,
+        excluding the subsampling layers.
+        channels (int): Number of output channels for each layer,
+        except the last layer.
+        subsampling_factor (int): The subsampling factor for the model.
+        skip_add (bool): Whether to use skip connection for each convolution layer.
+        dscnn (bool): Whether to use depthwise-separated convolution.
+        activation (str): Activation function type.
+    """
+
+    def __init__(
+        self,
+        output_dim: int,
+        input_dim: int = 80,
+        conv_layers: int = 10,
+        channels: int = 256,
+        subsampling_factor: int = 4,
+        skip_add: bool = False,
+        dscnn: bool = True,
+        activation: str = "relu",
+    ) -> None:
+        super().__init__()
+        assert subsampling_factor == 4, "Only support subsampling = 4"
+
+        self.conv_layers = conv_layers
+        self.skip_add = skip_add
+        # 80ms latency for subsample_layer
+        self.subsample_layer = nn.Sequential(
+            conv1d_bn_block(
+                input_dim, channels, 9, stride=2, activation=activation, dscnn=dscnn
+            ),
+            conv1d_bn_block(
+                channels, channels, 5, stride=2, activation=activation, dscnn=dscnn
+            ),
+        )
+
+        self.conv_blocks = nn.ModuleList()
+        cin = [channels] * conv_layers
+        cout = [channels] * (conv_layers - 1) + [output_dim]
+
+        # Use causal and standard convolution alternatively
+        for ly in range(conv_layers):
+            self.conv_blocks.append(
+                nn.Sequential(
+                    conv1d_bn_block(
+                        cin[ly],
+                        cout[ly],
+                        3,
+                        activation=activation,
+                        dscnn=dscnn,
+                        causal=ly % 2,
+                    ),
+                    CausalSqueezeExcite1d(cout[ly], 16, 30),
+                )
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, encoder_dims)
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+        x = self.subsample_layer(x)
+        for idx, layer in enumerate(self.conv_blocks):
+            if self.skip_add and 0 < idx < self.conv_layers - 1:
+                x = layer(x) + x
+            else:
+                x = layer(x)
+        x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        lengths = x_lens >> 2
+        return x, lengths
+
+
+def get_activation(
+    name: str,
+    channels: int,
+    channel_dim: int = -1,
+    min_val: int = 0,
+    max_val: int = 1,
+) -> nn.Module:
+    """
+    Get activation function from name in string.
+
+    Args:
+        name: activation function name
+        channels: only used for PReLU, should be equal to x.shape[1].
+        channel_dim: the axis/dimension corresponding to the channel,
+            interprted as an offset from the input's ndim if negative.
+            e.g. for NCHW tensor, channel_dim = 1
+        min_val: minimum value of hardtanh
+        max_val: maximum value of hardtanh
+
+    Returns:
+        The activation function module
+
+    """
+    act_layer = nn.Identity()
+    name = name.lower()
+    if name == "prelu":
+        act_layer = nn.PReLU(channels)
+    elif name == "relu":
+        act_layer = nn.ReLU()
+    elif name == "relu6":
+        act_layer = nn.ReLU6()
+    elif name == "hardtanh":
+        act_layer = nn.Hardtanh(min_val, max_val)
+    elif name in ["swish", "silu"]:
+        act_layer = nn.SiLU()
+    elif name == "elu":
+        act_layer = nn.ELU()
+    elif name == "doubleswish":
+        act_layer = nn.Sequential(
+            ActivationBalancer(num_channels=channels, channel_dim=channel_dim),
+            DoubleSwish(),
+        )
+    elif name == "":
+        act_layer = nn.Identity()
+    else:
+        raise Exception(f"Unknown activation function: {name}")
+
+    return act_layer
+
+
+class CausalSqueezeExcite1d(nn.Module):
+    """
+    Causal squeeze and excitation module with input and output shape
+    (batch, channels, time). The global average pooling in the original
+    SE module is replaced by a causal filter, so
+    the layer does not introduce any algorithmic latency.
+
+    Args:
+        channels (int): Number of channels
+        reduction (int): channel reduction rate
+        context_window (int): Context window size for the moving average operation.
+        For EMA, the smoothing factor is 1 / context_window.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        reduction: int = 16,
+        context_window: int = 10,
+    ) -> None:
+        super(CausalSqueezeExcite1d, self).__init__()
+
+        assert channels >= reduction
+
+        self.context_window = context_window
+        c_squeeze = channels // reduction
+        self.linear1 = nn.Linear(channels, c_squeeze, bias=True)
+        self.act1 = nn.ReLU(inplace=True)
+        self.linear2 = nn.Linear(c_squeeze, channels, bias=True)
+        self.act2 = nn.Sigmoid()
+
+        # EMA worked better than MA empirically
+        # self.avg_filter = self.moving_avg
+        self.avg_filter = self.exponential_moving_avg
+        self.ema_matrix = torch.tensor([0])
+        self.ema_matrix_size = 0
+
+    def _precompute_ema_matrix(self, N: int, device: torch.device):
+        a = 1.0 / self.context_window  # smoothing factor
+        w = [[(1 - a) ** k * a for k in range(n, n - N, -1)] for n in range(N)]
+        w = torch.tensor(w).to(device).tril()
+        w[:, 0] *= self.context_window
+        self.ema_matrix = w.T
+        self.ema_matrix_size = N
+
+    def exponential_moving_avg(self, x: Tensor) -> Tensor:
+        """
+        Exponential moving average filter, which is calculated as:
+            y[t] = (1-a) * y[t-1] + a * x[t]
+        where a = 1 / self.context_window is the smoothing factor.
+
+        For training, the iterative version is too slow. A better way is
+        to expand y[t] as a function of x[0..t] only and use matrix-vector multiplication.
+        The weight matrix can be precomputed if the smoothing factor is fixed.
+        """
+        if self.training:
+            # use matrix version to speed up training
+            N = x.shape[-1]
+            if N > self.ema_matrix_size:
+                self._precompute_ema_matrix(N, x.device)
+            y = torch.matmul(x, self.ema_matrix[:N, :N])
+        else:
+            # use iterative version to save memory
+            a = 1.0 / self.context_window
+            y = torch.empty_like(x)
+            y[:, :, 0] = x[:, :, 0]
+            for t in range(1, y.shape[-1]):
+                y[:, :, t] = (1 - a) * y[:, :, t - 1] + a * x[:, :, t]
+        return y
+
+    def moving_avg(self, x: Tensor) -> Tensor:
+        """
+        Simple moving average with context_window as window size.
+        """
+        y = torch.empty_like(x)
+        k = min(x.shape[2], self.context_window)
+        w = [[1 / n] * n + [0] * (k - n - 1) for n in range(1, k)]
+        w = torch.tensor(w, device=x.device)
+        y[:, :, : k - 1] = torch.matmul(x[:, :, : k - 1], w.T)
+        y[:, :, k - 1 :] = F.avg_pool1d(x, k, 1)
+        return y
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert len(x.shape) == 3, "Input is not a 3D tensor!"
+        y = self.exponential_moving_avg(x)
+        y = y.permute(0, 2, 1)  # make channel last for squeeze op
+        y = self.act1(self.linear1(y))
+        y = self.act2(self.linear2(y))
+        y = y.permute(0, 2, 1)  # back to original shape
+        y = x * y
+        return y
+
+
+def conv1d_bn_block(
+    in_channels: int,
+    out_channels: int,
+    kernel_size: int = 3,
+    stride: int = 1,
+    dilation: int = 1,
+    activation: str = "relu",
+    dscnn: bool = False,
+    causal: bool = False,
+) -> nn.Sequential:
+    """
+    Conv1d - batchnorm - activation block.
+    If kernel size is even, output length = input length + 1.
+    Otherwise, output and input lengths are equal.
+
+    Args:
+        in_channels (int): Number of input channels
+        out_channels (int): Number of output channels
+        kernel_size (int): kernel size
+        stride (int): convolution stride
+        dilation (int): convolution dilation rate
+        dscnn (bool): Use depthwise separated convolution.
+        causal (bool): Use causal convolution
+        activation (str): Activation function type.
+
+    """
+    if dscnn:
+        return nn.Sequential(
+            CausalConv1d(
+                in_channels,
+                in_channels,
+                kernel_size,
+                stride=stride,
+                dilation=dilation,
+                groups=in_channels,
+                bias=False,
+            )
+            if causal
+            else nn.Conv1d(
+                in_channels,
+                in_channels,
+                kernel_size,
+                stride=stride,
+                padding=(kernel_size // 2) * dilation,
+                dilation=dilation,
+                groups=in_channels,
+                bias=False,
+            ),
+            nn.BatchNorm1d(in_channels),
+            get_activation(activation, in_channels),
+            nn.Conv1d(in_channels, out_channels, 1, bias=False),
+            nn.BatchNorm1d(out_channels),
+            get_activation(activation, out_channels),
+        )
+    else:
+        return nn.Sequential(
+            CausalConv1d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                stride=stride,
+                dilation=dilation,
+                bias=False,
+            )
+            if causal
+            else nn.Conv1d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                stride=stride,
+                padding=(kernel_size // 2) * dilation,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm1d(out_channels),
+            get_activation(activation, out_channels),
+        )
+
+
+class CausalConv1d(nn.Module):
+    """
+    Causal convolution with padding automatically chosen to match input/output length.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = 1,
+        dilation: int = 1,
+        groups: int = 1,
+        bias: bool = True,
+    ) -> None:
+        super(CausalConv1d, self).__init__()
+        assert kernel_size > 2
+
+        self.padding = dilation * (kernel_size - 1)
+        self.stride = stride
+
+        self.conv = nn.Conv1d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            self.padding,
+            dilation,
+            groups,
+            bias=bias,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.conv(x)[:, :, : -self.padding // self.stride]
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/encoder_interface.py b/egs/librispeech/ASR/tiny_transducer_ctc/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/export.py b/egs/librispeech/ASR/tiny_transducer_ctc/export.py
new file mode 100755
index 000000000..334dd011e
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/export.py
@@ -0,0 +1,305 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./tiny_transducer_ctc/export.py \
+  --exp-dir ./tiny_transducer_ctc/exp \
+  --lang-dir data/lang_bpe_500 \
+  --epoch 30 \
+  --avg 2 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./tiny_transducer_ctc/export.py \
+  --exp-dir ./tiny_transducer_ctc/exp \
+  --lang-dir data/lang_bpe_500 \
+  --epoch 30 \
+  --avg 2
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `tiny_transducer_ctc/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./tiny_transducer_ctc/decode.py \
+        --exp-dir ./tiny_transducer_ctc/exp \
+        --epoch 9999 \
+        --use-averaged-model 0
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --lang-dir data/lang_bpe_500 \
+
+Check ./pretrained.py for its usage.
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="tiny_transducer_ctc/exp_4m_bpe500_halfdelay_specaug",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        logging.info("Using torch.jit.script()")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/jit_pretrained.py b/egs/librispeech/ASR/tiny_transducer_ctc/jit_pretrained.py
new file mode 100755
index 000000000..3888d3544
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/jit_pretrained.py
@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./tiny_transducer_ctc/export.py \
+  --exp-dir ./tiny_transducer_ctc/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./tiny_transducer_ctc/jit_pretrained.py \
+  --nn-model-filename ./tiny_transducer_ctc/exp/cpu_jit.pt \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = sp.decode(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/jit_pretrained_ctc.py b/egs/librispeech/ASR/tiny_transducer_ctc/jit_pretrained_ctc.py
new file mode 100755
index 000000000..6f2cbaabd
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/jit_pretrained_ctc.py
@@ -0,0 +1,426 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./tiny_transducer_ctc/export.py \
+  --exp-dir ./tiny_transducer_ctc/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+(1) ctc-decoding
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --model-filename ./tiny_transducer_ctc/exp/cpu_jit.pt \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --model-filename ./tiny_transducer_ctc/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(3) nbest-rescoring
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --model-filename ./tiny_transducer_ctc/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --model-filename ./tiny_transducer_ctc/exp/cpu_jit.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.model_filename)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+    nnet_output = model.ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/joiner.py b/egs/librispeech/ASR/tiny_transducer_ctc/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/model.py b/egs/librispeech/ASR/tiny_transducer_ctc/model.py
new file mode 120000
index 000000000..545af927f
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7_ctc/model.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/pretrained.py b/egs/librispeech/ASR/tiny_transducer_ctc/pretrained.py
new file mode 100755
index 000000000..981039b8f
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/pretrained.py
@@ -0,0 +1,357 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./tiny_transducer_ctc/export.py \
+  --exp-dir ./tiny_transducer_ctc/exp \
+  --lang-dir data/lang_bpe_500 \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./tiny_transducer_ctc/pretrained.py \
+    --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+    --lang-dir data/lang_bpe_500 \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./tiny_transducer_ctc/pretrained.py \
+    --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+    --lang-dir data/lang_bpe_500 \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./tiny_transducer_ctc/pretrained.py \
+    --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+    --lang-dir data/lang_bpe_500 \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./tiny_transducer_ctc/pretrained.py \
+    --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+    --lang-dir data/lang_bpe_500 \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./tiny_transducer_ctc/exp/epoch-xx.pt`.
+
+Note: ./tiny_transducer_ctc/exp/pretrained.pt is generated by
+./tiny_transducer_ctc/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.lang_dir + "/bpe.model")
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/pretrained_ctc.py b/egs/librispeech/ASR/tiny_transducer_ctc/pretrained_ctc.py
new file mode 100755
index 000000000..a06d6d684
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/pretrained_ctc.py
@@ -0,0 +1,444 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./tiny_transducer_ctc/export.py \
+  --exp-dir ./tiny_transducer_ctc/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) ctc-decoding
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+  --lang-dir data/lang_bpe_500 \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./tiny_transducer_ctc/jit_pretrained_ctc.py \
+  --checkpoint ./tiny_transducer_ctc/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+    params.vocab_size = params.num_classes
+    params.blank_id = 0
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+    nnet_output = model.ctc_output(encoder_out)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/scaling.py b/egs/librispeech/ASR/tiny_transducer_ctc/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/train.py b/egs/librispeech/ASR/tiny_transducer_ctc/train.py
new file mode 100644
index 000000000..8920764cd
--- /dev/null
+++ b/egs/librispeech/ASR/tiny_transducer_ctc/train.py
@@ -0,0 +1,1249 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+cd egs/librispeech/ASR/
+./prepare.sh
+
+Run below if you want to use the phone lexicon instead of BPE:
+python local/generate_unique_lexicon.py --lang-dir data/lang_phone 
+
+"""
+import argparse
+import copy
+import logging
+import pprint
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from encoder import Conv1dNet
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import AdamW
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics, is_module_available
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import UniqLexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = torch.optim.lr_scheduler._LRScheduler
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=256,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=256,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=256,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--conv-layers",
+        type=int,
+        default=10,
+        help="""Number of convolution layers for the encoder.
+        """,
+    )
+
+    parser.add_argument(
+        "--channels",
+        type=int,
+        default=256,
+        help="""Number of channels for the encoder.
+        """,
+    )
+
+    parser.add_argument(
+        "--skip-add",
+        type=str2bool,
+        default=False,
+        help="""Use skip connection in the encoder.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="tiny_transducer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="""Weight for CTC loss, between 0 and 1. 
+        When set to 0, only transducer loss is used. 
+        When set to 1, only CTC loss is used.""",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=10000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=5,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 500,
+            "reset_interval": 200,
+            "warm_step": 5000,
+            "beam_size": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_dscnn": True,
+            "activation": "doubleswish",
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conv1dNet(
+        output_dim=params.encoder_dim,
+        input_dim=params.feature_dim,
+        conv_layers=params.conv_layers,
+        channels=params.channels,
+        subsampling_factor=params.subsampling_factor,
+        skip_add=params.skip_add,
+        dscnn=params.use_dscnn,
+        activation=params.activation,
+    )
+
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> Transducer:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    if is_module_available("thop"):
+        if torch.cuda.is_available():
+            device = torch.device("cuda:0")
+        else:
+            device = torch.device("cpu")
+        # Assuming 10ms stride, 1000 frames is about 10 seconds.
+        x = torch.zeros((1, 1000, params.feature_dim)).to(device)
+        x_lens = torch.Tensor([1000]).int().to(device)
+        from thop import clever_format, profile
+
+        m = copy.deepcopy(encoder)
+        m = m.to(device)
+        ops, _ = clever_format(profile(m, (x, x_lens), verbose=False))
+        logging.info(f"Encoder MAC ops for 10 seconds of audio is {ops}")
+    else:
+        logging.info("You can install thop to calculate the number of ops.")
+        logging.info("Command: pip install thop")
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    phone_lexicon: UniqLexicon,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    if sp is not None:
+        token_ids = sp.encode(texts, out_type=int)
+        y = k2.RaggedTensor(token_ids).to(device)
+    else:
+        y = phone_lexicon.texts_to_token_ids(texts).to(device)
+        token_ids = y.tolist()
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_output = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    # Compute ctc loss
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        supervision_segments, token_ids = encode_supervisions(
+            supervisions,
+            subsampling_factor=params.subsampling_factor,
+            token_ids=token_ids,
+        )
+
+    decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device)
+    dense_fsa_vec = k2.DenseFsaVec(
+        ctc_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction="sum",
+        use_double_scores=params.use_double_scores,
+    )
+    assert ctc_loss.requires_grad == is_training
+    assert 0 <= params.ctc_loss_scale <= 1, "ctc_loss_scale must be between 0 and 1"
+    loss = params.ctc_loss_scale * ctc_loss + (1 - params.ctc_loss_scale) * loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    phone_lexicon: UniqLexicon,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            phone_lexicon=phone_lexicon,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    phone_lexicon: UniqLexicon,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    phone_lexicon=phone_lexicon,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            # scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(
+                batch, params=params, sp=sp, phone_lexicon=phone_lexicon
+            )
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+    logging.info("Computing validation loss")
+    valid_info = compute_validation_loss(
+        params=params,
+        model=model,
+        sp=sp,
+        phone_lexicon=phone_lexicon,
+        valid_dl=valid_dl,
+        world_size=world_size,
+    )
+    model.train()
+    logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+    logging.info(
+        f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+    )
+    if tb_writer is not None:
+        valid_info.write_summary(tb_writer, "train/valid_", params.batch_idx_train)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    if "lang_bpe" in str(params.lang_dir):
+        sp = spm.SentencePieceProcessor()
+        sp.load(params.lang_dir + "/bpe.model")
+        # <blk> is defined in local/train_bpe_model.py
+        params.blank_id = sp.piece_to_id("<blk>")
+        params.vocab_size = sp.get_piece_size()
+        phone_lexicon = None
+    else:
+        assert "lang_phone" in str(params.lang_dir)
+        phone_lexicon = UniqLexicon(params.lang_dir)
+        params.blank_id = 0
+        params.vocab_size = max(phone_lexicon.tokens) + 1
+        sp = None
+
+    logging.info(pprint.pformat(params, indent=2))
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if rank == 0:
+        num_param = sum([p.numel() for p in model.parameters()])
+        enc_param = sum([p.numel() for p in model.encoder.parameters()])
+        dec_param = sum([p.numel() for p in model.decoder.parameters()])
+        join_param = sum([p.numel() for p in model.joiner.parameters()])
+        ctc_param = sum([p.numel() for p in model.ctc_output.parameters()])
+
+        logging.info(f"Number of model parameters: {num_param}")
+        logging.info(f"Number of encoder parameters: {enc_param}")
+        logging.info(f"Number of decoder parameters: {dec_param}")
+        logging.info(f"Number of joiner parameters: {join_param}")
+        logging.info(f"Number of ctc parameters: {ctc_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = AdamW(
+        model.parameters(),
+        lr=params.initial_lr,
+        weight_decay=5e-4,
+    )
+
+    scheduler = StepLR(optimizer, step_size=2, gamma=0.8)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         sp=sp,
+    #         phone_lexicon=phone_lexicon,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        # scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            phone_lexicon=phone_lexicon,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+    phone_lexicon: UniqLexicon,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    if sp is not None:
+        y = sp.encode(supervisions["text"], out_type=int)
+    else:
+        y = phone_lexicon.texts_to_token_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    phone_lexicon: UniqLexicon,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    phone_lexicon=phone_lexicon,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(
+                batch, params=params, sp=sp, phone_lexicon=phone_lexicon
+            )
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/README.md b/egs/librispeech/ASR/transducer/README.md
new file mode 100644
index 000000000..a5b067f64
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/README.md
@@ -0,0 +1,19 @@
+## Introduction
+
+The encoder consists of Conformer layers in this folder. You can use the
+following command to start the training:
+
+```bash
+cd egs/librispeech/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer/exp \
+  --full-libri 1 \
+  --max-duration 250 \
+  --lr-factor 2.5
+```
diff --git a/egs/librispeech/ASR/transducer/__init__.py b/egs/librispeech/ASR/transducer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/transducer/asr_datamodule.py b/egs/librispeech/ASR/transducer/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer/beam_search.py b/egs/librispeech/ASR/transducer/beam_search.py
new file mode 100644
index 000000000..b45b6a9d8
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/beam_search.py
@@ -0,0 +1,217 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+import torch
+from model import Transducer
+
+
+def greedy_search(model: Transducer, encoder_out: torch.Tensor) -> List[int]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    device = model.device
+
+    sos = torch.tensor([blank_id], device=device, dtype=torch.int64).reshape(1, 1)
+    decoder_out, (h, c) = model.decoder(sos)
+    T = encoder_out.size(1)
+    t = 0
+    hyp = []
+
+    sym_per_frame = 0
+    sym_per_utt = 0
+
+    max_sym_per_utt = 1000
+    max_sym_per_frame = 3
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(current_encoder_out, decoder_out)
+        # logits is (1, 1, 1, vocab_size)
+
+        log_prob = logits.log_softmax(dim=-1)
+        # log_prob is (1, 1, 1, vocab_size)
+        # TODO: Use logits.argmax()
+        y = log_prob.argmax()
+        if y != blank_id:
+            hyp.append(y.item())
+            y = y.reshape(1, 1)
+            decoder_out, (h, c) = model.decoder(y, (h, c))
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+
+        if y == blank_id or sym_per_frame > max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+
+    return hyp
+
+
+@dataclass
+class Hypothesis:
+    ys: List[int]  # the predicted sequences so far
+    log_prob: float  # The log prob of ys
+
+    # Optional decoder state. We assume it is LSTM for now,
+    # so the state is a tuple (h, c)
+    decoder_state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 5,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    device = model.device
+
+    sos = torch.tensor([blank_id], device=device).reshape(1, 1)
+    decoder_out, (h, c) = model.decoder(sos)
+    T = encoder_out.size(1)
+    t = 0
+    B = [Hypothesis(ys=[blank_id], log_prob=0.0, decoder_state=None)]
+    max_u = 20000  # terminate after this number of steps
+    u = 0
+
+    cache: Dict[str, Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = {}
+
+    while t < T and u < max_u:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        A = B
+        B = []
+        #  for hyp in A:
+        #      for h in A:
+        #          if h.ys == hyp.ys[:-1]:
+        #              # update the score of hyp
+        #              decoder_input = torch.tensor(
+        #                  [h.ys[-1]], device=device
+        #              ).reshape(1, 1)
+        #              decoder_out, _ = model.decoder(
+        #                  decoder_input, h.decoder_state
+        #              )
+        #              logits = model.joiner(current_encoder_out, decoder_out)
+        #              log_prob = logits.log_softmax(dim=-1)
+        #              log_prob = log_prob.squeeze()
+        #              hyp.log_prob += h.log_prob + log_prob[hyp.ys[-1]].item()
+
+        while u < max_u:
+            y_star = max(A, key=lambda hyp: hyp.log_prob)
+            A.remove(y_star)
+
+            # Note: y_star.ys is unhashable, i.e., cannot be used
+            # as a key into a dict
+            cached_key = "_".join(map(str, y_star.ys))
+
+            if cached_key not in cache:
+                decoder_input = torch.tensor([y_star.ys[-1]], device=device).reshape(
+                    1, 1
+                )
+
+                decoder_out, decoder_state = model.decoder(
+                    decoder_input,
+                    y_star.decoder_state,
+                )
+                cache[cached_key] = (decoder_out, decoder_state)
+            else:
+                decoder_out, decoder_state = cache[cached_key]
+
+            logits = model.joiner(current_encoder_out, decoder_out)
+            log_prob = logits.log_softmax(dim=-1)
+            # log_prob is (1, 1, 1, vocab_size)
+            log_prob = log_prob.squeeze()
+            # Now log_prob is (vocab_size,)
+
+            # If we choose blank here, add the new hypothesis to B.
+            # Otherwise, add the new hypothesis to A
+
+            # First, choose blank
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob.item()
+
+            # ys[:] returns a copy of ys
+            new_y_star = Hypothesis(
+                ys=y_star.ys[:],
+                log_prob=new_y_star_log_prob,
+                # Caution: Use y_star.decoder_state here
+                decoder_state=y_star.decoder_state,
+            )
+            B.append(new_y_star)
+
+            # Second, choose other labels
+            for i, v in enumerate(log_prob.tolist()):
+                if i == blank_id:
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                new_hyp = Hypothesis(
+                    ys=new_ys,
+                    log_prob=new_log_prob,
+                    decoder_state=decoder_state,
+                )
+                A.append(new_hyp)
+            u += 1
+            # check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = max(A, key=lambda hyp: hyp.log_prob)
+            B = sorted(
+                [hyp for hyp in B if hyp.log_prob > A_most_probable.log_prob],
+                key=lambda hyp: hyp.log_prob,
+                reverse=True,
+            )
+            if len(B) >= beam:
+                B = B[:beam]
+                break
+        t += 1
+    best_hyp = max(B, key=lambda hyp: hyp.log_prob / len(hyp.ys[1:]))
+    ys = best_hyp.ys[1:]  # [1:] to remove the blank
+    return ys
diff --git a/egs/librispeech/ASR/transducer/conformer.py b/egs/librispeech/ASR/transducer/conformer.py
new file mode 120000
index 000000000..70a7ddf11
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/conformer.py
@@ -0,0 +1 @@
+../transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer/decode.py b/egs/librispeech/ASR/transducer/decode.py
new file mode 100755
index 000000000..8d379d1fa
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/decode.py
@@ -0,0 +1,450 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./transducer/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import beam_search, greedy_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=11,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=5,
+        help="Used only when --decoding-method is beam_search",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # decoder params
+            "decoder_embedding_dim": 1024,
+            "num_decoder_layers": 2,
+            "decoder_hidden_dim": 512,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.decoder_embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.decoder_hidden_dim,
+        output_dim=params.encoder_out_dim,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.decoding_method == "greedy_search":
+            hyp = greedy_search(model=model, encoder_out=encoder_out_i)
+        elif params.decoding_method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+        hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in ("greedy_search", "beam_search")
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if params.decoding_method == "beam_search":
+        params.suffix += f"-beam-{params.beam_size}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/decoder.py b/egs/librispeech/ASR/transducer/decoder.py
new file mode 100644
index 000000000..333fff300
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/decoder.py
@@ -0,0 +1,97 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+
+# TODO(fangjun): Support switching between LSTM and GRU
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        num_layers: int,
+        hidden_dim: int,
+        output_dim: int,
+        embedding_dropout: float = 0.0,
+        rnn_dropout: float = 0.0,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          num_layers:
+            Number of LSTM layers.
+          hidden_dim:
+            Hidden dimension of LSTM layers.
+          output_dim:
+            Output dimension of the decoder.
+          embedding_dropout:
+            Dropout rate for the embedding layer.
+          rnn_dropout:
+            Dropout for LSTM layers.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.embedding_dropout = nn.Dropout(embedding_dropout)
+        # TODO(fangjun): Use layer normalized LSTM
+        self.rnn = nn.LSTM(
+            input_size=embedding_dim,
+            hidden_size=hidden_dim,
+            num_layers=num_layers,
+            batch_first=True,
+            dropout=rnn_dropout,
+        )
+        self.blank_id = blank_id
+        self.output_linear = nn.Linear(hidden_dim, output_dim)
+
+    def forward(
+        self,
+        y: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U) with BOS prepended.
+          states:
+            A tuple of two tensors containing the states information of
+            LSTM layers in this decoder.
+        Returns:
+          Return a tuple containing:
+
+            - rnn_output, a tensor of shape (N, U, C)
+            - (h, c), containing the state information for LSTM layers.
+              Both are of shape (num_layers, N, C)
+        """
+        embedding_out = self.embedding(y)
+        embedding_out = self.embedding_dropout(embedding_out)
+        rnn_out, (h, c) = self.rnn(embedding_out, states)
+        out = self.output_linear(rnn_out)
+
+        return out, (h, c)
diff --git a/egs/librispeech/ASR/transducer/encoder_interface.py b/egs/librispeech/ASR/transducer/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer/export.py b/egs/librispeech/ASR/transducer/export.py
new file mode 100755
index 000000000..3b9e4a5dc
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/export.py
@@ -0,0 +1,246 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer/export.py \
+  --exp-dir ./transducer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 34 \
+  --avg 11
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer/decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./transducer/decode.py \
+        --exp-dir ./transducer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=11,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # decoder params
+            "decoder_embedding_dim": 1024,
+            "num_decoder_layers": 2,
+            "decoder_hidden_dim": 512,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.decoder_embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.decoder_hidden_dim,
+        output_dim=params.encoder_out_dim,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert args.jit is False, "Support torchscript will be added later"
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer/joiner.py b/egs/librispeech/ASR/transducer/joiner.py
new file mode 100644
index 000000000..2ef3f1de6
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/joiner.py
@@ -0,0 +1,54 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, C).
+        Returns:
+          Return a tensor of shape (N, T, U, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == decoder_out.size(2)
+
+        encoder_out = encoder_out.unsqueeze(2)
+        # Now encoder_out is (N, T, 1, C)
+
+        decoder_out = decoder_out.unsqueeze(1)
+        # Now decoder_out is (N, 1, U, C)
+
+        logit = encoder_out + decoder_out
+        logit = torch.tanh(logit)
+
+        output = self.output_linear(logit)
+
+        return output
diff --git a/egs/librispeech/ASR/transducer/model.py b/egs/librispeech/ASR/transducer/model.py
new file mode 100644
index 000000000..8305248c9
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/model.py
@@ -0,0 +1,126 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Note we use `rnnt_loss` from torchaudio, which exists only in
+torchaudio >= v0.10.0. It also means you have to use torch >= v1.10.0
+"""
+import k2
+import torch
+import torch.nn as nn
+import torchaudio
+import torchaudio.functional
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        decoder_out, _ = self.decoder(sos_y_padded)
+
+        logits = self.joiner(encoder_out, decoder_out)
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        assert hasattr(torchaudio.functional, "rnnt_loss"), (
+            f"Current torchaudio version: {torchaudio.__version__}\n"
+            "Please install a version >= 0.10.0"
+        )
+
+        loss = torchaudio.functional.rnnt_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+        )
+
+        return loss
diff --git a/egs/librispeech/ASR/transducer/pretrained.py b/egs/librispeech/ASR/transducer/pretrained.py
new file mode 100755
index 000000000..c2413f5de
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/pretrained.py
@@ -0,0 +1,295 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./transducer/pretrained.py \
+        --checkpoint ./transducer/exp/pretrained.pt \
+        --tokens data/lang_bpe_500/tokens.txt \
+        --method greedy_search \
+        /path/to/foo.wav \
+        /path/to/bar.wav \
+
+You can also use `./transducer/exp/epoch-xx.pt`.
+
+Note: ./transducer/exp/pretrained.pt is generated by
+./transducer/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import beam_search, greedy_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="Path to tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=5,
+        help="Used only when --method is beam_search",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # decoder params
+            "decoder_embedding_dim": 1024,
+            "num_decoder_layers": 2,
+            "decoder_hidden_dim": 512,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.decoder_embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.decoder_hidden_dim,
+        output_dim=params.encoder_out_dim,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    for i in range(num_waves):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.method == "greedy_search":
+            hyp = greedy_search(model=model, encoder_out=encoder_out_i)
+        elif params.method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported method: {params.method}")
+
+        hyps.append(token_ids_to_words(hyp))
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer/rnn.py b/egs/librispeech/ASR/transducer/rnn.py
new file mode 100644
index 000000000..fe8732301
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/rnn.py
@@ -0,0 +1,665 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Apply layer normalization to the output of each gate in LSTM/GRU.
+
+This file uses
+https://github.com/pytorch/pytorch/blob/master/benchmarks/fastrnns/custom_lstms.py
+as a reference.
+"""
+
+import math
+from typing import List, Optional, Tuple, Type
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typeguard import check_argument_types
+
+
+class LayerNormLSTMCell(nn.Module):
+    """This class places a `nn.LayerNorm` after the output of
+    each gate (right before the activation).
+
+    See the following paper for more details
+
+    'Improving RNN Transducer Modeling for End-to-End Speech Recognition'
+    https://arxiv.org/abs/1909.12415
+
+    Examples::
+
+        >>> cell = LayerNormLSTMCell(10, 20)
+        >>> input = torch.rand(5, 10)
+        >>> h0 = torch.rand(5, 20)
+        >>> c0 = torch.rand(5, 20)
+        >>> h1, c1 = cell(input, (h0, c0))
+        >>> output = h1
+        >>> h1.shape
+        torch.Size([5, 20])
+        >>> c1.shape
+        torch.Size([5, 20])
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        bias: bool = True,
+        ln: Type[nn.Module] = nn.LayerNorm,
+        proj_size: int = 0,
+        device=None,
+        dtype=None,
+    ):
+        """
+        Args:
+          input_size:
+            The number of expected features in the input `x`. `x` should
+            be of shape (batch_size, input_size).
+          hidden_size:
+            The number of features in the hidden state `h` and `c`.
+            Both `h` and `c` are of shape (batch_size, hidden_size) when
+            proj_size is 0. If proj_size is not zero, the shape of `h`
+            is (batch_size, proj_size).
+          bias:
+            If ``False``, then the cell does not use bias weights
+            `bias_ih` and `bias_hh`.
+          ln:
+            Defaults to `nn.LayerNorm`. The output of all gates are processed
+            by `ln`. We pass it as an argument so that we can replace it
+            with `nn.Identity` at the testing time.
+          proj_size:
+            If not zero, it applies an affine transform to the output. In this
+            case, the shape of `h` is (batch_size, proj_size).
+            See https://arxiv.org/pdf/1402.1128.pdf
+        """
+        assert check_argument_types()
+        super().__init__()
+        factory_kwargs = {"device": device, "dtype": dtype}
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.bias = bias
+        self.proj_size = proj_size
+
+        if proj_size < 0:
+            raise ValueError(
+                f"proj_size {proj_size} should be a positive integer "
+                "or zero to disable projections"
+            )
+
+        if proj_size >= hidden_size:
+            raise ValueError(
+                f"proj_size {proj_size} has to be smaller "
+                f"than hidden_size {hidden_size}"
+            )
+
+        real_hidden_size = proj_size if proj_size > 0 else hidden_size
+
+        self.weight_ih = nn.Parameter(
+            torch.empty((4 * hidden_size, input_size), **factory_kwargs)
+        )
+
+        self.weight_hh = nn.Parameter(
+            torch.empty((4 * hidden_size, real_hidden_size), **factory_kwargs)
+        )
+
+        if bias:
+            self.bias_ih = nn.Parameter(torch.empty(4 * hidden_size, **factory_kwargs))
+            self.bias_hh = nn.Parameter(torch.empty(4 * hidden_size, **factory_kwargs))
+        else:
+            self.register_parameter("bias_ih", None)
+            self.register_parameter("bias_hh", None)
+
+        if proj_size > 0:
+            self.weight_hr = nn.Parameter(
+                torch.empty((proj_size, hidden_size), **factory_kwargs)
+            )
+        else:
+            self.register_parameter("weight_hr", None)
+
+        self.layernorm_i = ln(hidden_size)
+        self.layernorm_f = ln(hidden_size)
+        self.layernorm_cx = ln(hidden_size)
+        self.layernorm_cy = ln(hidden_size)
+        self.layernorm_o = ln(hidden_size)
+
+        self.reset_parameters()
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          input:
+            A 2-D tensor of shape (batch_size, input_size).
+          state:
+            If not ``None``, it contains the hidden state (h, c) for each
+            element in the batch. Both are of shape (batch_size, hidden_size)
+            if proj_size is 0. If proj_size is not zero, the shape of `h` is
+            (batch_size, proj_size).
+            If ``None``, it uses zeros for `h` and `c`.
+        Returns:
+          Return two tensors:
+            - `next_h`: It is of shape (batch_size, hidden_size) if proj_size
+              is 0, else (batch_size, proj_size), containing the next hidden
+              state for each element in the batch.
+            - `next_c`: It is of shape (batch_size, hidden_size) containing the
+              next cell state for each element in the batch.
+        """
+        if state is None:
+            zeros = torch.zeros(
+                input.size(0),
+                self.hidden_size,
+                dtype=input.dtype,
+                device=input.device,
+            )
+            state = (zeros, zeros)
+
+        hx, cx = state
+        gates = F.linear(input, self.weight_ih, self.bias_ih) + F.linear(
+            hx, self.weight_hh, self.bias_hh
+        )
+
+        in_gate, forget_gate, cell_gate, out_gate = gates.chunk(chunks=4, dim=1)
+
+        in_gate = self.layernorm_i(in_gate)
+        forget_gate = self.layernorm_f(forget_gate)
+        cell_gate = self.layernorm_cx(cell_gate)
+        out_gate = self.layernorm_o(out_gate)
+
+        in_gate = torch.sigmoid(in_gate)
+        forget_gate = torch.sigmoid(forget_gate)
+        cell_gate = torch.tanh(cell_gate)
+        out_gate = torch.sigmoid(out_gate)
+
+        cy = (forget_gate * cx) + (in_gate * cell_gate)
+        cy = self.layernorm_cy(cy)
+        hy = out_gate * torch.tanh(cy)
+
+        if self.weight_hr is not None:
+            hy = torch.matmul(hy, self.weight_hr.t())
+
+        return hy, cy
+
+    def extra_repr(self) -> str:
+        s = "{input_size}, {hidden_size}"
+        if "bias" in self.__dict__ and self.bias is not True:
+            s += ", bias={bias}"
+        return s.format(**self.__dict__)
+
+    def reset_parameters(self) -> None:
+        stdv = 1.0 / math.sqrt(self.hidden_size)
+        for name, weight in self.named_parameters():
+            if "layernorm" not in name:
+                nn.init.uniform_(weight, -stdv, stdv)
+
+            if "bias_ih" in name or "bias_hh" in name:
+                # See the paper
+                # An Empirical Exploration of Recurrent Network Architectures
+                # https://proceedings.mlr.press/v37/jozefowicz15.pdf
+                #
+                # It recommends initializing the bias of the forget gate to
+                # a large value, such as 1 or 2. In PyTorch, there are two
+                # biases for the forget gate, we set both of them to 1 here.
+                #
+                # See also https://arxiv.org/pdf/1804.04849.pdf
+                assert weight.ndim == 1
+                # Layout of the bias:
+                # | in_gate | forget_gate | cell_gate | output_gate |
+                start = weight.numel() // 4
+                end = weight.numel() // 2
+                with torch.no_grad():
+                    weight[start:end].fill_(1.0)
+
+
+class LayerNormLSTMLayer(nn.Module):
+    """
+    Examples::
+
+        >>> layer = LayerNormLSTMLayer(10, 20)
+        >>> input = torch.rand(2, 5, 10)
+        >>> h0 = torch.rand(2, 20)
+        >>> c0 = torch.rand(2, 20)
+        >>> output, (hn, cn) = layer(input, (h0, c0))
+        >>> output.shape
+        torch.Size([2, 5, 20])
+        >>> hn.shape
+        torch.Size([2, 20])
+        >>> cn.shape
+        torch.Size([2, 20])
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        bias: bool = True,
+        ln: Type[nn.Module] = nn.LayerNorm,
+        proj_size: int = 0,
+        device=None,
+        dtype=None,
+    ):
+        """
+        See the args in LayerNormLSTMCell
+        """
+        assert check_argument_types()
+        super().__init__()
+        self.cell = LayerNormLSTMCell(
+            input_size=input_size,
+            hidden_size=hidden_size,
+            bias=bias,
+            ln=ln,
+            proj_size=proj_size,
+            device=device,
+            dtype=dtype,
+        )
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        state: Tuple[torch.Tensor, torch.Tensor],
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          input:
+            A 3-D tensor of shape (batch_size, seq_len, input_size).
+            Caution:
+              We use `batch_first=True` here.
+          state:
+            If not ``None``, it contains the hidden state (h, c) of this layer.
+            Both are of shape (batch_size, hidden_size) if proj_size is 0.
+            If proj_size is not 0, the shape of `h` is (batch_size, proj_size).
+            Note:
+              We did not annotate `state` with `Optional[Tuple[...]]` since
+              torchscript will complain.
+        Return:
+          - output, a tensor of shape (batch_size, seq_len, hidden_size)
+          - (next_h, next_c) containing the next hidden state
+        """
+        inputs = input.unbind(1)
+        outputs = torch.jit.annotate(List[torch.Tensor], [])
+        for i in range(len(inputs)):
+            state = self.cell(inputs[i], state)
+            outputs.append(state[0])
+        return torch.stack(outputs, dim=1), state
+
+
+class LayerNormLSTM(nn.Module):
+    """
+    Examples::
+
+        >>> lstm = LayerNormLSTM(10, 20, 8)
+        >>> input = torch.rand(2, 3, 10)
+        >>> h0 = torch.rand(8, 2, 20).unbind(0)
+        >>> c0 = torch.rand(8, 2, 20).unbind(0)
+        >>> states = list(zip(h0, c0))
+        >>> output, next_states = lstm(input, states)
+        >>> output.shape
+        torch.Size([2, 3, 20])
+        >>> hn = torch.stack([s[0] for s in next_states])
+        >>> cn = torch.stack([s[1] for s in next_states])
+        >>> hn.shape
+        torch.Size([8, 2, 20])
+        >>> cn.shape
+        torch.Size([8, 2, 20])
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        num_layers: int,
+        bias: bool = True,
+        proj_size: int = 0,
+        ln: Type[nn.Module] = nn.LayerNorm,
+        device=None,
+        dtype=None,
+    ):
+        """
+        See the args in LayerNormLSTMLayer.
+        """
+        assert check_argument_types()
+        super().__init__()
+        assert num_layers >= 1
+        factory_kwargs = dict(
+            hidden_size=hidden_size,
+            bias=bias,
+            ln=ln,
+            proj_size=proj_size,
+            device=device,
+            dtype=dtype,
+        )
+        first_layer = LayerNormLSTMLayer(input_size=input_size, **factory_kwargs)
+        layers = [first_layer]
+        for i in range(1, num_layers):
+            layers.append(
+                LayerNormLSTMLayer(
+                    input_size=proj_size if proj_size > 0 else hidden_size,
+                    **factory_kwargs,
+                )
+            )
+        self.layers = nn.ModuleList(layers)
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        states: List[Tuple[torch.Tensor, torch.Tensor]],
+    ) -> Tuple[torch.Tensor, List[Tuple[torch.Tensor, torch.Tensor]]]:
+        """
+        Args:
+          input:
+            A 3-D tensor of shape (batch_size, seq_len, input_size).
+            Caution:
+              We use `batch_first=True` here.
+          states:
+            One state per layer. Each entry contains the hidden state (h, c)
+            for a layer. Both are of shape (batch_size, hidden_size) if
+            proj_size is 0. If proj_size is not 0, the shape of `h` is
+            (batch_size, proj_size).
+        Returns:
+          Return a tuple containing:
+
+            - output: A tensor of shape (batch_size, seq_len, hidden_size)
+            - List[(next_h, next_c)] containing the hidden states for all layers
+
+        """
+        output_states = torch.jit.annotate(List[Tuple[torch.Tensor, torch.Tensor]], [])
+        output = input
+        for i, rnn_layer in enumerate(self.layers):
+            state = states[i]
+            output, out_state = rnn_layer(output, state)
+            output_states += [out_state]
+        return output, output_states
+
+
+class LayerNormGRUCell(nn.Module):
+    """This class places a `nn.LayerNorm` after the output of
+    each gate (right before the activation).
+
+    See the following paper for more details
+
+    'Improving RNN Transducer Modeling for End-to-End Speech Recognition'
+    https://arxiv.org/abs/1909.12415
+
+    Examples::
+
+        >>> cell = LayerNormGRUCell(10, 20)
+        >>> input = torch.rand(2, 10)
+        >>> h0 = torch.rand(2, 20)
+        >>> hn = cell(input, h0)
+        >>> hn.shape
+        torch.Size([2, 20])
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        bias: bool = True,
+        ln: Type[nn.Module] = nn.LayerNorm,
+        device=None,
+        dtype=None,
+    ):
+        """
+        Args:
+          input_size:
+            The number of expected features in the input `x`. `x` should
+            be of shape (batch_size, input_size).
+          hidden_size:
+            The number of features in the hidden state `h` and `c`.
+            Both `h` and `c` are of shape (batch_size, hidden_size).
+          bias:
+            If ``False``, then the cell does not use bias weights
+            `bias_ih` and `bias_hh`.
+          ln:
+            Defaults to `nn.LayerNorm`. The output of all gates are processed
+            by `ln`. We pass it as an argument so that we can replace it
+            with `nn.Identity` at the testing time.
+        """
+        assert check_argument_types()
+        super().__init__()
+        factory_kwargs = {"device": device, "dtype": dtype}
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.bias = bias
+
+        self.weight_ih = nn.Parameter(
+            torch.empty((3 * hidden_size, input_size), **factory_kwargs)
+        )
+
+        self.weight_hh = nn.Parameter(
+            torch.empty((3 * hidden_size, hidden_size), **factory_kwargs)
+        )
+
+        if bias:
+            self.bias_ih = nn.Parameter(torch.empty(3 * hidden_size, **factory_kwargs))
+            self.bias_hh = nn.Parameter(torch.empty(3 * hidden_size, **factory_kwargs))
+        else:
+            self.register_parameter("bias_ih", None)
+            self.register_parameter("bias_hh", None)
+
+        self.layernorm_r = ln(hidden_size)
+        self.layernorm_i = ln(hidden_size)
+        self.layernorm_n = ln(hidden_size)
+
+        self.reset_parameters()
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        hx: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+          input:
+            A 2-D tensor of shape (batch_size, input_size) containing
+            input features.
+          hx:
+            If not `None`, it is a tensor of shape (batch_size, hidden_size)
+            containing the initial hidden state for each element in the batch.
+            If `None`, it uses zeros for the hidden state.
+        Returns:
+          Return a tensor of shape (batch_size, hidden_size) containing the
+          next hidden state for each element in the batch
+        """
+        if hx is None:
+            hx = torch.zeros(
+                input.size(0),
+                self.hidden_size,
+                dtype=input.dtype,
+                device=input.device,
+            )
+
+        i_r, i_i, i_n = F.linear(input, self.weight_ih, self.bias_ih).chunk(
+            chunks=3, dim=1
+        )
+
+        h_r, h_i, h_n = F.linear(hx, self.weight_hh, self.bias_hh).chunk(
+            chunks=3, dim=1
+        )
+
+        reset_gate = torch.sigmoid(self.layernorm_r(i_r + h_r))
+        input_gate = torch.sigmoid(self.layernorm_i(i_i + h_i))
+        new_gate = torch.tanh(self.layernorm_n(i_n + reset_gate * h_n))
+
+        # hy = (1 - input_gate) * new_gate + input_gate * hx
+        #    = new_gate - input_gate * new_gate + input_gate * hx
+        #    = new_gate + input_gate * (hx - new_gate)
+        hy = new_gate + input_gate * (hx - new_gate)
+
+        return hy
+
+    def extra_repr(self) -> str:
+        s = "{input_size}, {hidden_size}"
+        if "bias" in self.__dict__ and self.bias is not True:
+            s += ", bias={bias}"
+        return s.format(**self.__dict__)
+
+    def reset_parameters(self) -> None:
+        stdv = 1.0 / math.sqrt(self.hidden_size)
+        for weight in self.parameters():
+            nn.init.uniform_(weight, -stdv, stdv)
+
+
+class LayerNormGRULayer(nn.Module):
+    """
+    Examples::
+
+        >>> layer = LayerNormGRULayer(10, 20)
+        >>> input = torch.rand(2, 3, 10)
+        >>> hx = torch.rand(2, 20)
+        >>> output, hn = layer(input, hx)
+        >>> output.shape
+        torch.Size([2, 3, 20])
+        >>> hn.shape
+        torch.Size([2, 20])
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        bias: bool = True,
+        ln: Type[nn.Module] = nn.LayerNorm,
+        device=None,
+        dtype=None,
+    ):
+        """
+        See the args in LayerNormGRUCell
+        """
+        assert check_argument_types()
+        super().__init__()
+        self.cell = LayerNormGRUCell(
+            input_size=input_size,
+            hidden_size=hidden_size,
+            bias=bias,
+            ln=ln,
+            device=device,
+            dtype=dtype,
+        )
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        hx: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          input:
+            A 3-D tensor of shape (batch_size, seq_len, input_size).
+            Caution:
+              We use `batch_first=True` here.
+          hx:
+            If not ``None``, it is a tensor of shape (batch_size, hidden_size)
+            containing the hidden state for each element in the batch.
+        Return:
+          - output, a tensor of shape (batch_size, seq_len, hidden_size)
+          - next_h, a tensor of shape (batch_size, hidden_size) containing the
+            final hidden state for each element in the batch.
+        """
+        inputs = input.unbind(1)
+        outputs = torch.jit.annotate(List[torch.Tensor], [])
+        next_h = hx
+        for i in range(len(inputs)):
+            next_h = self.cell(inputs[i], next_h)
+            outputs.append(next_h)
+        return torch.stack(outputs, dim=1), next_h
+
+
+class LayerNormGRU(nn.Module):
+    """
+    Examples::
+
+        >>> input = torch.rand(2, 3, 10)
+        >>> h0 = torch.rand(8, 2, 20)
+        >>> states = h0.unbind(0)
+        >>> output, next_states = gru(input, states)
+        >>> output.shape
+        torch.Size([2, 3, 20])
+        >>> hn = torch.stack(next_states)
+        >>> hn.shape
+        torch.Size([8, 2, 20])
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        num_layers: int,
+        bias: bool = True,
+        ln: Type[nn.Module] = nn.LayerNorm,
+        device=None,
+        dtype=None,
+    ):
+        """
+        See the args in LayerNormGRULayer.
+        """
+        assert check_argument_types()
+        super().__init__()
+        assert num_layers >= 1
+        factory_kwargs = dict(
+            hidden_size=hidden_size,
+            bias=bias,
+            ln=ln,
+            device=device,
+            dtype=dtype,
+        )
+        first_layer = LayerNormGRULayer(input_size=input_size, **factory_kwargs)
+        layers = [first_layer]
+        for i in range(1, num_layers):
+            layers.append(
+                LayerNormGRULayer(
+                    input_size=hidden_size,
+                    **factory_kwargs,
+                )
+            )
+        self.layers = nn.ModuleList(layers)
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          input:
+            A tensor of shape (batch_size, seq_len, input_size) containing
+            input features.
+            Caution:
+              We use `batch_first=True` here.
+          states:
+            One state per layer. Each entry contains the hidden state for each
+            element in the batch. Each hidden state is of shape
+            (batch_size, hidden_size)
+        Returns:
+          Return a tuple containing:
+
+            - output: A tensor of shape (batch_size, seq_len, hidden_size)
+            - List[next_state] containing the final hidden states for each
+              element in the batch
+
+        """
+        output_states = torch.jit.annotate(List[torch.Tensor], [])
+        output = input
+        for i, rnn_layer in enumerate(self.layers):
+            state = states[i]
+            output, out_state = rnn_layer(output, state)
+            output_states += [out_state]
+        return output, output_states
diff --git a/egs/librispeech/ASR/transducer/subsampling.py b/egs/librispeech/ASR/transducer/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer/test_conformer.py b/egs/librispeech/ASR/transducer/test_conformer.py
new file mode 100755
index 000000000..9529e9c59
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/test_conformer.py
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer/test_conformer.py
+"""
+
+import torch
+from conformer import Conformer
+
+
+def test_conformer():
+    output_dim = 1024
+    conformer = Conformer(
+        num_features=80,
+        output_dim=output_dim,
+        subsampling_factor=4,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=12,
+    )
+    N = 3
+    T = 100
+    C = 80
+    x = torch.randn(N, T, C)
+    x_lens = torch.tensor([50, 100, 80])
+    logits, logit_lens = conformer(x, x_lens)
+
+    expected_T = ((T - 1) // 2 - 1) // 2
+    assert logits.shape == (N, expected_T, output_dim)
+    assert logit_lens.max().item() == expected_T
+    print(logits.shape)
+    print(logit_lens)
+
+
+def main():
+    test_conformer()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/test_decoder.py b/egs/librispeech/ASR/transducer/test_decoder.py
new file mode 100755
index 000000000..f0a7aa9cc
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/test_decoder.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    num_layers = 2
+    hidden_dim = 6
+    output_dim = 8
+    N = 3
+    U = 5
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        num_layers=num_layers,
+        hidden_dim=hidden_dim,
+        output_dim=output_dim,
+        embedding_dropout=0.0,
+        rnn_dropout=0.0,
+    )
+    x = torch.randint(1, vocab_size, (N, U))
+    decoder_out, (h, c) = decoder(x)
+
+    assert decoder_out.shape == (N, U, output_dim)
+    assert h.shape == (num_layers, N, hidden_dim)
+    assert c.shape == (num_layers, N, hidden_dim)
+
+    decoder_out, (h, c) = decoder(x, (h, c))
+    assert decoder_out.shape == (N, U, output_dim)
+    assert h.shape == (num_layers, N, hidden_dim)
+    assert c.shape == (num_layers, N, hidden_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/test_joiner.py b/egs/librispeech/ASR/transducer/test_joiner.py
new file mode 100755
index 000000000..23948bbf6
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/test_joiner.py
@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer/test_joiner.py
+"""
+
+
+import torch
+from joiner import Joiner
+
+
+def test_joiner():
+    N = 2
+    T = 3
+    C = 4
+    U = 5
+
+    joiner = Joiner(C, 10)
+
+    encoder_out = torch.rand(N, T, C)
+    decoder_out = torch.rand(N, U, C)
+
+    joint = joiner(encoder_out, decoder_out)
+    assert joint.shape == (N, T, U, 10)
+
+
+def main():
+    test_joiner()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/test_rnn.py b/egs/librispeech/ASR/transducer/test_rnn.py
new file mode 100755
index 000000000..74c94cc70
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/test_rnn.py
@@ -0,0 +1,787 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer/test_rnn.py
+"""
+import torch
+import torch.nn as nn
+from rnn import (
+    LayerNormGRU,
+    LayerNormGRUCell,
+    LayerNormGRULayer,
+    LayerNormLSTM,
+    LayerNormLSTMCell,
+    LayerNormLSTMLayer,
+)
+
+
+def get_devices():
+    devices = [torch.device("cpu")]
+    if torch.cuda.is_available():
+        devices.append(torch.device("cuda", 0))
+    return devices
+
+
+def assert_allclose(a: torch.Tensor, b: torch.Tensor, atol=1e-6, **kwargs):
+    assert torch.allclose(
+        a, b, atol=atol, **kwargs
+    ), f"{(a - b).abs().max()}, {a.numel()}"
+
+
+def test_layernorm_lstm_cell_jit(device="cpu"):
+    input_size = 10
+    hidden_size = 20
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+    cell = LayerNormLSTMCell(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        bias=bias,
+        device=device,
+    )
+
+    torch.jit.script(cell)
+
+
+def test_layernorm_lstm_cell_constructor(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+
+    self_cell = LayerNormLSTMCell(
+        input_size,
+        hidden_size,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_cell = nn.LSTMCell(
+        input_size,
+        hidden_size,
+    ).to(device)
+
+    for name, param in self_cell.named_parameters():
+        assert param.shape == getattr(torch_cell, name).shape
+
+    assert len(self_cell.state_dict()) == len(torch_cell.state_dict())
+
+
+def test_layernorm_lstm_cell_with_projection_jit(device="cpu"):
+    input_size = 10
+    hidden_size = 20
+    proj_size = 5
+    self_cell = LayerNormLSTMCell(
+        input_size,
+        hidden_size,
+        proj_size=proj_size,
+        device=device,
+    )
+    torch.jit.script(self_cell)
+
+
+def test_layernorm_lstm_cell_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+
+    self_cell = LayerNormLSTMCell(
+        input_size,
+        hidden_size,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_cell = nn.LSTMCell(
+        input_size,
+        hidden_size,
+        bias=bias,
+    ).to(device)
+    with torch.no_grad():
+        for name, torch_param in torch_cell.named_parameters():
+            self_param = getattr(self_cell, name)
+            torch_param.copy_(self_param)
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    x = torch.rand(N, input_size, device=device).requires_grad_()
+    h = torch.rand(N, hidden_size, device=device)
+    c = torch.rand(N, hidden_size, device=device)
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_h, self_c = self_cell(x.clone(), (h, c))
+    torch_h, torch_c = torch_cell(x_clone, (h, c))
+
+    assert_allclose(self_h, torch_h)
+    assert_allclose(self_c, torch_c)
+
+    self_hc = self_h * self_c
+    torch_hc = torch_h * torch_c
+    (
+        self_hc.reshape(-1) * torch.arange(self_hc.numel(), device=device)
+    ).sum().backward()
+    (
+        torch_hc.reshape(-1) * torch.arange(torch_hc.numel(), device=device)
+    ).sum().backward()
+
+    assert_allclose(x.grad, x_clone.grad, atol=1e-3)
+
+
+def test_layernorm_lstm_cell_with_projection_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=10, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+    proj_size = torch.randint(low=2, high=hidden_size, size=(1,)).item()
+
+    self_cell = LayerNormLSTMCell(
+        input_size,
+        hidden_size,
+        bias=bias,
+        ln=nn.Identity,
+        proj_size=proj_size,
+        device=device,
+    )
+    torch_cell = nn.LSTM(
+        input_size,
+        hidden_size,
+        bias=bias,
+        proj_size=proj_size,
+        batch_first=True,
+    ).to(device)
+    with torch.no_grad():
+        for name, self_param in self_cell.named_parameters():
+            getattr(torch_cell, f"{name}_l0").copy_(self_param)
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    x = torch.rand(N, input_size, device=device).requires_grad_()
+    h = torch.rand(N, proj_size, device=device)
+    c = torch.rand(N, hidden_size, device=device)
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_h, self_c = self_cell(x.clone(), (h, c))
+    _, (torch_h, torch_c) = torch_cell(
+        x_clone.unsqueeze(1), (h.unsqueeze(0), c.unsqueeze(0))
+    )
+
+    torch_h = torch_h.squeeze(0)
+    torch_c = torch_c.squeeze(0)
+
+    assert_allclose(self_h, torch_h)
+    assert_allclose(self_c, torch_c)
+
+    (self_h.sum() * self_c.sum()).backward()
+    (torch_h.sum() * torch_c.sum()).backward()
+
+    assert_allclose(x.grad, x_clone.grad, atol=1e-5)
+
+
+def test_layernorm_lstm_layer_jit(device="cpu"):
+    input_size = 10
+    hidden_size = 20
+    layer = LayerNormLSTMLayer(
+        input_size,
+        hidden_size=hidden_size,
+        device=device,
+    )
+    torch.jit.script(layer)
+
+
+def test_layernorm_lstm_layer_with_project_jit(device="cpu"):
+    input_size = 10
+    hidden_size = 20
+    proj_size = 5
+    layer = LayerNormLSTMLayer(
+        input_size,
+        hidden_size=hidden_size,
+        proj_size=proj_size,
+        device=device,
+    )
+    torch.jit.script(layer)
+
+
+def test_layernorm_lstm_layer_with_projection_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=10, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+    proj_size = torch.randint(low=2, high=hidden_size, size=(1,)).item()
+
+    self_layer = LayerNormLSTMLayer(
+        input_size,
+        hidden_size,
+        bias=bias,
+        proj_size=proj_size,
+        ln=nn.Identity,
+        device=device,
+    )
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    T = torch.randint(low=2, high=100, size=(1,))
+
+    x = torch.rand(N, T, input_size, device=device).requires_grad_()
+    h = torch.rand(N, proj_size, device=device)
+    c = torch.rand(N, hidden_size, device=device)
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_y, (self_h, self_c) = self_layer(x, (h, c))
+
+    torch_layer = nn.LSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=1,
+        bias=bias,
+        proj_size=proj_size,
+        batch_first=True,
+        dropout=0,
+        bidirectional=False,
+    ).to(device)
+    with torch.no_grad():
+        for name, self_param in self_layer.cell.named_parameters():
+            getattr(torch_layer, f"{name}_l0").copy_(self_param)
+
+    torch_y, (torch_h, torch_c) = torch_layer(x_clone, (h.unsqueeze(0), c.unsqueeze(0)))
+    assert_allclose(self_y, torch_y)
+    assert_allclose(self_h, torch_h)
+    assert_allclose(self_c, torch_c)
+
+    self_y.sum().backward()
+    torch_y.sum().backward()
+
+    assert_allclose(x.grad, x_clone.grad, atol=1e-5)
+
+
+def test_layernorm_lstm_layer_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+    self_layer = LayerNormLSTMLayer(
+        input_size,
+        hidden_size,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    T = torch.randint(low=2, high=100, size=(1,))
+
+    x = torch.rand(N, T, input_size, device=device).requires_grad_()
+    h = torch.rand(N, hidden_size, device=device)
+    c = torch.rand(N, hidden_size, device=device)
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_y, (self_h, self_c) = self_layer(x, (h, c))
+
+    torch_layer = nn.LSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=1,
+        bias=bias,
+        batch_first=True,
+        dropout=0,
+        bidirectional=False,
+    ).to(device)
+    with torch.no_grad():
+        for name, self_param in self_layer.cell.named_parameters():
+            getattr(torch_layer, f"{name}_l0").copy_(self_param)
+
+    torch_y, (torch_h, torch_c) = torch_layer(x_clone, (h.unsqueeze(0), c.unsqueeze(0)))
+    assert_allclose(self_y, torch_y)
+    assert_allclose(self_h, torch_h)
+    assert_allclose(self_c, torch_c)
+
+    self_hc = self_h * self_c
+    torch_hc = torch_h * torch_c
+    self_hc_sum = (
+        self_hc.reshape(-1) * torch.arange(self_hc.numel(), device=device)
+    ).sum()
+    torch_hc_sum = (
+        torch_hc.reshape(-1) * torch.arange(torch_hc.numel(), device=device)
+    ).sum()
+
+    self_y_sum = (
+        self_y.reshape(-1) * torch.arange(self_y.numel(), device=device)
+    ).sum()
+    torch_y_sum = (
+        torch_y.reshape(-1) * torch.arange(torch_y.numel(), device=device)
+    ).sum()
+
+    (self_hc_sum + self_y_sum).backward()
+    (torch_hc_sum + torch_y_sum).backward()
+
+    assert_allclose(x.grad, x_clone.grad, atol=0.1)
+
+
+def test_layernorm_lstm_jit(device="cpu"):
+    input_size = 2
+    hidden_size = 3
+    num_layers = 4
+    bias = True
+
+    lstm = LayerNormLSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch.jit.script(lstm)
+
+
+def test_layernorm_lstm_with_projection_jit(device="cpu"):
+    input_size = 2
+    hidden_size = 5
+    proj_size = 3
+    num_layers = 4
+    bias = True
+
+    lstm = LayerNormLSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        proj_size=proj_size,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch.jit.script(lstm)
+
+
+def test_layernorm_lstm_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+    num_layers = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+
+    self_lstm = LayerNormLSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_lstm = nn.LSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        batch_first=True,
+        bidirectional=False,
+    ).to(device)
+    assert len(self_lstm.state_dict()) == len(torch_lstm.state_dict())
+    with torch.no_grad():
+        for name, param in self_lstm.named_parameters():
+            # name has the form layers.0.cell.weight_hh
+            parts = name.split(".")
+            layer_num = parts[1]
+            getattr(torch_lstm, f"{parts[-1]}_l{layer_num}").copy_(param)
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    T = torch.randint(low=2, high=100, size=(1,))
+
+    x = torch.rand(N, T, input_size, device=device).requires_grad_()
+    hs = [torch.rand(N, hidden_size, device=device) for _ in range(num_layers)]
+    cs = [torch.rand(N, hidden_size, device=device) for _ in range(num_layers)]
+    states = list(zip(hs, cs))
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_y, self_states = self_lstm(x, states)
+
+    h = torch.stack(hs)
+    c = torch.stack(cs)
+    torch_y, (torch_h, torch_c) = torch_lstm(x_clone, (h, c))
+
+    assert_allclose(self_y, torch_y)
+
+    self_h = torch.stack([s[0] for s in self_states])
+    self_c = torch.stack([s[1] for s in self_states])
+
+    assert_allclose(self_h, torch_h)
+    assert_allclose(self_c, torch_c)
+
+    s = self_y.reshape(-1)
+    t = torch_y.reshape(-1)
+
+    s_sum = (s * torch.arange(s.numel(), device=device)).sum()
+    t_sum = (t * torch.arange(t.numel(), device=device)).sum()
+    shc_sum = s_sum + self_h.sum() + self_c.sum()
+    thc_sum = t_sum + torch_h.sum() + torch_c.sum()
+
+    shc_sum.backward()
+    thc_sum.backward()
+    assert_allclose(x.grad, x_clone.grad)
+
+
+def test_layernorm_lstm_with_projection_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=10, high=100, size=(1,)).item()
+    proj_size = torch.randint(low=2, high=hidden_size, size=(1,)).item()
+    num_layers = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+
+    self_lstm = LayerNormLSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        proj_size=proj_size,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_lstm = nn.LSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        proj_size=proj_size,
+        batch_first=True,
+        bidirectional=False,
+    ).to(device)
+    assert len(self_lstm.state_dict()) == len(torch_lstm.state_dict())
+    with torch.no_grad():
+        for name, param in self_lstm.named_parameters():
+            # name has the form layers.0.cell.weight_hh
+            parts = name.split(".")
+            layer_num = parts[1]
+            getattr(torch_lstm, f"{parts[-1]}_l{layer_num}").copy_(param)
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    T = torch.randint(low=2, high=100, size=(1,))
+
+    x = torch.rand(N, T, input_size, device=device).requires_grad_()
+    hs = [torch.rand(N, proj_size, device=device) for _ in range(num_layers)]
+    cs = [torch.rand(N, hidden_size, device=device) for _ in range(num_layers)]
+    states = list(zip(hs, cs))
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_y, self_states = self_lstm(x, states)
+
+    h = torch.stack(hs)
+    c = torch.stack(cs)
+    torch_y, (torch_h, torch_c) = torch_lstm(x_clone, (h, c))
+
+    assert_allclose(self_y, torch_y)
+
+    self_h = torch.stack([s[0] for s in self_states])
+    self_c = torch.stack([s[1] for s in self_states])
+
+    assert_allclose(self_h, torch_h)
+    assert_allclose(self_c, torch_c)
+
+    s = self_y.reshape(-1)
+    t = torch_y.reshape(-1)
+
+    s_sum = (s * torch.arange(s.numel(), device=device)).sum()
+    t_sum = (t * torch.arange(t.numel(), device=device)).sum()
+    shc_sum = s_sum + self_h.sum() + self_c.sum()
+    thc_sum = t_sum + torch_h.sum() + torch_c.sum()
+
+    shc_sum.backward()
+    thc_sum.backward()
+    assert_allclose(x.grad, x_clone.grad)
+
+
+def test_lstm_forget_gate_bias(device="cpu"):
+    input_size = 2
+    hidden_size = 3
+    num_layers = 4
+    bias = True
+
+    lstm = LayerNormLSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    for name, weight in lstm.named_parameters():
+        if "bias_hh" in name or "bias_ih" in name:
+            start = weight.numel() // 4
+            end = weight.numel() // 2
+            expected = torch.ones(hidden_size).to(weight)
+            assert torch.all(torch.eq(weight[start:end], expected))
+
+
+def test_layernorm_gru_cell_jit(device="cpu"):
+    input_size = 10
+    hidden_size = 20
+    cell = LayerNormGRUCell(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        bias=True,
+        device=device,
+    )
+
+    torch.jit.script(cell)
+
+
+def test_layernorm_gru_cell_constructor(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+
+    self_cell = LayerNormGRUCell(
+        input_size,
+        hidden_size,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_cell = nn.GRUCell(
+        input_size,
+        hidden_size,
+    ).to(device)
+
+    for name, param in self_cell.named_parameters():
+        assert param.shape == getattr(torch_cell, name).shape
+
+    assert len(self_cell.state_dict()) == len(torch_cell.state_dict())
+
+
+def test_layernorm_gru_cell_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+
+    self_cell = LayerNormGRUCell(
+        input_size,
+        hidden_size,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_cell = nn.GRUCell(
+        input_size,
+        hidden_size,
+        bias=bias,
+    ).to(device)
+    with torch.no_grad():
+        for name, torch_param in torch_cell.named_parameters():
+            self_param = getattr(self_cell, name)
+            torch_param.copy_(self_param)
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    x = torch.rand(N, input_size, device=device).requires_grad_()
+    h = torch.rand(N, hidden_size, device=device)
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_h = self_cell(x.clone(), h)
+    torch_h = torch_cell(x_clone, h)
+
+    assert_allclose(self_h, torch_h, atol=1e-5)
+
+    (self_h.reshape(-1) * torch.arange(self_h.numel(), device=device)).sum().backward()
+    (
+        torch_h.reshape(-1) * torch.arange(torch_h.numel(), device=device)
+    ).sum().backward()
+
+    assert_allclose(x.grad, x_clone.grad, atol=1e-3)
+
+
+def test_layernorm_gru_layer_jit(device="cpu"):
+    input_size = 10
+    hidden_size = 20
+    layer = LayerNormGRULayer(
+        input_size,
+        hidden_size=hidden_size,
+        device=device,
+    )
+    torch.jit.script(layer)
+
+
+def test_layernorm_gru_layer_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+    self_layer = LayerNormGRULayer(
+        input_size,
+        hidden_size,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    T = torch.randint(low=2, high=100, size=(1,))
+
+    x = torch.rand(N, T, input_size, device=device).requires_grad_()
+    h = torch.rand(N, hidden_size, device=device)
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_y, self_h = self_layer(x, h.clone())
+
+    torch_layer = nn.GRU(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=1,
+        bias=bias,
+        batch_first=True,
+        dropout=0,
+        bidirectional=False,
+    ).to(device)
+    with torch.no_grad():
+        for name, self_param in self_layer.cell.named_parameters():
+            getattr(torch_layer, f"{name}_l0").copy_(self_param)
+
+    torch_y, torch_h = torch_layer(x_clone, h.unsqueeze(0))
+    assert_allclose(self_y, torch_y)
+    assert_allclose(self_h, torch_h)
+
+    self_y_sum = (
+        self_y.reshape(-1) * torch.arange(self_y.numel(), device=device)
+    ).sum()
+    torch_y_sum = (
+        torch_y.reshape(-1) * torch.arange(torch_y.numel(), device=device)
+    ).sum()
+
+    self_y_sum.backward()
+    torch_y_sum.backward()
+
+    assert_allclose(x.grad, x_clone.grad, atol=0.1)
+
+
+def test_layernorm_gru_jit(device="cpu"):
+    input_size = 2
+    hidden_size = 3
+    num_layers = 4
+    bias = True
+
+    gru = LayerNormGRU(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch.jit.script(gru)
+
+
+def test_layernorm_gru_forward(device="cpu"):
+    input_size = torch.randint(low=2, high=100, size=(1,)).item()
+    hidden_size = torch.randint(low=2, high=100, size=(1,)).item()
+    num_layers = torch.randint(low=2, high=100, size=(1,)).item()
+    bias = torch.randint(low=0, high=1000, size=(1,)).item() & 2 == 0
+
+    self_gru = LayerNormGRU(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        ln=nn.Identity,
+        device=device,
+    )
+    torch_gru = nn.GRU(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        bias=bias,
+        batch_first=True,
+        bidirectional=False,
+    ).to(device)
+    assert len(self_gru.state_dict()) == len(torch_gru.state_dict())
+    with torch.no_grad():
+        for name, param in self_gru.named_parameters():
+            # name has the form layers.0.cell.weight_hh
+            parts = name.split(".")
+            layer_num = parts[1]
+            getattr(torch_gru, f"{parts[-1]}_l{layer_num}").copy_(param)
+
+    N = torch.randint(low=2, high=100, size=(1,))
+    T = torch.randint(low=2, high=100, size=(1,))
+
+    x = torch.rand(N, T, input_size, device=device).requires_grad_()
+    states = [torch.rand(N, hidden_size, device=device) for _ in range(num_layers)]
+
+    x_clone = x.detach().clone().requires_grad_()
+
+    self_y, self_states = self_gru(x, states)
+
+    torch_y, torch_states = torch_gru(x_clone, torch.stack(states))
+
+    assert_allclose(self_y, torch_y)
+
+    self_states = torch.stack(self_states)
+
+    assert_allclose(self_states, torch_states)
+
+    s = self_y.reshape(-1)
+    t = torch_y.reshape(-1)
+
+    s_sum = (s * torch.arange(s.numel(), device=device)).sum()
+    t_sum = (t * torch.arange(t.numel(), device=device)).sum()
+    s_state_sum = s_sum + self_states.sum()
+    t_state_sum = t_sum + torch_states.sum()
+
+    s_state_sum.backward()
+    t_state_sum.backward()
+    assert_allclose(x.grad, x_clone.grad, atol=1e-2)
+
+
+def _test_lstm(device):
+    test_layernorm_lstm_cell_jit(device)
+    test_layernorm_lstm_cell_constructor(device)
+    test_layernorm_lstm_cell_with_projection_jit(device)
+    test_layernorm_lstm_cell_forward(device)
+    test_layernorm_lstm_cell_with_projection_forward(device)
+    #
+    test_layernorm_lstm_layer_jit(device)
+    test_layernorm_lstm_layer_with_project_jit(device)
+    test_layernorm_lstm_layer_forward(device)
+    test_layernorm_lstm_layer_with_projection_forward(device)
+
+    test_layernorm_lstm_jit(device)
+    test_layernorm_lstm_with_projection_jit(device)
+    test_layernorm_lstm_forward(device)
+    test_layernorm_lstm_with_projection_forward(device)
+    #
+    test_lstm_forget_gate_bias(device)
+
+
+def _test_gru(device):
+    test_layernorm_gru_cell_jit(device)
+    test_layernorm_gru_cell_constructor(device)
+    test_layernorm_gru_cell_forward(device)
+    #
+    test_layernorm_gru_layer_jit(device)
+    test_layernorm_gru_layer_forward(device)
+    #
+    test_layernorm_gru_jit(device)
+    test_layernorm_gru_forward(device)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def main():
+    for device in get_devices():
+        print("device", device)
+        _test_lstm(device)
+        _test_gru(device)
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20211202)
+    main()
diff --git a/egs/librispeech/ASR/transducer/test_transducer.py b/egs/librispeech/ASR/transducer/test_transducer.py
new file mode 100755
index 000000000..15aa3b330
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/test_transducer.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer/test_transducer.py
+"""
+
+
+import k2
+import torch
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+
+def test_transducer():
+    # encoder params
+    input_dim = 10
+    output_dim = 20
+
+    # decoder params
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    num_layers = 2
+
+    encoder = Conformer(
+        num_features=input_dim,
+        output_dim=output_dim,
+        subsampling_factor=4,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=12,
+    )
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        num_layers=num_layers,
+        hidden_dim=output_dim,
+        output_dim=output_dim,
+        embedding_dropout=0.0,
+        rnn_dropout=0.0,
+    )
+
+    joiner = Joiner(output_dim, vocab_size)
+    transducer = Transducer(encoder=encoder, decoder=decoder, joiner=joiner)
+
+    y = k2.RaggedTensor([[1, 2, 1], [1, 1, 1, 2, 1]])
+    N = y.dim0
+    T = 50
+
+    x = torch.rand(N, T, input_dim)
+    x_lens = torch.randint(low=30, high=T, size=(N,), dtype=torch.int32)
+    x_lens[0] = T
+
+    loss = transducer(x, x_lens, y)
+    print(loss)
+
+
+def main():
+    test_transducer()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/test_transformer.py b/egs/librispeech/ASR/transducer/test_transformer.py
new file mode 100755
index 000000000..bb68c22be
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/test_transformer.py
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer/test_transformer.py
+"""
+
+import torch
+from transformer import Transformer
+
+
+def test_transformer():
+    output_dim = 1024
+    transformer = Transformer(
+        num_features=80,
+        output_dim=output_dim,
+        subsampling_factor=4,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=12,
+    )
+    N = 3
+    T = 100
+    C = 80
+    x = torch.randn(N, T, C)
+    x_lens = torch.tensor([50, 100, 80])
+    logits, logit_lens = transformer(x, x_lens)
+
+    expected_T = ((T - 1) // 2 - 1) // 2
+    assert logits.shape == (N, expected_T, output_dim)
+    assert logit_lens.max().item() == expected_T
+    print(logits.shape)
+    print(logit_lens)
+
+
+def main():
+    test_transformer()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/train.py b/egs/librispeech/ASR/transducer/train.py
new file mode 100755
index 000000000..9ac6b7d03
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/train.py
@@ -0,0 +1,739 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer/train.py \
+  --world-size 4 \
+  --num-epochs 35 \
+  --start-epoch 0 \
+  --exp-dir transducer/exp \
+  --full-libri 1 \
+  --max-duration 250 \
+  --lr-factor 2.5
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=35,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=3.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # decoder params
+            "decoder_embedding_dim": 1024,
+            "num_decoder_layers": 2,
+            "decoder_hidden_dim": 512,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.decoder_embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.decoder_hidden_dim,
+        output_dim=params.encoder_out_dim,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(x=feature, x_lens=feature_lens, y=y)
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 8000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        return 1.0 <= c.duration <= 20.0
+
+    num_in_total = len(train_cuts)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer/transformer.py b/egs/librispeech/ASR/transducer/transformer.py
new file mode 120000
index 000000000..e43f520f9
--- /dev/null
+++ b/egs/librispeech/ASR/transducer/transformer.py
@@ -0,0 +1 @@
+../transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_lstm/README.md b/egs/librispeech/ASR/transducer_lstm/README.md
new file mode 100644
index 000000000..38c3d2bfd
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/README.md
@@ -0,0 +1,19 @@
+## Introduction
+
+The encoder consists of LSTM layers in this folder. You can use the
+following command to start the training:
+
+```bash
+cd egs/librispeech/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_lstm/train.py \
+  --world-size 3 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_lstm/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --lr-factor 3
+```
diff --git a/egs/librispeech/ASR/transducer_lstm/__init__.py b/egs/librispeech/ASR/transducer_lstm/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/transducer_lstm/asr_datamodule.py b/egs/librispeech/ASR/transducer_lstm/asr_datamodule.py
new file mode 120000
index 000000000..07f39b451
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_lstm/beam_search.py b/egs/librispeech/ASR/transducer_lstm/beam_search.py
new file mode 100644
index 000000000..5342c3e8c
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/beam_search.py
@@ -0,0 +1,218 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+import torch
+from model import Transducer
+
+
+def greedy_search(model: Transducer, encoder_out: torch.Tensor) -> List[int]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    device = model.device
+
+    sos = torch.tensor([blank_id], device=device, dtype=torch.int64).reshape(1, 1)
+    decoder_out, (h, c) = model.decoder(sos)
+    T = encoder_out.size(1)
+    t = 0
+    hyp = []
+
+    sym_per_frame = 0
+    sym_per_utt = 0
+
+    max_sym_per_utt = 1000
+    max_sym_per_frame = 3
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(current_encoder_out, decoder_out)
+        # logits is (1, 1, 1, vocab_size)
+
+        log_prob = logits.log_softmax(dim=-1)
+        # log_prob is (1, 1, 1, vocab_size)
+        # TODO: Use logits.argmax()
+        y = log_prob.argmax()
+        if y != blank_id:
+            hyp.append(y.item())
+            y = y.reshape(1, 1)
+            decoder_out, (h, c) = model.decoder(y, (h, c))
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+
+        if y == blank_id or sym_per_frame > max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+
+    return hyp
+
+
+@dataclass
+class Hypothesis:
+    ys: List[int]  # the predicted sequences so far
+    log_prob: float  # The log prob of ys
+
+    # Optional decoder state. We assume it is LSTM for now,
+    # so the state is a tuple (h, c)
+    decoder_state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 5,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    sos_id = model.decoder.sos_id
+    device = model.device
+
+    sos = torch.tensor([blank_id], device=device).reshape(1, 1)
+    decoder_out, (h, c) = model.decoder(sos)
+    T = encoder_out.size(1)
+    t = 0
+    B = [Hypothesis(ys=[blank_id], log_prob=0.0, decoder_state=None)]
+    max_u = 20000  # terminate after this number of steps
+    u = 0
+
+    cache: Dict[str, Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = {}
+
+    while t < T and u < max_u:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        A = B
+        B = []
+        #  for hyp in A:
+        #      for h in A:
+        #          if h.ys == hyp.ys[:-1]:
+        #              # update the score of hyp
+        #              decoder_input = torch.tensor(
+        #                  [h.ys[-1]], device=device
+        #              ).reshape(1, 1)
+        #              decoder_out, _ = model.decoder(
+        #                  decoder_input, h.decoder_state
+        #              )
+        #              logits = model.joiner(current_encoder_out, decoder_out)
+        #              log_prob = logits.log_softmax(dim=-1)
+        #              log_prob = log_prob.squeeze()
+        #              hyp.log_prob += h.log_prob + log_prob[hyp.ys[-1]].item()
+
+        while u < max_u:
+            y_star = max(A, key=lambda hyp: hyp.log_prob)
+            A.remove(y_star)
+
+            # Note: y_star.ys is unhashable, i.e., cannot be used
+            # as a key into a dict
+            cached_key = "_".join(map(str, y_star.ys))
+
+            if cached_key not in cache:
+                decoder_input = torch.tensor([y_star.ys[-1]], device=device).reshape(
+                    1, 1
+                )
+
+                decoder_out, decoder_state = model.decoder(
+                    decoder_input,
+                    y_star.decoder_state,
+                )
+                cache[cached_key] = (decoder_out, decoder_state)
+            else:
+                decoder_out, decoder_state = cache[cached_key]
+
+            logits = model.joiner(current_encoder_out, decoder_out)
+            log_prob = logits.log_softmax(dim=-1)
+            # log_prob is (1, 1, 1, vocab_size)
+            log_prob = log_prob.squeeze()
+            # Now log_prob is (vocab_size,)
+
+            # If we choose blank here, add the new hypothesis to B.
+            # Otherwise, add the new hypothesis to A
+
+            # First, choose blank
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob.item()
+
+            # ys[:] returns a copy of ys
+            new_y_star = Hypothesis(
+                ys=y_star.ys[:],
+                log_prob=new_y_star_log_prob,
+                # Caution: Use y_star.decoder_state here
+                decoder_state=y_star.decoder_state,
+            )
+            B.append(new_y_star)
+
+            # Second, choose other labels
+            for i, v in enumerate(log_prob.tolist()):
+                if i in (blank_id, sos_id):
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                new_hyp = Hypothesis(
+                    ys=new_ys,
+                    log_prob=new_log_prob,
+                    decoder_state=decoder_state,
+                )
+                A.append(new_hyp)
+            u += 1
+            # check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = max(A, key=lambda hyp: hyp.log_prob)
+            B = sorted(
+                [hyp for hyp in B if hyp.log_prob > A_most_probable.log_prob],
+                key=lambda hyp: hyp.log_prob,
+                reverse=True,
+            )
+            if len(B) >= beam:
+                B = B[:beam]
+                break
+        t += 1
+    best_hyp = max(B, key=lambda hyp: hyp.log_prob / len(hyp.ys[1:]))
+    ys = best_hyp.ys[1:]  # [1:] to remove the blank
+    return ys
diff --git a/egs/librispeech/ASR/transducer_lstm/decode.py b/egs/librispeech/ASR/transducer_lstm/decode.py
new file mode 100755
index 000000000..806b68f40
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/decode.py
@@ -0,0 +1,448 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_lstm/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_lstm/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+(2) beam search
+
+./transducer_lstm/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_lstm/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import beam_search, greedy_search
+from decoder import Decoder
+from encoder import LstmEncoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=77,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_lstm/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=5,
+        help="Used only when --decoding-method is beam_search",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "encoder_hidden_size": 1024,
+            "num_encoder_layers": 4,
+            "proj_size": 512,
+            "vgg_frontend": False,
+            # decoder params
+            "decoder_embedding_dim": 1024,
+            "num_decoder_layers": 4,
+            "decoder_hidden_dim": 512,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    encoder = LstmEncoder(
+        num_features=params.feature_dim,
+        hidden_size=params.encoder_hidden_size,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.decoder_embedding_dim,
+        blank_id=params.blank_id,
+        sos_id=params.sos_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.decoder_hidden_dim,
+        output_dim=params.encoder_out_dim,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.decoding_method == "greedy_search":
+            hyp = greedy_search(model=model, encoder_out=encoder_out_i)
+        elif params.decoding_method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+        hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in ("greedy_search", "beam_search")
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if params.decoding_method == "beam_search":
+        params.suffix += f"-beam-{params.beam_size}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.sos_id = sp.piece_to_id("<sos/eos>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_lstm/decoder.py b/egs/librispeech/ASR/transducer_lstm/decoder.py
new file mode 100644
index 000000000..4d531bde1
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/decoder.py
@@ -0,0 +1,101 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+
+# TODO(fangjun): Support switching between LSTM and GRU
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        sos_id: int,
+        num_layers: int,
+        hidden_dim: int,
+        output_dim: int,
+        embedding_dropout: float = 0.0,
+        rnn_dropout: float = 0.0,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          sos_id:
+            The ID of the SOS symbol.
+          num_layers:
+            Number of LSTM layers.
+          hidden_dim:
+            Hidden dimension of LSTM layers.
+          output_dim:
+            Output dimension of the decoder.
+          embedding_dropout:
+            Dropout rate for the embedding layer.
+          rnn_dropout:
+            Dropout for LSTM layers.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.embedding_dropout = nn.Dropout(embedding_dropout)
+        # TODO(fangjun): Use layer normalized LSTM
+        self.rnn = nn.LSTM(
+            input_size=embedding_dim,
+            hidden_size=hidden_dim,
+            num_layers=num_layers,
+            batch_first=True,
+            dropout=rnn_dropout,
+        )
+        self.blank_id = blank_id
+        self.sos_id = sos_id
+        self.output_linear = nn.Linear(hidden_dim, output_dim)
+
+    def forward(
+        self,
+        y: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U) with BOS prepended.
+          states:
+            A tuple of two tensors containing the states information of
+            LSTM layers in this decoder.
+        Returns:
+          Return a tuple containing:
+
+            - rnn_output, a tensor of shape (N, U, C)
+            - (h, c), containing the state information for LSTM layers.
+              Both are of shape (num_layers, N, C)
+        """
+        embedding_out = self.embedding(y)
+        embedding_out = self.embedding_dropout(embedding_out)
+        rnn_out, (h, c) = self.rnn(embedding_out, states)
+        out = self.output_linear(rnn_out)
+
+        return out, (h, c)
diff --git a/egs/librispeech/ASR/transducer_lstm/encoder.py b/egs/librispeech/ASR/transducer_lstm/encoder.py
new file mode 100644
index 000000000..038d80077
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/encoder.py
@@ -0,0 +1,113 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
+
+
+class LstmEncoder(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        hidden_size: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        vgg_frontend: bool = False,
+        proj_size: int = 0,
+    ):
+        super().__init__()
+        real_hidden_size = proj_size if proj_size > 0 else hidden_size
+        assert (
+            subsampling_factor == 4
+        ), "Only subsampling_factor==4 is supported at present"
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, real_hidden_size)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, real_hidden_size)
+
+        self.rnn = nn.LSTM(
+            input_size=hidden_size,
+            hidden_size=hidden_size,
+            num_layers=num_encoder_layers,
+            bias=True,
+            proj_size=proj_size,
+            batch_first=True,
+            dropout=dropout,
+            bidirectional=False,
+        )
+
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout),
+            nn.Linear(real_hidden_size, output_dim),
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(1) == lengths.max().item(), (
+            x.size(1),
+            lengths.max(),
+        )
+
+        if False:
+            # It is commented out as DDP complains that not all parameters are
+            # used. Need more checks later for the reason.
+            #
+            # Caution: We assume the dataloader returns utterances with
+            # duration being sorted in decreasing order
+            packed_x = pack_padded_sequence(
+                input=x,
+                lengths=lengths.cpu(),
+                batch_first=True,
+                enforce_sorted=True,
+            )
+
+            packed_rnn_out, _ = self.rnn(packed_x)
+            rnn_out, _ = pad_packed_sequence(packed_rnn_out, batch_first=True)
+        else:
+            rnn_out, _ = self.rnn(x)
+
+        logits = self.encoder_output_layer(rnn_out)
+        return logits, lengths
diff --git a/egs/librispeech/ASR/transducer_lstm/encoder_interface.py b/egs/librispeech/ASR/transducer_lstm/encoder_interface.py
new file mode 100644
index 000000000..257facce4
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/encoder_interface.py
@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")
diff --git a/egs/librispeech/ASR/transducer_lstm/joiner.py b/egs/librispeech/ASR/transducer_lstm/joiner.py
new file mode 100644
index 000000000..0422f8a6f
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/joiner.py
@@ -0,0 +1,55 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, C).
+        Returns:
+          Return a tensor of shape (N, T, U, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == decoder_out.size(2)
+
+        encoder_out = encoder_out.unsqueeze(2)
+        # Now encoder_out is (N, T, 1, C)
+
+        decoder_out = decoder_out.unsqueeze(1)
+        # Now decoder_out is (N, 1, U, C)
+
+        logit = encoder_out + decoder_out
+        logit = F.relu(logit)
+
+        output = self.output_linear(logit)
+
+        return output
diff --git a/egs/librispeech/ASR/transducer_lstm/model.py b/egs/librispeech/ASR/transducer_lstm/model.py
new file mode 100644
index 000000000..e37558a98
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/model.py
@@ -0,0 +1,127 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Note we use `rnnt_loss` from torchaudio, which exists only in
+torchaudio >= v0.10.0. It also means you have to use torch >= v1.10.0
+"""
+import k2
+import torch
+import torch.nn as nn
+import torchaudio
+import torchaudio.functional
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            two attributes: `blank_id` and `sos_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface)
+        assert hasattr(decoder, "blank_id")
+        assert hasattr(decoder, "sos_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        decoder_out, _ = self.decoder(sos_y_padded)
+
+        logits = self.joiner(encoder_out, decoder_out)
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        assert hasattr(torchaudio.functional, "rnnt_loss"), (
+            f"Current torchaudio version: {torchaudio.__version__}\n"
+            "Please install a version >= 0.10.0"
+        )
+
+        loss = torchaudio.functional.rnnt_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+        )
+
+        return loss
diff --git a/egs/librispeech/ASR/transducer_lstm/noam.py b/egs/librispeech/ASR/transducer_lstm/noam.py
new file mode 100644
index 000000000..e46bf35fb
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/noam.py
@@ -0,0 +1,104 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
diff --git a/egs/librispeech/ASR/transducer_lstm/subsampling.py b/egs/librispeech/ASR/transducer_lstm/subsampling.py
new file mode 120000
index 000000000..73068da26
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/subsampling.py
@@ -0,0 +1 @@
+../transducer/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_lstm/test_encoder.py b/egs/librispeech/ASR/transducer_lstm/test_encoder.py
new file mode 100755
index 000000000..cad5f1148
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/test_encoder.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+#
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_lstm/test_encoder.py
+"""
+
+from encoder import LstmEncoder
+
+
+def test_encoder():
+    encoder = LstmEncoder(
+        num_features=80,
+        hidden_size=1024,
+        proj_size=512,
+        output_dim=512,
+        subsampling_factor=4,
+        num_encoder_layers=12,
+    )
+    num_params = sum(p.numel() for p in encoder.parameters() if p.requires_grad)
+    print(num_params)
+    # 93979284
+    # 66427392
+
+
+def main():
+    test_encoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_lstm/train.py b/egs/librispeech/ASR/transducer_lstm/train.py
new file mode 100755
index 000000000..92134116c
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_lstm/train.py
@@ -0,0 +1,743 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2"
+
+./transducer_lstm/train.py \
+  --world-size 3 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_lstm/exp \
+  --full-libri 1 \
+  --max-duration 400 \
+  --lr-factor 3
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from encoder import LstmEncoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from noam import Noam
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_lstm/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_lstm/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=3.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Whether to do batch normalization for the
+                              input features.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "encoder_hidden_size": 1024,
+            "num_encoder_layers": 4,
+            "proj_size": 512,
+            "vgg_frontend": False,
+            # decoder params
+            "decoder_embedding_dim": 1024,
+            "num_decoder_layers": 4,
+            "decoder_hidden_dim": 512,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    encoder = LstmEncoder(
+        num_features=params.feature_dim,
+        hidden_size=params.encoder_hidden_size,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.decoder_embedding_dim,
+        blank_id=params.blank_id,
+        sos_id=params.sos_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.decoder_hidden_dim,
+        output_dim=params.encoder_out_dim,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(x=feature, x_lens=feature_lens, y=y)
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 8000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <sos/eos> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.sos_id = sp.piece_to_id("<sos/eos>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    num_param = sum([p.numel() for p in model.parameters() if p.requires_grad])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.encoder_hidden_size,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/README.md b/egs/librispeech/ASR/transducer_stateless/README.md
new file mode 100644
index 000000000..978fa2ada
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/README.md
@@ -0,0 +1,139 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+You can use the following command to start the training:
+
+```bash
+cd egs/librispeech/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 250 \
+  --lr-factor 2.5
+```
+
+## How to get framewise token alignment
+
+Assume that you already have a trained model. If not, you can either
+train one by yourself or download a pre-trained model from hugging face:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-multi-datasets-bpe-500-2022-03-01>
+
+**Caution**: If you are going to use your own trained model, remember
+to set `--modified-transducer-prob` to a nonzero value since the
+force alignment code assumes that `--max-sym-per-frame` is 1.
+
+
+The following shows how to get framewise token alignment using the above
+pre-trained model.
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall/egs/librispeech/ASR
+mkdir tmp
+sudo apt-get install git-lfs
+git lfs install
+git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-multi-datasets-bpe-500-2022-03-01 ./tmp/
+
+ln -s $PWD/tmp/exp/pretrained.pt $PWD/tmp/epoch-999.pt
+
+./transducer_stateless/compute_ali.py \
+        --exp-dir ./tmp/exp \
+        --bpe-model ./tmp/data/lang_bpe_500/bpe.model \
+        --epoch 999 \
+        --avg 1 \
+        --max-duration 100 \
+        --dataset dev-clean \
+        --out-dir data/ali
+```
+
+After running the above commands, you will find the following two files
+in the folder `./data/ali`:
+
+```
+-rw-r--r-- 1 xxx xxx 412K Mar  7 15:45 cuts_dev-clean.json.gz
+-rw-r--r-- 1 xxx xxx 2.9M Mar  7 15:45 token_ali_dev-clean.h5
+```
+
+You can find usage examples in `./test_compute_ali.py` about
+extracting framewise token alignment information from the above
+two files.
+
+## How to get word starting time from framewise token alignment
+
+Assume you have run the above commands to get framewise token alignment
+using a pre-trained model from `tmp/exp/epoch-999.pt`. You can use the following
+commands to obtain word starting time.
+
+```bash
+./transducer_stateless/test_compute_ali.py \
+        --bpe-model ./tmp/data/lang_bpe_500/bpe.model \
+        --ali-dir data/ali \
+        --dataset dev-clean
+```
+
+**Caution**: Since the frame shift is 10ms and the subsampling factor
+of the model is 4, the time resolution is 0.04 second.
+
+**Note**: The script `test_compute_ali.py` is for illustration only
+and it processes only one batch and then exits.
+
+You will get the following output:
+
+```
+5694-64029-0022-1998-0
+[('THE', '0.20'), ('LEADEN', '0.36'), ('HAIL', '0.72'), ('STORM', '1.00'), ('SWEPT', '1.48'), ('THEM', '1.88'), ('OFF', '2.00'), ('THE', '2.24'), ('FIELD', '2.36'), ('THEY', '3.20'), ('FELL', '3.36'), ('BACK', '3.64'), ('AND', '3.92'), ('RE', '4.04'), ('FORMED', '4.20')]
+
+3081-166546-0040-308-0
+[('IN', '0.32'), ('OLDEN', '0.60'), ('DAYS', '1.00'), ('THEY', '1.40'), ('WOULD', '1.56'), ('HAVE', '1.76'), ('SAID', '1.92'), ('STRUCK', '2.60'), ('BY', '3.16'), ('A', '3.36'), ('BOLT', '3.44'), ('FROM', '3.84'), ('HEAVEN', '4.04')]
+
+2035-147960-0016-1283-0
+[('A', '0.44'), ('SNAKE', '0.52'), ('OF', '0.84'), ('HIS', '0.96'), ('SIZE', '1.12'), ('IN', '1.60'), ('FIGHTING', '1.72'), ('TRIM', '2.12'), ('WOULD', '2.56'), ('BE', '2.76'), ('MORE', '2.88'), ('THAN', '3.08'), ('ANY', '3.28'), ('BOY', '3.56'), ('COULD', '3.88'), ('HANDLE', '4.04')]
+
+2428-83699-0020-1734-0
+[('WHEN', '0.28'), ('THE', '0.48'), ('TRAP', '0.60'), ('DID', '0.88'), ('APPEAR', '1.08'), ('IT', '1.80'), ('LOOKED', '1.96'), ('TO',
+'2.24'), ('ME', '2.36'), ('UNCOMMONLY', '2.52'), ('LIKE', '3.16'), ('AN', '3.40'), ('OPEN', '3.56'), ('SPRING', '3.92'), ('CART', '4.28')]
+
+8297-275154-0026-2108-0
+[('LET', '0.44'), ('ME', '0.72'), ('REST', '0.92'), ('A', '1.32'), ('LITTLE', '1.40'), ('HE', '1.80'), ('PLEADED', '2.00'), ('IF', '3.04'), ("I'M", '3.28'), ('NOT', '3.52'), ('IN', '3.76'), ('THE', '3.88'), ('WAY', '4.00')]
+
+652-129742-0007-1002-0
+[('SURROUND', '0.28'), ('WITH', '0.80'), ('A', '0.92'), ('GARNISH', '1.00'), ('OF', '1.44'), ('COOKED', '1.56'), ('AND', '1.88'), ('DICED', '4.16'), ('CARROTS', '4.28'), ('TURNIPS', '4.44'), ('GREEN', '4.60'), ('PEAS', '4.72')]
+```
+
+
+For the row:
+```
+5694-64029-0022-1998-0
+[('THE', '0.20'), ('LEADEN', '0.36'), ('HAIL', '0.72'), ('STORM', '1.00'), ('SWEPT', '1.48'),
+('THEM', '1.88'), ('OFF', '2.00'), ('THE', '2.24'), ('FIELD', '2.36'), ('THEY', '3.20'), ('FELL', '3.36'),
+('BACK', '3.64'), ('AND', '3.92'), ('RE', '4.04'), ('FORMED', '4.20')]
+```
+
+- `5694-64029-0022-1998-0` is the cut ID.
+- `('THE', '0.20')` means the word `THE` starts at 0.20 second.
+- `('LEADEN', '0.36')` means the word `LEADEN` starts at 0.36 second.
+
+
+You can compare the above word starting time with the one
+from <https://github.com/CorentinJ/librispeech-alignments>
+
+```
+5694-64029-0022 ",THE,LEADEN,HAIL,STORM,SWEPT,THEM,OFF,THE,FIELD,,THEY,FELL,BACK,AND,RE,FORMED," "0.230,0.360,0.670,1.010,1.440,1.860,1.990,2.230,2.350,2.870,3.230,3.390,3.660,3.960,4.060,4.160,4.850,4.9"
+```
+
+We reformat it below for readability:
+
+```
+5694-64029-0022 ",THE,LEADEN,HAIL,STORM,SWEPT,THEM,OFF,THE,FIELD,,THEY,FELL,BACK,AND,RE,FORMED,"
+"0.230,0.360,0.670,1.010,1.440,1.860,1.990,2.230,2.350,2.870,3.230,3.390,3.660,3.960,4.060,4.160,4.850,4.9"
+  the  leaden hail storm swept them  off   the   field  sil   they  fell  back  and   re   formed  sil
+```
diff --git a/egs/librispeech/ASR/transducer_stateless/__init__.py b/egs/librispeech/ASR/transducer_stateless/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/transducer_stateless/alignment.py b/egs/librispeech/ASR/transducer_stateless/alignment.py
new file mode 100644
index 000000000..65f2c58d8
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/alignment.py
@@ -0,0 +1,266 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Iterator, List, Optional
+
+import sentencepiece as spm
+import torch
+from model import Transducer
+
+# The force alignment problem can be formulated as finding
+# a path in a rectangular lattice, where the path starts
+# from the lower left corner and ends at the upper right
+# corner. The horizontal axis of the lattice is `t` (representing
+# acoustic frame indexes) and the vertical axis is `u` (representing
+# BPE tokens of the transcript).
+#
+# The notations `t` and `u` are from the paper
+# https://arxiv.org/pdf/1211.3711.pdf
+#
+# Beam search is used to find the path with the
+# highest log probabilities.
+#
+# It assumes the maximum number of symbols that can be
+# emitted per frame is 1. You can use `--modified-transducer-prob`
+# from `./train.py` to train a model that satisfies this assumption.
+
+
+# AlignItem is the ending node of a path originated from the starting node.
+# len(ys) equals to `t` and pos_u is the u coordinate
+# in the lattice.
+@dataclass
+class AlignItem:
+    # total log prob of the path that ends at this item.
+    # The path is originated from the starting node.
+    log_prob: float
+
+    # It contains framewise token alignment
+    ys: List[int]
+
+    # It equals to the number of non-zero entries in ys
+    pos_u: int
+
+
+class AlignItemList:
+    def __init__(self, items: Optional[List[AlignItem]] = None):
+        """
+        Args:
+          items:
+            A list of AlignItem
+        """
+        if items is None:
+            items = []
+        self.data = items
+
+    def __iter__(self) -> Iterator:
+        return iter(self.data)
+
+    def __len__(self) -> int:
+        """Return the number of AlignItem in this object."""
+        return len(self.data)
+
+    def __getitem__(self, i: int) -> AlignItem:
+        """Return the i-th item in this object."""
+        return self.data[i]
+
+    def append(self, item: AlignItem) -> None:
+        """Append an item to the end of this object."""
+        self.data.append(item)
+
+    def get_decoder_input(
+        self,
+        ys: List[int],
+        context_size: int,
+        blank_id: int,
+    ) -> List[List[int]]:
+        """Get input for the decoder for each item in this object.
+
+        Args:
+          ys:
+            The transcript of the utterance in BPE tokens.
+          context_size:
+            Context size of the NN decoder model.
+          blank_id:
+            The ID of the blank symbol.
+        Returns:
+          Return a list-of-list int. `ans[i]` contains the decoder
+          input for the i-th item in this object and its lengths
+          is `context_size`.
+        """
+        ans: List[List[int]] = []
+        buf = [blank_id] * context_size + ys
+        for item in self:
+            # fmt: off
+            ans.append(buf[item.pos_u:(item.pos_u + context_size)])
+            # fmt: on
+        return ans
+
+    def topk(self, k: int) -> "AlignItemList":
+        """Return the top-k items.
+
+        Items are ordered by their log probs in descending order
+        and the top-k items are returned.
+
+        Args:
+          k:
+            Size of top-k.
+        Returns:
+          Return a new AlignItemList that contains the top-k items
+          in this object. Caution: It uses shallow copy.
+        """
+        items = list(self)
+        items = sorted(items, key=lambda i: i.log_prob, reverse=True)
+        return AlignItemList(items[:k])
+
+
+def force_alignment(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    ys: List[int],
+    beam_size: int = 4,
+) -> List[int]:
+    """Compute the force alignment of an utterance given its transcript
+    in BPE tokens and the corresponding acoustic output from the encoder.
+
+    Caution:
+      We assume that the maximum number of sybmols per frame is 1.
+      That is, the model should be trained using a nonzero value
+      for the option `--modified-transducer-prob` in train.py.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A tensor of shape (N, T, C). Support only for N==1 at present.
+      ys:
+        A list of BPE token IDs. We require that len(ys) <= T.
+      beam_size:
+        Size of the beam used in beam search.
+    Returns:
+      Return a list of int such that
+        - len(ans) == T
+        - After removing blanks from ans, we have ans == ys.
+    """
+    assert encoder_out.ndim == 3, encoder_out.ndim
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    assert 0 < len(ys) <= encoder_out.size(1), (len(ys), encoder_out.size(1))
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    T = encoder_out.size(1)
+    U = len(ys)
+    assert 0 < U <= T
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = encoder_out_len
+
+    start = AlignItem(log_prob=0.0, ys=[], pos_u=0)
+    B = AlignItemList([start])
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # current_encoder_out is of shape (1, 1, encoder_out_dim)
+        # fmt: on
+
+        A = B  # shallow copy
+        B = AlignItemList()
+
+        decoder_input = A.get_decoder_input(
+            ys=ys, context_size=context_size, blank_id=blank_id
+        )
+        decoder_input = torch.tensor(decoder_input, device=device)
+        # decoder_input is of shape (num_active_items, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        # decoder_output is of shape (num_active_items, 1, decoder_output_dim)
+
+        current_encoder_out = current_encoder_out.expand(decoder_out.size(0), 1, -1)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(decoder_out.size(0)),
+            decoder_out_len.expand(decoder_out.size(0)),
+        )
+
+        # logits is of shape (num_active_items, vocab_size)
+        log_probs = logits.log_softmax(dim=-1).tolist()
+
+        for i, item in enumerate(A):
+            if (T - 1 - t) >= (U - item.pos_u):
+                # horizontal transition (left -> right)
+                new_item = AlignItem(
+                    log_prob=item.log_prob + log_probs[i][blank_id],
+                    ys=item.ys + [blank_id],
+                    pos_u=item.pos_u,
+                )
+                B.append(new_item)
+
+            if item.pos_u < U:
+                # diagonal transition (lower left -> upper right)
+                u = ys[item.pos_u]
+                new_item = AlignItem(
+                    log_prob=item.log_prob + log_probs[i][u],
+                    ys=item.ys + [u],
+                    pos_u=item.pos_u + 1,
+                )
+                B.append(new_item)
+
+        if len(B) > beam_size:
+            B = B.topk(beam_size)
+
+    ans = B.topk(1)[0].ys
+
+    assert len(ans) == T
+    assert list(filter(lambda i: i != blank_id, ans)) == ys
+
+    return ans
+
+
+def get_word_starting_frames(
+    ali: List[int], sp: spm.SentencePieceProcessor
+) -> List[int]:
+    """Get the starting frame of each word from the given token alignments.
+
+    When a word is encoded into BPE tokens, the first token starts
+    with underscore "_", which can be used to identify the starting frame
+    of a word.
+
+    Args:
+      ali:
+        Framewise token alignment. It can be the return value of
+        :func:`force_alignment`.
+      sp:
+        The sentencepiece model.
+    Returns:
+      Return a list of int representing the starting frame of each word
+      in the alignment.
+      Caution:
+        You have to take into account the model subsampling factor when
+        converting the starting frame into time.
+    """
+    underscore = b"\xe2\x96\x81".decode()  # '_'
+    ans = []
+    for i in range(len(ali)):
+        if sp.id_to_piece(ali[i]).startswith(underscore):
+            ans.append(i)
+    return ans
diff --git a/egs/librispeech/ASR/transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/transducer_stateless/asr_datamodule.py
new file mode 120000
index 000000000..07f39b451
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless/beam_search.py b/egs/librispeech/ASR/transducer_stateless/beam_search.py
new file mode 100644
index 000000000..1d79eef9d
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/beam_search.py
@@ -0,0 +1,1042 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import k2
+import torch
+from model import Transducer
+
+from icefall.decode import Nbest, one_best_decoding
+from icefall.utils import get_texts
+
+
+def fast_beam_search_one_best(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    the shortest path within the lattice is used as the final output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+    return hyps
+
+
+def fast_beam_search_nbest_oracle(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    ref_texts: List[List[int]],
+    use_double_scores: bool = True,
+    nbest_scale: float = 0.5,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first obtained using modified beam search, and then
+    we select `num_paths` linear paths from the lattice. The path
+    that has the minimum edit distance with the given reference transcript
+    is used as the output.
+
+    This is the best result we can achieve for any nbest based rescoring
+    methods.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      ref_texts:
+        A list-of-list of integers containing the reference transcripts.
+        If the decoding_graph is a trivial_graph, the integer ID is the
+        BPE token ID.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    hyps = nbest.build_levenshtein_graphs()
+    refs = k2.levenshtein_graph(ref_texts, device=hyps.device)
+
+    levenshtein_alignment = k2.levenshtein_alignment(
+        refs=refs,
+        hyps=hyps,
+        hyp_to_ref_map=nbest.shape.row_ids(1),
+        sorted_match_ref=True,
+    )
+
+    tot_scores = levenshtein_alignment.get_tot_scores(
+        use_double_scores=False, log_semiring=False
+    )
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+
+    max_indexes = ragged_tot_scores.argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    hyps = get_texts(best_path)
+    return hyps
+
+
+def fast_beam_search(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> k2.Fsa:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    Returns:
+      Return an FsaVec with axes [utt][state][arc] containing the decoded
+      lattice. Note: When the input graph is a TrivialGraph, the returned
+      lattice is actually an acceptor.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(k2.RnntDecodingStream(decoding_graph))
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out_len = torch.ones(1, dtype=torch.int32)
+    decoder_out_len = torch.ones(1, dtype=torch.int32)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(decoder_out.size(0)),
+            decoder_out_len.expand(decoder_out.size(0)),
+        )  # (N, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+    decoding_streams.terminate_and_flush_to_streams()
+    lattice = decoding_streams.format_output(encoder_out_lens.tolist())
+
+    return lattice
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """Greedy search for a single utterance.
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size, device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out, encoder_out_len, decoder_out_len
+        )
+        # logits is (1, vocab_size)
+
+        y = logits.argmax().item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+def greedy_search_batch(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+    Returns:
+      Return a list-of-list of token IDs containing the decoded results.
+      len(ans) equals to encoder_out.size(0).
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = next(model.parameters()).device
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    # decoder_out: (N, 1, decoder_out_dim)
+
+    encoder_out_len = torch.ones(1, dtype=torch.int32)
+    decoder_out_len = torch.ones(1, dtype=torch.int32)
+
+    encoder_out = packed_encoder_out.data
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(batch_size),
+            decoder_out_len.expand(batch_size),
+        )  # (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )  # (batch_size, context_size)
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )  # (batch_size, 1, decoder_out_dim)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def run_decoder(
+    ys: List[int],
+    model: Transducer,
+    decoder_cache: Dict[str, torch.Tensor],
+) -> torch.Tensor:
+    """Run the neural decoder model for a given hypothesis.
+
+    Args:
+      ys:
+        The current hypothesis.
+      model:
+        The transducer model.
+      decoder_cache:
+        Cache to save computations.
+    Returns:
+      Return a 1-D tensor of shape (decoder_out_dim,) containing
+      output of `model.decoder`.
+    """
+    context_size = model.decoder.context_size
+    key = "_".join(map(str, ys[-context_size:]))
+    if key in decoder_cache:
+        return decoder_cache[key]
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [ys[-context_size:]],
+        device=device,
+        dtype=torch.int64,
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_cache[key] = decoder_out
+
+    return decoder_out
+
+
+def run_joiner(
+    key: str,
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    decoder_out: torch.Tensor,
+    encoder_out_len: torch.Tensor,
+    decoder_out_len: torch.Tensor,
+    joint_cache: Dict[str, torch.Tensor],
+):
+    """Run the joint network given outputs from the encoder and decoder.
+
+    Args:
+      key:
+        A key into the `joint_cache`.
+      model:
+        The transducer model.
+      encoder_out:
+        A tensor of shape (1, 1, encoder_out_dim).
+      decoder_out:
+        A tensor of shape (1, 1, decoder_out_dim).
+      encoder_out_len:
+        A tensor with value [1].
+      decoder_out_len:
+        A tensor with value [1].
+      joint_cache:
+        A dict to save computations.
+    Returns:
+      Return a tensor from the output of log-softmax.
+      Its shape is (vocab_size,).
+    """
+    if key in joint_cache:
+        return joint_cache[key]
+
+    logits = model.joiner(
+        encoder_out,
+        decoder_out,
+        encoder_out_len,
+        decoder_out_len,
+    )
+
+    # TODO(fangjun): Scale the blank posterior
+    log_prob = logits.log_softmax(dim=-1)
+    # log_prob is (1, 1, 1, vocab_size)
+
+    log_prob = log_prob.squeeze()
+    # Now log_prob is (vocab_size,)
+
+    joint_cache[key] = log_prob
+
+    return log_prob
+
+
+def _get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
+def modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: int = 4,
+) -> List[List[int]]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcodded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+            )
+        )
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    encoder_out = packed_encoder_out.data
+    offset = 0
+    finalized_B = []
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1)
+        # current_encoder_out's shape is: (batch_size, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = _get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        # decoder_output is of shape (num_hyps, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(decoder_out.size(0)),
+            decoder_out_len.expand(decoder_out.size(0)),
+        )
+        # logits is of shape (num_hyps, vocab_size)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    B = B + finalized_B
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+
+    sorted_ans = [h.ys[context_size:] for h in best_hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+def _deprecated_modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    It decodes only one utterance at a time. We keep it only for reference.
+    The function :func:`modified_beam_search` should be preferred as it
+    supports batch decoding.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # current_encoder_out is of shape (1, 1, encoder_out_dim)
+        # fmt: on
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        # decoder_output is of shape (num_hyps, 1, decoder_output_dim)
+
+        current_encoder_out = current_encoder_out.expand(decoder_out.size(0), 1, -1)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(decoder_out.size(0)),
+            decoder_out_len.expand(decoder_out.size(0)),
+        )
+        # logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            topk_hyp_indexes = topk_hyp_indexes.tolist()
+            topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
+            new_ys = hyp.ys[:]
+            new_token = topk_token_indexes[i]
+            if new_token != blank_id:
+                new_ys.append(new_token)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+            B.add(new_hyp)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    return ys
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            decoder_out = run_decoder(
+                ys=y_star.ys, model=model, decoder_cache=decoder_cache
+            )
+
+            key = "_".join(map(str, y_star.ys[-context_size:]))
+            key += f"-t-{t}"
+            log_prob = run_joiner(
+                key=key,
+                model=model,
+                encoder_out=current_encoder_out,
+                decoder_out=decoder_out,
+                encoder_out_len=encoder_out_len,
+                decoder_out_len=decoder_out_len,
+                joint_cache=joint_cache,
+            )
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for idx in range(values.size(0)):
+                i = indices[idx].item()
+                if i == blank_id:
+                    continue
+
+                new_ys = y_star.ys + [i]
+
+                new_log_prob = y_star.log_prob + values[idx]
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys
diff --git a/egs/librispeech/ASR/transducer_stateless/compute_ali.py b/egs/librispeech/ASR/transducer_stateless/compute_ali.py
new file mode 100755
index 000000000..f479389df
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/compute_ali.py
@@ -0,0 +1,321 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+    ./transducer_stateless/compute_ali.py \
+            --exp-dir ./transducer_stateless/exp \
+            --bpe-model ./data/lang_bpe_500/bpe.model \
+            --epoch 20 \
+            --avg 10 \
+            --max-duration 300 \
+            --dataset train-clean-100 \
+            --out-dir data/ali
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List
+
+import numpy as np
+import sentencepiece as spm
+import torch
+from alignment import force_alignment
+from asr_datamodule import LibriSpeechAsrDataModule
+from lhotse import CutSet
+from lhotse.features.io import FeaturesWriter, NumpyHdf5Writer
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import AttributeDict, setup_logger
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--out-dir",
+        type=str,
+        required=True,
+        help="""Output directory.
+        It contains 2 generated files:
+
+        - token_ali_xxx.h5
+        - cuts_xxx.json.gz
+
+        where xxx is the value of `--dataset`. For instance, if
+        `--dataset` is `train-clean-100`, it will contain 2 files:
+
+        - `token_ali_train-clean-100.h5`
+        - `cuts_train-clean-100.json.gz`
+        """,
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="""The name of the dataset to compute alignments for.
+        Possible values are:
+            - test-clean.
+            - test-other
+            - train-clean-100
+            - train-clean-360
+            - train-other-500
+            - dev-clean
+            - dev-other
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def compute_alignments(
+    model: torch.nn.Module,
+    dl: torch.utils.data,
+    ali_writer: FeaturesWriter,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+):
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+    num_cuts = 0
+
+    device = model.device
+    cuts = []
+
+    for batch_idx, batch in enumerate(dl):
+        feature = batch["inputs"]
+
+        # at entry, feature is [N, T, C]
+        assert feature.ndim == 3
+        feature = feature.to(device)
+
+        supervisions = batch["supervisions"]
+
+        cut_list = supervisions["cut"]
+        for cut in cut_list:
+            assert len(cut.supervisions) == 1, f"{len(cut.supervisions)}"
+
+        feature_lens = supervisions["num_frames"].to(device)
+
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+        batch_size = encoder_out.size(0)
+
+        texts = supervisions["text"]
+
+        ys_list: List[List[int]] = sp.encode(texts, out_type=int)
+
+        ali_list = []
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+
+            ali = force_alignment(
+                model=model,
+                encoder_out=encoder_out_i,
+                ys=ys_list[i],
+                beam_size=params.beam_size,
+            )
+            ali_list.append(ali)
+        assert len(ali_list) == len(cut_list)
+
+        for cut, ali in zip(cut_list, ali_list):
+            cut.token_alignment = ali_writer.store_array(
+                key=cut.id,
+                value=np.asarray(ali, dtype=np.int32),
+                # frame shift is 0.01s, subsampling_factor is 4
+                frame_shift=0.04,
+                temporal_dim=0,
+                start=0,
+            )
+
+        cuts += cut_list
+
+        num_cuts += len(cut_list)
+
+        if batch_idx % 2 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return CutSet.from_cuts(cuts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    args.enable_spec_aug = False
+    args.enable_musan = False
+    args.return_cuts = True
+    args.concatenate_cuts = False
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-ali")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"Computing alignments for {params.dataset} - started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    out_dir = Path(params.out_dir)
+    out_dir.mkdir(exist_ok=True)
+
+    out_ali_filename = out_dir / f"token_ali_{params.dataset}.h5"
+    out_manifest_filename = out_dir / f"cuts_{params.dataset}.json.gz"
+
+    done_file = out_dir / f".{params.dataset}.done"
+    if done_file.is_file():
+        logging.info(f"{done_file} exists - skipping")
+        exit()
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+    if params.dataset == "test-clean":
+        test_clean_cuts = librispeech.test_clean_cuts()
+        dl = librispeech.test_dataloaders(test_clean_cuts)
+    elif params.dataset == "test-other":
+        test_other_cuts = librispeech.test_other_cuts()
+        dl = librispeech.test_dataloaders(test_other_cuts)
+    elif params.dataset == "train-clean-100":
+        train_clean_100_cuts = librispeech.train_clean_100_cuts()
+        dl = librispeech.train_dataloaders(train_clean_100_cuts)
+    elif params.dataset == "train-clean-360":
+        train_clean_360_cuts = librispeech.train_clean_360_cuts()
+        dl = librispeech.train_dataloaders(train_clean_360_cuts)
+    elif params.dataset == "train-other-500":
+        train_other_500_cuts = librispeech.train_other_500_cuts()
+        dl = librispeech.train_dataloaders(train_other_500_cuts)
+    elif params.dataset == "dev-clean":
+        dev_clean_cuts = librispeech.dev_clean_cuts()
+        dl = librispeech.valid_dataloaders(dev_clean_cuts)
+    else:
+        assert params.dataset == "dev-other", f"{params.dataset}"
+        dev_other_cuts = librispeech.dev_other_cuts()
+        dl = librispeech.valid_dataloaders(dev_other_cuts)
+
+    logging.info(f"Processing {params.dataset}")
+
+    with NumpyHdf5Writer(out_ali_filename) as ali_writer:
+        cut_set = compute_alignments(
+            model=model,
+            dl=dl,
+            ali_writer=ali_writer,
+            params=params,
+            sp=sp,
+        )
+
+    cut_set.to_file(out_manifest_filename)
+
+    logging.info(
+        f"For dataset {params.dataset}, its framewise token alignments are "
+        f"saved to {out_ali_filename} and the cut manifest "
+        f"file is {out_manifest_filename}. Number of cuts: {len(cut_set)}"
+    )
+    done_file.touch()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/conformer.py b/egs/librispeech/ASR/transducer_stateless/conformer.py
new file mode 100644
index 000000000..90b722bde
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/conformer.py
@@ -0,0 +1,1445 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Transformer
+
+from icefall.utils import is_jit_tracing, make_pad_mask, subsequent_chunk_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        output_dim (int): Number of output dimension
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+        dynamic_chunk_training (bool): whether to use dynamic chunk training, if
+            you want to train a streaming model, this is expected to be True.
+            When setting True, it will use a masking strategy to make the attention
+            see only limited left and right context.
+        short_chunk_threshold (float): a threshold to determinize the chunk size
+            to be used in masking training, if the randomly generated chunk size
+            is greater than ``max_len * short_chunk_threshold`` (max_len is the
+            max sequence length of current batch) then it will use
+            full context in training (i.e. with chunk size equals to max_len).
+            This will be used only when dynamic_chunk_training is True.
+        short_chunk_size (int): see docs above, if the randomly generated chunk
+            size equals to or less than ``max_len * short_chunk_threshold``, the
+            chunk size will be sampled uniformly from 1 to short_chunk_size.
+            This also will be used only when dynamic_chunk_training is True.
+        num_left_chunks (int): the left context (in chunks) attention can see, the
+            chunk size is decided by short_chunk_threshold and short_chunk_size.
+            A minus value means seeing full left context.
+            This also will be used only when dynamic_chunk_training is True.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        dynamic_chunk_training: bool = False,
+        short_chunk_threshold: float = 0.75,
+        short_chunk_size: int = 25,
+        num_left_chunks: int = -1,
+        causal: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            output_dim=output_dim,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+        )
+
+        self.encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.cnn_module_kernel = cnn_module_kernel
+        self.causal = causal
+
+        self.dynamic_chunk_training = dynamic_chunk_training
+        self.short_chunk_threshold = short_chunk_threshold
+        self.short_chunk_size = short_chunk_size
+        self.num_left_chunks = num_left_chunks
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+            causal,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+        self._init_state: List[torch.Tensor] = [torch.empty(0)]
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        if not is_jit_tracing():
+            assert x.size(0) == lengths.max().item()
+
+        src_key_padding_mask = make_pad_mask(lengths)
+
+        if self.dynamic_chunk_training:
+            assert (
+                self.causal
+            ), "Causal convolution is required for streaming conformer."
+            max_len = x.size(0)
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * self.short_chunk_threshold):
+                chunk_size = max_len
+            else:
+                chunk_size = chunk_size % self.short_chunk_size + 1
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=self.num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x, pos_emb, mask=mask, src_key_padding_mask=src_key_padding_mask
+            )  # (T, N, C)
+        else:
+            x = self.encoder(
+                x, pos_emb, mask=None, src_key_padding_mask=src_key_padding_mask
+            )  # (T, N, C)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+    @torch.jit.export
+    def get_init_state(
+        self, left_context: int, device: torch.device
+    ) -> List[torch.Tensor]:
+        """Return the initial cache state of the model.
+
+        Args:
+          left_context: The left context size (in frames after subsampling).
+
+        Returns:
+          Return the initial state of the model, it is a list containing two
+          tensors, the first one is the cache for attentions which has a shape
+          of (num_encoder_layers, left_context, encoder_dim), the second one
+          is the cache of conv_modules which has a shape of
+          (num_encoder_layers, cnn_module_kernel - 1, encoder_dim).
+
+          NOTE: the returned tensors are on the given device.
+        """
+        if len(self._init_state) == 2 and self._init_state[0].size(1) == left_context:
+            # Note: It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        init_states: List[torch.Tensor] = [
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    left_context,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+        ]
+
+        self._init_state = init_states
+
+        return init_states
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[torch.Tensor]] = None,
+        processed_lens: Optional[Tensor] = None,
+        left_context: int = 64,
+        right_context: int = 0,
+        chunk_size: int = 16,
+        simulate_streaming: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states:
+            The decode states for previous frames which contains the cached data.
+            It has two elements, the first element is the attn_cache which has
+            a shape of (encoder_layers, left_context, batch, attention_dim),
+            the second element is the conv_cache which has a shape of
+            (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+            Note: states will be modified in this function.
+          processed_lens:
+            How many frames (after subsampling) have been processed for each sequence.
+          left_context:
+            How many previous frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `left_context` frames
+            of left context, some have more.
+          right_context:
+            How many future frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `right_context` frames
+            of right context, some have more.
+          chunk_size:
+            The chunk size for decoding, this will be used to simulate streaming
+            decoding using masking.
+          simulate_streaming:
+            If setting True, it will use a masking strategy to simulate streaming
+            fashion (i.e. every chunk data only see limited left context and
+            right context). The whole sequence is supposed to be send at a time
+            When using simulate_streaming.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+            - states, the updated states(i.e. caches) including the information
+              of current chunk.
+        """
+
+        # x: [N, T, C]
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        if not simulate_streaming:
+            assert states is not None
+            assert processed_lens is not None
+            assert (
+                len(states) == 2
+                and states[0].shape
+                == (self.encoder_layers, left_context, x.size(0), self.d_model)
+                and states[1].shape
+                == (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    x.size(0),
+                    self.d_model,
+                )
+            ), f"""The length of states MUST be equal to 2, and the shape of
+             first element should be {(self.encoder_layers, left_context, x.size(0), self.d_model)},
+             given {states[0].shape}. the shape of second element should be
+             {(self.encoder_layers, self.cnn_module_kernel - 1, x.size(0), self.d_model)},
+             given {states[1].shape}."""
+
+            lengths -= 2  # we will cut off 1 frame on each side of encoder_embed output
+            src_key_padding_mask = make_pad_mask(lengths)
+
+            processed_mask = torch.arange(left_context, device=x.device).expand(
+                x.size(0), left_context
+            )
+            processed_lens = processed_lens.view(x.size(0), 1)
+            processed_mask = (processed_lens <= processed_mask).flip(1)
+
+            src_key_padding_mask = torch.cat(
+                [processed_mask, src_key_padding_mask], dim=1
+            )
+
+            embed = self.encoder_embed(x)
+
+            # cut off 1 frame on each size of embed as they see the padding
+            # value which causes a training and decoding mismatch.
+            embed = embed[:, 1:-1, :]
+
+            embed, pos_enc = self.encoder_pos(embed, left_context)
+            embed = embed.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+
+            x, states = self.encoder.chunk_forward(
+                embed,
+                pos_enc,
+                src_key_padding_mask=src_key_padding_mask,
+                states=states,
+                left_context=left_context,
+                right_context=right_context,
+            )  # (T, B, F)
+        else:
+            assert states is None
+            states = []  # just to make torch.script.jit happy
+            src_key_padding_mask = make_pad_mask(lengths)
+            x = self.encoder_embed(x)
+            x, pos_emb = self.encoder_pos(x)
+            x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+            assert x.size(0) == lengths.max().item()
+
+            if chunk_size < 0:
+                # use full attention
+                chunk_size = x.size(0)
+                left_context = -1
+
+            num_left_chunks = -1
+            if left_context >= 0:
+                assert left_context % chunk_size == 0
+                num_left_chunks = left_context // chunk_size
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )  # (T, N, C)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths, states
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before (bool): whether to use layer_norm before the first block.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training and streaming decoding.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        causal: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel, causal=causal)
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+
+        src, _ = self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        src = residual + self.dropout(src)
+
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(src))
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+
+        # We put the attention cache this level (i.e. before linear transformation)
+        # to save memory consumption, when decoding in streaming fashion, the
+        # batch size would be thousands (for 32GB machine), if we cache key & val
+        # separately, it needs extra several GB memory.
+        # TODO(WeiKang): Move cache to self_attn level (i.e. cache key & val
+        # separately) if needed.
+        key = torch.cat([states[0], src], dim=0)
+        val = key
+        if right_context > 0:
+            states[0] = key[
+                -(left_context + right_context) : -right_context, ...  # noqa
+            ]
+        else:
+            states[0] = key[-left_context:, ...]
+
+        src_att = self.self_attn(
+            src,
+            key,
+            val,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+            left_context=left_context,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+
+        src, conv_cache = self.conv_module(src, states[1], right_context=right_context)
+        states[1] = conv_cache
+        src = residual + self.dropout(src)
+
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src, states
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+        Shape:
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+        return output
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (encoder_layers, left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        assert not self.training
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            cache = [states[0][layer_index], states[1][layer_index]]
+            output, cache = mod.chunk_forward(
+                output,
+                pos_emb,
+                states=cache,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                left_context=left_context,
+                right_context=right_context,
+            )
+            states[0][layer_index] = cache[0]
+            states[1][layer_index] = cache[1]
+
+        return output, states
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        if is_jit_tracing():
+            # 10k frames correspond to ~100k ms, e.g., 100 seconds, i.e.,
+            # It assumes that the maximum input won't have more than
+            # 10k frames.
+            #
+            # TODO(fangjun): Use torch.jit.script() for this module
+            max_len = 10000
+
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor, left_context: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_1 = x.size(1) + left_context
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_1 * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_1, self.d_model)
+        pe_negative = torch.zeros(x_size_1, self.d_model)
+        position = torch.arange(0, x_size_1, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor, left_context: int = 0) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x, left_context)
+        x = x * self.xscale
+        x_size_1 = x.size(1) + left_context
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_1
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            left_context=left_context,
+        )
+
+    def rel_shift(self, x: Tensor, left_context: int = 0) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+
+        time2 = time1 + left_context
+        if not is_jit_tracing():
+            assert (
+                n == left_context + 2 * time1 - 1
+            ), f"{n} == {left_context} + 2 * {time1} - 1"
+
+        if is_jit_tracing():
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(time2)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, time1, time2)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time1_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, time1, time2),
+                (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+                storage_offset=n_stride * (time1 - 1),
+            )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        if not is_jit_tracing():
+            assert embed_dim == embed_dim_to_check
+            assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        if not is_jit_tracing():
+            assert (
+                head_dim * num_heads == embed_dim
+            ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        if not is_jit_tracing():
+            assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+
+        # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1)
+        p = p.permute(0, 2, 3, 1)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(q_with_bias_v, p)  # (batch, head, time1, 2*time1-1)
+
+        matrix_bd = self.rel_shift(matrix_bd, left_context=left_context)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        if not is_jit_tracing():
+            assert list(attn_output_weights.size()) == [
+                bsz * num_heads,
+                tgt_len,
+                src_len,
+            ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+
+        # If we are using dynamic_chunk_training and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`, at this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax), so, we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                1
+            ).unsqueeze(2)
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+
+        if not is_jit_tracing():
+            assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+        causal (bool): Whether to use causal convolution.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        causal: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+        self.causal = causal
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        self.lorder = kernel_size - 1
+        padding = (kernel_size - 1) // 2
+        if self.causal:
+            padding = 0
+
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.LayerNorm(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        cache: Optional[Tensor] = None,
+        right_context: int = 0,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: The cache of depthwise_conv, only used in real streaming
+                decoding.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        if self.causal and self.lorder > 0:
+            if cache is None:
+                # Make depthwise_conv causal by
+                # manualy padding self.lorder zeros to the left
+                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+            else:
+                assert not self.training, "Cache should be None in training time"
+                assert cache.size(0) == self.lorder
+                x = torch.cat([cache.permute(1, 2, 0), x], dim=2)
+                if right_context > 0:
+                    cache = x.permute(2, 0, 1)[
+                        -(self.lorder + right_context) : (-right_context),  # noqa
+                        ...,
+                    ]
+                else:
+                    cache = x.permute(2, 0, 1)[-self.lorder :, ...]  # noqa
+
+        x = self.depthwise_conv(x)
+        # x is (batch, channels, time)
+        x = x.permute(0, 2, 1)
+        x = self.norm(x)
+        x = x.permute(0, 2, 1)
+
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        if cache is None:
+            cache = torch.empty(0)
+
+        return x.permute(2, 0, 1), cache
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x
diff --git a/egs/librispeech/ASR/transducer_stateless/decode.py b/egs/librispeech/ASR/transducer_stateless/decode.py
new file mode 100755
index 000000000..42125e19f
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/decode.py
@@ -0,0 +1,520 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_stateless/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./transducer_stateless/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./transducer_stateless/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./transducer_stateless/decode.py \
+    --epoch 14 \
+    --avg 7 \
+    --exp-dir ./transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/decoder.py b/egs/librispeech/ASR/transducer_stateless/decoder.py
new file mode 100644
index 000000000..ac6292f63
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/decoder.py
@@ -0,0 +1,101 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+        self.vocab_size = vocab_size
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+        else:
+            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
+            # when inference with torch.jit.script and context_size == 1
+            self.conv = nn.Identity()
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, embedding_dim).
+        """
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        return embedding_out
diff --git a/egs/librispeech/ASR/transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/transducer_stateless/encoder_interface.py
new file mode 100644
index 000000000..257facce4
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/encoder_interface.py
@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")
diff --git a/egs/librispeech/ASR/transducer_stateless/export.py b/egs/librispeech/ASR/transducer_stateless/export.py
new file mode 100755
index 000000000..c397eb171
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/export.py
@@ -0,0 +1,251 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless/export.py \
+  --exp-dir ./transducer_stateless/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless/decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./transducer_stateless/decode.py \
+        --exp-dir ./transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/joiner.py b/egs/librispeech/ASR/transducer_stateless/joiner.py
new file mode 100644
index 000000000..e1625992d
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/joiner.py
@@ -0,0 +1,79 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List
+
+import torch
+import torch.nn as nn
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        encoder_out_len: torch.Tensor,
+        decoder_out_len: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, self.input_dim).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, self.input_dim).
+          encoder_out_len:
+            A 1-D tensor of shape (N,) containing valid number of frames
+            before padding in `encoder_out`.
+          decoder_out_len:
+            A 1-D tensor of shape (N,) containing valid number of frames
+            before padding in `decoder_out`.
+        Returns:
+          Return a tensor of shape (sum_all_TU, self.output_dim).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == self.input_dim
+        assert decoder_out.size(2) == self.input_dim
+
+        N = encoder_out.size(0)
+
+        encoder_out_len: List[int] = encoder_out_len.tolist()
+        decoder_out_len: List[int] = decoder_out_len.tolist()
+
+        encoder_out_list = [encoder_out[i, : encoder_out_len[i], :] for i in range(N)]
+
+        decoder_out_list = [decoder_out[i, : decoder_out_len[i], :] for i in range(N)]
+
+        x = [
+            e.unsqueeze(1) + d.unsqueeze(0)
+            for e, d in zip(encoder_out_list, decoder_out_list)
+        ]
+
+        x = [p.reshape(-1, self.input_dim) for p in x]
+        x = torch.cat(x)
+
+        activations = torch.tanh(x)
+
+        logits = self.output_linear(activations)
+
+        return logits
diff --git a/egs/librispeech/ASR/transducer_stateless/model.py b/egs/librispeech/ASR/transducer_stateless/model.py
new file mode 100644
index 000000000..8281e1fb5
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/model.py
@@ -0,0 +1,143 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        modified_transducer_prob: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          modified_transducer_prob:
+            The probability to use modified transducer loss.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        decoder_out = self.decoder(sos_y_padded)
+
+        # +1 here since a blank is prepended to each utterance.
+        logits = self.joiner(
+            encoder_out=encoder_out,
+            decoder_out=decoder_out,
+            encoder_out_len=x_lens,
+            decoder_out_len=y_lens + 1,
+        )
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        # We don't put this `import` at the beginning of the file
+        # as it is required only in the training, not during the
+        # reference stage
+        import optimized_transducer
+
+        assert 0 <= modified_transducer_prob <= 1
+
+        if modified_transducer_prob == 0:
+            one_sym_per_frame = False
+        elif random.random() < modified_transducer_prob:
+            # random.random() returns a float in the range [0, 1)
+            one_sym_per_frame = True
+        else:
+            one_sym_per_frame = False
+
+        loss = optimized_transducer.transducer_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+            one_sym_per_frame=one_sym_per_frame,
+            from_log_softmax=False,
+        )
+
+        return loss
diff --git a/egs/librispeech/ASR/transducer_stateless/pretrained.py b/egs/librispeech/ASR/transducer_stateless/pretrained.py
new file mode 100755
index 000000000..5898dd0f5
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/pretrained.py
@@ -0,0 +1,344 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless/pretrained.py \
+    --checkpoint ./transducer_stateless/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    --max-sym-per-frame 1 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless/pretrained.py \
+    --checkpoint ./transducer_stateless/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./transducer_stateless/pretrained.py \
+    --checkpoint ./transducer_stateless/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./transducer_stateless/pretrained.py \
+    --checkpoint ./transducer_stateless/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./transducer_stateless/exp/pretrained.pt is generated by
+./transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = [token_ids_to_words(hyp) for hyp in hyp_list]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/subsampling.py b/egs/librispeech/ASR/transducer_stateless/subsampling.py
new file mode 120000
index 000000000..73068da26
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/subsampling.py
@@ -0,0 +1 @@
+../transducer/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless/test_compute_ali.py b/egs/librispeech/ASR/transducer_stateless/test_compute_ali.py
new file mode 100755
index 000000000..9af46846a
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/test_compute_ali.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script shows how to get word starting time
+from framewise token alignment.
+
+Usage:
+    ./transducer_stateless/compute_ali.py \
+            --exp-dir ./transducer_stateless/exp \
+            --bpe-model ./data/lang_bpe_500/bpe.model \
+            --epoch 20 \
+            --avg 10 \
+            --max-duration 300 \
+            --dataset train-clean-100 \
+            --out-dir data/ali
+
+And the you can run:
+
+    ./transducer_stateless/test_compute_ali.py \
+            --bpe-model ./data/lang_bpe_500/bpe.model \
+            --ali-dir data/ali \
+            --dataset train-clean-100
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from alignment import get_word_starting_frames
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.dataset import DynamicBucketingSampler, K2SpeechRecognitionDataset
+from lhotse.dataset.collation import collate_custom_field
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--ali-dir",
+        type=Path,
+        default="./data/ali",
+        help="It specifies the directory where alignments can be found.",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="""The name of the dataset:
+        Possible values are:
+            - test-clean.
+            - test-other
+            - train-clean-100
+            - train-clean-360
+            - train-other-500
+            - dev-clean
+            - dev-other
+        """,
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    cuts_jsonl = args.ali_dir / f"librispeech_cuts_{args.dataset}.jsonl.gz"
+
+    logging.info(f"Loading {cuts_jsonl}")
+    cuts = load_manifest_lazy(cuts_jsonl)
+
+    sampler = DynamicBucketingSampler(
+        cuts,
+        max_duration=30,
+        num_buckets=30,
+        shuffle=False,
+    )
+
+    dataset = K2SpeechRecognitionDataset(return_cuts=True)
+
+    dl = torch.utils.data.DataLoader(
+        dataset,
+        sampler=sampler,
+        batch_size=None,
+        num_workers=1,
+        persistent_workers=False,
+    )
+
+    frame_shift = 10  # ms
+    subsampling_factor = 4
+
+    frame_shift_in_second = frame_shift * subsampling_factor / 1000.0
+
+    # key: cut.id
+    # value: a list of pairs (word, time_in_second)
+    word_starting_time_dict = {}
+    for batch in dl:
+        supervisions = batch["supervisions"]
+        cuts = supervisions["cut"]
+
+        token_alignment, token_alignment_length = collate_custom_field(
+            CutSet.from_cuts(cuts), "token_alignment"
+        )
+
+        for i in range(len(cuts)):
+            assert (
+                (cuts[i].features.num_frames - 1) // 2 - 1
+            ) // 2 == token_alignment_length[i]
+
+            word_starting_frames = get_word_starting_frames(
+                token_alignment[i, : token_alignment_length[i]].tolist(), sp=sp
+            )
+            word_starting_time = [
+                "{:.2f}".format(i * frame_shift_in_second) for i in word_starting_frames
+            ]
+
+            words = supervisions["text"][i].split()
+
+            assert len(word_starting_frames) == len(words)
+            word_starting_time_dict[cuts[i].id] = list(zip(words, word_starting_time))
+
+        # This is a demo script and we exit here after processing
+        # one batch.
+        # You can find word starting time in the dict "word_starting_time_dict"
+        for cut_id, word_time in word_starting_time_dict.items():
+            print(f"{cut_id}\n{word_time}\n")
+        break
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/test_conformer.py b/egs/librispeech/ASR/transducer_stateless/test_conformer.py
new file mode 100755
index 000000000..65b08d425
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/test_conformer.py
@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey
+#                                                  Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless/test_conformer.py
+"""
+
+import torch
+from conformer import Conformer
+
+
+def test_conformer():
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, output_dim=256, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    logits, lengths = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    print(logits.shape)
+    print(lengths.shape)
+
+
+def main():
+    test_conformer()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/test_decoder.py b/egs/librispeech/ASR/transducer_stateless/test_decoder.py
new file mode 100755
index 000000000..3a653c1b7
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/test_decoder.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, embedding_dim)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, embedding_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/test_joiner.py b/egs/librispeech/ASR/transducer_stateless/test_joiner.py
new file mode 100755
index 000000000..593577c7c
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/test_joiner.py
@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless/test_joiner.py
+"""
+
+import torch
+from joiner import Joiner
+
+
+def test_joiner():
+    device = torch.device("cpu")
+    input_dim = 3
+    output_dim = 5
+    joiner = Joiner(input_dim, output_dim)
+    joiner.to(device)
+
+    encoder_out = torch.rand(3, 10, input_dim, device=device)
+    decoder_out = torch.rand(3, 8, input_dim, device=device)
+
+    encoder_out_len = torch.tensor([5, 10, 3], device=device)
+    decoder_out_len = torch.tensor([6, 8, 7], device=device)
+
+    out = joiner(
+        encoder_out=encoder_out,
+        decoder_out=decoder_out,
+        encoder_out_len=encoder_out_len,
+        decoder_out_len=decoder_out_len,
+    )
+    assert out.size(0) == (encoder_out_len * decoder_out_len).sum()
+    assert out.size(1) == output_dim
+
+
+def main():
+    test_joiner()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/test_model.py b/egs/librispeech/ASR/transducer_stateless/test_model.py
new file mode 100755
index 000000000..9e500f477
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/test_model.py
@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless/test_model.py
+"""
+
+import torch
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/train.py b/egs/librispeech/ASR/transducer_stateless/train.py
new file mode 100755
index 000000000..baeff6016
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/train.py
@@ -0,0 +1,772 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --full-libri 1 \
+  --max-duration 250 \
+  --lr-factor 2.5
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--modified-transducer-prob",
+        type=float,
+        default=0.25,
+        help="""The probability to use modified transducer loss.
+        In modified transduer, it limits the maximum number of symbols
+        per frame to 1. See also the option --max-sym-per-frame in
+        transducer_stateless/decode.py
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            modified_transducer_prob=params.modified_transducer_prob,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 8000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        return 1.0 <= c.duration <= 20.0
+
+    num_in_total = len(train_cuts)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless/transformer.py b/egs/librispeech/ASR/transducer_stateless/transformer.py
new file mode 100644
index 000000000..b3ff153c1
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless/transformer.py
@@ -0,0 +1,416 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+from icefall.utils import make_pad_mask
+
+
+class Transformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          dropout:
+            Dropout in encoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.output_dim = output_dim
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(0) == lengths.max().item()
+
+        mask = make_pad_mask(lengths)
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
diff --git a/egs/librispeech/ASR/transducer_stateless2/__init__.py b/egs/librispeech/ASR/transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/transducer_stateless2/asr_datamodule.py b/egs/librispeech/ASR/transducer_stateless2/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless2/beam_search.py b/egs/librispeech/ASR/transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..08cb32ef7
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless2/conformer.py b/egs/librispeech/ASR/transducer_stateless2/conformer.py
new file mode 120000
index 000000000..70a7ddf11
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless2/decode.py b/egs/librispeech/ASR/transducer_stateless2/decode.py
new file mode 100755
index 000000000..b05fe2a4d
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/decode.py
@@ -0,0 +1,520 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_stateless2/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless2/exp \
+        --max-duration 600 \
+        --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./transducer_stateless2/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless2/exp \
+        --max-duration 600 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./transducer_stateless2/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless2/exp \
+        --max-duration 600 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./transducer_stateless2/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless2/exp \
+        --max-duration 600 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless2/decoder.py b/egs/librispeech/ASR/transducer_stateless2/decoder.py
new file mode 120000
index 000000000..eada91097
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless2/encoder_interface.py b/egs/librispeech/ASR/transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless2/export.py b/egs/librispeech/ASR/transducer_stateless2/export.py
new file mode 100755
index 000000000..f4b6f5554
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/export.py
@@ -0,0 +1,183 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless2/export.py \
+  --exp-dir ./transducer_stateless2/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless2/decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./transducer_stateless2/decode.py \
+        --exp-dir ./transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless2/joiner.py b/egs/librispeech/ASR/transducer_stateless2/joiner.py
new file mode 100644
index 000000000..4882f9971
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/joiner.py
@@ -0,0 +1,72 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        unused_encoder_out_len: Optional[torch.Tensor] = None,
+        unused_decoder_out_len: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, self.input_dim).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, self.input_dim).
+          unused_encoder_out_len:
+            This is a placeholder so that we can reuse
+            transducer_stateless/beam_search.py in this folder as that
+            script assumes the joiner networks accepts 4 inputs.
+          unused_decoder_out_len:
+            Just a placeholder.
+        Returns:
+          Return a tensor of shape (N, T, U, self.output_dim).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == self.input_dim
+        assert decoder_out.size(2) == self.input_dim
+
+        encoder_out = encoder_out.unsqueeze(2)  # (N, T, 1, C)
+        decoder_out = decoder_out.unsqueeze(1)  # (N, 1, U, C)
+        x = encoder_out + decoder_out  # (N, T, U, C)
+
+        activations = torch.tanh(x)
+
+        logits = self.output_linear(activations)
+
+        if not self.training:
+            # We reuse the beam_search.py from transducer_stateless,
+            # which expects that the joiner network outputs
+            # a 2-D tensor.
+            logits = logits.squeeze(2).squeeze(1)
+
+        return logits
diff --git a/egs/librispeech/ASR/transducer_stateless2/model.py b/egs/librispeech/ASR/transducer_stateless2/model.py
new file mode 100644
index 000000000..d04716706
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/model.py
@@ -0,0 +1,130 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Note we use `rnnt_loss` from torchaudio, which exists only in
+torchaudio >= v0.10.0. It also means you have to use torch >= v1.10.0
+"""
+
+
+import k2
+import torch
+import torch.nn as nn
+import torchaudio
+import torchaudio.functional
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        decoder_out = self.decoder(sos_y_padded)
+
+        logits = self.joiner(
+            encoder_out=encoder_out,
+            decoder_out=decoder_out,
+        )
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        assert hasattr(torchaudio.functional, "rnnt_loss"), (
+            f"Current torchaudio version: {torchaudio.__version__}\n"
+            "Please install a version >= 0.10.0"
+        )
+
+        loss = torchaudio.functional.rnnt_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+        )
+
+        return loss
diff --git a/egs/librispeech/ASR/transducer_stateless2/pretrained.py b/egs/librispeech/ASR/transducer_stateless2/pretrained.py
new file mode 100755
index 000000000..b69b347ef
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/pretrained.py
@@ -0,0 +1,344 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless2/pretrained.py \
+    --checkpoint ./transducer_stateless2/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    --max-sym-per-frame 1 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless2/pretrained.py \
+    --checkpoint ./transducer_stateless2/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./transducer_stateless2/pretrained.py \
+    --checkpoint ./transducer_stateless2/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./transducer_stateless2/pretrained.py \
+    --checkpoint ./transducer_stateless2/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./transducer_stateless2/exp/pretrained.pt is generated by
+./transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = [token_ids_to_words(hyp) for hyp in hyp_list]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless2/subsampling.py b/egs/librispeech/ASR/transducer_stateless2/subsampling.py
new file mode 120000
index 000000000..af74db6e3
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/subsampling.py
@@ -0,0 +1 @@
+../transducer_stateless/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless2/test_model.py b/egs/librispeech/ASR/transducer_stateless2/test_model.py
new file mode 100755
index 000000000..bd2230a45
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/test_model.py
@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless2/test_model.py
+"""
+
+import torch
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless2/train.py b/egs/librispeech/ASR/transducer_stateless2/train.py
new file mode 100755
index 000000000..cca0d0e27
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/train.py
@@ -0,0 +1,760 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 250 \
+  --lr-factor 2.5
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 8000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        return 1.0 <= c.duration <= 20.0
+
+    num_in_total = len(train_cuts)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless2/transformer.py b/egs/librispeech/ASR/transducer_stateless2/transformer.py
new file mode 120000
index 000000000..e43f520f9
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless2/transformer.py
@@ -0,0 +1 @@
+../transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/README.md b/egs/librispeech/ASR/transducer_stateless_multi_datasets/README.md
new file mode 100644
index 000000000..574fbf78e
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/README.md
@@ -0,0 +1,27 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+You can use the following command to start the training:
+
+```bash
+cd egs/librispeech/ASR
+./prepare.sh
+./prepare_giga_speech.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./transducer_stateless_multi_datasets/train.py \
+  --world-size 2 \
+  --num-epochs 60 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_multi_datasets/exp-100 \
+  --full-libri 0 \
+  --max-duration 300 \
+  --lr-factor 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --modified-transducer-prob 0.25
+  --giga-prob 0.2
+```
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/__init__.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/asr_datamodule.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/asr_datamodule.py
new file mode 120000
index 000000000..3ba9ada4f
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/beam_search.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/beam_search.py
new file mode 120000
index 000000000..08cb32ef7
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/beam_search.py
@@ -0,0 +1 @@
+../transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/conformer.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/conformer.py
new file mode 120000
index 000000000..70a7ddf11
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/conformer.py
@@ -0,0 +1 @@
+../transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/decode.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/decode.py
new file mode 100755
index 000000000..5570b30ae
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/decode.py
@@ -0,0 +1,524 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_stateless_multi_datasets/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless_multi_datasets/exp \
+        --max-duration 600 \
+        --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./transducer_stateless_multi_datasets/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless_multi_datasets/exp \
+        --max-duration 600 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./transducer_stateless_multi_datasets/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless_multi_datasets/exp \
+        --max-duration 600 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./transducer_stateless_multi_datasets/decode.py \
+        --epoch 14 \
+        --avg 7 \
+        --exp-dir ./transducer_stateless_multi_datasets/exp \
+        --max-duration 600 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from librispeech import LibriSpeech
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_multi_datasets/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/decoder.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/decoder.py
new file mode 120000
index 000000000..eada91097
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/decoder.py
@@ -0,0 +1 @@
+../transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/encoder_interface.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/export.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/export.py
new file mode 100755
index 000000000..6d31dfe34
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/export.py
@@ -0,0 +1,254 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless_multi_datasets/export.py \
+  --exp-dir ./transducer_stateless_multi_datasets/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless_multi_datasets/decode.py`,
+you can do::
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./transducer_stateless_multi_datasets/decode.py \
+        --exp-dir ./transducer_stateless_multi_datasets/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_multi_datasets/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/gigaspeech.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/gigaspeech.py
new file mode 120000
index 000000000..5242c652a
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/gigaspeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/gigaspeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/joiner.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/joiner.py
new file mode 120000
index 000000000..cfc14f0a9
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/joiner.py
@@ -0,0 +1 @@
+../transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/librispeech.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/librispeech.py
new file mode 120000
index 000000000..b76723bf5
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/librispeech.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless3/librispeech.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/model.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/model.py
new file mode 100644
index 000000000..8141f9a83
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/model.py
@@ -0,0 +1,168 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+from typing import Optional
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        decoder_giga: Optional[nn.Module] = None,
+        joiner_giga: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          decoder_giga:
+            The decoder for the GigaSpeech dataset.
+          joiner_giga:
+            The joiner for the GigaSpeech dataset.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        if decoder_giga is not None:
+            assert hasattr(decoder_giga, "blank_id")
+
+        self.encoder = encoder
+
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_giga = decoder_giga
+        self.joiner_giga = joiner_giga
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        libri: bool = True,
+        modified_transducer_prob: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          libri:
+            True to use the decoder and joiner for the LibriSpeech dataset.
+            False to use the decoder and joiner for the GigaSpeech dataset.
+          modified_transducer_prob:
+            The probability to use modified transducer loss.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        if libri:
+            decoder = self.decoder
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_giga
+            joiner = self.joiner_giga
+
+        decoder_out = decoder(sos_y_padded)
+
+        # +1 here since a blank is prepended to each utterance.
+        logits = joiner(
+            encoder_out=encoder_out,
+            decoder_out=decoder_out,
+            encoder_out_len=x_lens,
+            decoder_out_len=y_lens + 1,
+        )
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        # We don't put this `import` at the beginning of the file
+        # as it is required only in the training, not during the
+        # reference stage
+        import optimized_transducer
+
+        assert 0 <= modified_transducer_prob <= 1
+
+        if modified_transducer_prob == 0:
+            one_sym_per_frame = False
+        elif random.random() < modified_transducer_prob:
+            # random.random() returns a float in the range [0, 1)
+            one_sym_per_frame = True
+        else:
+            one_sym_per_frame = False
+
+        loss = optimized_transducer.transducer_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+            one_sym_per_frame=one_sym_per_frame,
+            from_log_softmax=False,
+        )
+
+        return loss
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/pretrained.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/pretrained.py
new file mode 100755
index 000000000..4f29d6f1f
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/pretrained.py
@@ -0,0 +1,344 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless_multi_datasets/pretrained.py \
+    --checkpoint ./transducer_stateless_multi_datasets/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method greedy_search \
+    --max-sym-per-frame 1 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless_multi_datasets/pretrained.py \
+    --checkpoint ./transducer_stateless_multi_datasets/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./transducer_stateless_multi_datasets/pretrained.py \
+    --checkpoint ./transducer_stateless_multi_datasets/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./transducer_stateless_multi_datasets/pretrained.py \
+    --checkpoint ./transducer_stateless_multi_datasets/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./transducer_stateless_multi_datasets/exp/epoch-xx.pt`.
+
+Note: ./transducer_stateless_multi_datasets/exp/pretrained.pt is generated by
+./transducer_stateless_multi_datasets/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.utils import num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+            hyp_list.append(hyp)
+
+    hyps = [token_ids_to_words(hyp) for hyp in hyp_list]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/subsampling.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/subsampling.py
new file mode 120000
index 000000000..73068da26
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/subsampling.py
@@ -0,0 +1 @@
+../transducer/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/test_asr_datamodule.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/test_asr_datamodule.py
new file mode 100755
index 000000000..1e1188ca6
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/test_asr_datamodule.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless_multi_datasets/test_asr_datamodule.py
+"""
+
+import argparse
+import random
+from pathlib import Path
+
+from asr_datamodule import AsrDataModule
+from gigaspeech import GigaSpeech
+from lhotse import load_manifest
+from librispeech import LibriSpeech
+
+
+def test_dataset():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    print(args)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+
+    train_clean_100 = librispeech.train_clean_100_cuts()
+    train_S = gigaspeech.train_S_cuts()
+
+    asr_datamodule = AsrDataModule(args)
+
+    libri_train_dl = asr_datamodule.train_dataloaders(
+        train_clean_100,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    giga_train_dl = asr_datamodule.train_dataloaders(
+        train_S,
+        on_the_fly_feats=True,
+        cuts_musan=cuts_musan,
+    )
+
+    seed = 20220216
+    rng = random.Random(seed)
+
+    for epoch in range(2):
+        print("epoch", epoch)
+        batch_idx = 0
+        libri_train_dl.sampler.set_epoch(epoch)
+        giga_train_dl.sampler.set_epoch(epoch)
+
+        iter_libri = iter(libri_train_dl)
+        iter_giga = iter(giga_train_dl)
+        while True:
+            idx = rng.choices((0, 1), weights=[0.8, 0.2], k=1)[0]
+            dl = iter_libri if idx == 0 else iter_giga
+            batch_idx += 1
+
+            print("dl idx", idx, "batch_idx", batch_idx)
+            try:
+                _ = next(dl)
+            except StopIteration:
+                print("dl idx", idx)
+                print("Go to the next epoch")
+                break
+
+
+def main():
+    test_dataset()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/test_decoder.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/test_decoder.py
new file mode 100755
index 000000000..9ee197ee8
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/test_decoder.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./transducer_stateless_multi_datasets/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, embedding_dim)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, embedding_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/train.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/train.py
new file mode 100755
index 000000000..dafccd088
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/train.py
@@ -0,0 +1,914 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+cd egs/librispeech/ASR/
+./prepare.sh
+./prepare_giga_speech.sh
+
+# 100-hours
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./transducer_stateless_multi_datasets/train.py \
+  --world-size 2 \
+  --num-epochs 60 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_multi_datasets/exp-100-2 \
+  --full-libri 0 \
+  --max-duration 300 \
+  --lr-factor 1 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --modified-transducer-prob 0.25
+  --giga-prob 0.2
+
+# 960-hours
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless_multi_datasets/train.py \
+  --world-size 4 \
+  --num-epochs 40 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless_multi_datasets/exp-full-2 \
+  --full-libri 1 \
+  --max-duration 300 \
+  --lr-factor 5 \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --modified-transducer-prob 0.25 \
+  --giga-prob 0.2
+"""
+
+
+import argparse
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from gigaspeech import GigaSpeech
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from librispeech import LibriSpeech
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--full-libri",
+        type=str2bool,
+        default=True,
+        help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless_multi_datasets/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--modified-transducer-prob",
+        type=float,
+        default=0.25,
+        help="""The probability to use modified transducer loss.
+        In modified transduer, it limits the maximum number of symbols
+        per frame to 1. See also the option --max-sym-per-frame in
+        transducer_stateless/decode.py
+        """,
+    )
+
+    parser.add_argument(
+        "--giga-prob",
+        type=float,
+        default=0.2,
+        help="The probability to select a batch from the GigaSpeech dataset",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    decoder_giga = get_decoder_model(params)
+    joiner_giga = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        decoder_giga=decoder_giga,
+        joiner_giga=joiner_giga,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_libri(c: Cut) -> bool:
+    """Return True if this cut is from the LibriSpeech dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of GigaSpeech to dict(origin='giga')
+      See ../local/preprocess_gigaspeech.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    libri = is_libri(supervisions["cut"][0])
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            libri=libri,
+            modified_transducer_prob=params.modified_transducer_prob,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    giga_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      rng:
+        For select which dataset to use.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    libri_tot_loss = MetricsTracker()
+    giga_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.giga_prob, params.giga_prob]
+
+    iter_libri = iter(train_dl)
+    iter_giga = iter(giga_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_libri if idx == 0 else iter_giga
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            break
+
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        libri = is_libri(batch["supervisions"]["cut"][0])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+        if libri:
+            libri_tot_loss = (
+                libri_tot_loss * (1 - 1 / params.reset_interval)
+            ) + loss_info
+            prefix = "libri"  # for logging only
+        else:
+            giga_tot_loss = (
+                giga_tot_loss * (1 - 1 / params.reset_interval)
+            ) + loss_info
+            prefix = "giga"
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], "
+                f"libri_tot_loss[{libri_tot_loss}], "
+                f"giga_tot_loss[{giga_tot_loss}], "
+                f"batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                libri_tot_loss.write_summary(
+                    tb_writer, "train/libri_tot_", params.batch_idx_train
+                )
+                giga_tot_loss.write_summary(
+                    tb_writer, "train/giga_tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(cuts: CutSet) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    cuts = cuts.filter(remove_short_and_long_utt)
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 8000
+
+    seed = 42
+    fix_random_seed(seed)
+    rng = random.Random(seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    train_cuts = filter_short_and_long_utterances(train_cuts)
+
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+    # XL 10k hours
+    # L  2.5k hours
+    # M  1k hours
+    # S  250 hours
+    # XS 10 hours
+    # DEV 12 hours
+    # Test 40 hours
+    if params.full_libri:
+        logging.info("Using the XL subset of GigaSpeech (10k hours)")
+        train_giga_cuts = gigaspeech.train_XL_cuts()
+    else:
+        logging.info("Using the S subset of GigaSpeech (250 hours)")
+        train_giga_cuts = gigaspeech.train_S_cuts()
+
+    train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts)
+    train_giga_cuts = train_giga_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(Path(args.manifest_dir) / "musan_cuts.jsonl.gz")
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    giga_train_dl = asr_datamodule.train_dataloaders(
+        train_giga_cuts,
+        on_the_fly_feats=True,
+        cuts_musan=cuts_musan,
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    # It's time consuming to include `giga_train_dl` here
+    #  for dl in [train_dl, giga_train_dl]:
+    for dl in [train_dl]:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+        giga_train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            giga_train_dl=giga_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert 0 <= args.giga_prob < 1, args.giga_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/transducer_stateless_multi_datasets/transformer.py b/egs/librispeech/ASR/transducer_stateless_multi_datasets/transformer.py
new file mode 120000
index 000000000..e43f520f9
--- /dev/null
+++ b/egs/librispeech/ASR/transducer_stateless_multi_datasets/transformer.py
@@ -0,0 +1 @@
+../transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer/.gitignore b/egs/librispeech/ASR/zipformer/.gitignore
new file mode 100644
index 000000000..e47ac1582
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/.gitignore
@@ -0,0 +1 @@
+swoosh.pdf
diff --git a/egs/librispeech/ASR/zipformer/asr_datamodule.py b/egs/librispeech/ASR/zipformer/asr_datamodule.py
new file mode 120000
index 000000000..07f39b451
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer/beam_search.py b/egs/librispeech/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer/ctc_decode.py b/egs/librispeech/ASR/zipformer/ctc_decode.py
new file mode 100755
index 000000000..4db50b981
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/ctc_decode.py
@@ -0,0 +1,847 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) ctc-decoding
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) 1best
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method 1best
+
+(3) nbest
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method nbest
+
+(4) nbest-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+
+(5) whole-lattice-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.6,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/decode.py b/egs/librispeech/ASR/zipformer/decode.py
new file mode 100755
index 000000000..339e253e6
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/decode.py
@@ -0,0 +1,1052 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_rescore,
+    modified_beam_search_lm_rescore_LODR,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+)
+from train import add_model_arguments, get_model, get_params
+
+from icefall import ContextGraph, LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - modified_beam_search_LODR
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding-method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding-method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=2,
+        help="""The order of the ngram lm.
+        """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="ID of the backoff symbol in the ngram LM",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding-method is modified_beam_search and
+        modified_beam_search_LODR.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding-method is modified_beam_search and
+        modified_beam_search_LODR.
+        """,
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding-method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network language model.
+      ngram_lm:
+        A ngram language model
+      ngram_lm_scale:
+        The scale for the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+            context_graph=context_graph,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_rescore":
+        lm_scale_list = [0.01 * i for i in range(10, 50)]
+        ans_dict = modified_beam_search_lm_rescore(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        lm_scale_list = [0.02 * i for i in range(2, 30)]
+        ans_dict = modified_beam_search_lm_rescore_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            LODR_lm=ngram_lm,
+            sp=sp,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    elif "modified_beam_search" in params.decoding_method:
+        prefix = f"beam_size_{params.beam_size}"
+        if params.decoding_method in (
+            "modified_beam_search_lm_rescore",
+            "modified_beam_search_lm_rescore_LODR",
+        ):
+            ans = dict()
+            assert ans_dict is not None
+            for key, hyps in ans_dict.items():
+                hyps = [sp.decode(hyp).split() for hyp in hyps]
+                ans[f"{prefix}_{key}"] = hyps
+            return ans
+        else:
+            if params.has_contexts:
+                prefix += f"-context-score-{params.context_score}"
+            return {prefix: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding-method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            word_table=word_table,
+            batch=batch,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.decoding_method in (
+            "modified_beam_search",
+            "modified_beam_search_LODR",
+        ):
+            if params.has_contexts:
+                params.suffix += f"-context-score-{params.context_score}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_shallow_fusion:
+        params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load the neural network LM if required
+    if params.use_shallow_fusion or params.decoding_method in (
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    ):
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+    else:
+        LM = None
+
+    # only load N-gram LM when needed
+    if params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        try:
+            import kenlm
+        except ImportError:
+            print("Please install kenlm first. You can use")
+            print(" pip install https://github.com/kpu/kenlm/archive/master.zip")
+            print("to install it")
+            import sys
+
+            sys.exit(-1)
+        ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa")
+        logging.info(f"lm filename: {ngram_file_name}")
+        ngram_lm = kenlm.Model(ngram_file_name)
+        ngram_lm_scale = None  # use a list to search
+
+    elif params.decoding_method == "modified_beam_search_LODR":
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"Loading token level lm: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    if "modified_beam_search" in params.decoding_method:
+        if os.path.exists(params.context_file):
+            contexts = []
+            for line in open(params.context_file).readlines():
+                contexts.append((sp.encode(line.strip()), 0.0))
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/decode_stream.py b/egs/librispeech/ASR/zipformer/decode_stream.py
new file mode 100644
index 000000000..d6918bf32
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/decode_stream.py
@@ -0,0 +1,148 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class DecodeStream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        initial_states: List[torch.Tensor],
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+    ) -> None:
+        """
+        Args:
+          initial_states:
+            Initial decode states of the model, e.g. the return value of
+            `get_init_state` in conformer.py
+          decoding_graph:
+            Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+            Used only when decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+        """
+        if params.decoding_method == "fast_beam_search":
+            assert decoding_graph is not None
+            assert device == decoding_graph.device
+
+        self.params = params
+        self.cut_id = cut_id
+        self.LOG_EPS = math.log(1e-10)
+
+        self.states = initial_states
+
+        # It contains a 2-D tensors representing the feature frames.
+        self.features: torch.Tensor = None
+
+        self.num_frames: int = 0
+        # how many frames have been processed. (before subsampling).
+        # we only modify this value in `func:get_feature_frames`.
+        self.num_processed_frames: int = 0
+
+        self._done: bool = False
+
+        # The transcript of current utterance.
+        self.ground_truth: str = ""
+
+        # The decoding result (partial or final) of current utterance.
+        self.hyp: List = []
+
+        # how many frames have been processed, at encoder output
+        self.done_frames: int = 0
+
+        # The encoder_embed subsample features (T - 7) // 2
+        # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+        self.pad_length = 7 + 2 * 3
+
+        if params.decoding_method == "greedy_search":
+            self.hyp = [-1] * (params.context_size - 1) + [params.blank_id]
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[-1] * (params.context_size - 1) + [params.blank_id],
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    @property
+    def done(self) -> bool:
+        """Return True if all the features are processed."""
+        return self._done
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    def set_features(
+        self,
+        features: torch.Tensor,
+        tail_pad_len: int = 0,
+    ) -> None:
+        """Set features tensor of current utterance."""
+        assert features.dim() == 2, features.dim()
+        self.features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, self.pad_length + tail_pad_len),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+        self.num_frames = self.features.size(0)
+
+    def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]:
+        """Consume chunk_size frames of features"""
+        chunk_length = chunk_size + self.pad_length
+
+        ret_length = min(self.num_frames - self.num_processed_frames, chunk_length)
+
+        ret_features = self.features[
+            self.num_processed_frames : self.num_processed_frames + ret_length  # noqa
+        ]
+
+        self.num_processed_frames += chunk_size
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_features, ret_length
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.params.decoding_method == "greedy_search":
+            return self.hyp[self.params.context_size :]  # noqa
+        elif self.params.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.params.context_size :]  # noqa
+        else:
+            assert self.params.decoding_method == "fast_beam_search"
+            return self.hyp
diff --git a/egs/librispeech/ASR/zipformer/decoder.py b/egs/librispeech/ASR/zipformer/decoder.py
new file mode 100644
index 000000000..7ce44495b
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/decoder.py
@@ -0,0 +1,134 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import Balancer
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+        )
+        # the balancers are to avoid any drift in the magnitude of the
+        # embeddings, which would interact badly with parameter averaging.
+        self.balancer = Balancer(
+            decoder_dim,
+            channel_dim=-1,
+            min_positive=0.0,
+            max_positive=1.0,
+            min_abs=0.5,
+            max_abs=1.0,
+            prob=0.05,
+        )
+
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim // 4,  # group size == 4
+                bias=False,
+            )
+            self.balancer2 = Balancer(
+                decoder_dim,
+                channel_dim=-1,
+                min_positive=0.0,
+                max_positive=1.0,
+                min_abs=0.5,
+                max_abs=1.0,
+                prob=0.05,
+            )
+        else:
+            # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'`
+            # when inference with torch.jit.script and context_size == 1
+            self.conv = nn.Identity()
+            self.balancer2 = nn.Identity()
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+
+        embedding_out = self.balancer(embedding_out)
+
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+            embedding_out = F.relu(embedding_out)
+            embedding_out = self.balancer2(embedding_out)
+
+        return embedding_out
diff --git a/egs/librispeech/ASR/zipformer/encoder_interface.py b/egs/librispeech/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..aa5d0217a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer/export-onnx-ctc.py b/egs/librispeech/ASR/zipformer/export-onnx-ctc.py
new file mode 100755
index 000000000..3345d20d3
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/export-onnx-ctc.py
@@ -0,0 +1,436 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a CTC model from PyTorch to ONNX.
+
+Note that the model is trained using both transducer and CTC loss. This script
+exports only the CTC head.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx-ctc.py \
+  --use-transducer 0 \
+  --use-ctc 1 \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal False \
+  --chunk-size 16 \
+  --left-context-frames 128
+
+It will generate the following 2 files inside $repo/exp:
+
+  - model.onnx
+  - model.int8.onnx
+
+See ./onnx_pretrained_ctc.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxModel(nn.Module):
+    """A wrapper for encoder_embed, Zipformer, and ctc_output layer"""
+
+    def __init__(
+        self,
+        encoder: Zipformer2,
+        encoder_embed: nn.Module,
+        ctc_output: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_embed:
+            The first downsampling layer for zipformer.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.ctc_output = ctc_output
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - log_probs, a 3-D tensor of shape (N, T', vocab_size)
+            - log_probs_len, a 1-D int64 tensor of shape (N,)
+        """
+        x, x_lens = self.encoder_embed(x, x_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)
+        encoder_out, log_probs_len = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)
+        log_probs = self.ctc_output(encoder_out)
+
+        return log_probs, log_probs_len
+
+
+def export_ctc_model_onnx(
+    model: OnnxModel,
+    filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - log_probs, a tensor of shape (N, T', joiner_dim)
+        - log_probs_len, a tensor of shape (N,)
+
+    Args:
+      model:
+        The input model
+      filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    model = torch.jit.trace(model, (x, x_lens))
+
+    torch.onnx.export(
+        model,
+        (x, x_lens),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["log_probs", "log_probs_len"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "log_probs": {0: "N", 1: "T"},
+            "log_probs_len": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer2_ctc",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "non-streaming zipformer2 CTC",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(
+                f"{params.exp_dir}/epoch-{params.epoch}.pt", model, strict=False
+            )
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    model = OnnxModel(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        ctc_output=model.ctc_output,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"num parameters: {num_param}")
+
+    opset_version = 13
+
+    logging.info("Exporting ctc model")
+    filename = params.exp_dir / f"model.onnx"
+    export_ctc_model_onnx(
+        model,
+        filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported to {filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    filename_int8 = params.exp_dir / f"model.int8.onnx"
+    quantize_dynamic(
+        model_input=filename,
+        model_output=filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/export-onnx-streaming-ctc.py b/egs/librispeech/ASR/zipformer/export-onnx-streaming-ctc.py
new file mode 100755
index 000000000..3c0f74005
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/export-onnx-streaming-ctc.py
@@ -0,0 +1,570 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang, Zengrui Jin)
+# Copyright 2023 Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script exports a CTC model from PyTorch to ONNX.
+
+
+1. Download the pre-trained streaming model with CTC head
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx-streaming-ctc.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal True \
+  --chunk-size 16 \
+  --left-context-frames 64 \
+  --use-ctc 1
+
+The --chunk-size in training is "16,32,64,-1", so we select one of them
+(excluding -1) during streaming export. The same applies to `--left-context`,
+whose value is "64,128,256,-1".
+
+It will generate the following file inside $repo/exp:
+
+  - ctc-epoch-99-avg-1-chunk-16-left-64.onnx
+
+See ./onnx_pretrained-streaming-ctc.py for how to use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxModel(nn.Module):
+    """A wrapper for Zipformer and the ctc_head"""
+
+    def __init__(
+        self,
+        encoder: Zipformer2,
+        encoder_embed: nn.Module,
+        ctc_output: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+          ctc_output:
+            The ctc head.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.ctc_output = ctc_output
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+        self.pad_length = 7 + 2 * 3
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        N = x.size(0)
+        T = self.chunk_size * 2 + self.pad_length
+        x_lens = torch.tensor([T] * N, device=x.device)
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == self.chunk_size, (x.size(1), self.chunk_size)
+
+        src_key_padding_mask = torch.zeros(N, self.chunk_size, dtype=torch.bool)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)
+        encoder_states = states[:-2]
+        logging.info(f"len_encoder_states={len(encoder_states)}")
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)
+        encoder_out = self.ctc_output(encoder_out)
+        # Now encoder_out is of shape (N, T, ctc_output_dim)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+
+        return encoder_out, new_states
+
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int64, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+def export_streaming_ctc_model_onnx(
+    model: OnnxModel,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    model.encoder.__class__.forward = model.encoder.__class__.streaming_forward
+
+    decode_chunk_len = model.chunk_size * 2
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    T = decode_chunk_len + model.pad_length
+
+    x = torch.rand(1, T, 80, dtype=torch.float32)
+    init_state = model.get_init_states()
+    num_encoders = len(model.encoder.encoder_dim)
+    logging.info(f"num_encoders: {num_encoders}")
+    logging.info(f"len(init_state): {len(init_state)}")
+
+    inputs = {}
+    input_names = ["x"]
+
+    outputs = {}
+    output_names = ["log_probs"]
+
+    def build_inputs_outputs(tensors, i):
+        assert len(tensors) == 6, len(tensors)
+
+        # (downsample_left, batch_size, key_dim)
+        name = f"cached_key_{i}"
+        logging.info(f"{name}.shape: {tensors[0].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (1, batch_size, downsample_left, nonlin_attn_head_dim)
+        name = f"cached_nonlin_attn_{i}"
+        logging.info(f"{name}.shape: {tensors[1].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (downsample_left, batch_size, value_dim)
+        name = f"cached_val1_{i}"
+        logging.info(f"{name}.shape: {tensors[2].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (downsample_left, batch_size, value_dim)
+        name = f"cached_val2_{i}"
+        logging.info(f"{name}.shape: {tensors[3].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (batch_size, embed_dim, conv_left_pad)
+        name = f"cached_conv1_{i}"
+        logging.info(f"{name}.shape: {tensors[4].shape}")
+        inputs[name] = {0: "N"}
+        outputs[f"new_{name}"] = {0: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (batch_size, embed_dim, conv_left_pad)
+        name = f"cached_conv2_{i}"
+        logging.info(f"{name}.shape: {tensors[5].shape}")
+        inputs[name] = {0: "N"}
+        outputs[f"new_{name}"] = {0: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+    num_encoder_layers = ",".join(map(str, model.encoder.num_encoder_layers))
+    encoder_dims = ",".join(map(str, model.encoder.encoder_dim))
+    cnn_module_kernels = ",".join(map(str, model.encoder.cnn_module_kernel))
+    ds = model.encoder.downsampling_factor
+    left_context_len = model.left_context_len
+    left_context_len = [left_context_len // k for k in ds]
+    left_context_len = ",".join(map(str, left_context_len))
+    query_head_dims = ",".join(map(str, model.encoder.query_head_dim))
+    value_head_dims = ",".join(map(str, model.encoder.value_head_dim))
+    num_heads = ",".join(map(str, model.encoder.num_heads))
+
+    meta_data = {
+        "model_type": "zipformer2",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "streaming ctc zipformer2",
+        "decode_chunk_len": str(decode_chunk_len),  # 32
+        "T": str(T),  # 32+7+2*3=45
+        "num_encoder_layers": num_encoder_layers,
+        "encoder_dims": encoder_dims,
+        "cnn_module_kernels": cnn_module_kernels,
+        "left_context_len": left_context_len,
+        "query_head_dims": query_head_dims,
+        "value_head_dims": value_head_dims,
+        "num_heads": num_heads,
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    for i in range(len(init_state[:-2]) // 6):
+        build_inputs_outputs(init_state[i * 6 : (i + 1) * 6], i)
+
+    # (batch_size, channels, left_pad, freq)
+    embed_states = init_state[-2]
+    name = "embed_states"
+    logging.info(f"{name}.shape: {embed_states.shape}")
+    inputs[name] = {0: "N"}
+    outputs[f"new_{name}"] = {0: "N"}
+    input_names.append(name)
+    output_names.append(f"new_{name}")
+
+    # (batch_size,)
+    processed_lens = init_state[-1]
+    name = "processed_lens"
+    logging.info(f"{name}.shape: {processed_lens.shape}")
+    inputs[name] = {0: "N"}
+    outputs[f"new_{name}"] = {0: "N"}
+    input_names.append(name)
+    output_names.append(f"new_{name}")
+
+    logging.info(inputs)
+    logging.info(outputs)
+    logging.info(input_names)
+    logging.info(output_names)
+
+    torch.onnx.export(
+        model,
+        (x, init_state),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes={
+            "x": {0: "N"},
+            "log_probs": {0: "N"},
+            **inputs,
+            **outputs,
+        },
+    )
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    model = OnnxModel(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        ctc_output=model.ctc_output,
+    )
+
+    total_num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+    suffix += f"-chunk-{params.chunk_size}"
+    suffix += f"-left-{params.left_context_frames}"
+
+    opset_version = 13
+
+    logging.info("Exporting model")
+    model_filename = params.exp_dir / f"ctc-{suffix}.onnx"
+    export_streaming_ctc_model_onnx(
+        model,
+        model_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported model to {model_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    model_filename_int8 = params.exp_dir / f"ctc-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=model_filename,
+        model_output=model_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py
new file mode 100755
index 000000000..6bc9b1858
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1,781 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang)
+# Copyright 2023 Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx-streaming.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal True \
+  --chunk-size 16 \
+  --left-context-frames 64
+
+The --chunk-size in training is "16,32,64,-1", so we select one of them
+(excluding -1) during streaming export. The same applies to `--left-context`,
+whose value is "64,128,256,-1".
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1-chunk-16-left-64.onnx
+  - decoder-epoch-99-avg-1-chunk-16-left-64.onnx
+  - joiner-epoch-99-avg-1-chunk-16-left-64.onnx
+
+See ./onnx_pretrained-streaming.py for how to use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(
+        self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.encoder_proj = encoder_proj
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+        self.pad_length = 7 + 2 * 3
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        states: List[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        N = x.size(0)
+        T = self.chunk_size * 2 + self.pad_length
+        x_lens = torch.tensor([T] * N, device=x.device)
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == self.chunk_size, (x.size(1), self.chunk_size)
+
+        src_key_padding_mask = torch.zeros(N, self.chunk_size, dtype=torch.bool)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)
+        encoder_states = states[:-2]
+        logging.info(f"len_encoder_states={len(encoder_states)}")
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+
+        return encoder_out, new_states
+
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int64, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    encoder_model.encoder.__class__.forward = (
+        encoder_model.encoder.__class__.streaming_forward
+    )
+
+    decode_chunk_len = encoder_model.chunk_size * 2
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    T = decode_chunk_len + encoder_model.pad_length
+
+    x = torch.rand(1, T, 80, dtype=torch.float32)
+    init_state = encoder_model.get_init_states()
+    num_encoders = len(encoder_model.encoder.encoder_dim)
+    logging.info(f"num_encoders: {num_encoders}")
+    logging.info(f"len(init_state): {len(init_state)}")
+
+    inputs = {}
+    input_names = ["x"]
+
+    outputs = {}
+    output_names = ["encoder_out"]
+
+    def build_inputs_outputs(tensors, i):
+        assert len(tensors) == 6, len(tensors)
+
+        # (downsample_left, batch_size, key_dim)
+        name = f"cached_key_{i}"
+        logging.info(f"{name}.shape: {tensors[0].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (1, batch_size, downsample_left, nonlin_attn_head_dim)
+        name = f"cached_nonlin_attn_{i}"
+        logging.info(f"{name}.shape: {tensors[1].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (downsample_left, batch_size, value_dim)
+        name = f"cached_val1_{i}"
+        logging.info(f"{name}.shape: {tensors[2].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (downsample_left, batch_size, value_dim)
+        name = f"cached_val2_{i}"
+        logging.info(f"{name}.shape: {tensors[3].shape}")
+        inputs[name] = {1: "N"}
+        outputs[f"new_{name}"] = {1: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (batch_size, embed_dim, conv_left_pad)
+        name = f"cached_conv1_{i}"
+        logging.info(f"{name}.shape: {tensors[4].shape}")
+        inputs[name] = {0: "N"}
+        outputs[f"new_{name}"] = {0: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+        # (batch_size, embed_dim, conv_left_pad)
+        name = f"cached_conv2_{i}"
+        logging.info(f"{name}.shape: {tensors[5].shape}")
+        inputs[name] = {0: "N"}
+        outputs[f"new_{name}"] = {0: "N"}
+        input_names.append(name)
+        output_names.append(f"new_{name}")
+
+    num_encoder_layers = ",".join(map(str, encoder_model.encoder.num_encoder_layers))
+    encoder_dims = ",".join(map(str, encoder_model.encoder.encoder_dim))
+    cnn_module_kernels = ",".join(map(str, encoder_model.encoder.cnn_module_kernel))
+    ds = encoder_model.encoder.downsampling_factor
+    left_context_len = encoder_model.left_context_len
+    left_context_len = [left_context_len // k for k in ds]
+    left_context_len = ",".join(map(str, left_context_len))
+    query_head_dims = ",".join(map(str, encoder_model.encoder.query_head_dim))
+    value_head_dims = ",".join(map(str, encoder_model.encoder.value_head_dim))
+    num_heads = ",".join(map(str, encoder_model.encoder.num_heads))
+
+    meta_data = {
+        "model_type": "zipformer2",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "streaming zipformer2",
+        "decode_chunk_len": str(decode_chunk_len),  # 32
+        "T": str(T),  # 32+7+2*3=45
+        "num_encoder_layers": num_encoder_layers,
+        "encoder_dims": encoder_dims,
+        "cnn_module_kernels": cnn_module_kernels,
+        "left_context_len": left_context_len,
+        "query_head_dims": query_head_dims,
+        "value_head_dims": value_head_dims,
+        "num_heads": num_heads,
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    for i in range(len(init_state[:-2]) // 6):
+        build_inputs_outputs(init_state[i * 6 : (i + 1) * 6], i)
+
+    # (batch_size, channels, left_pad, freq)
+    embed_states = init_state[-2]
+    name = "embed_states"
+    logging.info(f"{name}.shape: {embed_states.shape}")
+    inputs[name] = {0: "N"}
+    outputs[f"new_{name}"] = {0: "N"}
+    input_names.append(name)
+    output_names.append(f"new_{name}")
+
+    # (batch_size,)
+    processed_lens = init_state[-1]
+    name = "processed_lens"
+    logging.info(f"{name}.shape: {processed_lens.shape}")
+    inputs[name] = {0: "N"}
+    outputs[f"new_{name}"] = {0: "N"}
+    input_names.append(name)
+    output_names.append(f"new_{name}")
+
+    logging.info(inputs)
+    logging.info(outputs)
+    logging.info(input_names)
+    logging.info(output_names)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, init_state),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes={
+            "x": {0: "N"},
+            "encoder_out": {0: "N"},
+            **inputs,
+            **outputs,
+        },
+    )
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+    suffix += f"-chunk-{params.chunk_size}"
+    suffix += f"-left-{params.left_context_frames}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/export-onnx.py b/egs/librispeech/ASR/zipformer/export-onnx.py
new file mode 100755
index 000000000..3682f0b62
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/export-onnx.py
@@ -0,0 +1,620 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang)
+# Copyright 2023 Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal False \
+  --chunk-size "16,32,64,-1" \
+  --left-context-frames "64,128,256,-1"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(
+        self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        x, x_lens = self.encoder_embed(x, x_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    encoder_model = torch.jit.trace(encoder_model, (x, x_lens))
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer2",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "non-streaming zipformer2",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/export.py b/egs/librispeech/ASR/zipformer/export.py
new file mode 100755
index 000000000..2b8d1aaf3
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/export.py
@@ -0,0 +1,526 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+Note: This is a example for librispeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+(1) Export to torchscript model using torch.jit.script()
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `jit_script.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("jit_script.pt")`.
+
+Check ./jit_pretrained.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`.
+You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`.
+
+Check ./jit_pretrained_streaming.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+- For non-streaming model:
+
+To use the generated file with `zipformer/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+- For streaming model:
+
+To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+
+    # simulated streaming decoding
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+    # chunk-wise streaming decoding
+    ./zipformer/streaming_decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+- non-streaming model:
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+
+- streaming model:
+https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+    # You will find the pre-trained models in exp dir
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor, nn
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class EncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+        """
+        x, x_lens = self.encoder_embed(features, feature_lengths)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+class StreamingEncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        assert len(encoder.chunk_size) == 1, encoder.chunk_size
+        assert len(encoder.left_context_frames) == 1, encoder.left_context_frames
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+
+        # The encoder_embed subsample features (T - 7) // 2
+        # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+        self.pad_length = 7 + 2 * 3
+
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor, states: List[Tensor]
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """Streaming forward for encoder_embed and encoder.
+
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+            states: a list of Tensors
+
+        Returns encoder outputs, output lengths, and updated states.
+        """
+        chunk_size = self.chunk_size
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        encoder_states = states[:-2]
+
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+        return encoder_out, encoder_out_lens, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+
+        # Wrap encoder and encoder_embed as a module
+        if params.causal:
+            model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed)
+            chunk_size = model.encoder.chunk_size
+            left_context_len = model.encoder.left_context_len
+            filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt"
+        else:
+            model.encoder = EncoderModel(model.encoder, model.encoder_embed)
+            filename = "jit_script.pt"
+
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        model.save(str(params.exp_dir / filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/generate_averaged_model.py b/egs/librispeech/ASR/zipformer/generate_averaged_model.py
new file mode 100755
index 000000000..68111fad7
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/generate_averaged_model.py
@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) use the checkpoint exp_dir/epoch-xxx.pt
+./zipformer/generate_averaged_model.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp
+
+It will generate a file `epoch-28-avg-15.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15.pt")`.
+
+(2) use the checkpoint exp_dir/checkpoint-iter.pt
+./zipformer/generate_averaged_model.py \
+    --iter 22000 \
+    --avg 5 \
+    --exp-dir ./zipformer/exp
+
+It will generate a file `iter-22000-avg-5.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5.pt")`.
+"""
+
+
+import argparse
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import average_checkpoints_with_averaged_model, find_checkpoints
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    print("Script started")
+
+    device = torch.device("cpu")
+    print(f"Device: {device}")
+
+    symbol_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = symbol_table["<blk>"]
+    params.unk_id = symbol_table["<unk>"]
+    params.vocab_size = len(symbol_table)
+
+    print("About to create model")
+    model = get_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg + 1
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg + 1:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        filename_start = filenames[-1]
+        filename_end = filenames[0]
+        print(
+            "Calculating the averaged model over iteration checkpoints"
+            f" from {filename_start} (excluded) to {filename_end}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+        filename = params.exp_dir / f"iter-{params.iter}-avg-{params.avg}.pt"
+        torch.save({"model": model.state_dict()}, filename)
+    else:
+        assert params.avg > 0, params.avg
+        start = params.epoch - params.avg
+        assert start >= 1, start
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        print(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+        filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+        torch.save({"model": model.state_dict()}, filename)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    print("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/jit_pretrained.py b/egs/librispeech/ASR/zipformer/jit_pretrained.py
new file mode 100755
index 000000000..a41fbc1c9
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1,280 @@
+#!/usr/bin/env python3
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+Usage of this script:
+
+./zipformer/jit_pretrained.py \
+  --nn-model-filename ./zipformer/exp/cpu_jit.pt \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = model.decoder.blank_id
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        features=features,
+        feature_lengths=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+
+    s = "\n"
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = token_ids_to_words(hyp)
+        s += f"{filename}:\n{words}\n"
+
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py b/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py
new file mode 100755
index 000000000..660a4bfc6
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/jit_pretrained_ctc.py
@@ -0,0 +1,436 @@
+#!/usr/bin/env python3
+# Copyright      2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                         Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --causal 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+Usage of this script:
+
+(1) ctc-decoding
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) whole-lattice-rescoring
+./zipformer/jit_pretrained_ctc.py \
+  --model-filename ./zipformer/exp/jit_script.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from export import num_tokens
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a token table,
+            i.e., lang_dir/token.txt, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + nbest n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + whole-lattice n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.model_filename)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(features, feature_lengths)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    batch_size = ctc_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i].item() // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        max_token_id = params.vocab_size - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = [[token_table[i] for i in ids] for ids in token_ids]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    if params.method == "ctc-decoding":
+        for filename, hyp in zip(params.sound_files, hyps):
+            words = "".join(hyp)
+            words = words.replace("▁", " ").strip()
+            s += f"{filename}:\n{words}\n\n"
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        for filename, hyp in zip(params.sound_files, hyps):
+            words = " ".join(hyp)
+            words = words.replace("▁", " ").strip()
+            s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py b/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py
new file mode 100755
index 000000000..d4ceacefd
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1,273 @@
+#!/usr/bin/env python3
+# flake8: noqa
+# Copyright      2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+Usage of this script:
+
+./zipformer/jit_pretrained_streaming.py \
+  --nn-model-filename ./zipformer/exp-causal/jit_script_chunk_16_left_128.pt \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  /path/to/foo.wav \
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Optional
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model jit_script.pt",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+    device: torch.device = torch.device("cpu"),
+):
+    assert encoder_out.ndim == 2
+    context_size = decoder.context_size
+    blank_id = decoder.blank_id
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(hyp, dtype=torch.int32, device=device).unsqueeze(0)
+        # decoder_input.shape (1,, 1 context_size)
+        decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1)
+    else:
+        assert decoder_out.ndim == 2
+        assert hyp is not None, hyp
+
+    T = encoder_out.size(0)
+    for i in range(T):
+        cur_encoder_out = encoder_out[i : i + 1]
+        joiner_out = joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+
+            decoder_input = torch.tensor(
+                decoder_input, dtype=torch.int32, device=device
+            ).unsqueeze(0)
+            decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1)
+
+    return hyp, decoder_out
+
+
+def create_streaming_feature_extractor(sample_rate) -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+    model.eval()
+    model.to(device)
+
+    encoder = model.encoder
+    decoder = model.decoder
+    joiner = model.joiner
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+    context_size = decoder.context_size
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor(args.sample_rate)
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=args.sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    logging.info("Decoding started")
+
+    chunk_length = encoder.chunk_size * 2
+    T = chunk_length + encoder.pad_length
+
+    logging.info(f"chunk_length: {chunk_length}")
+    logging.info(f"T: {T}")
+
+    states = encoder.get_init_states(device=device)
+
+    tail_padding = torch.zeros(int(0.3 * args.sample_rate), dtype=torch.float32)
+
+    wave_samples = torch.cat([wave_samples, tail_padding])
+
+    chunk = int(0.25 * args.sample_rate)  # 0.2 second
+    num_processed_frames = 0
+
+    hyp = None
+    decoder_out = None
+
+    start = 0
+    while start < wave_samples.numel():
+        logging.info(f"{start}/{wave_samples.numel()}")
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+        online_fbank.accept_waveform(
+            sampling_rate=args.sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= T:
+            frames = []
+            for i in range(T):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            frames = torch.cat(frames, dim=0).to(device).unsqueeze(0)
+            x_lens = torch.tensor([T], dtype=torch.int32, device=device)
+            encoder_out, out_lens, states = encoder(
+                features=frames,
+                feature_lengths=x_lens,
+                states=states,
+            )
+            num_processed_frames += chunk_length
+
+            hyp, decoder_out = greedy_search(
+                decoder, joiner, encoder_out.squeeze(0), decoder_out, hyp, device=device
+            )
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += token_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+torch.set_num_threads(4)
+torch.set_num_interop_threads(1)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/joiner.py b/egs/librispeech/ASR/zipformer/joiner.py
new file mode 100644
index 000000000..dfb0a0057
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/joiner.py
@@ -0,0 +1,67 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim, initial_scale=0.25)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim, initial_scale=0.25)
+        self.output_linear = nn.Linear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim, (
+            encoder_out.shape,
+            decoder_out.shape,
+        )
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit
diff --git a/egs/librispeech/ASR/zipformer/model.py b/egs/librispeech/ASR/zipformer/model.py
new file mode 100644
index 000000000..f2f86af47
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/model.py
@@ -0,0 +1,358 @@
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos, make_pad_mask
+from scaling import ScaledLinear
+
+
+class AsrModel(nn.Module):
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        decoder: Optional[nn.Module] = None,
+        joiner: Optional[nn.Module] = None,
+        encoder_dim: int = 384,
+        decoder_dim: int = 512,
+        vocab_size: int = 500,
+        use_transducer: bool = True,
+        use_ctc: bool = False,
+    ):
+        """A joint CTC & Transducer ASR model.
+
+        - Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks (http://imagine.enpc.fr/~obozinsg/teaching/mva_gm/papers/ctc.pdf)
+        - Sequence Transduction with Recurrent Neural Networks (https://arxiv.org/pdf/1211.3711.pdf)
+        - Pruned RNN-T for fast, memory-efficient ASR training (https://arxiv.org/pdf/2206.13236.pdf)
+
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dim) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+            It is used when use_transducer is True.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+            It is used when use_transducer is True.
+          use_transducer:
+            Whether use transducer head. Default: True.
+          use_ctc:
+            Whether use CTC head. Default: False.
+        """
+        super().__init__()
+
+        assert (
+            use_transducer or use_ctc
+        ), f"At least one of them should be True, but got use_transducer={use_transducer}, use_ctc={use_ctc}"
+
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+
+        self.use_transducer = use_transducer
+        if use_transducer:
+            # Modules for Transducer head
+            assert decoder is not None
+            assert hasattr(decoder, "blank_id")
+            assert joiner is not None
+
+            self.decoder = decoder
+            self.joiner = joiner
+
+            self.simple_am_proj = ScaledLinear(
+                encoder_dim, vocab_size, initial_scale=0.25
+            )
+            self.simple_lm_proj = ScaledLinear(
+                decoder_dim, vocab_size, initial_scale=0.25
+            )
+        else:
+            assert decoder is None
+            assert joiner is None
+
+        self.use_ctc = use_ctc
+        if use_ctc:
+            # Modules for CTC head
+            self.ctc_output = nn.Sequential(
+                nn.Dropout(p=0.1),
+                nn.Linear(encoder_dim, vocab_size),
+                nn.LogSoftmax(dim=-1),
+            )
+
+    def forward_encoder(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute encoder outputs.
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+
+        Returns:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+        """
+        # logging.info(f"Memory allocated at entry: {torch.cuda.memory_allocated() // 1000000}M")
+        x, x_lens = self.encoder_embed(x, x_lens)
+        # logging.info(f"Memory allocated after encoder_embed: {torch.cuda.memory_allocated() // 1000000}M")
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        assert torch.all(encoder_out_lens > 0), (x_lens, encoder_out_lens)
+
+        return encoder_out, encoder_out_lens
+
+    def forward_ctc(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute CTC loss.
+        Args:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+          targets:
+            Target Tensor of shape (sum(target_lengths)). The targets are assumed
+            to be un-padded and concatenated within 1 dimension.
+        """
+        # Compute CTC log-prob
+        ctc_output = self.ctc_output(encoder_out)  # (N, T, C)
+
+        ctc_loss = torch.nn.functional.ctc_loss(
+            log_probs=ctc_output.permute(1, 0, 2),  # (T, N, C)
+            targets=targets,
+            input_lengths=encoder_out_lens,
+            target_lengths=target_lengths,
+            reduction="sum",
+        )
+        return ctc_loss
+
+    def forward_transducer(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        y_lens: torch.Tensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute Transducer loss.
+        Args:
+          encoder_out:
+            Encoder output, of shape (N, T, C).
+          encoder_out_lens:
+            Encoder output lengths, of shape (N,).
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        """
+        # Now for the decoder, i.e., the prediction network
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = encoder_out_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        # if self.training and random.random() < 0.25:
+        #    lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04)
+        # if self.training and random.random() < 0.25:
+        #    am = penalize_abs_values_gt(am, 30.0, 1.0e-04)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return simple_loss, pruned_loss
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer losses and CTC loss,
+          in form of (simple_loss, pruned_loss, ctc_loss)
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0, (x.shape, x_lens.shape, y.dim0)
+
+        # Compute encoder outputs
+        encoder_out, encoder_out_lens = self.forward_encoder(x, x_lens)
+
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        if self.use_transducer:
+            # Compute transducer loss
+            simple_loss, pruned_loss = self.forward_transducer(
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                y=y.to(x.device),
+                y_lens=y_lens,
+                prune_range=prune_range,
+                am_scale=am_scale,
+                lm_scale=lm_scale,
+            )
+        else:
+            simple_loss = torch.empty(0)
+            pruned_loss = torch.empty(0)
+
+        if self.use_ctc:
+            # Compute CTC loss
+            targets = y.values
+            ctc_loss = self.forward_ctc(
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                targets=targets,
+                target_lengths=y_lens,
+            )
+        else:
+            ctc_loss = torch.empty(0)
+
+        return simple_loss, pruned_loss, ctc_loss
diff --git a/egs/librispeech/ASR/zipformer/onnx_check.py b/egs/librispeech/ASR/zipformer/onnx_check.py
new file mode 100755
index 000000000..93bd3a211
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_check.py
@@ -0,0 +1,240 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script checks that exported onnx models produce the same output
+with the given torchscript model for the same input.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model via torchscript (torch.jit.script())
+
+./zipformer/export.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp/ \
+  --jit 1
+
+It will generate the following file in $repo/exp:
+    - jit_script.pt
+
+3. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+4. Run this file
+
+./zipformer/onnx_check.py \
+  --jit-filename $repo/exp/jit_script.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+"""
+
+import argparse
+import logging
+
+import torch
+from onnx_pretrained import OnnxModel
+
+from icefall import is_module_available
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript model",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    C = 80
+    for i in range(3):
+        N = torch.randint(low=1, high=20, size=(1,)).item()
+        T = torch.randint(low=30, high=50, size=(1,)).item()
+        logging.info(f"test_encoder: iter {i}, N={N}, T={T}")
+
+        x = torch.rand(N, T, C)
+        x_lens = torch.randint(low=30, high=T + 1, size=(N,))
+        x_lens[0] = T
+
+        torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens)
+        torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out)
+
+        onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens)
+
+        assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), (
+            (torch_encoder_out - onnx_encoder_out).abs().max()
+        )
+
+
+def test_decoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    context_size = onnx_model.context_size
+    vocab_size = onnx_model.vocab_size
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_decoder: iter {i}, N={N}")
+        x = torch.randint(
+            low=1,
+            high=vocab_size,
+            size=(N, context_size),
+            dtype=torch.int64,
+        )
+        torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False]))
+        torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out)
+        torch_decoder_out = torch_decoder_out.squeeze(1)
+
+        onnx_decoder_out = onnx_model.run_decoder(x)
+        assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
+            (torch_decoder_out - onnx_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1]
+    decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1]
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_joiner: iter {i}, N={N}")
+        encoder_out = torch.rand(N, encoder_dim)
+        decoder_out = torch.rand(N, decoder_dim)
+
+        projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out)
+        projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out)
+
+        torch_joiner_out = torch_model.joiner(encoder_out, decoder_out)
+        onnx_joiner_out = onnx_model.run_joiner(
+            projected_encoder_out, projected_decoder_out
+        )
+
+        assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
+            (torch_joiner_out - onnx_joiner_out).abs().max()
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_model = torch.jit.load(args.jit_filename)
+
+    onnx_model = OnnxModel(
+        encoder_model_filename=args.onnx_encoder_filename,
+        decoder_model_filename=args.onnx_decoder_filename,
+        joiner_model_filename=args.onnx_joiner_filename,
+    )
+
+    logging.info("Test encoder")
+    test_encoder(torch_model, onnx_model)
+
+    logging.info("Test decoder")
+    test_decoder(torch_model, onnx_model)
+
+    logging.info("Test joiner")
+    test_joiner(torch_model, onnx_model)
+    logging.info("Finished checking ONNX models")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# See https://github.com/pytorch/pytorch/issues/38342
+# and https://github.com/pytorch/pytorch/issues/33354
+#
+# If we don't do this, the delay increases whenever there is
+# a new request that changes the actual batch size.
+# If you use `py-spy dump --pid <server-pid> --native`, you will
+# see a lot of time is spent in re-compiling the torch script model.
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    torch.manual_seed(20220727)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_decode.py b/egs/librispeech/ASR/zipformer/onnx_decode.py
new file mode 100755
index 000000000..356c2a830
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_decode.py
@@ -0,0 +1,325 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --causal False
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+2. Run this file
+
+./zipformer/onnx_decode.py \
+  --exp-dir $repo/exp \
+  --max-duration 600 \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+"""
+
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+
+from onnx_pretrained import greedy_search, OnnxModel
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+from k2 import SymbolTable
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, token_table: SymbolTable, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      token_table:
+        The token table.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    hyps = [token_ids_to_words(h).split() for h in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    token_table: SymbolTable,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      token_table:
+        The token table.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, token_table=token_table, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    token_table = SymbolTable.from_file(args.tokens)
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(
+            dl=test_dl, model=model, token_table=token_table
+        )
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming-ctc.py b/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming-ctc.py
new file mode 100755
index 000000000..44546cae5
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming-ctc.py
@@ -0,0 +1,426 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+# Copyright      2023  Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script loads ONNX models exported by ./export-onnx-streaming-ctc.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-streaming-2023-11-05
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-streaming-2023-11-05
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx-streaming-ctc.py \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --causal True \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --use-ctc 1
+
+It will generate the following 2 files inside $repo/exp:
+
+ - ctc-epoch-99-avg-1-chunk-16-left-128.int8.onnx
+ - ctc-epoch-99-avg-1-chunk-16-left-128.onnx
+
+You can use either the ``int8.onnx`` model or just the ``.onnx`` model.
+
+3. Run this file with the exported ONNX models
+
+./zipformer/onnx_pretrained-streaming-ctc.py \
+  --model-filename $repo/exp/ctc-epoch-99-avg-1-chunk-16-left-128.onnx \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  $repo/test_wavs/DEV_T0000000001.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Tuple
+
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(model_filename)
+
+    def init_model(self, model_filename: str):
+        self.model = ort.InferenceSession(
+            model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_states()
+
+    def init_states(self, batch_size: int = 1):
+        meta = self.model.get_modelmeta().custom_metadata_map
+        logging.info(f"meta={meta}")
+
+        model_type = meta["model_type"]
+        assert model_type == "zipformer2", model_type
+
+        decode_chunk_len = int(meta["decode_chunk_len"])
+        T = int(meta["T"])
+
+        num_encoder_layers = meta["num_encoder_layers"]
+        encoder_dims = meta["encoder_dims"]
+        cnn_module_kernels = meta["cnn_module_kernels"]
+        left_context_len = meta["left_context_len"]
+        query_head_dims = meta["query_head_dims"]
+        value_head_dims = meta["value_head_dims"]
+        num_heads = meta["num_heads"]
+
+        def to_int_list(s):
+            return list(map(int, s.split(",")))
+
+        num_encoder_layers = to_int_list(num_encoder_layers)
+        encoder_dims = to_int_list(encoder_dims)
+        cnn_module_kernels = to_int_list(cnn_module_kernels)
+        left_context_len = to_int_list(left_context_len)
+        query_head_dims = to_int_list(query_head_dims)
+        value_head_dims = to_int_list(value_head_dims)
+        num_heads = to_int_list(num_heads)
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+        logging.info(f"encoder_dims: {encoder_dims}")
+        logging.info(f"cnn_module_kernels: {cnn_module_kernels}")
+        logging.info(f"left_context_len: {left_context_len}")
+        logging.info(f"query_head_dims: {query_head_dims}")
+        logging.info(f"value_head_dims: {value_head_dims}")
+        logging.info(f"num_heads: {num_heads}")
+
+        num_encoders = len(num_encoder_layers)
+
+        self.states = []
+        for i in range(num_encoders):
+            num_layers = num_encoder_layers[i]
+            key_dim = query_head_dims[i] * num_heads[i]
+            embed_dim = encoder_dims[i]
+            nonlin_attn_head_dim = 3 * embed_dim // 4
+            value_dim = value_head_dims[i] * num_heads[i]
+            conv_left_pad = cnn_module_kernels[i] // 2
+
+            for layer in range(num_layers):
+                cached_key = torch.zeros(
+                    left_context_len[i], batch_size, key_dim
+                ).numpy()
+                cached_nonlin_attn = torch.zeros(
+                    1, batch_size, left_context_len[i], nonlin_attn_head_dim
+                ).numpy()
+                cached_val1 = torch.zeros(
+                    left_context_len[i], batch_size, value_dim
+                ).numpy()
+                cached_val2 = torch.zeros(
+                    left_context_len[i], batch_size, value_dim
+                ).numpy()
+                cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).numpy()
+                cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).numpy()
+                self.states += [
+                    cached_key,
+                    cached_nonlin_attn,
+                    cached_val1,
+                    cached_val2,
+                    cached_conv1,
+                    cached_conv2,
+                ]
+        embed_states = torch.zeros(batch_size, 128, 3, 19).numpy()
+        self.states.append(embed_states)
+        processed_lens = torch.zeros(batch_size, dtype=torch.int64).numpy()
+        self.states.append(processed_lens)
+
+        self.num_encoders = num_encoders
+
+        self.segment = T
+        self.offset = decode_chunk_len
+
+    def _build_model_input_output(
+        self,
+        x: torch.Tensor,
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        model_input = {"x": x.numpy()}
+        model_output = ["log_probs"]
+
+        def build_inputs_outputs(tensors, i):
+            assert len(tensors) == 6, len(tensors)
+
+            # (downsample_left, batch_size, key_dim)
+            name = f"cached_key_{i}"
+            model_input[name] = tensors[0]
+            model_output.append(f"new_{name}")
+
+            # (1, batch_size, downsample_left, nonlin_attn_head_dim)
+            name = f"cached_nonlin_attn_{i}"
+            model_input[name] = tensors[1]
+            model_output.append(f"new_{name}")
+
+            # (downsample_left, batch_size, value_dim)
+            name = f"cached_val1_{i}"
+            model_input[name] = tensors[2]
+            model_output.append(f"new_{name}")
+
+            # (downsample_left, batch_size, value_dim)
+            name = f"cached_val2_{i}"
+            model_input[name] = tensors[3]
+            model_output.append(f"new_{name}")
+
+            # (batch_size, embed_dim, conv_left_pad)
+            name = f"cached_conv1_{i}"
+            model_input[name] = tensors[4]
+            model_output.append(f"new_{name}")
+
+            # (batch_size, embed_dim, conv_left_pad)
+            name = f"cached_conv2_{i}"
+            model_input[name] = tensors[5]
+            model_output.append(f"new_{name}")
+
+        for i in range(len(self.states[:-2]) // 6):
+            build_inputs_outputs(self.states[i * 6 : (i + 1) * 6], i)
+
+        # (batch_size, channels, left_pad, freq)
+        name = "embed_states"
+        embed_states = self.states[-2]
+        model_input[name] = embed_states
+        model_output.append(f"new_{name}")
+
+        # (batch_size,)
+        name = "processed_lens"
+        processed_lens = self.states[-1]
+        model_input[name] = processed_lens
+        model_output.append(f"new_{name}")
+
+        return model_input, model_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        self.states = states
+
+    def __call__(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+        Returns:
+          Return a 3-D tensor containing log_probs. Its shape is (N, T, vocab_size)
+          where T' is usually equal to ((T-7)//2 - 3)//2
+        """
+        model_input, model_output_names = self._build_model_input_output(x)
+
+        out = self.model.run(model_output_names, model_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    log_probs: torch.Tensor,
+) -> List[int]:
+    """Greedy search for a single utterance.
+    Args:
+      log_probs:
+        A 3-D tensor of shape (1, T, vocab_size)
+    Returns:
+      Return the decoded result.
+    """
+    assert log_probs.ndim == 3, log_probs.shape
+    assert log_probs.shape[0] == 1, log_probs.shape
+
+    max_indexes = log_probs[0].argmax(dim=1)
+    unique_indexes = torch.unique_consecutive(max_indexes)
+
+    blank_id = 0
+    unique_indexes = unique_indexes[unique_indexes != blank_id]
+    return unique_indexes.tolist()
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(model_filename=args.model_filename)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    hyp = []
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            log_probs = model(frames)
+
+            hyp += greedy_search(log_probs)
+
+    # To handle byte-level BPE, we convert string tokens to utf-8 encoded bytes
+    id2token = {}
+    with open(args.tokens, encoding="utf-8") as f:
+        for line in f:
+            token, idx = line.split()
+            if token[:3] == "<0x" and token[-1] == ">":
+                token = int(token[1:-1], base=16)
+                assert 0 <= token < 256, token
+                token = token.to_bytes(1, byteorder="little")
+            else:
+                token = token.encode(encoding="utf-8")
+
+            id2token[int(idx)] = token
+
+    text = b""
+    for i in hyp:
+        text += id2token[i]
+    text = text.decode(encoding="utf-8")
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py b/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 100755
index 000000000..e7c4f40ee
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1,546 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+# Copyright      2023  Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script loads ONNX models exported by ./export-onnx-streaming.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx-streaming.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal True \
+  --chunk-size 16 \
+  --left-context-frames 64
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file with the exported ONNX models
+
+./zipformer/onnx_pretrained-streaming.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_encoder_states()
+
+    def init_encoder_states(self, batch_size: int = 1):
+        encoder_meta = self.encoder.get_modelmeta().custom_metadata_map
+        logging.info(f"encoder_meta={encoder_meta}")
+
+        model_type = encoder_meta["model_type"]
+        assert model_type == "zipformer2", model_type
+
+        decode_chunk_len = int(encoder_meta["decode_chunk_len"])
+        T = int(encoder_meta["T"])
+
+        num_encoder_layers = encoder_meta["num_encoder_layers"]
+        encoder_dims = encoder_meta["encoder_dims"]
+        cnn_module_kernels = encoder_meta["cnn_module_kernels"]
+        left_context_len = encoder_meta["left_context_len"]
+        query_head_dims = encoder_meta["query_head_dims"]
+        value_head_dims = encoder_meta["value_head_dims"]
+        num_heads = encoder_meta["num_heads"]
+
+        def to_int_list(s):
+            return list(map(int, s.split(",")))
+
+        num_encoder_layers = to_int_list(num_encoder_layers)
+        encoder_dims = to_int_list(encoder_dims)
+        cnn_module_kernels = to_int_list(cnn_module_kernels)
+        left_context_len = to_int_list(left_context_len)
+        query_head_dims = to_int_list(query_head_dims)
+        value_head_dims = to_int_list(value_head_dims)
+        num_heads = to_int_list(num_heads)
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+        logging.info(f"encoder_dims: {encoder_dims}")
+        logging.info(f"cnn_module_kernels: {cnn_module_kernels}")
+        logging.info(f"left_context_len: {left_context_len}")
+        logging.info(f"query_head_dims: {query_head_dims}")
+        logging.info(f"value_head_dims: {value_head_dims}")
+        logging.info(f"num_heads: {num_heads}")
+
+        num_encoders = len(num_encoder_layers)
+
+        self.states = []
+        for i in range(num_encoders):
+            num_layers = num_encoder_layers[i]
+            key_dim = query_head_dims[i] * num_heads[i]
+            embed_dim = encoder_dims[i]
+            nonlin_attn_head_dim = 3 * embed_dim // 4
+            value_dim = value_head_dims[i] * num_heads[i]
+            conv_left_pad = cnn_module_kernels[i] // 2
+
+            for layer in range(num_layers):
+                cached_key = torch.zeros(
+                    left_context_len[i], batch_size, key_dim
+                ).numpy()
+                cached_nonlin_attn = torch.zeros(
+                    1, batch_size, left_context_len[i], nonlin_attn_head_dim
+                ).numpy()
+                cached_val1 = torch.zeros(
+                    left_context_len[i], batch_size, value_dim
+                ).numpy()
+                cached_val2 = torch.zeros(
+                    left_context_len[i], batch_size, value_dim
+                ).numpy()
+                cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).numpy()
+                cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).numpy()
+                self.states += [
+                    cached_key,
+                    cached_nonlin_attn,
+                    cached_val1,
+                    cached_val2,
+                    cached_conv1,
+                    cached_conv2,
+                ]
+        embed_states = torch.zeros(batch_size, 128, 3, 19).numpy()
+        self.states.append(embed_states)
+        processed_lens = torch.zeros(batch_size, dtype=torch.int64).numpy()
+        self.states.append(processed_lens)
+
+        self.num_encoders = num_encoders
+
+        self.segment = T
+        self.offset = decode_chunk_len
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def _build_encoder_input_output(
+        self,
+        x: torch.Tensor,
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        encoder_input = {"x": x.numpy()}
+        encoder_output = ["encoder_out"]
+
+        def build_inputs_outputs(tensors, i):
+            assert len(tensors) == 6, len(tensors)
+
+            # (downsample_left, batch_size, key_dim)
+            name = f"cached_key_{i}"
+            encoder_input[name] = tensors[0]
+            encoder_output.append(f"new_{name}")
+
+            # (1, batch_size, downsample_left, nonlin_attn_head_dim)
+            name = f"cached_nonlin_attn_{i}"
+            encoder_input[name] = tensors[1]
+            encoder_output.append(f"new_{name}")
+
+            # (downsample_left, batch_size, value_dim)
+            name = f"cached_val1_{i}"
+            encoder_input[name] = tensors[2]
+            encoder_output.append(f"new_{name}")
+
+            # (downsample_left, batch_size, value_dim)
+            name = f"cached_val2_{i}"
+            encoder_input[name] = tensors[3]
+            encoder_output.append(f"new_{name}")
+
+            # (batch_size, embed_dim, conv_left_pad)
+            name = f"cached_conv1_{i}"
+            encoder_input[name] = tensors[4]
+            encoder_output.append(f"new_{name}")
+
+            # (batch_size, embed_dim, conv_left_pad)
+            name = f"cached_conv2_{i}"
+            encoder_input[name] = tensors[5]
+            encoder_output.append(f"new_{name}")
+
+        for i in range(len(self.states[:-2]) // 6):
+            build_inputs_outputs(self.states[i * 6 : (i + 1) * 6], i)
+
+        # (batch_size, channels, left_pad, freq)
+        name = "embed_states"
+        embed_states = self.states[-2]
+        encoder_input[name] = embed_states
+        encoder_output.append(f"new_{name}")
+
+        # (batch_size,)
+        name = "processed_lens"
+        processed_lens = self.states[-1]
+        encoder_input[name] = processed_lens
+        encoder_output.append(f"new_{name}")
+
+        return encoder_input, encoder_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        self.states = states
+
+    def run_encoder(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+        Returns:
+          Return a 3-D tensor of shape (N, T', joiner_dim) where
+          T' is usually equal to ((T-7)//2-3)//2
+        """
+        encoder_input, encoder_output_names = self._build_encoder_input_output(x)
+
+        out = self.encoder.run(encoder_output_names, encoder_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    context_size: int,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+) -> List[int]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (1, T, joiner_dim)
+      context_size:
+        The context size of the decoder model.
+      decoder_out:
+        Optional. Decoder output of the previous chunk.
+      hyp:
+        Decoding results for previous chunks.
+    Returns:
+      Return the decoded results so far.
+    """
+
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert hyp is not None, hyp
+
+    encoder_out = encoder_out.squeeze(0)
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t : t + 1]
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+            decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    context_size = model.context_size
+    hyp = None
+    decoder_out = None
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            encoder_out = model.run_encoder(frames)
+            hyp, decoder_out = greedy_search(
+                model,
+                encoder_out,
+                context_size,
+                decoder_out,
+                hyp,
+            )
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += token_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained.py b/egs/librispeech/ASR/zipformer/onnx_pretrained.py
new file mode 100755
index 000000000..334376093
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1,421 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --causal False
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file
+
+./zipformer/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 4
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 2-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, its shape is (N, T', joiner_dim)
+            - encoder_out_lens, its shape is (N,)
+        """
+        out = self.encoder.run(
+            [
+                self.encoder.get_outputs()[0].name,
+                self.encoder.get_outputs()[1].name,
+            ],
+            {
+                self.encoder.get_inputs()[0].name: x.numpy(),
+                self.encoder.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, joiner_dim)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.run_decoder(decoder_input)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        # current_encoder_out's shape: (batch_size, joiner_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+        logits = model.run_joiner(current_encoder_out, decoder_out)
+
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                dtype=torch.int64,
+            )
+            decoder_out = model.run_decoder(decoder_input)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = token_ids_to_words(hyp)
+        s += f"{filename}:\n{words}\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc.py
new file mode 100755
index 000000000..eb5cee9cd
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc.py
@@ -0,0 +1,213 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc.py \
+  --nn-model /path/to/model.onnx \
+  --tokens /path/to/data/lang_bpe_500/tokens.txt \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    blank_id = 0
+    s = "\n"
+    for i in range(log_probs.size(0)):
+        # greedy search
+        indexes = log_probs[i, : log_probs_len[i]].argmax(dim=-1)
+        token_ids = torch.unique_consecutive(indexes)
+
+        token_ids = token_ids[token_ids != blank_id]
+        words = token_ids_to_words(token_ids.tolist())
+        s += f"{args.sound_files[i]}:\n{words}\n\n"
+
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py
new file mode 100755
index 000000000..683a7dc20
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py
@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc_H.py \
+  --nn-model /path/to/model.onnx \
+  --tokens /path/to/data/lang_bpe_500/tokens.txt \
+  --H /path/to/H.fst \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+
+You can find exported ONNX models at
+https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+from typing import Dict
+import kaldifst
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--H",
+        type=str,
+        help="""Path to H.fst.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    log_probs: torch.Tensor,
+    H: kaldifst,
+    id2token: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      log_probs:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      H:
+        The H graph.
+      id2word:
+        A map mapping token ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {log_probs.shape}")
+    decodable = DecodableCtc(log_probs.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(H, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # tokens are incremented during graph construction
+    # are shifted by 1 during graph construction
+    hyps = [id2token[i - 1] for i in osymbols_out if i != 1]
+    hyps = "".join(hyps).split("\u2581")  # unicode codepoint of ▁
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    logging.info(f"Loading H from {args.H}")
+    H = kaldifst.StdVectorFst.read(args.H)
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    hyps = []
+    for i in range(log_probs.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            log_probs=log_probs[i, : log_probs_len[i]],
+            H=H,
+            id2token=token_table,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
new file mode 100755
index 000000000..0b94bfa65
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc_HL.py \
+  --nn-model /path/to/model.onnx \
+  --words /path/to/data/lang_bpe_500/words.txt \
+  --HL /path/to/HL.fst \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+
+You can find exported ONNX models at
+https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+from typing import Dict
+import kaldifst
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        help="""Path to words.txt.""",
+    )
+
+    parser.add_argument(
+        "--HL",
+        type=str,
+        help="""Path to HL.fst.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    log_probs: torch.Tensor,
+    HL: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      log_probs:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HL:
+        The HL graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {log_probs.shape}")
+    decodable = DecodableCtc(log_probs.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HL, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    logging.info(f"Loading HL from {args.HL}")
+    HL = kaldifst.StdVectorFst.read(args.HL)
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    word_table = k2.SymbolTable.from_file(args.words)
+
+    hyps = []
+    for i in range(log_probs.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            log_probs=log_probs[i, : log_probs_len[i]],
+            HL=HL,
+            id2word=word_table,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
new file mode 100755
index 000000000..93569142a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc_HLG.py \
+  --nn-model /path/to/model.onnx \
+  --words /path/to/data/lang_bpe_500/words.txt \
+  --HLG /path/to/HLG.fst \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+
+You can find exported ONNX models at
+https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+from typing import Dict
+import kaldifst
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        help="""Path to words.txt.""",
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.fst.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    log_probs: torch.Tensor,
+    HLG: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      log_probs:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HLG:
+        The HLG graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {log_probs.shape}")
+    decodable = DecodableCtc(log_probs.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HLG, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    logging.info(f"Loading HLG from {args.HLG}")
+    HLG = kaldifst.StdVectorFst.read(args.HLG)
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    word_table = k2.SymbolTable.from_file(args.words)
+
+    hyps = []
+    for i in range(log_probs.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            log_probs=log_probs[i, : log_probs_len[i]],
+            HLG=HLG,
+            id2word=word_table,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/optim.py b/egs/librispeech/ASR/zipformer/optim.py
new file mode 100644
index 000000000..714d8db9a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/optim.py
@@ -0,0 +1,1248 @@
+#      Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import contextlib
+import logging
+import random
+from collections import defaultdict
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from lhotse.utils import fix_random_seed
+from torch import Tensor, nn
+from torch.optim import Optimizer
+
+
+class BatchedOptimizer(Optimizer):
+    """
+    This class adds to class Optimizer the capability to optimize parameters in batches:
+    it will stack the parameters and their grads for you so the optimizer can work
+    on tensors with an extra leading dimension.  This is intended for speed with GPUs,
+    as it reduces the number of kernels launched in the optimizer.
+
+    Args:
+      params:
+    """
+
+    def __init__(self, params, defaults):
+        super(BatchedOptimizer, self).__init__(params, defaults)
+
+    @contextlib.contextmanager
+    def batched_params(self, param_group, group_params_names):
+        """
+        This function returns (technically, yields) a list of
+          of tuples (p, state), where
+        p is a `fake` parameter that is stacked (over axis 0) from real parameters
+        that share the same shape, and its gradient is also stacked;
+        `state` is the state corresponding to this batch of parameters
+        (it will be physically located in the "state" for one of the real
+        parameters, the last one that has any particular shape and dtype).
+
+        This function is decorated as a context manager so that it can
+        write parameters back to their "real" locations.
+
+        The idea is, instead of doing:
+        <code>
+          for p in group["params"]:
+             state = self.state[p]
+             ...
+        </code>
+        you can do:
+        <code>
+          with self.batched_params(group["params"]) as batches:
+             for p, state, p_names in batches:
+                 ...
+        </code>
+
+        Args:
+          group: a parameter group, which is a list of parameters; should be
+                one of self.param_groups.
+          group_params_names: name for each parameter in group,
+                which is List[str].
+        """
+        batches = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
+        batches_names = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of str
+
+        assert len(param_group) == len(group_params_names)
+        for p, named_p in zip(param_group, group_params_names):
+            key = (str(p.dtype), *p.shape)
+            batches[key].append(p)
+            batches_names[key].append(named_p)
+
+        batches_names_keys = list(batches_names.keys())
+        sorted_idx = sorted(
+            range(len(batches_names)), key=lambda i: batches_names_keys[i]
+        )
+        batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx]
+        batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
+
+        stacked_params_dict = dict()
+
+        # turn batches into a list, in deterministic order.
+        # tuples will contain tuples of (stacked_param, state, stacked_params_names),
+        # one for each batch in `batches`.
+        tuples = []
+
+        for batch, batch_names in zip(batches, batches_names):
+            p = batch[0]
+            # we arbitrarily store the state in the
+            # state corresponding to the 1st parameter in the
+            # group.  class Optimizer will take care of saving/loading state.
+            state = self.state[p]
+            p_stacked = torch.stack(batch)
+            grad = torch.stack(
+                [torch.zeros_like(p) if p.grad is None else p.grad for p in batch]
+            )
+            p_stacked.grad = grad
+            stacked_params_dict[key] = p_stacked
+            tuples.append((p_stacked, state, batch_names))
+
+        yield tuples  # <-- calling code will do the actual optimization here!
+
+        for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
+            for i, p in enumerate(batch):  # batch is list of Parameter
+                p.copy_(stacked_params[i])
+
+
+class ScaledAdam(BatchedOptimizer):
+    """
+     Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
+     proportional to the norm of that parameter; and also learn the scale of the parameter,
+     in log space, subject to upper and lower limits (as if we had factored each parameter as
+     param = underlying_param * log_scale.exp())
+
+
+     Args:
+          params:  The parameters or param_groups to optimize (like other Optimizer subclasses)
+                   Unlike common optimizers, which accept model.parameters() or groups of parameters(),
+                   this optimizer could accept model.named_parameters() or groups of named_parameters().
+                   See comments of function _get_names_of_parameters for its 4 possible cases.
+              lr:  The learning rate.  We will typically use a learning rate schedule that starts
+                   at 0.03 and decreases over time, i.e. much higher than other common
+                   optimizers.
+     clipping_scale: (e.g. 2.0)
+                   A scale for gradient-clipping: if specified, the normalized gradients
+                   over the whole model will be clipped to have 2-norm equal to
+                   `clipping_scale` times the median 2-norm over the most recent period
+                   of `clipping_update_period` minibatches.  By "normalized gradients",
+                   we mean after multiplying by the rms parameter value for this tensor
+                   [for non-scalars]; this is appropriate because our update is scaled
+                   by this quantity.
+            betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
+                   Must satisfy 0 < beta <= beta2 < 1.
+     scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
+                   scale of each parameter tensor and scalar parameters of the mode..
+                   If each parameter were decomposed
+                   as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
+                   would be a the scaling factor on the learning rate of p_scale.
+              eps:  A general-purpose epsilon to prevent division by zero
+    param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be >= this value)
+    param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be <= this value)
+       scalar_max: Maximum absolute value for scalar parameters (applicable if your
+                   model has any parameters with numel() == 1).
+    size_update_period: The periodicity, in steps, with which we update the size (scale)
+                   of the parameter tensor.  This is provided to save a little time
+                   in the update.
+     clipping_update_period: if clipping_scale is specified, this is the period
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=3e-02,
+        clipping_scale=None,
+        betas=(0.9, 0.98),
+        scalar_lr_scale=0.1,
+        eps=1.0e-08,
+        param_min_rms=1.0e-05,
+        param_max_rms=3.0,
+        scalar_max=10.0,
+        size_update_period=4,
+        clipping_update_period=100,
+    ):
+
+        defaults = dict(
+            lr=lr,
+            clipping_scale=clipping_scale,
+            betas=betas,
+            scalar_lr_scale=scalar_lr_scale,
+            eps=eps,
+            param_min_rms=param_min_rms,
+            param_max_rms=param_max_rms,
+            scalar_max=scalar_max,
+            size_update_period=size_update_period,
+            clipping_update_period=clipping_update_period,
+        )
+
+        # If params only contains parameters or group of parameters,
+        # i.e when parameter names are not given,
+        # this flag will be set to False in funciton _get_names_of_parameters.
+        self.show_dominant_parameters = True
+        param_groups, parameters_names = self._get_names_of_parameters(params)
+        super(ScaledAdam, self).__init__(param_groups, defaults)
+        assert len(self.param_groups) == len(parameters_names)
+        self.parameters_names = parameters_names
+
+    def _get_names_of_parameters(
+        self, params_or_named_params
+    ) -> Tuple[List[Dict], List[List[str]]]:
+        """
+        Args:
+          params_or_named_params: according to the way ScaledAdam is initialized in train.py,
+            this argument could be one of following 4 cases,
+            case 1, a generator of parameter, e.g.:
+              optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=3.0)
+
+            case 2, a list of parameter groups with different config, e.g.:
+              model_param_groups = [
+                      {'params': model.encoder.parameters(), 'lr': 0.05},
+                      {'params': model.decoder.parameters(), 'lr': 0.01},
+                      {'params': model.joiner.parameters(), 'lr': 0.03},
+                      ]
+              optimizer = ScaledAdam(model_param_groups, lr=params.base_lr, clipping_scale=3.0)
+
+            case 3, a generator of named_parameter, e.g.:
+              optimizer = ScaledAdam(model.named_parameters(), lr=params.base_lr, clipping_scale=3.0)
+
+            case 4, a list of named_parameter groups with different config, e.g.:
+              model_named_param_groups = [
+                      {'named_params': model.encoder.named_parameters(), 'lr': 0.05},
+                      {'named_params': model.decoder.named_parameters(), 'lr': 0.01},
+                      {'named_params': model.joiner.named_parameters(), 'lr': 0.03},
+                      ]
+              optimizer = ScaledAdam(model_named_param_groups, lr=params.base_lr, clipping_scale=3.0)
+
+          For case 1 and case 2, input params is used to initialize the underlying torch.optimizer.
+          For case 3 and case 4, firstly, names and params are extracted from input named_params,
+            then, these extracted params are used to initialize the underlying torch.optimizer,
+            and these extracted names are mainly used by function
+            `_show_gradient_dominating_parameter`
+
+        Returns:
+          Returns a tuple containing 2 elements:
+            - `param_groups` with type List[Dict], each Dict element is a parameter group.
+              An example of `param_groups` could be:
+              [
+                  {'params': `one iterable of Parameter`, 'lr': 0.05},
+                  {'params': `another iterable of Parameter`, 'lr': 0.08},
+                  {'params': `a third iterable of Parameter`, 'lr': 0.1},
+              ]
+            - `param_gruops_names` with type List[List[str]],
+               each `List[str]` is for a group['params'] in param_groups,
+               and each `str` is the name of a parameter.
+               A dummy name "foo" is related to each parameter,
+               if input are params without names, i.e. case 1 or case 2.
+        """
+        # variable naming convention in this function:
+        #   p is short for param.
+        #   np is short for named_param.
+        #   p_or_np is short for param_or_named_param.
+        #   cur is short for current.
+        #   group is a dict, e.g. {'params': iterable of parameter, 'lr': 0.05, other fields}.
+        #   groups is a List[group]
+
+        iterable_or_groups = list(params_or_named_params)
+        if len(iterable_or_groups) == 0:
+            raise ValueError("optimizer got an empty parameter list")
+
+        # The first value of returned tuple.  A list of dicts containing at
+        # least 'params' as a key.
+        param_groups = []
+
+        # The second value of returned tuple,
+        # a List[List[str]], each sub-List is for a group.
+        param_groups_names = []
+
+        if not isinstance(iterable_or_groups[0], dict):
+            # case 1 or case 3,
+            # the input is an iterable of parameter or named parameter.
+            param_iterable_cur_group = []
+            param_names_cur_group = []
+            for p_or_np in iterable_or_groups:
+                if isinstance(p_or_np, tuple):
+                    # case 3
+                    name, param = p_or_np
+                else:
+                    # case 1
+                    assert isinstance(p_or_np, torch.Tensor)
+                    param = p_or_np
+                    # Assign a dummy name as a placeholder
+                    name = "foo"
+                    self.show_dominant_parameters = False
+                param_iterable_cur_group.append(param)
+                param_names_cur_group.append(name)
+            param_groups.append({"params": param_iterable_cur_group})
+            param_groups_names.append(param_names_cur_group)
+        else:
+            # case 2 or case 4
+            # the input is groups of parameter or named parameter.
+            for cur_group in iterable_or_groups:
+                assert "named_params" in cur_group
+                name_list = [x[0] for x in cur_group["named_params"]]
+                p_list = [x[1] for x in cur_group["named_params"]]
+                del cur_group["named_params"]
+                cur_group["params"] = p_list
+                param_groups.append(cur_group)
+                param_groups_names.append(name_list)
+
+        return param_groups, param_groups_names
+
+    def __setstate__(self, state):
+        super(ScaledAdam, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        batch = True
+
+        for group, group_params_names in zip(self.param_groups, self.parameters_names):
+
+            with self.batched_params(group["params"], group_params_names) as batches:
+
+                # batches is list of pairs (stacked_param, state).  stacked_param is like
+                # a regular parameter, and will have a .grad, but the 1st dim corresponds to
+                # a stacking dim, it is not a real dim.
+
+                if (
+                    len(batches[0][1]) == 0
+                ):  # if len(first state) == 0: not yet initialized
+                    clipping_scale = 1
+                else:
+                    clipping_scale = self._get_clipping_scale(group, batches)
+
+                for p, state, _ in batches:
+                    # Perform optimization step.
+                    # grad is not going to be None, we handled that when creating the batches.
+                    grad = p.grad
+                    if grad.is_sparse:
+                        raise RuntimeError(
+                            "ScaledAdam optimizer does not support sparse gradients"
+                        )
+                    # State initialization
+                    if len(state) == 0:
+                        self._init_state(group, p, state)
+
+                    self._step_one_batch(group, p, state, clipping_scale)
+
+        return loss
+
+    def _init_state(self, group: dict, p: Tensor, state: dict):
+        """
+        Initializes state dict for parameter 'p'.  Assumes that dim 0 of tensor p
+        is actually the batch dimension, corresponding to batched-together
+        parameters of a given shape.
+
+
+        Args:
+           group:   Dict to look up configuration values.
+               p: The parameter that we are initializing the state for
+           state: Dict from string to whatever state we are initializing
+        """
+        size_update_period = group["size_update_period"]
+
+        state["step"] = 0
+
+        kwargs = {"device": p.device, "dtype": p.dtype}
+
+        # 'delta' implements conventional momentum.  There are
+        # several different kinds of update going on, so rather than
+        # compute "exp_avg" like in Adam, we store and decay a
+        # parameter-change "delta", which combines all forms of
+        # update.  this is equivalent to how it's done in Adam,
+        # except for the first few steps.
+        state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+
+        if numel > 1:
+            # "param_rms" just periodically records the scalar root-mean-square value of
+            # the parameter tensor.
+            # it has a shape like (batch_size, 1, 1, 1, 1)
+            param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+            state["param_rms"] = param_rms
+
+            state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
+            state["scale_grads"] = torch.zeros(
+                size_update_period, *param_rms.shape, **kwargs
+            )
+
+        # exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
+        state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+    def _get_clipping_scale(
+        self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]]
+    ) -> float:
+        """
+        Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
+        by this amount before applying the rest of the update.
+
+        Args:
+           group: the parameter group, an item in self.param_groups
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+        """
+        assert len(tuples) >= 1
+        clipping_scale = group["clipping_scale"]
+        (first_p, first_state, _) = tuples[0]
+        step = first_state["step"]
+        if clipping_scale is None or step == 0:
+            # no clipping.  return early on step == 0 because the other
+            # parameters' state won't have been initialized yet.
+            return 1.0
+        clipping_update_period = group["clipping_update_period"]
+        scalar_lr_scale = group["scalar_lr_scale"]
+
+        tot_sumsq = torch.tensor(0.0, device=first_p.device)
+        for (p, state, param_names) in tuples:
+            grad = p.grad
+            if grad.is_sparse:
+                raise RuntimeError(
+                    "ScaledAdam optimizer does not support sparse gradients"
+                )
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                tot_sumsq += (grad**2).sum() * (
+                    scalar_lr_scale**2
+                )  # sum() to change shape [1] to []
+            else:
+                tot_sumsq += ((grad * state["param_rms"]) ** 2).sum()
+
+        tot_norm = tot_sumsq.sqrt()
+        if "model_norms" not in first_state:
+            first_state["model_norms"] = torch.zeros(
+                clipping_update_period, device=p.device
+            )
+        first_state["model_norms"][step % clipping_update_period] = tot_norm
+
+        irregular_estimate_steps = [
+            i for i in [10, 20, 40] if i < clipping_update_period
+        ]
+        if step % clipping_update_period == 0 or step in irregular_estimate_steps:
+            # Print some stats.
+            # We don't reach here if step == 0 because we would have returned
+            # above.
+            sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
+            if step in irregular_estimate_steps:
+                sorted_norms = sorted_norms[-step:]
+            num_norms = sorted_norms.numel()
+            quartiles = []
+            for n in range(0, 5):
+                index = min(num_norms - 1, (num_norms // 4) * n)
+                quartiles.append(sorted_norms[index].item())
+
+            median = quartiles[2]
+            if median - median != 0:
+                raise RuntimeError("Too many grads were not finite")
+            threshold = clipping_scale * median
+            if step in irregular_estimate_steps:
+                # use larger thresholds on first few steps of estimating threshold,
+                # as norm may be changing rapidly.
+                threshold = threshold * 2.0
+            first_state["model_norm_threshold"] = threshold
+            percent_clipped = (
+                first_state["num_clipped"] * 100.0 / num_norms
+                if "num_clipped" in first_state
+                else 0.0
+            )
+            first_state["num_clipped"] = 0
+            quartiles = " ".join(["%.3e" % x for x in quartiles])
+            logging.warn(
+                f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
+                f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
+            )
+
+        try:
+            model_norm_threshold = first_state["model_norm_threshold"]
+        except KeyError:
+            return 1.0  # threshold has not yet been set.
+
+        ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
+        if ans != ans:  # e.g. ans is nan
+            ans = 0.0
+        if ans < 1.0:
+            first_state["num_clipped"] += 1
+        if ans < 0.1:
+            logging.warn(
+                f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
+            )
+            if self.show_dominant_parameters:
+                assert p.shape[0] == len(param_names)
+                self._show_gradient_dominating_parameter(
+                    tuples, tot_sumsq, group["scalar_lr_scale"]
+                )
+
+        if ans == 0.0:
+            for (p, state, param_names) in tuples:
+                p.grad.zero_()  # get rid of infinity()
+
+        return ans
+
+    def _show_gradient_dominating_parameter(
+        self,
+        tuples: List[Tuple[Tensor, dict, List[str]]],
+        tot_sumsq: Tensor,
+        scalar_lr_scale: float,
+    ):
+        """
+        Show information of parameter which dominates tot_sumsq.
+
+        Args:
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+            tot_sumsq: sumsq of all parameters. Though it's could be calculated
+                from tuples, we still pass it to save some time.
+        """
+        all_sumsq_orig = {}
+        for (p, state, batch_param_names) in tuples:
+            # p is a stacked batch parameters.
+            batch_grad = p.grad
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                # Dummy values used by following `zip` statement.
+                batch_rms_orig = torch.full(
+                    p.shape, scalar_lr_scale, device=batch_grad.device
+                )
+            else:
+                batch_rms_orig = state["param_rms"]
+            batch_sumsq_orig = (batch_grad * batch_rms_orig) ** 2
+            if batch_grad.ndim > 1:
+                # need to guard it with if-statement because sum() sums over
+                # all dims if dim == ().
+                batch_sumsq_orig = batch_sumsq_orig.sum(
+                    dim=list(range(1, batch_grad.ndim))
+                )
+            for name, sumsq_orig, rms, grad in zip(
+                batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
+            ):
+
+                proportion_orig = sumsq_orig / tot_sumsq
+                all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
+
+        sorted_by_proportion = {
+            k: v
+            for k, v in sorted(
+                all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True
+            )
+        }
+        dominant_param_name = next(iter(sorted_by_proportion))
+        (
+            dominant_proportion,
+            dominant_sumsq,
+            dominant_rms,
+            dominant_grad,
+        ) = sorted_by_proportion[dominant_param_name]
+        logging.warn(
+            f"Parameter dominating tot_sumsq {dominant_param_name}"
+            f" with proportion {dominant_proportion:.2f},"
+            f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
+            f"={dominant_sumsq:.3e},"
+            f" grad_sumsq={(dominant_grad**2).sum():.3e},"
+            f" orig_rms_sq={(dominant_rms**2).item():.3e}"
+        )
+
+    def _step_one_batch(
+        self, group: dict, p: Tensor, state: dict, clipping_scale: float
+    ):
+        """
+        Do the step for one parameter, which is actually going to be a batch of
+        `real` parameters, with dim 0 as the batch dim.
+        Args:
+                  group:  dict to look up configuration values
+                    p: parameter to update (actually multiple parameters stacked together
+                       as a batch)
+                  state: state-dict for p, to look up the optimizer state
+        """
+        lr = group["lr"]
+        size_update_period = group["size_update_period"]
+        beta1 = group["betas"][0]
+
+        grad = p.grad
+        if clipping_scale != 1.0:
+            grad *= clipping_scale
+        step = state["step"]
+        delta = state["delta"]
+
+        delta.mul_(beta1)
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        if numel > 1:
+            # Update the size/scale of p, and set param_rms
+            scale_grads = state["scale_grads"]
+            scale_grads[step % size_update_period] = (p * grad).sum(
+                dim=list(range(1, p.ndim)), keepdim=True
+            )
+            if step % size_update_period == size_update_period - 1:
+                param_rms = state["param_rms"]  # shape: (batch_size, 1, 1, ..)
+                param_rms.copy_(
+                    (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+                )
+                if step > 0:
+                    # self._size_update() learns the overall scale on the
+                    # parameter, by shrinking or expanding it.
+                    self._size_update(group, scale_grads, p, state)
+
+        if numel == 1:
+            # For parameters with 1 element we just use regular Adam.
+            # Updates delta.
+            self._step_scalar(group, p, state)
+        else:
+            self._step(group, p, state)
+
+        state["step"] = step + 1
+
+    def _size_update(
+        self, group: dict, scale_grads: Tensor, p: Tensor, state: dict
+    ) -> None:
+        """
+               Called only where p.numel() > 1, this updates the scale of the parameter.
+               If we imagine: p =  underlying_param * scale.exp(), and we are doing
+               gradient descent on underlying param and on scale, this function does the update
+               on `scale`.
+
+               Args:
+              group: dict to look up configuration values
+        scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
+                      grads w.r.t. the scales.
+                  p:  The parameter to update
+               state: The state-dict of p
+        """
+
+        param_rms = state["param_rms"]
+        beta1, beta2 = group["betas"]
+        size_lr = group["lr"] * group["scalar_lr_scale"]
+        param_min_rms = group["param_min_rms"]
+        param_max_rms = group["param_max_rms"]
+        eps = group["eps"]
+        step = state["step"]
+        batch_size = p.shape[0]
+
+        size_update_period = scale_grads.shape[0]
+        # correct beta2 for the size update period: we will have
+        # faster decay at this level.
+        beta2_corr = beta2**size_update_period
+
+        scale_exp_avg_sq = state["scale_exp_avg_sq"]  # shape: (batch_size, 1, 1, ..)
+        scale_exp_avg_sq.mul_(beta2_corr).add_(
+            (scale_grads**2).mean(dim=0),  # mean over dim `size_update_period`
+            alpha=1 - beta2_corr,
+        )  # shape is (batch_size, 1, 1, ...)
+
+        # The 1st time we reach here is when size_step == 1.
+        size_step = (step + 1) // size_update_period
+        bias_correction2 = 1 - beta2_corr**size_step
+        # we don't bother with bias_correction1; this will help prevent divergence
+        # at the start of training.
+
+        denom = scale_exp_avg_sq.sqrt() + eps
+
+        scale_step = (
+            -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom
+        )
+
+        is_too_small = param_rms < param_min_rms
+
+        # when the param gets too small, just don't shrink it any further.
+        scale_step.masked_fill_(is_too_small, 0.0)
+
+        # and ensure the parameter rms after update never exceeds param_max_rms.
+        # We have to look at the trained model for parameters at or around the
+        # param_max_rms, because sometimes they can indicate a problem with the
+        # topology or settings.
+        scale_step = torch.minimum(scale_step, (param_max_rms - param_rms) / param_rms)
+
+        delta = state["delta"]
+        # the factor of (1-beta1) relates to momentum.
+        delta.add_(p * scale_step, alpha=(1 - beta1))
+
+    def _step(self, group: dict, p: Tensor, state: dict):
+        """
+        This function does the core update of self.step(), in the case where the members of
+        the batch have more than 1 element.
+
+        Args:
+            group: A dict which will be used to look up configuration values
+                p: The parameter to be updated
+             grad: The grad of p
+            state: The state-dict corresponding to parameter p
+
+        This function modifies p.
+        """
+        grad = p.grad
+        lr = group["lr"]
+        beta1, beta2 = group["betas"]
+        eps = group["eps"]
+        param_min_rms = group["param_min_rms"]
+        step = state["step"]
+
+        exp_avg_sq = state["exp_avg_sq"]
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
+
+        this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0)
+        bias_correction2 = 1 - beta2 ** (this_step + 1)
+        if bias_correction2 < 0.99:
+            # note: not in-place.
+            exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
+
+        denom = exp_avg_sq.sqrt()
+        denom += eps
+        grad = grad / denom
+
+        alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
+
+        delta = state["delta"]
+        delta.add_(grad * alpha)
+        p.add_(delta)
+
+    def _step_scalar(self, group: dict, p: Tensor, state: dict):
+        """
+        A simplified form of the core update for scalar tensors, where we cannot get a good
+        estimate of the parameter rms.
+        """
+        beta1, beta2 = group["betas"]
+        scalar_max = group["scalar_max"]
+        eps = group["eps"]
+        lr = group["lr"] * group["scalar_lr_scale"]
+        grad = p.grad
+
+        exp_avg_sq = state["exp_avg_sq"]  # shape: (batch_size,)
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+        # bias_correction2 is like in Adam.  Don't bother with bias_correction1;
+        # slower update at the start will help stability anyway.
+        bias_correction2 = 1 - beta2 ** (state["step"] + 1)
+        denom = (exp_avg_sq / bias_correction2).sqrt() + eps
+
+        delta = state["delta"]
+        delta.add_(grad / denom, alpha=-lr * (1 - beta1))
+        p.clamp_(min=-scalar_max, max=scalar_max)
+        p.add_(delta)
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("base_lr", group["lr"])
+
+        self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            logging.warn(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+    The basic formula (before warmup) is:
+      lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                     (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup
+    where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
+    and then stays constant at 1.
+
+    If you don't have the concept of epochs, or one epoch takes a very long time,
+    you can replace the notion of 'epoch' with some measure of the amount of data
+    processed, e.g. hours of data or frames of data, with 'lr_epochs' being set to
+    some measure representing "quite a lot of data": say, one fifth or one third
+    of an entire training run, but it doesn't matter much.  You could also use
+    Eden2 which has only the notion of batches.
+
+    We suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        warmup_batches: Union[int, float] = 500.0,
+        warmup_start: float = 0.5,
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+        self.warmup_batches = warmup_batches
+
+        assert 0.0 <= warmup_start <= 1.0, warmup_start
+        self.warmup_start = warmup_start
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        warmup_factor = (
+            1.0
+            if self.batch >= self.warmup_batches
+            else self.warmup_start
+            + (1.0 - self.warmup_start) * (self.batch / self.warmup_batches)
+            # else 0.5 + 0.5 * (self.batch / self.warmup_batches)
+        )
+
+        return [x * factor * warmup_factor for x in self.base_lrs]
+
+
+class Eden2(LRScheduler):
+    """
+    Eden2 scheduler, simpler than Eden because it does not use the notion of epoch,
+    only batches.
+
+    The basic formula (before warmup) is:
+      lr = base_lr * ((batch**2 + lr_batches**2) / lr_batches**2) ** -0.5) * warmup
+
+    where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
+    and then stays constant at 1.
+
+
+     E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        warmup_batches: Union[int, float] = 500.0,
+        warmup_start: float = 0.5,
+        verbose: bool = False,
+    ):
+        super().__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.warmup_batches = warmup_batches
+
+        assert 0.0 <= warmup_start <= 1.0, warmup_start
+        self.warmup_start = warmup_start
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.5
+        warmup_factor = (
+            1.0
+            if self.batch >= self.warmup_batches
+            else self.warmup_start
+            + (1.0 - self.warmup_start) * (self.batch / self.warmup_batches)
+            # else 0.5 + 0.5 * (self.batch / self.warmup_batches)
+        )
+
+        return [x * factor * warmup_factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = ScaledAdam(m.parameters(), lr=0.03)
+
+    scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+
+    logging.info(f"last lr = {scheduler.get_last_lr()}")
+    logging.info(f"state dict = {scheduler.state_dict()}")
+
+
+# This is included mostly as a baseline for ScaledAdam.
+class Eve(Optimizer):
+    """
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                if random.random() < 0.0005:
+                    step = (exp_avg / denom) * step_size
+                    logging.info(
+                        f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}"
+                    )
+
+        return loss
+
+
+def _test_scaled_adam(hidden_dim: int):
+    import timeit
+
+    from scaling import ScaledLinear
+
+    E = 100
+    B = 4
+    T = 2
+    logging.info("in test_eve_cain")
+    # device = torch.device('cuda')
+    device = torch.device("cpu")
+    dtype = torch.float32
+
+    fix_random_seed(42)
+    # these input_magnitudes and output_magnitudes are to test that
+    # Abel is working as we expect and is able to adjust scales of
+    # different dims differently.
+    input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+    output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+
+    for iter in [1, 0]:
+        fix_random_seed(42)
+        Linear = torch.nn.Linear if iter == 0 else ScaledLinear
+
+        m = torch.nn.Sequential(
+            Linear(E, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, E),
+        ).to(device)
+
+        train_pairs = [
+            (
+                100.0
+                * torch.randn(B, T, E, device=device, dtype=dtype)
+                * input_magnitudes,
+                torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes,
+            )
+            for _ in range(20)
+        ]
+
+        if iter == 0:
+            optim = Eve(m.parameters(), lr=0.003)
+        elif iter == 1:
+            optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0)
+        scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False)
+
+        start = timeit.default_timer()
+        avg_loss = 0.0
+        for epoch in range(180):
+            scheduler.step_epoch()
+            # if epoch == 100 and iter in [2,3]:
+            #    optim.reset_speedup()  # check it doesn't crash.
+
+            # if epoch == 130:
+            #    opts = diagnostics.TensorDiagnosticOptions(
+            #        512
+            #    )  # allow 4 megabytes per sub-module
+            #    diagnostic = diagnostics.attach_diagnostics(m, opts)
+
+            for n, (x, y) in enumerate(train_pairs):
+                y_out = m(x)
+                loss = ((y_out - y) ** 2).mean() * 100.0
+                if epoch == 0 and n == 0:
+                    avg_loss = loss.item()
+                else:
+                    avg_loss = 0.98 * avg_loss + 0.02 * loss.item()
+                if n == 0 and epoch % 5 == 0:
+                    # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item()
+                    # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item()
+                    # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item()
+                    # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item()
+                    # scale1 = '%.2e' % (m[0].weight_scale.exp().item())
+                    # scale1b = '%.2e' % (m[0].bias_scale.exp().item())
+                    # scale2 = '%.2e' % (m[2].weight_scale.exp().item())
+                    # scale2b = '%.2e' % (m[2].bias_scale.exp().item())
+                    lr = scheduler.get_last_lr()[0]
+                    logging.info(
+                        f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}"
+                    )  # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
+                loss.log().backward()
+                optim.step()
+                optim.zero_grad()
+                scheduler.step_batch()
+
+        # diagnostic.print_diagnostics()
+
+        stop = timeit.default_timer()
+        logging.info(f"Iter={iter}, Time taken: {stop - start}")
+
+        logging.info(f"last lr = {scheduler.get_last_lr()}")
+        # logging.info("state dict = ", scheduler.state_dict())
+        # logging.info("optim state_dict = ", optim.state_dict())
+        logging.info(f"input_magnitudes = {input_magnitudes}")
+        logging.info(f"output_magnitudes = {output_magnitudes}")
+
+
+if __name__ == "__main__":
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    logging.getLogger().setLevel(logging.INFO)
+    import subprocess
+
+    s = subprocess.check_output(
+        "git status -uno .; git log -1; git diff HEAD .", shell=True
+    )
+    logging.info(s)
+    import sys
+
+    if len(sys.argv) > 1:
+        hidden_dim = int(sys.argv[1])
+    else:
+        hidden_dim = 200
+
+    _test_scaled_adam(hidden_dim)
+    _test_eden()
diff --git a/egs/librispeech/ASR/zipformer/pretrained.py b/egs/librispeech/ASR/zipformer/pretrained.py
new file mode 100755
index 000000000..3104b6084
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/pretrained.py
@@ -0,0 +1,381 @@
+#!/usr/bin/env python3
+# Copyright      2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+Note: This is a example for librispeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+Usage of this script:
+
+- For non-streaming model:
+
+(1) greedy search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) modified beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --method modified_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) fast beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --method fast_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+- For streaming model:
+
+(1) greedy search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) modified beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --method modified_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) fast beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bpe_500/tokens.txt \
+  --method fast_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+You can also use `./zipformer/exp/epoch-xx.pt`.
+
+Note: ./zipformer/exp/pretrained.pt is generated by ./zipformer/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    fast_beam_search_one_best,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from export import num_tokens
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall.utils import make_pad_mask
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+
+    logging.info("Creating model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    # model forward
+    encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths)
+
+    hyps = []
+    msg = f"Using {params.method}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        raise ValueError(f"Unsupported method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/pretrained_ctc.py b/egs/librispeech/ASR/zipformer/pretrained_ctc.py
new file mode 100755
index 000000000..9dff2e6fc
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/pretrained_ctc.py
@@ -0,0 +1,455 @@
+#!/usr/bin/env python3
+# Copyright      2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                         Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --use-ctc 1 \
+  --causal 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+Usage of this script:
+
+(1) ctc-decoding
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --method ctc-decoding \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) 1best
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --method 1best \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) nbest-rescoring
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method nbest-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+(4) whole-lattice-rescoring
+./zipformer/pretrained_ctc.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --HLG data/lang_bpe_500/HLG.pt \
+  --words-file data/lang_bpe_500/words.txt  \
+  --G data/lm/G_4_gram.pt \
+  --method whole-lattice-rescoring \
+  --sample-rate 16000 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from ctc_decode import get_decoding_params
+from export import num_tokens
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a token table,
+            i.e., lang_dir/tokens.txt, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) nbest-rescoring. Extract n paths from the decoding lattice,
+            rescore them with an LM, the path with
+            the highest score is the decoding result.
+            We call it HLG decoding + nbest n-gram LM rescoring.
+        (3) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + whole-lattice n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or nbest-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and nbest-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""
+        Used only when method is nbest-rescoring.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for blank
+    params.blank_id = token_table["<blk>"]
+    assert params.blank_id == 0
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    batch_size = ctc_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i].item() // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        max_token_id = params.vocab_size - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = [[token_table[i] for i in ids] for ids in token_ids]
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "nbest-rescoring",
+            "whole-lattice-rescoring",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            G = G.to(device)
+            if params.method == "whole-lattice-rescoring":
+                # Add epsilon self-loops to G as we will compose
+                # it with the whole lattice later
+                G = k2.add_epsilon_self_loops(G)
+                G = k2.arc_sort(G)
+
+            # G.lm_scores is used to replace HLG.lm_scores during
+            # LM rescoring.
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=ctc_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        if params.method == "nbest-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_n_best_list(
+                lattice=lattice,
+                G=G,
+                num_paths=params.num_paths,
+                lm_scale_list=[params.ngram_lm_scale],
+                nbest_scale=params.nbest_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    if params.method == "ctc-decoding":
+        for filename, hyp in zip(params.sound_files, hyps):
+            words = "".join(hyp)
+            words = words.replace("▁", " ").strip()
+            s += f"{filename}:\n{words}\n\n"
+    elif params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]:
+        for filename, hyp in zip(params.sound_files, hyps):
+            words = " ".join(hyp)
+            words = words.replace("▁", " ").strip()
+            s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/profile.py b/egs/librispeech/ASR/zipformer/profile.py
new file mode 100755
index 000000000..57f44a90a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/profile.py
@@ -0,0 +1,170 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage: ./zipformer/profile.py
+"""
+
+import argparse
+import logging
+import sentencepiece as spm
+import torch
+
+from typing import Tuple
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask
+from icefall.profiler import get_model_profile
+from scaling import BiasNorm
+from train import (
+    get_encoder_embed,
+    get_encoder_model,
+    get_joiner_model,
+    add_model_arguments,
+    get_params,
+)
+from zipformer import BypassModule
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _bias_norm_flops_compute(module, input, output):
+    assert len(input) == 1, len(input)
+    # estimate as layer_norm, see icefall/profiler.py
+    flops = input[0].numel() * 5
+    module.__flops__ += int(flops)
+
+
+def _swoosh_module_flops_compute(module, input, output):
+    # For SwooshL and SwooshR modules
+    assert len(input) == 1, len(input)
+    # estimate as swish/silu, see icefall/profiler.py
+    flops = input[0].numel()
+    module.__flops__ += int(flops)
+
+
+def _bypass_module_flops_compute(module, input, output):
+    # For Bypass module
+    assert len(input) == 2, len(input)
+    flops = input[0].numel() * 2
+    module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    BiasNorm: _bias_norm_flops_compute,
+    BypassModule: _bypass_module_flops_compute,
+}
+
+
+class Model(nn.Module):
+    """A Wrapper for encoder, encoder_embed, and encoder_proj"""
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        encoder_embed: nn.Module,
+        encoder_proj: nn.Module,
+    ) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.encoder_proj = encoder_proj
+
+    def forward(self, feature: Tensor, feature_lens: Tensor) -> Tuple[Tensor, Tensor]:
+        x, x_lens = self.encoder_embed(feature, feature_lens)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+
+        encoder_out = encoder_out.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        logits = self.encoder_proj(encoder_out)
+
+        return logits, encoder_out_lens
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    # We only profile the encoder part
+    model = Model(
+        encoder=get_encoder_model(params),
+        encoder_embed=get_encoder_embed(params),
+        encoder_proj=get_joiner_model(params).encoder_proj,
+    )
+    model.eval()
+    model.to(device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # for 30-second input
+    B, T, D = 1, 3000, 80
+    feature = torch.ones(B, T, D, dtype=torch.float32).to(device)
+    feature_lens = torch.full((B,), T, dtype=torch.int64).to(device)
+
+    flops, params = get_model_profile(
+        model=model,
+        args=(feature, feature_lens),
+        module_hoop_mapping=MODULE_HOOK_MAPPING,
+    )
+    logging.info(f"For the encoder part, params: {params}, flops: {flops}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/zipformer/scaling.py b/egs/librispeech/ASR/zipformer/scaling.py
new file mode 100644
index 000000000..c0f1e3087
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/scaling.py
@@ -0,0 +1,1907 @@
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional, Tuple, Union
+import logging
+import k2
+from torch.cuda.amp import custom_fwd, custom_bwd
+import random
+import torch
+import math
+import torch.nn as nn
+from torch import Tensor
+
+
+def logaddexp_onnx(x: Tensor, y: Tensor) -> Tensor:
+    max_value = torch.max(x, y)
+    diff = torch.abs(x - y)
+    return max_value + torch.log1p(torch.exp(-diff))
+
+
+# RuntimeError: Exporting the operator logaddexp to ONNX opset version
+# 14 is not supported. Please feel free to request support or submit
+# a pull request on PyTorch GitHub.
+#
+# The following function is to solve the above error when exporting
+# models to ONNX via torch.jit.trace()
+def logaddexp(x: Tensor, y: Tensor) -> Tensor:
+    # Caution(fangjun): Put torch.jit.is_scripting() before
+    # torch.onnx.is_in_onnx_export();
+    # otherwise, it will cause errors for torch.jit.script().
+    #
+    # torch.logaddexp() works for both torch.jit.script() and
+    # torch.jit.trace() but it causes errors for ONNX export.
+    #
+    if torch.jit.is_scripting():
+        # Note: We cannot use torch.jit.is_tracing() here as it also
+        # matches torch.onnx.export().
+        return torch.logaddexp(x, y)
+    elif torch.onnx.is_in_onnx_export():
+        return logaddexp_onnx(x, y)
+    else:
+        # for torch.jit.trace()
+        return torch.logaddexp(x, y)
+
+
+class PiecewiseLinear(object):
+    """
+    Piecewise linear function, from float to float, specified as nonempty list of (x,y) pairs with
+    the x values in order.  x values <[initial x] or >[final x] are map to [initial y], [final y]
+    respectively.
+    """
+
+    def __init__(self, *args):
+        assert len(args) >= 1, len(args)
+        if len(args) == 1 and isinstance(args[0], PiecewiseLinear):
+            self.pairs = list(args[0].pairs)
+        else:
+            self.pairs = [(float(x), float(y)) for x, y in args]
+        for x, y in self.pairs:
+            assert isinstance(x, (float, int)), type(x)
+            assert isinstance(y, (float, int)), type(y)
+
+        for i in range(len(self.pairs) - 1):
+            assert self.pairs[i + 1][0] > self.pairs[i][0], (
+                i,
+                self.pairs[i],
+                self.pairs[i + 1],
+            )
+
+    def __str__(self):
+        # e.g. 'PiecewiseLinear((0., 10.), (100., 0.))'
+        return f"PiecewiseLinear({str(self.pairs)[1:-1]})"
+
+    def __call__(self, x):
+        if x <= self.pairs[0][0]:
+            return self.pairs[0][1]
+        elif x >= self.pairs[-1][0]:
+            return self.pairs[-1][1]
+        else:
+            cur_x, cur_y = self.pairs[0]
+            for i in range(1, len(self.pairs)):
+                next_x, next_y = self.pairs[i]
+                if x >= cur_x and x <= next_x:
+                    return cur_y + (next_y - cur_y) * (x - cur_x) / (next_x - cur_x)
+                cur_x, cur_y = next_x, next_y
+            assert False
+
+    def __mul__(self, alpha):
+        return PiecewiseLinear(*[(x, y * alpha) for x, y in self.pairs])
+
+    def __add__(self, x):
+        if isinstance(x, (float, int)):
+            return PiecewiseLinear(*[(p[0], p[1] + x) for p in self.pairs])
+        s, x = self.get_common_basis(x)
+        return PiecewiseLinear(
+            *[(sp[0], sp[1] + xp[1]) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def max(self, x):
+        if isinstance(x, (float, int)):
+            x = PiecewiseLinear((0, x))
+        s, x = self.get_common_basis(x, include_crossings=True)
+        return PiecewiseLinear(
+            *[(sp[0], max(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def min(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            x = PiecewiseLinear((0, x))
+        s, x = self.get_common_basis(x, include_crossings=True)
+        return PiecewiseLinear(
+            *[(sp[0], min(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def __eq__(self, other):
+        return self.pairs == other.pairs
+
+    def get_common_basis(self, p: "PiecewiseLinear", include_crossings: bool = False):
+        """
+        Returns (self_mod, p_mod) which are equivalent piecewise linear
+        functions to self and p, but with the same x values.
+
+          p: the other piecewise linear function
+          include_crossings: if true, include in the x values positions
+              where the functions indicate by this and p crosss.
+        """
+        assert isinstance(p, PiecewiseLinear), type(p)
+
+        # get sorted x-values without repetition.
+        x_vals = sorted(set([x for x, _ in self.pairs] + [x for x, _ in p.pairs]))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+
+        if include_crossings:
+            extra_x_vals = []
+            for i in range(len(x_vals) - 1):
+                if (y_vals1[i] > y_vals2[i]) != (y_vals1[i + 1] > y_vals2[i + 1]):
+                    # if the two lines in this subsegment potentially cross each other..
+                    diff_cur = abs(y_vals1[i] - y_vals2[i])
+                    diff_next = abs(y_vals1[i + 1] - y_vals2[i + 1])
+                    # `pos`, between 0 and 1, gives the relative x position,
+                    # with 0 being x_vals[i] and 1 being x_vals[i+1].
+                    pos = diff_cur / (diff_cur + diff_next)
+                    extra_x_val = x_vals[i] + pos * (x_vals[i + 1] - x_vals[i])
+                    extra_x_vals.append(extra_x_val)
+            if len(extra_x_vals) > 0:
+                x_vals = sorted(set(x_vals + extra_x_vals))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+        return (
+            PiecewiseLinear(*zip(x_vals, y_vals1)),
+            PiecewiseLinear(*zip(x_vals, y_vals2)),
+        )
+
+
+class ScheduledFloat(torch.nn.Module):
+    """
+    This object is a torch.nn.Module only because we want it to show up in [top_level module].modules();
+    it does not have a working forward() function.  You are supposed to cast it to float, as
+    in, float(parent_module.whatever), and use it as something like a dropout prob.
+
+    It is a floating point value whose value changes depending on the batch count of the
+    training loop.  It is a piecewise linear function where you specify the (x,y) pairs
+    in sorted order on x; x corresponds to the batch index.  For batch-index values before the
+    first x or after the last x, we just use the first or last y value.
+
+    Example:
+       self.dropout = ScheduledFloat((0.0, 0.2), (4000.0, 0.0), default=0.0)
+
+    `default` is used when self.batch_count is not set or not in training mode or in
+     torch.jit scripting mode.
+    """
+
+    def __init__(self, *args, default: float = 0.0):
+        super().__init__()
+        # self.batch_count and self.name will be written to in the training loop.
+        self.batch_count = None
+        self.name = None
+        self.default = default
+        self.schedule = PiecewiseLinear(*args)
+
+    def extra_repr(self) -> str:
+        return (
+            f"batch_count={self.batch_count}, schedule={str(self.schedule.pairs[1:-1])}"
+        )
+
+    def __float__(self):
+        batch_count = self.batch_count
+        if (
+            batch_count is None
+            or not self.training
+            or torch.jit.is_scripting()
+            or torch.jit.is_tracing()
+        ):
+            return float(self.default)
+        else:
+            ans = self.schedule(self.batch_count)
+            if random.random() < 0.0002:
+                logging.info(
+                    f"ScheduledFloat: name={self.name}, batch_count={self.batch_count}, ans={ans}"
+                )
+            return ans
+
+    def __add__(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return ScheduledFloat(self.schedule + x, default=self.default)
+        else:
+            return ScheduledFloat(
+                self.schedule + x.schedule, default=self.default + x.default
+            )
+
+    def max(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return ScheduledFloat(self.schedule.max(x), default=self.default)
+        else:
+            return ScheduledFloat(
+                self.schedule.max(x.schedule), default=max(self.default, x.default)
+            )
+
+
+FloatLike = Union[float, ScheduledFloat]
+
+
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
+    """
+    A randomized way of casting a floating point value to half precision.
+    """
+    if x.dtype == torch.float16:
+        return x
+    x_abs = x.abs()
+    is_too_small = x_abs < min_abs
+    # for elements where is_too_small is true, random_val will contain +-min_abs with
+    # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
+    # for those elements].
+    random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs)
+    return torch.where(is_too_small, random_val, x).to(torch.float16)
+
+
+class CutoffEstimator:
+    """
+    Estimates cutoffs of an arbitrary numerical quantity such that a specified
+    proportion of items will be above the cutoff on average.
+
+      p is the proportion of items that should be above the cutoff.
+    """
+
+    def __init__(self, p: float):
+        self.p = p
+        # total count of items
+        self.count = 0
+        # total count of items that were above the cutoff
+        self.count_above = 0
+        # initial cutoff value
+        self.cutoff = 0
+
+    def __call__(self, x: float) -> bool:
+        """
+        Returns true if x is above the cutoff.
+        """
+        ans = x > self.cutoff
+        self.count += 1
+        if ans:
+            self.count_above += 1
+        cur_p = self.count_above / self.count
+        delta_p = cur_p - self.p
+        if (delta_p > 0) == ans:
+            q = abs(delta_p)
+            self.cutoff = x * q + self.cutoff * (1 - q)
+        return ans
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    """
+    Tries to handle half-precision derivatives in a randomized way that should
+    be more accurate for training than the default behavior.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, dim: int):
+        ans = x.softmax(dim=dim)
+        # if x dtype is float16, x.softmax() returns a float32 because
+        # (presumably) that op does not support float16, and autocast
+        # is enabled.
+        if torch.is_autocast_enabled():
+            ans = ans.to(torch.float16)
+        ctx.save_for_backward(ans)
+        ctx.x_dtype = x.dtype
+        ctx.dim = dim
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        (ans,) = ctx.saved_tensors
+        with torch.cuda.amp.autocast(enabled=False):
+            ans_grad = ans_grad.to(torch.float32)
+            ans = ans.to(torch.float32)
+            x_grad = ans_grad * ans
+            x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True)
+            return x_grad, None
+
+
+def softmax(x: Tensor, dim: int):
+    if not x.requires_grad or torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x.softmax(dim=dim)
+
+    return SoftmaxFunction.apply(x, dim)
+
+
+class MaxEigLimiterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.channel_dim = channel_dim
+        ctx.grad_scale = grad_scale
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        with torch.enable_grad():
+            (x_orig, coeffs, new_direction) = ctx.saved_tensors
+            x_orig.requires_grad = True
+            num_channels = x_orig.shape[ctx.channel_dim]
+            x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
+            new_direction.requires_grad = False
+            x = x - x.mean(dim=0)
+            x_var = (x**2).mean()
+            x_residual = x - coeffs * new_direction
+            x_residual_var = (x_residual**2).mean()
+            # `variance_proportion` is the proportion of the variance accounted for
+            # by the top eigen-direction.  This is to be minimized.
+            variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+            variance_proportion.backward()
+        x_orig_grad = x_orig.grad
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
+        return x_grad + x_extra_grad.detach(), None, None, None, None
+
+
+class BiasNormFunction(torch.autograd.Function):
+    # This computes:
+    #   scales = (torch.mean((x - bias) ** 2, keepdim=True)) ** -0.5 * log_scale.exp()
+    #   return x * scales
+    # (after unsqueezing the bias), but it does it in a memory-efficient way so that
+    # it can just store the returned value (chances are, this will also be needed for
+    # some other reason, related to the next operation, so we can save memory).
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        bias: Tensor,
+        log_scale: Tensor,
+        channel_dim: int,
+        store_output_for_backprop: bool,
+    ) -> Tensor:
+        assert bias.ndim == 1
+        if channel_dim < 0:
+            channel_dim = channel_dim + x.ndim
+        ctx.store_output_for_backprop = store_output_for_backprop
+        ctx.channel_dim = channel_dim
+        for _ in range(channel_dim + 1, x.ndim):
+            bias = bias.unsqueeze(-1)
+        scales = (
+            torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+        ) * log_scale.exp()
+        ans = x * scales
+        ctx.save_for_backward(
+            ans.detach() if store_output_for_backprop else x,
+            scales.detach(),
+            bias.detach(),
+            log_scale.detach(),
+        )
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tensor:
+        ans_or_x, scales, bias, log_scale = ctx.saved_tensors
+        if ctx.store_output_for_backprop:
+            x = ans_or_x / scales
+        else:
+            x = ans_or_x
+        x = x.detach()
+        x.requires_grad = True
+        bias.requires_grad = True
+        log_scale.requires_grad = True
+        with torch.enable_grad():
+            # recompute scales from x, bias and log_scale.
+            scales = (
+                torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) ** -0.5
+            ) * log_scale.exp()
+            ans = x * scales
+            ans.backward(gradient=ans_grad)
+        return x.grad, bias.grad.flatten(), log_scale.grad, None, None
+
+
+class BiasNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    Instead, we give the BiasNorm a trainable bias that it can use when
+    computing the scale for normalization.  We also give it a (scalar)
+    trainable scale on the output.
+
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+       channel_dim: the axis/dimension corresponding to the channel,
+         interpreted as an offset from the input's ndim if negative.
+         This is NOT the num_channels; it should typically be one of
+         {-2, -1, 0, 1, 2, 3}.
+      log_scale: the initial log-scale that we multiply the output by; this
+         is learnable.
+      log_scale_min: FloatLike, minimum allowed value of log_scale
+      log_scale_max: FloatLike, maximum allowed value of log_scale
+      store_output_for_backprop: only possibly affects memory use; recommend
+         to set to True if you think the output of this module is more likely
+         than the input of this module to be required to be stored for the
+         backprop.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        log_scale: float = 1.0,
+        log_scale_min: float = -1.5,
+        log_scale_max: float = 1.5,
+        store_output_for_backprop: bool = False,
+    ) -> None:
+        super(BiasNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.log_scale = nn.Parameter(torch.tensor(log_scale))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+
+        self.log_scale_min = log_scale_min
+        self.log_scale_max = log_scale_max
+
+        self.store_output_for_backprop = store_output_for_backprop
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            channel_dim = self.channel_dim
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            bias = self.bias
+            for _ in range(channel_dim + 1, x.ndim):
+                bias = bias.unsqueeze(-1)
+            scales = (
+                torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+            ) * self.log_scale.exp()
+            return x * scales
+
+        log_scale = limit_param_value(
+            self.log_scale,
+            min=float(self.log_scale_min),
+            max=float(self.log_scale_max),
+            training=self.training,
+        )
+
+        return BiasNormFunction.apply(
+            x, self.bias, log_scale, self.channel_dim, self.store_output_for_backprop
+        )
+
+
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
+    """
+    Behaves like a constructor of a modified version of nn.Linear
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Linear(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv1d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv1d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv2d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv2d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv2d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False, but:
+    NO PADDING-RELATED ARGS.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv2d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+class ChunkCausalDepthwiseConv1d(torch.nn.Module):
+    """
+    Behaves like a depthwise 1d convolution, except that it is causal in
+    a chunkwise way, as if we had a block-triangular attention mask.
+    The chunk size is provided at test time (it should probably be
+    kept in sync with the attention mask).
+
+    This has a little more than twice the parameters of a conventional
+    depthwise conv1d module: we implement it by having one
+    depthwise convolution, of half the width, that is causal (via
+    right-padding); and one depthwise convolution that is applied only
+    within chunks, that we multiply by a scaling factor which depends
+    on the position within the chunk.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        initial_scale: float = 1.0,
+        bias: bool = True,
+    ):
+        super().__init__()
+        assert kernel_size % 2 == 1
+
+        half_kernel_size = (kernel_size + 1) // 2
+        # will pad manually, on one side.
+        self.causal_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=half_kernel_size,
+            padding=0,
+            bias=True,
+        )
+
+        self.chunkwise_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            bias=bias,
+        )
+
+        # first row is correction factors added to the scale near the left edge of the chunk,
+        # second row is correction factors added to the scale near the right edge of the chunk,
+        # both of these are added to a default scale of 1.0.
+        self.chunkwise_conv_scale = nn.Parameter(torch.zeros(2, channels, kernel_size))
+        self.kernel_size = kernel_size
+
+        with torch.no_grad():
+            self.causal_conv.weight[:] *= initial_scale
+            self.chunkwise_conv.weight[:] *= initial_scale
+            if bias:
+                torch.nn.init.uniform_(
+                    self.causal_conv.bias, -0.1 * initial_scale, 0.1 * initial_scale
+                )
+
+    def forward(self, x: Tensor, chunk_size: int = -1) -> Tensor:
+        """
+             Forward function.  Args:
+               x: a Tensor of shape (batch_size, channels, seq_len)
+        chunk_size: the chunk size, in frames; does not have to divide seq_len exactly.
+        """
+        (batch_size, num_channels, seq_len) = x.shape
+
+        # half_kernel_size = self.kernel_size + 1 // 2
+        # left_pad is half_kernel_size - 1 where half_kernel_size is the size used
+        # in the causal conv.  It's the amount by which we must pad on the left,
+        # to make the convolution causal.
+        left_pad = self.kernel_size // 2
+
+        if chunk_size < 0 or chunk_size > seq_len:
+            chunk_size = seq_len
+        right_pad = -seq_len % chunk_size
+
+        x = torch.nn.functional.pad(x, (left_pad, right_pad))
+
+        x_causal = self.causal_conv(x[..., : left_pad + seq_len])
+        assert x_causal.shape == (batch_size, num_channels, seq_len)
+
+        x_chunk = x[..., left_pad:]
+        num_chunks = x_chunk.shape[2] // chunk_size
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks, chunk_size)
+        x_chunk = x_chunk.permute(0, 2, 1, 3).reshape(
+            batch_size * num_chunks, num_channels, chunk_size
+        )
+        x_chunk = self.chunkwise_conv(x_chunk)  # does not change shape
+
+        chunk_scale = self._get_chunk_scale(chunk_size)
+
+        x_chunk = x_chunk * chunk_scale
+        x_chunk = x_chunk.reshape(
+            batch_size, num_chunks, num_channels, chunk_size
+        ).permute(0, 2, 1, 3)
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks * chunk_size)[
+            ..., :seq_len
+        ]
+
+        return x_chunk + x_causal
+
+    def _get_chunk_scale(self, chunk_size: int):
+        """Returns tensor of shape (num_channels, chunk_size) that will be used to
+        scale the output of self.chunkwise_conv."""
+        left_edge = self.chunkwise_conv_scale[0]
+        right_edge = self.chunkwise_conv_scale[1]
+        if chunk_size < self.kernel_size:
+            left_edge = left_edge[:, :chunk_size]
+            right_edge = right_edge[:, -chunk_size:]
+        else:
+            t = chunk_size - self.kernel_size
+            channels = left_edge.shape[0]
+            pad = torch.zeros(
+                channels, t, device=left_edge.device, dtype=left_edge.dtype
+            )
+            left_edge = torch.cat((left_edge, pad), dim=-1)
+            right_edge = torch.cat((pad, right_edge), dim=-1)
+        return 1.0 + (left_edge + right_edge)
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cache: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Streaming Forward function.
+
+        Args:
+            x: a Tensor of shape (batch_size, channels, seq_len)
+            cache: cached left context of shape (batch_size, channels, left_pad)
+        """
+        (batch_size, num_channels, seq_len) = x.shape
+
+        # left_pad is half_kernel_size - 1 where half_kernel_size is the size used
+        # in the causal conv.  It's the amount by which we must pad on the left,
+        # to make the convolution causal.
+        left_pad = self.kernel_size // 2
+
+        # Pad cache
+        assert cache.shape[-1] == left_pad, (cache.shape[-1], left_pad)
+        x = torch.cat([cache, x], dim=2)
+        # Update cache
+        cache = x[..., -left_pad:]
+
+        x_causal = self.causal_conv(x)
+        assert x_causal.shape == (batch_size, num_channels, seq_len)
+
+        x_chunk = x[..., left_pad:]
+        x_chunk = self.chunkwise_conv(x_chunk)  # does not change shape
+
+        chunk_scale = self._get_chunk_scale(chunk_size=seq_len)
+        x_chunk = x_chunk * chunk_scale
+
+        return x_chunk + x_causal, cache
+
+
+class BalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        min_mean: float,
+        max_mean: float,
+        min_rms: float,
+        max_rms: float,
+        grad_scale: float,
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        ctx.save_for_backward(x)
+        ctx.config = (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None, None, None]:
+        (x,) = ctx.saved_tensors
+        (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim) = ctx.config
+
+        try:
+            with torch.enable_grad():
+                with torch.cuda.amp.autocast(enabled=False):
+                    x = x.to(torch.float32)
+                    x = x.detach()
+                    x.requires_grad = True
+                    mean_dims = [i for i in range(x.ndim) if i != channel_dim]
+                    uncentered_var = (x**2).mean(dim=mean_dims, keepdim=True)
+                    mean = x.mean(dim=mean_dims, keepdim=True)
+                    stddev = (uncentered_var - (mean * mean)).clamp(min=1.0e-20).sqrt()
+                    rms = uncentered_var.clamp(min=1.0e-20).sqrt()
+
+                    m = mean / stddev
+                    # part of loss that relates to mean / stddev
+                    m_loss = (m - m.clamp(min=min_mean, max=max_mean)).abs()
+
+                    # put a much larger scale on the RMS-max-limit loss, so that if both it and the
+                    # m_loss are violated we fix the RMS loss first.
+                    rms_clamped = rms.clamp(min=min_rms, max=max_rms)
+                    r_loss = (rms_clamped / rms).log().abs()
+
+                    loss = m_loss + r_loss
+
+                    loss.backward(gradient=torch.ones_like(loss))
+                    loss_grad = x.grad
+                    loss_grad_rms = (
+                        (loss_grad**2)
+                        .mean(dim=mean_dims, keepdim=True)
+                        .sqrt()
+                        .clamp(min=1.0e-20)
+                    )
+
+                    loss_grad = loss_grad * (grad_scale / loss_grad_rms)
+
+                    x_grad_float = x_grad.to(torch.float32)
+                    # scale each element of loss_grad by the absolute value of the corresponding
+                    # element of x_grad, which we view as a noisy estimate of its magnitude for that
+                    # (frame and dimension).  later we can consider factored versions.
+                    x_grad_mod = x_grad_float + (x_grad_float.abs() * loss_grad)
+                    x_grad = x_grad_mod.to(x_grad.dtype)
+        except Exception as e:
+            logging.info(
+                f"Caught exception in Balancer backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
+
+        return x_grad, None, None, None, None, None, None
+
+
+class Balancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+         prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: FloatLike = 0.05,
+        max_positive: FloatLike = 0.95,
+        min_abs: FloatLike = 0.2,
+        max_abs: FloatLike = 100.0,
+        grad_scale: FloatLike = 0.04,
+        prob: Optional[FloatLike] = None,
+    ):
+        super().__init__()
+
+        if prob is None:
+            prob = ScheduledFloat((0.0, 0.5), (8000.0, 0.125), default=0.4)
+        self.prob = prob
+        # 5% of the time we will return and do nothing because memory usage is
+        # too high.
+        self.mem_cutoff = CutoffEstimator(0.05)
+
+        # actually self.num_channels is no longer needed except for an assertion.
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.grad_scale = grad_scale
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))
+        ):
+            return _no_op(x)
+
+        prob = float(self.prob)
+        if random.random() < prob:
+            # The following inner-functions convert from the way we historically specified
+            # these limitations, as limits on the absolute value and the proportion of positive
+            # values, to limits on the RMS value and the (mean / stddev).
+            def _abs_to_rms(x):
+                # for normally distributed data, if the expected absolute value is x, the
+                # expected rms value will be sqrt(pi/2) * x.
+                return 1.25331413732 * x
+
+            def _proportion_positive_to_mean(x):
+                def _atanh(x):
+                    eps = 1.0e-10
+                    # eps is to prevent crashes if x is exactly 0 or 1.
+                    # we'll just end up returning a fairly large value.
+                    return (math.log(1 + x + eps) - math.log(1 - x + eps)) / 2.0
+
+                def _approx_inverse_erf(x):
+                    # 1 / (sqrt(pi) * ln(2)),
+                    # see https://math.stackexchange.com/questions/321569/approximating-the-error-function-erf-by-analytical-functions
+                    # this approximation is extremely crude and gets progressively worse for
+                    # x very close to -1 or +1, but we mostly care about the "middle" region
+                    # e.g. _approx_inverse_erf(0.05) = 0.0407316414078772,
+                    # and math.erf(0.0407316414078772) = 0.045935330944660666,
+                    # which is pretty close to 0.05.
+                    return 0.8139535143 * _atanh(x)
+
+                # first convert x from the range 0..1 to the range -1..1 which the error
+                # function returns
+                x = -1 + (2 * x)
+                return _approx_inverse_erf(x)
+
+            min_mean = _proportion_positive_to_mean(float(self.min_positive))
+            max_mean = _proportion_positive_to_mean(float(self.max_positive))
+            min_rms = _abs_to_rms(float(self.min_abs))
+            max_rms = _abs_to_rms(float(self.max_abs))
+            grad_scale = float(self.grad_scale)
+
+            assert x.shape[self.channel_dim] == self.num_channels
+
+            return BalancerFunction.apply(
+                x, min_mean, max_mean, min_rms, max_rms, grad_scale, self.channel_dim
+            )
+        else:
+            return _no_op(x)
+
+
+def penalize_abs_values_gt(
+    x: Tensor, limit: float, penalty: float, name: str = None
+) -> Tensor:
+    """
+    Returns x unmodified, but in backprop will put a penalty for the excess of
+    the absolute values of elements of x over the limit "limit".  E.g. if
+    limit == 10.0, then if x has any values over 10 it will get a penalty.
+
+    Caution: the value of this penalty will be affected by grad scaling used
+    in automatic mixed precision training.  For this reasons we use this,
+    it shouldn't really matter, or may even be helpful; we just use this
+    to disallow really implausible values of scores to be given to softmax.
+
+    The name is for randomly printed debug info.
+    """
+    x_sign = x.sign()
+    over_limit = (x.abs() - limit) > 0
+    # The following is a memory efficient way to penalize the absolute values of
+    # x that's over the limit.  (The memory efficiency comes when you think
+    # about which items torch needs to cache for the autograd, and which ones it
+    # can throw away).  The numerical value of aux_loss as computed here will
+    # actually be larger than it should be, by limit * over_limit.sum(), but it
+    # has the same derivative as the real aux_loss which is penalty * (x.abs() -
+    # limit).relu().
+    aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x)
+    # note: we don't do sum() here on aux)_loss, but it's as if we had done
+    # sum() due to how with_loss() works.
+    x = with_loss(x, aux_loss, name)
+    # you must use x for something, or this will be ineffective.
+    return x
+
+
+def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
+    if x.ndim == 2:
+        return x.diag()
+    else:
+        (batch, dim, dim) = x.shape
+        x = x.reshape(batch, dim * dim)
+        x = x[:, :: dim + 1]
+        assert x.shape == (batch, dim)
+        return x
+
+
+def _whitening_metric(x: Tensor, num_groups: int):
+    """
+    Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
+    of the centered feature covariance are the same within each group's covariance matrix
+    and also between groups.
+    Args:
+        x: a Tensor of shape (*, num_channels)
+     num_groups:  the number of groups of channels, a number >=1 that divides num_channels
+    Returns:
+        Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and
+    greater than 1.0 otherwise.
+    """
+    assert x.dtype != torch.float16
+    x = x.reshape(-1, x.shape[-1])
+    (num_frames, num_channels) = x.shape
+    assert num_channels % num_groups == 0
+    channels_per_group = num_channels // num_groups
+    x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1)
+    # x now has shape (num_groups, num_frames, channels_per_group)
+    # subtract the mean so we use the centered, not uncentered, covariance.
+    # My experience has been that when we "mess with the gradients" like this,
+    # it's better not do anything that tries to move the mean around, because
+    # that can easily cause instability.
+    x = x - x.mean(dim=1, keepdim=True)
+    # x_covar: (num_groups, channels_per_group, channels_per_group)
+    x_covar = torch.matmul(x.transpose(1, 2), x)
+    x_covar_mean_diag = _diag(x_covar).mean()
+    # the following expression is what we'd get if we took the matrix product
+    # of each covariance and measured the mean of its trace, i.e.
+    # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
+    x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group)
+    # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20)
+    return metric
+
+
+class WhiteningPenaltyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, module: nn.Module) -> Tensor:
+        ctx.save_for_backward(x)
+        ctx.module = module
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
+        w = ctx.module
+
+        try:
+            with torch.enable_grad():
+                with torch.cuda.amp.autocast(enabled=False):
+                    x_detached = x_orig.to(torch.float32).detach()
+                    x_detached.requires_grad = True
+
+                    metric = _whitening_metric(x_detached, w.num_groups)
+
+                    if random.random() < 0.005 or __name__ == "__main__":
+                        logging.info(
+                            f"Whitening: name={w.name}, num_groups={w.num_groups}, num_channels={x_orig.shape[-1]}, "
+                            f"metric={metric.item():.2f} vs. limit={float(w.whitening_limit)}"
+                        )
+
+                    if metric < float(w.whitening_limit):
+                        w.prob = w.min_prob
+                        return x_grad, None
+                    else:
+                        w.prob = w.max_prob
+                        metric.backward()
+                        penalty_grad = x_detached.grad
+                        scale = w.grad_scale * (
+                            x_grad.to(torch.float32).norm()
+                            / (penalty_grad.norm() + 1.0e-20)
+                        )
+                        penalty_grad = penalty_grad * scale
+                        return x_grad + penalty_grad.to(x_grad.dtype), None
+        except Exception as e:
+            logging.info(
+                f"Caught exception in Whiten backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
+        return x_grad, None
+
+
+class Whiten(nn.Module):
+    def __init__(
+        self,
+        num_groups: int,
+        whitening_limit: FloatLike,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: FloatLike,
+    ):
+        """
+        Args:
+          num_groups: the number of groups to divide the channel dim into before
+            whitening.  We will attempt to make the feature covariance
+            within each group, after mean subtraction, as "white" as possible,
+            while having the same trace across all groups.
+         whitening_limit: a value greater than 1.0, that dictates how much
+           freedom we have to violate the constraints.  1.0 would mean perfectly
+           white, with exactly the same trace across groups; larger values
+           give more freedom.  E.g. 2.0.
+         prob: the probability with which we apply the gradient modification
+           (also affects the grad scale).  May be supplied as a float,
+           or as a pair (min_prob, max_prob)
+
+          grad_scale: determines the scale on the gradient term from this object,
+            relative to the rest of the gradient on the attention weights.
+            E.g. 0.02 (you may want to use smaller values than this if prob is large)
+        """
+        super(Whiten, self).__init__()
+        assert num_groups >= 1
+        assert float(whitening_limit) >= 1
+        assert grad_scale >= 0
+        self.num_groups = num_groups
+        self.whitening_limit = whitening_limit
+        self.grad_scale = grad_scale
+
+        if isinstance(prob, float):
+            prob = (prob, prob)
+        (self.min_prob, self.max_prob) = prob
+        assert 0 < self.min_prob <= self.max_prob <= 1
+        self.prob = self.max_prob
+        self.name = None  # will be set in training loop
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        In the forward pass, this function just returns the input unmodified.
+        In the backward pass, it will modify the gradients to ensure that the
+        distribution in each group has close to (lambda times I) as the covariance
+        after mean subtraction, with the same lambda across groups.
+        For whitening_limit > 1, there will be more freedom to violate this
+        constraint.
+
+        Args:
+           x: the input of shape (*, num_channels)
+
+        Returns:
+            x, unmodified.   You should make sure
+        you use the returned value, or the graph will be freed
+        and nothing will happen in backprop.
+        """
+        grad_scale = float(self.grad_scale)
+        if not x.requires_grad or random.random() > self.prob or grad_scale == 0:
+            return _no_op(x)
+        else:
+            return WhiteningPenaltyFunction.apply(x, self)
+
+
+class WithLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, y: Tensor, name: str):
+        ctx.y_shape = y.shape
+        if random.random() < 0.002 and name is not None:
+            loss_sum = y.sum().item()
+            logging.info(f"WithLoss: name={name}, loss-sum={loss_sum:.3e}")
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        return (
+            ans_grad,
+            torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
+            None,
+        )
+
+
+def with_loss(x, y, name):
+    # returns x but adds y.sum() to the loss function.
+    return WithLoss.apply(x, y, name)
+
+
+class ScaleGradFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, alpha: float) -> Tensor:
+        ctx.alpha = alpha
+        return x
+
+    @staticmethod
+    def backward(ctx, grad: Tensor):
+        return grad * ctx.alpha, None
+
+
+def scale_grad(x: Tensor, alpha: float):
+    return ScaleGradFunction.apply(x, alpha)
+
+
+class ScaleGrad(nn.Module):
+    def __init__(self, alpha: float):
+        super().__init__()
+        self.alpha = alpha
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training:
+            return x
+        return scale_grad(x, self.alpha)
+
+
+class LimitParamValue(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, min: float, max: float):
+        ctx.save_for_backward(x)
+        assert max >= min
+        ctx.min = min
+        ctx.max = max
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x,) = ctx.saved_tensors
+        # where x < ctx.min, ensure all grads are negative (this will tend to make
+        # x more positive).
+        x_grad = x_grad * torch.where(
+            torch.logical_and(x_grad > 0, x < ctx.min), -1.0, 1.0
+        )
+        # where x > ctx.max, ensure all grads are positive (this will tend to make
+        # x more negative).
+        x_grad *= torch.where(torch.logical_and(x_grad < 0, x > ctx.max), -1.0, 1.0)
+        return x_grad, None, None
+
+
+def limit_param_value(
+    x: Tensor, min: float, max: float, prob: float = 0.6, training: bool = True
+):
+    # You apply this to (typically) an nn.Parameter during training to ensure that its
+    # (elements mostly) stays within a supplied range.  This is done by modifying the
+    # gradients in backprop.
+    # It's not necessary to do this on every batch: do it only some of the time,
+    # to save a little time.
+    if training and random.random() < prob:
+        return LimitParamValue.apply(x, min, max)
+    else:
+        return x
+
+
+def _no_op(x: Tensor) -> Tensor:
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    else:
+        # a no-op function that will have a node in the autograd graph,
+        # to avoid certain bugs relating to backward hooks
+        return x.chunk(1, dim=-1)[0]
+
+
+class Identity(torch.nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, x):
+        return _no_op(x)
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.044 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.044
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+# Dropout2 is just like normal dropout, except it supports schedules on the dropout rates.
+class Dropout2(nn.Module):
+    def __init__(self, p: FloatLike):
+        super().__init__()
+        self.p = p
+
+    def forward(self, x: Tensor) -> Tensor:
+        return torch.nn.functional.dropout(x, p=float(self.p), training=self.training)
+
+
+class MulForDropout3(torch.autograd.Function):
+    # returns (x * y * alpha) where alpha is a float and y doesn't require
+    # grad and is zero-or-one.
+    @staticmethod
+    @custom_fwd
+    def forward(ctx, x, y, alpha):
+        assert not y.requires_grad
+        ans = x * y * alpha
+        ctx.save_for_backward(ans)
+        ctx.alpha = alpha
+        return ans
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad):
+        (ans,) = ctx.saved_tensors
+        x_grad = ctx.alpha * ans_grad * (ans != 0)
+        return x_grad, None, None
+
+
+# Dropout3 is just like normal dropout, except it supports schedules on the dropout rates,
+# and it lets you choose one dimension to share the dropout mask over
+class Dropout3(nn.Module):
+    def __init__(self, p: FloatLike, shared_dim: int):
+        super().__init__()
+        self.p = p
+        self.shared_dim = shared_dim
+
+    def forward(self, x: Tensor) -> Tensor:
+        p = float(self.p)
+        if not self.training or p == 0:
+            return _no_op(x)
+        scale = 1.0 / (1 - p)
+        rand_shape = list(x.shape)
+        rand_shape[self.shared_dim] = 1
+        mask = torch.rand(*rand_shape, device=x.device) > p
+        ans = MulForDropout3.apply(x, mask, scale)
+        return ans
+
+
+class SwooshLFunction(torch.autograd.Function):
+    """
+    swoosh_l(x) =  log(1 + exp(x-4)) - 0.08*x - 0.035
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+
+        coeff = -0.08
+
+        with torch.cuda.amp.autocast(enabled=False):
+            with torch.enable_grad():
+                x = x.detach()
+                x.requires_grad = True
+                y = torch.logaddexp(zero, x - 4.0) + coeff * x - 0.035
+
+                if not requires_grad:
+                    return y
+
+                y.backward(gradient=torch.ones_like(y))
+
+                grad = x.grad
+                floor = coeff
+                ceil = 1.0 + coeff + 0.005
+
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
+                if __name__ == "__main__":
+                    # for self-testing only.
+                    assert d_scaled.min() >= 0.0
+                    assert d_scaled.max() < 256.0
+
+                d_int = d_scaled.to(torch.uint8)
+                ctx.save_for_backward(d_int)
+                if x.dtype == torch.float16 or torch.is_autocast_enabled():
+                    y = y.to(torch.float16)
+                return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+
+        coeff = -0.08
+        floor = coeff
+        ceil = 1.0 + coeff + 0.005
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class SwooshL(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-L activation."""
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+            return logaddexp(zero, x - 4.0) - 0.08 * x - 0.035
+        if not x.requires_grad:
+            return k2.swoosh_l_forward(x)
+        else:
+            return k2.swoosh_l(x)
+        # return SwooshLFunction.apply(x)
+
+
+class SwooshLOnnx(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-L activation."""
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+        return logaddexp_onnx(zero, x - 4.0) - 0.08 * x - 0.035
+
+
+class SwooshRFunction(torch.autograd.Function):
+    """
+     swoosh_r(x) =  log(1 + exp(x-1)) - 0.08*x - 0.313261687
+
+    derivatives are between -0.08 and 0.92.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            with torch.enable_grad():
+                x = x.detach()
+                x.requires_grad = True
+                y = torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
+
+                if not requires_grad:
+                    return y
+                y.backward(gradient=torch.ones_like(y))
+
+                grad = x.grad
+                floor = -0.08
+                ceil = 0.925
+
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
+                if __name__ == "__main__":
+                    # for self-testing only.
+                    assert d_scaled.min() >= 0.0
+                    assert d_scaled.max() < 256.0
+
+                d_int = d_scaled.to(torch.uint8)
+                ctx.save_for_backward(d_int)
+                if x.dtype == torch.float16 or torch.is_autocast_enabled():
+                    y = y.to(torch.float16)
+                return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.08
+        ceil = 0.925
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class SwooshR(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-R activation."""
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+            return logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
+        if not x.requires_grad:
+            return k2.swoosh_r_forward(x)
+        else:
+            return k2.swoosh_r(x)
+        # return SwooshRFunction.apply(x)
+
+
+class SwooshROnnx(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-R activation."""
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+        return logaddexp_onnx(zero, x - 1.0) - 0.08 * x - 0.313261687
+
+
+# simple version of SwooshL that does not redefine the backprop, used in
+# ActivationDropoutAndLinearFunction.
+def SwooshLForward(x: Tensor):
+    x_offset = x - 4.0
+    log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
+    return log_sum - 0.08 * x - 0.035
+
+
+# simple version of SwooshR that does not redefine the backprop, used in
+# ActivationDropoutAndLinearFunction.
+def SwooshRForward(x: Tensor):
+    x_offset = x - 1.0
+    log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
+    return log_sum - 0.08 * x - 0.313261687
+
+
+class ActivationDropoutAndLinearFunction(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(
+        ctx,
+        x: Tensor,
+        weight: Tensor,
+        bias: Optional[Tensor],
+        activation: str,
+        dropout_p: float,
+        dropout_shared_dim: Optional[int],
+    ):
+        if dropout_p != 0.0:
+            dropout_shape = list(x.shape)
+            if dropout_shared_dim is not None:
+                dropout_shape[dropout_shared_dim] = 1
+            # else it won't be very memory efficient.
+            dropout_mask = (1.0 / (1.0 - dropout_p)) * (
+                torch.rand(*dropout_shape, device=x.device, dtype=x.dtype) > dropout_p
+            )
+        else:
+            dropout_mask = None
+
+        ctx.save_for_backward(x, weight, bias, dropout_mask)
+
+        ctx.activation = activation
+
+        forward_activation_dict = {
+            "SwooshL": k2.swoosh_l_forward,
+            "SwooshR": k2.swoosh_r_forward,
+        }
+        # it will raise a KeyError if this fails.  This will be an error.  We let it
+        # propagate to the user.
+        activation_func = forward_activation_dict[activation]
+        x = activation_func(x)
+        if dropout_mask is not None:
+            x = x * dropout_mask
+        x = torch.nn.functional.linear(x, weight, bias)
+        return x
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad: Tensor):
+        saved = ctx.saved_tensors
+        (x, weight, bias, dropout_mask) = saved
+
+        forward_and_deriv_activation_dict = {
+            "SwooshL": k2.swoosh_l_forward_and_deriv,
+            "SwooshR": k2.swoosh_r_forward_and_deriv,
+        }
+        # the following lines a KeyError if the activation is unrecognized.
+        # This will be an error.  We let it propagate to the user.
+        func = forward_and_deriv_activation_dict[ctx.activation]
+
+        y, func_deriv = func(x)
+        if dropout_mask is not None:
+            y = y * dropout_mask
+        # now compute derivative of y w.r.t. weight and bias..
+        # y: (..., in_channels), ans_grad: (..., out_channels),
+        (out_channels, in_channels) = weight.shape
+
+        in_channels = y.shape[-1]
+        g = ans_grad.reshape(-1, out_channels)
+        weight_deriv = torch.matmul(g.t(), y.reshape(-1, in_channels))
+        y_deriv = torch.matmul(ans_grad, weight)
+        bias_deriv = None if bias is None else g.sum(dim=0)
+        x_deriv = y_deriv * func_deriv
+        if dropout_mask is not None:
+            # order versus func_deriv does not matter
+            x_deriv = x_deriv * dropout_mask
+
+        return x_deriv, weight_deriv, bias_deriv, None, None, None
+
+
+class ActivationDropoutAndLinear(torch.nn.Module):
+    """
+     This merges an activation function followed by dropout and then a nn.Linear module;
+     it does so in a memory efficient way so that it only stores the input to the whole
+     module.  If activation == SwooshL and dropout_shared_dim != None, this will be
+     equivalent to:
+       nn.Sequential(SwooshL(),
+                     Dropout3(dropout_p, shared_dim=dropout_shared_dim),
+                     ScaledLinear(in_channels, out_channels, bias=bias,
+                                  initial_scale=initial_scale))
+    If dropout_shared_dim is None, the dropout would be equivalent to
+    Dropout2(dropout_p).  Note: Dropout3 will be more memory efficient as the dropout
+    mask is smaller.
+
+     Args:
+        in_channels: number of input channels, e.g. 256
+        out_channels: number of output channels, e.g. 256
+        bias: if true, have a bias
+        activation: the activation function, for now just support SwooshL.
+        dropout_p: the dropout probability or schedule (happens after nonlinearity).
+        dropout_shared_dim: the dimension, if any, across which the dropout mask is
+             shared (e.g. the time dimension).  If None, this may be less memory
+             efficient if there are modules before this one that cache the input
+             for their backprop (e.g. Balancer or Whiten).
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        bias: bool = True,
+        activation: str = "SwooshL",
+        dropout_p: FloatLike = 0.0,
+        dropout_shared_dim: Optional[int] = -1,
+        initial_scale: float = 1.0,
+    ):
+        super().__init__()
+        # create a temporary module of nn.Linear that we'll steal the
+        # weights and bias from
+        l = ScaledLinear(
+            in_channels, out_channels, bias=bias, initial_scale=initial_scale
+        )
+
+        self.weight = l.weight
+        # register_parameter properly handles making it a parameter when l.bias
+        # is None. I think there is some reason for doing it this way rather
+        # than just setting it to None but I don't know what it is, maybe
+        # something to do with exporting the module..
+        self.register_parameter("bias", l.bias)
+
+        self.activation = activation
+        self.dropout_p = dropout_p
+        self.dropout_shared_dim = dropout_shared_dim
+
+    def forward(self, x: Tensor):
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            if self.activation == "SwooshL":
+                x = SwooshLForward(x)
+            elif self.activation == "SwooshR":
+                x = SwooshRForward(x)
+            else:
+                assert False, self.activation
+            return torch.nn.functional.linear(x, self.weight, self.bias)
+
+        return ActivationDropoutAndLinearFunction.apply(
+            x,
+            self.weight,
+            self.bias,
+            self.activation,
+            float(self.dropout_p),
+            self.dropout_shared_dim,
+        )
+
+
+def convert_num_channels(x: Tensor, num_channels: int) -> Tensor:
+    if num_channels <= x.shape[-1]:
+        return x[..., :num_channels]
+    else:
+        shape = list(x.shape)
+        shape[-1] = num_channels - shape[-1]
+        zeros = torch.zeros(shape, dtype=x.dtype, device=x.device)
+        return torch.cat((x, zeros), dim=-1)
+
+
+def _test_whiten():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"_test_whiten(): proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0)
+    x = x.detach()
+    x.requires_grad = True
+    m = Balancer(
+        probs.numel(),
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        min_abs=0.0,
+        prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_balancer_sign: x = ", x)
+    print("_test_balancer_sign: y grad = ", y_grad)
+    print("_test_balancer_sign: x grad = ", x.grad)
+
+
+def _test_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = Balancer(
+        magnitudes.numel(),
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        min_abs=0.2,
+        max_abs=0.7,
+        prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_balancer_magnitude: x = ", x)
+    print("_test_balancer_magnitude: y grad = ", y_grad)
+    print("_test_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = DoubleSwish()
+
+    tol = (1.2 - (-0.043637)) / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_swooshl_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = SwooshL()
+
+    tol = 1.0 / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_swooshr_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = SwooshR()
+
+    tol = 1.0 / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_softmax():
+    a = torch.randn(2, 10, dtype=torch.float64)
+    b = a.clone()
+    a.requires_grad = True
+    b.requires_grad = True
+    a.softmax(dim=1)[:, 0].sum().backward()
+    print("a grad = ", a.grad)
+    softmax(b, dim=1)[:, 0].sum().backward()
+    print("b grad = ", b.grad)
+    assert torch.allclose(a.grad, b.grad)
+
+
+def _test_piecewise_linear():
+    p = PiecewiseLinear((0, 10.0))
+    for x in [-100, 0, 100]:
+        assert p(x) == 10.0
+    p = PiecewiseLinear((0, 10.0), (1, 0.0))
+    for x, y in [(-100, 10.0), (0, 10.0), (0.5, 5.0), (1, 0.0), (2, 0.0)]:
+        print("x, y = ", x, y)
+        assert p(x) == y, (x, p(x), y)
+
+    q = PiecewiseLinear((0.5, 15.0), (0.6, 1.0))
+    x_vals = [-1.0, 0.0, 0.1, 0.2, 0.5, 0.6, 0.7, 0.9, 1.0, 2.0]
+    pq = p.max(q)
+    for x in x_vals:
+        y1 = max(p(x), q(x))
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+    pq = p.min(q)
+    for x in x_vals:
+        y1 = min(p(x), q(x))
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+    pq = p + q
+    for x in x_vals:
+        y1 = p(x) + q(x)
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+
+
+def _test_activation_dropout_and_linear():
+    in_channels = 20
+    out_channels = 30
+
+    for bias in [True, False]:
+        # actually we don't test for dropout_p != 0.0 because forward functions will give
+        # different answers.  This is because we are using the k2 implementation of
+        # swoosh_l an swoosh_r inside SwooshL() and SwooshR(), and they call randn()
+        # internally, messing up the random state.
+        for dropout_p in [0.0]:
+            for activation in ["SwooshL", "SwooshR"]:
+                m1 = nn.Sequential(
+                    SwooshL() if activation == "SwooshL" else SwooshR(),
+                    Dropout3(p=dropout_p, shared_dim=-1),
+                    ScaledLinear(
+                        in_channels, out_channels, bias=bias, initial_scale=0.5
+                    ),
+                )
+                m2 = ActivationDropoutAndLinear(
+                    in_channels,
+                    out_channels,
+                    bias=bias,
+                    initial_scale=0.5,
+                    activation=activation,
+                    dropout_p=dropout_p,
+                )
+                with torch.no_grad():
+                    m2.weight[:] = m1[2].weight
+                    if bias:
+                        m2.bias[:] = m1[2].bias
+                # make sure forward gives same result.
+                x1 = torch.randn(10, in_channels)
+                x1.requires_grad = True
+
+                # TEMP.
+                assert torch.allclose(
+                    SwooshRFunction.apply(x1), SwooshRForward(x1), atol=1.0e-03
+                )
+
+                x2 = x1.clone().detach()
+                x2.requires_grad = True
+                seed = 10
+                torch.manual_seed(seed)
+                y1 = m1(x1)
+                y_grad = torch.randn_like(y1)
+                y1.backward(gradient=y_grad)
+                torch.manual_seed(seed)
+                y2 = m2(x2)
+                y2.backward(gradient=y_grad)
+
+                print(
+                    f"bias = {bias}, dropout_p = {dropout_p}, activation = {activation}"
+                )
+                print("y1 = ", y1)
+                print("y2 = ", y2)
+                assert torch.allclose(y1, y2, atol=0.02)
+                assert torch.allclose(m1[2].weight.grad, m2.weight.grad, atol=1.0e-05)
+                if bias:
+                    assert torch.allclose(m1[2].bias.grad, m2.bias.grad, atol=1.0e-05)
+                print("x1.grad = ", x1.grad)
+                print("x2.grad = ", x2.grad)
+
+                def isclose(a, b):
+                    # return true if cosine similarity is > 0.9.
+                    return (a * b).sum() > 0.9 * (
+                        (a**2).sum() * (b**2).sum()
+                    ).sqrt()
+
+                # the SwooshL() implementation has a noisy gradient due to 1-byte
+                # storage of it.
+                assert isclose(x1.grad, x2.grad)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_piecewise_linear()
+    _test_softmax()
+    _test_whiten()
+    _test_balancer_sign()
+    _test_balancer_magnitude()
+    _test_double_swish_deriv()
+    _test_swooshr_deriv()
+    _test_swooshl_deriv()
+    _test_activation_dropout_and_linear()
diff --git a/egs/librispeech/ASR/zipformer/scaling_converter.py b/egs/librispeech/ASR/zipformer/scaling_converter.py
new file mode 100644
index 000000000..76622fa12
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/scaling_converter.py
@@ -0,0 +1,104 @@
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file replaces various modules in a model.
+Specifically, ActivationBalancer is replaced with an identity operator;
+Whiten is also replaced with an identity operator;
+BasicNorm is replaced by a module with `exp` removed.
+"""
+
+import copy
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+from scaling import (
+    Balancer,
+    Dropout3,
+    ScaleGrad,
+    SwooshL,
+    SwooshLOnnx,
+    SwooshR,
+    SwooshROnnx,
+    Whiten,
+)
+from zipformer import CompactRelPositionalEncoding
+
+
+# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule  # noqa
+# get_submodule was added to nn.Module at v1.9.0
+def get_submodule(model, target):
+    if target == "":
+        return model
+    atoms: List[str] = target.split(".")
+    mod: torch.nn.Module = model
+    for item in atoms:
+        if not hasattr(mod, item):
+            raise AttributeError(
+                mod._get_name() + " has no " "attribute `" + item + "`"
+            )
+        mod = getattr(mod, item)
+        if not isinstance(mod, torch.nn.Module):
+            raise AttributeError("`" + item + "` is not " "an nn.Module")
+    return mod
+
+
+def convert_scaled_to_non_scaled(
+    model: nn.Module,
+    inplace: bool = False,
+    is_pnnx: bool = False,
+    is_onnx: bool = False,
+):
+    """
+    Args:
+      model:
+        The model to be converted.
+      inplace:
+        If True, the input model is modified inplace.
+        If False, the input model is copied and we modify the copied version.
+      is_pnnx:
+        True if we are going to export the model for PNNX.
+      is_onnx:
+        True if we are going to export the model for ONNX.
+    Return:
+      Return a model without scaled layers.
+    """
+    if not inplace:
+        model = copy.deepcopy(model)
+
+    d = {}
+    for name, m in model.named_modules():
+        if isinstance(m, (Balancer, Dropout3, ScaleGrad, Whiten)):
+            d[name] = nn.Identity()
+        elif is_onnx and isinstance(m, SwooshR):
+            d[name] = SwooshROnnx()
+        elif is_onnx and isinstance(m, SwooshL):
+            d[name] = SwooshLOnnx()
+        elif is_onnx and isinstance(m, CompactRelPositionalEncoding):
+            # We want to recreate the positional encoding vector when
+            # the input changes, so we have to use torch.jit.script()
+            # to replace torch.jit.trace()
+            d[name] = torch.jit.script(m)
+
+    for k, v in d.items():
+        if "." in k:
+            parent, child = k.rsplit(".", maxsplit=1)
+            setattr(get_submodule(model, parent), child, v)
+        else:
+            setattr(model, k, v)
+
+    return model
diff --git a/egs/librispeech/ASR/zipformer/streaming_beam_search.py b/egs/librispeech/ASR/zipformer/streaming_beam_search.py
new file mode 100644
index 000000000..3c8565b33
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1,295 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from decode_stream import DecodeStream
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+    blank_penalty: float = 0.0,
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    T = encoder_out.size(1)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (N, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        if blank_penalty != 0.0:
+            logits[:, 0] -= blank_penalty
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+    num_active_paths: int = 4,
+    blank_penalty: float = 0.0,
+) -> None:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      num_active_paths:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        if blank_penalty != 0.0:
+            logits[:, 0] -= blank_penalty
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(num_active_paths)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    streams: List[DecodeStream],
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    blank_penalty: float = 0.0,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first generated by Fsa-based beam search, then we get the
+    recognition by applying shortest path on the lattice.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      streams:
+        A list of stream objects.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+
+        if blank_penalty != 0.0:
+            logits[:, 0] -= blank_penalty
+
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+    best_path = one_best_decoding(lattice)
+    hyp_tokens = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyp_tokens[i]
diff --git a/egs/librispeech/ASR/zipformer/streaming_decode.py b/egs/librispeech/ASR/zipformer/streaming_decode.py
new file mode 100755
index 000000000..904caf8af
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/streaming_decode.py
@@ -0,0 +1,853 @@
+#!/usr/bin/env python3
+# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
+#                                                  Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 32 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch import Tensor, nn
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_init_states(
+    model: nn.Module,
+    batch_size: int = 1,
+    device: torch.device = torch.device("cpu"),
+) -> List[torch.Tensor]:
+    """
+    Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+    is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+    states[-2] is the cached left padding for ConvNeXt module,
+    of shape (batch_size, num_channels, left_pad, num_freqs)
+    states[-1] is processed_lens of shape (batch,), which records the number
+    of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+    """
+    states = model.encoder.get_init_states(batch_size, device)
+
+    embed_states = model.encoder_embed.get_init_states(batch_size, device)
+    states.append(embed_states)
+
+    processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+    states.append(processed_lens)
+
+    return states
+
+
+def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance. For element-n,
+        state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
+        state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
+        cached_val2, cached_conv1, cached_conv2).
+        state_list[n][-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+        state_list[n][-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+    """
+    batch_size = len(state_list)
+    assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
+    tot_num_layers = (len(state_list[0]) - 2) // 6
+
+    batch_states = []
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key = torch.cat(
+            [state_list[i][layer_offset] for i in range(batch_size)], dim=1
+        )
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn = torch.cat(
+            [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1 = torch.cat(
+            [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2 = torch.cat(
+            [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1 = torch.cat(
+            [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2 = torch.cat(
+            [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
+        )
+        batch_states += [
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ]
+
+    cached_embed_left_pad = torch.cat(
+        [state_list[i][-2] for i in range(batch_size)], dim=0
+    )
+    batch_states.append(cached_embed_left_pad)
+
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
+    batch_states.append(processed_lens)
+
+    return batch_states
+
+
+def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+        batch_states: A list of cached tensors of all encoder layers. For layer-i,
+          states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+          cached_conv1, cached_conv2).
+          state_list[-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+          states[-1] is processed_lens of shape (batch,), which records the number
+          of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Returns:
+        state_list: A list of list. Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+    """
+    assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
+    tot_num_layers = (len(batch_states) - 2) // 6
+
+    processed_lens = batch_states[-1]
+    batch_size = processed_lens.shape[0]
+
+    state_list = [[] for _ in range(batch_size)]
+
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1_list = batch_states[layer_offset + 2].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2_list = batch_states[layer_offset + 3].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1_list = batch_states[layer_offset + 4].chunk(
+            chunks=batch_size, dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2_list = batch_states[layer_offset + 5].chunk(
+            chunks=batch_size, dim=0
+        )
+        for i in range(batch_size):
+            state_list[i] += [
+                cached_key_list[i],
+                cached_nonlin_attn_list[i],
+                cached_val1_list[i],
+                cached_val2_list[i],
+                cached_conv1_list[i],
+                cached_conv2_list[i],
+            ]
+
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(cached_embed_left_pad_list[i])
+
+    processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(processed_lens_list[i])
+
+    return state_list
+
+
+def streaming_forward(
+    features: Tensor,
+    feature_lens: Tensor,
+    model: nn.Module,
+    states: List[Tensor],
+    chunk_size: int,
+    left_context_len: int,
+) -> Tuple[Tensor, Tensor, List[Tensor]]:
+    """
+    Returns encoder outputs, output lengths, and updated states.
+    """
+    cached_embed_left_pad = states[-2]
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        cached_left_pad=cached_embed_left_pad,
+    )
+    assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    # processed_mask is used to mask out initial states
+    processed_mask = torch.arange(left_context_len, device=x.device).expand(
+        x.size(0), left_context_len
+    )
+    processed_lens = states[-1]  # (batch,)
+    # (batch, left_context_size)
+    processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+    # Update processed lengths
+    new_processed_lens = processed_lens + x_lens
+
+    # (batch, left_context_size + chunk_size)
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    encoder_states = states[:-2]
+    (
+        encoder_out,
+        encoder_out_lens,
+        new_encoder_states,
+    ) = model.encoder.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=encoder_states,
+        src_key_padding_mask=src_key_padding_mask,
+    )
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    new_states = new_encoder_states + [
+        new_cached_embed_left_pad,
+        new_processed_lens,
+    ]
+    return encoder_out, encoder_out_lens, new_states
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+    chunk_size = int(params.chunk_size)
+    left_context_len = int(params.left_context_frames)
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []  # Used in fast-beam-search
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(chunk_size * 2)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # Make sure the length after encoder_embed is at least 1.
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    tail_length = chunk_size * 2 + 7 + 2 * 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+
+    encoder_out, encoder_out_lens, new_states = streaming_forward(
+        features=features,
+        feature_lens=feature_lens,
+        model=model,
+        states=states,
+        chunk_size=chunk_size,
+        left_context_len=left_context_len,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = torch.tensor(processed_lens, device=device)
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = get_init_states(model=model, batch_size=1, device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=30)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    assert params.causal, params.causal
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
+    assert (
+        "," not in params.left_context_frames
+    ), "left_context_frames should be one value in decoding."
+    params.suffix += f"-chunk-{params.chunk_size}"
+    params.suffix += f"-left-context-{params.left_context_frames}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/subsampling.py b/egs/librispeech/ASR/zipformer/subsampling.py
new file mode 100644
index 000000000..d16d87bac
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/subsampling.py
@@ -0,0 +1,406 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Daniel Povey,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+import warnings
+
+import torch
+from torch import Tensor, nn
+from scaling import (
+    Balancer,
+    BiasNorm,
+    Dropout3,
+    FloatLike,
+    Optional,
+    ScaledConv2d,
+    ScaleGrad,
+    ScheduledFloat,
+    SwooshL,
+    SwooshR,
+    Whiten,
+)
+
+
+class ConvNeXt(nn.Module):
+    """
+    Our interpretation of the ConvNeXt module as used in https://arxiv.org/pdf/2206.14747.pdf
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        hidden_ratio: int = 3,
+        kernel_size: Tuple[int, int] = (7, 7),
+        layerdrop_rate: FloatLike = None,
+    ):
+        super().__init__()
+        self.padding = ((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2)
+        hidden_channels = channels * hidden_ratio
+        if layerdrop_rate is None:
+            layerdrop_rate = ScheduledFloat((0.0, 0.2), (20000.0, 0.015))
+        self.layerdrop_rate = layerdrop_rate
+
+        self.depthwise_conv = nn.Conv2d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=self.padding,
+        )
+
+        self.pointwise_conv1 = nn.Conv2d(
+            in_channels=channels, out_channels=hidden_channels, kernel_size=1
+        )
+
+        self.hidden_balancer = Balancer(
+            hidden_channels,
+            channel_dim=1,
+            min_positive=0.3,
+            max_positive=1.0,
+            min_abs=0.75,
+            max_abs=5.0,
+        )
+
+        self.activation = SwooshL()
+        self.pointwise_conv2 = ScaledConv2d(
+            in_channels=hidden_channels,
+            out_channels=channels,
+            kernel_size=1,
+            initial_scale=0.01,
+        )
+
+        self.out_balancer = Balancer(
+            channels,
+            channel_dim=1,
+            min_positive=0.4,
+            max_positive=0.6,
+            min_abs=1.0,
+            max_abs=6.0,
+        )
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=5.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training:
+            return self.forward_internal(x)
+        layerdrop_rate = float(self.layerdrop_rate)
+
+        if layerdrop_rate != 0.0:
+            batch_size = x.shape[0]
+            mask = (
+                torch.rand((batch_size, 1, 1, 1), dtype=x.dtype, device=x.device)
+                > layerdrop_rate
+            )
+        else:
+            mask = None
+        # turns out this caching idea does not work with --world-size > 1
+        # return caching_eval(self.forward_internal, x, mask)
+        return self.forward_internal(x, mask)
+
+    def forward_internal(
+        self, x: Tensor, layer_skip_mask: Optional[Tensor] = None
+    ) -> Tensor:
+        """
+        x layout: (N, C, H, W), i.e. (batch_size, num_channels, num_frames, num_freqs)
+
+        The returned value has the same shape as x.
+        """
+        bypass = x
+        x = self.depthwise_conv(x)
+        x = self.pointwise_conv1(x)
+        x = self.hidden_balancer(x)
+        x = self.activation(x)
+        x = self.pointwise_conv2(x)
+
+        if layer_skip_mask is not None:
+            x = x * layer_skip_mask
+
+        x = bypass + x
+        x = self.out_balancer(x)
+
+        if x.requires_grad:
+            x = x.transpose(1, 3)  # (N, W, H, C); need channel dim to be last
+            x = self.out_whiten(x)
+            x = x.transpose(1, 3)  # (N, C, H, W)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cached_left_pad: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            x layout: (N, C, H, W), i.e. (batch_size, num_channels, num_frames, num_freqs)
+            cached_left_pad: (batch_size, num_channels, left_pad, num_freqs)
+
+        Returns:
+            - The returned value has the same shape as x.
+            - Updated cached_left_pad.
+        """
+        padding = self.padding
+
+        # The length without right padding for depth-wise conv
+        T = x.size(2) - padding[0]
+
+        bypass = x[:, :, :T, :]
+
+        # Pad left side
+        assert cached_left_pad.size(2) == padding[0], (
+            cached_left_pad.size(2),
+            padding[0],
+        )
+        x = torch.cat([cached_left_pad, x], dim=2)
+        # Update cached left padding
+        cached_left_pad = x[:, :, T : padding[0] + T, :]
+
+        # depthwise_conv
+        x = torch.nn.functional.conv2d(
+            x,
+            weight=self.depthwise_conv.weight,
+            bias=self.depthwise_conv.bias,
+            padding=(0, padding[1]),
+            groups=self.depthwise_conv.groups,
+        )
+        x = self.pointwise_conv1(x)
+        x = self.hidden_balancer(x)
+        x = self.activation(x)
+        x = self.pointwise_conv2(x)
+
+        x = bypass + x
+        return x, cached_left_pad
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/2 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = (T-3)//2 - 2 == (T-7)//2
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        dropout: FloatLike = 0.1,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, (T-3)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+          bottleneck:
+            bottleneck dimension for 1d squeeze-excite
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        # The ScaleGrad module is there to prevent the gradients
+        # w.r.t. the weight or bias of the first Conv2d module in self.conv from
+        # exceeding the range of fp16 when using automatic mixed precision (amp)
+        # training.  (The second one is necessary to stop its bias from getting
+        # a too-large gradient).
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=(0, 1),  # (time, freq)
+            ),
+            ScaleGrad(0.2),
+            Balancer(layer1_channels, channel_dim=1, max_abs=1.0),
+            SwooshR(),
+            nn.Conv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+                padding=0,
+            ),
+            Balancer(layer2_channels, channel_dim=1, max_abs=4.0),
+            SwooshR(),
+            nn.Conv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=(1, 2),  # (time, freq)
+            ),
+            Balancer(layer3_channels, channel_dim=1, max_abs=4.0),
+            SwooshR(),
+        )
+
+        # just one convnext layer
+        self.convnext = ConvNeXt(layer3_channels, kernel_size=(7, 7))
+
+        # (in_channels-3)//4
+        self.out_width = (((in_channels - 1) // 2) - 1) // 2
+        self.layer3_channels = layer3_channels
+
+        self.out = nn.Linear(self.out_width * layer3_channels, out_channels)
+        # use a larger than normal grad_scale on this whitening module; there is
+        # only one such module, so there is not a concern about adding together
+        # many copies of this extra gradient term.
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=ScheduledFloat((0.0, 4.0), (20000.0, 8.0), default=4.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.02,
+        )
+
+        # max_log_eps=0.0 is to prevent both eps and the output of self.out from
+        # getting large, there is an unnecessary degree of freedom.
+        self.out_norm = BiasNorm(out_channels)
+        self.dropout = Dropout3(dropout, shared_dim=1)
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+
+        Returns:
+          - a tensor of shape (N, (T-7)//2, odim)
+          - output lengths, of shape (batch_size,)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        # scaling x by 0.1 allows us to use a larger grad-scale in fp16 "amp" (automatic mixed precision)
+        # training, since the weights in the first convolution are otherwise the limiting factor for getting infinite
+        # gradients.
+        x = self.conv(x)
+        x = self.convnext(x)
+
+        # Now x is of shape (N, odim, (T-7)//2, (idim-3)//4)
+        b, c, t, f = x.size()
+
+        x = x.transpose(1, 2).reshape(b, t, c * f)
+        # now x: (N, (T-7)//2, out_width * layer3_channels))
+
+        x = self.out(x)
+        # Now x is of shape (N, (T-7)//2, odim)
+        x = self.out_whiten(x)
+        x = self.out_norm(x)
+        x = self.dropout(x)
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            x_lens = (x_lens - 7) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                x_lens = (x_lens - 7) // 2
+        assert x.size(1) == x_lens.max().item(), (x.size(1), x_lens.max())
+
+        return x, x_lens
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        cached_left_pad: Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+
+        Returns:
+          - a tensor of shape (N, (T-7)//2, odim)
+          - output lengths, of shape (batch_size,)
+          - updated cache
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+
+        # T' = (T-7)//2
+        x = self.conv(x)
+
+        # T' = (T-7)//2-3
+        x, cached_left_pad = self.convnext.streaming_forward(
+            x, cached_left_pad=cached_left_pad
+        )
+
+        # Now x is of shape (N, odim, T', ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+
+        x = x.transpose(1, 2).reshape(b, t, c * f)
+        # now x: (N, T', out_width * layer3_channels))
+
+        x = self.out(x)
+        # Now x is of shape (N, T', odim)
+        x = self.out_norm(x)
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            assert self.convnext.padding[0] == 3
+            # The ConvNeXt module needs 3 frames of right padding after subsampling
+            x_lens = (x_lens - 7) // 2 - 3
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                # The ConvNeXt module needs 3 frames of right padding after subsampling
+                assert self.convnext.padding[0] == 3
+                x_lens = (x_lens - 7) // 2 - 3
+
+        assert x.size(1) == x_lens.max().item(), (x.shape, x_lens.max())
+
+        return x, x_lens, cached_left_pad
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> Tensor:
+        """Get initial states for Conv2dSubsampling module.
+        It is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        """
+        left_pad = self.convnext.padding[0]
+        freq = self.out_width
+        channels = self.layer3_channels
+        cached_embed_left_pad = torch.zeros(batch_size, channels, left_pad, freq).to(
+            device
+        )
+
+        return cached_embed_left_pad
diff --git a/egs/librispeech/ASR/zipformer/test_scaling.py b/egs/librispeech/ASR/zipformer/test_scaling.py
new file mode 100755
index 000000000..5c04291e7
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/test_scaling.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+
+import matplotlib.pyplot as plt
+import torch
+from scaling import PiecewiseLinear, ScheduledFloat, SwooshL, SwooshR
+
+
+def test_piecewise_linear():
+    # An identity map in the range [0, 1].
+    # 1 - identity map in the range [1, 2]
+    # x1=0, y1=0
+    # x2=1, y2=1
+    # x3=2, y3=0
+    pl = PiecewiseLinear((0, 0), (1, 1), (2, 0))
+    assert pl(0.25) == 0.25, pl(0.25)
+    assert pl(0.625) == 0.625, pl(0.625)
+    assert pl(1.25) == 0.75, pl(1.25)
+
+    assert pl(-10) == pl(0), pl(-10)  # out of range
+    assert pl(10) == pl(2), pl(10)  # out of range
+
+    # multiplication
+    pl10 = pl * 10
+    assert pl10(1) == 10 * pl(1)
+    assert pl10(0.5) == 10 * pl(0.5)
+
+
+def test_scheduled_float():
+    # Initial value is 0.2 and it decreases linearly towards 0 at 4000
+    dropout = ScheduledFloat((0, 0.2), (4000, 0.0), default=0.0)
+    dropout.batch_count = 0
+    assert float(dropout) == 0.2, (float(dropout), dropout.batch_count)
+
+    dropout.batch_count = 1000
+    assert abs(float(dropout) - 0.15) < 1e-5, (float(dropout), dropout.batch_count)
+
+    dropout.batch_count = 2000
+    assert float(dropout) == 0.1, (float(dropout), dropout.batch_count)
+
+    dropout.batch_count = 3000
+    assert abs(float(dropout) - 0.05) < 1e-5, (float(dropout), dropout.batch_count)
+
+    dropout.batch_count = 4000
+    assert float(dropout) == 0.0, (float(dropout), dropout.batch_count)
+
+    dropout.batch_count = 5000  # out of range
+    assert float(dropout) == 0.0, (float(dropout), dropout.batch_count)
+
+
+def test_swoosh():
+    x1 = torch.linspace(start=-10, end=0, steps=100, dtype=torch.float32)
+    x2 = torch.linspace(start=0, end=10, steps=100, dtype=torch.float32)
+    x = torch.cat([x1, x2[1:]])
+
+    left = SwooshL()(x)
+    r = SwooshR()(x)
+
+    relu = torch.nn.functional.relu(x)
+    print(left[x == 0], r[x == 0])
+    plt.plot(x, left, "k")
+    plt.plot(x, r, "r")
+    plt.plot(x, relu, "b")
+    plt.axis([-10, 10, -1, 10])  # [xmin, xmax, ymin, ymax]
+    plt.legend(
+        [
+            "SwooshL(x) = log(1 + exp(x-4)) - 0.08x - 0.035 ",
+            "SwooshR(x) = log(1 + exp(x-1)) - 0.08x - 0.313261687",
+            "ReLU(x) = max(0, x)",
+        ]
+    )
+    plt.grid()
+    plt.savefig("swoosh.pdf")
+
+
+def main():
+    test_piecewise_linear()
+    test_scheduled_float()
+    test_swoosh()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/test_subsampling.py b/egs/librispeech/ASR/zipformer/test_subsampling.py
new file mode 100755
index 000000000..078227fb6
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/test_subsampling.py
@@ -0,0 +1,152 @@
+#!/usr/bin/env python3
+
+import torch
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+
+
+def test_conv2d_subsampling():
+    layer1_channels = 8
+    layer2_channels = 32
+    layer3_channels = 128
+
+    out_channels = 192
+    encoder_embed = Conv2dSubsampling(
+        in_channels=80,
+        out_channels=out_channels,
+        layer1_channels=layer1_channels,
+        layer2_channels=layer2_channels,
+        layer3_channels=layer3_channels,
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    N = 2
+    T = 200
+    num_features = 80
+    x = torch.rand(N, T, num_features)
+    x_copy = x.clone()
+
+    x = x.unsqueeze(1)  # (N, 1, T, num_features)
+
+    x = encoder_embed.conv[0](x)  # conv2d, in 1, out 8, kernel 3, padding (0,1)
+    assert x.shape == (N, layer1_channels, T - 2, num_features)
+    # (2, 8, 198, 80)
+
+    x = encoder_embed.conv[1](x)  # scale grad
+    x = encoder_embed.conv[2](x)  # balancer
+    x = encoder_embed.conv[3](x)  # swooshR
+
+    x = encoder_embed.conv[4](x)  # conv2d, in 8, out 32, kernel 3, stride 2
+    assert x.shape == (
+        N,
+        layer2_channels,
+        ((T - 2) - 3) // 2 + 1,
+        (num_features - 3) // 2 + 1,
+    )
+    # (2, 32, 98, 39)
+
+    x = encoder_embed.conv[5](x)  # balancer
+    x = encoder_embed.conv[6](x)  # swooshR
+
+    # conv2d:
+    # in 32, out 128, kernel 3, stride (1, 2)
+    x = encoder_embed.conv[7](x)
+    assert x.shape == (
+        N,
+        layer3_channels,
+        (((T - 2) - 3) // 2 + 1) - 2,
+        (((num_features - 3) // 2 + 1) - 3) // 2 + 1,
+    )
+    # (2, 128, 96, 19)
+
+    x = encoder_embed.conv[8](x)  # balancer
+    x = encoder_embed.conv[9](x)  # swooshR
+
+    # (((T - 2) - 3) // 2 + 1) - 2
+    # = (T - 2) - 3) // 2 + 1 - 2
+    # = ((T - 2) - 3) // 2 - 1
+    # = (T - 2 - 3) // 2 - 1
+    # = (T - 5) // 2 - 1
+    # = (T - 7) // 2
+    assert x.shape[2] == (x_copy.shape[1] - 7) // 2
+
+    # (((num_features - 3) // 2 + 1) - 3) // 2 + 1,
+    # = ((num_features - 3) // 2 + 1 - 3) // 2 + 1,
+    # = ((num_features - 3) // 2 - 2) // 2 + 1,
+    # = (num_features - 3 - 4) // 2 // 2 + 1,
+    # = (num_features - 7) // 2 // 2 + 1,
+    # = (num_features - 7) // 4 + 1,
+    # = (num_features - 3) // 4
+    assert x.shape[3] == (x_copy.shape[2] - 3) // 4
+
+    assert x.shape == (N, layer3_channels, (T - 7) // 2, (num_features - 3) // 4)
+
+    # Input shape to convnext is
+    #
+    # (N, layer3_channels, (T-7)//2, (num_features - 3)//4)
+
+    # conv2d: in layer3_channels, out layer3_channels, groups layer3_channels
+    # kernel_size 7, padding 3
+    x = encoder_embed.convnext.depthwise_conv(x)
+    assert x.shape == (N, layer3_channels, (T - 7) // 2, (num_features - 3) // 4)
+
+    # conv2d: in layer3_channels, out hidden_ratio * layer3_channels, kernel_size 1
+    x = encoder_embed.convnext.pointwise_conv1(x)
+    assert x.shape == (N, layer3_channels * 3, (T - 7) // 2, (num_features - 3) // 4)
+
+    x = encoder_embed.convnext.hidden_balancer(x)  # balancer
+    x = encoder_embed.convnext.activation(x)  # swooshL
+
+    # conv2d: in hidden_ratio * layer3_channels, out layer3_channels, kernel 1
+    x = encoder_embed.convnext.pointwise_conv2(x)
+    assert x.shape == (N, layer3_channels, (T - 7) // 2, (num_features - 3) // 4)
+
+    # bypass and layer drop, omitted here.
+    x = encoder_embed.convnext.out_balancer(x)
+
+    # Note: the input and output shape of ConvNeXt are the same
+
+    x = x.transpose(1, 2).reshape(N, (T - 7) // 2, -1)
+    assert x.shape == (N, (T - 7) // 2, layer3_channels * ((num_features - 3) // 4))
+
+    x = encoder_embed.out(x)
+    assert x.shape == (N, (T - 7) // 2, out_channels)
+
+    x = encoder_embed.out_whiten(x)
+    x = encoder_embed.out_norm(x)
+    # final layer is dropout
+
+    # test streaming forward
+
+    subsampling_factor = 2
+    cached_left_padding = encoder_embed.get_init_states(batch_size=N)
+    depthwise_conv_kernel_size = 7
+    pad_size = (depthwise_conv_kernel_size - 1) // 2
+
+    assert cached_left_padding.shape == (
+        N,
+        layer3_channels,
+        pad_size,
+        (num_features - 3) // 4,
+    )
+
+    chunk_size = 16
+    right_padding = pad_size * subsampling_factor
+    T = chunk_size * subsampling_factor + 7 + right_padding
+    x = torch.rand(N, T, num_features)
+    x_lens = torch.tensor([T] * N)
+    y, y_lens, next_cached_left_padding = encoder_embed.streaming_forward(
+        x, x_lens, cached_left_padding
+    )
+
+    assert y.shape == (N, chunk_size, out_channels), y.shape
+    assert next_cached_left_padding.shape == cached_left_padding.shape
+
+    assert y.shape[1] == y_lens[0] == y_lens[1]
+
+
+def main():
+    test_conv2d_subsampling()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/train.py b/egs/librispeech/ASR/zipformer/train.py
new file mode 100755
index 000000000..7009f3346
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/train.py
@@ -0,0 +1,1389 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --full-libri 1 \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --full-libri 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}"
+    )
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        loss = 0.0
+
+        if params.use_transducer:
+            s = params.simple_loss_scale
+            # take down the scale on the simple loss from 1.0 at the start
+            # to params.simple_loss scale by warm_step.
+            simple_loss_scale = (
+                s
+                if batch_idx_train >= warm_step
+                else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+            )
+            pruned_loss_scale = (
+                1.0
+                if batch_idx_train >= warm_step
+                else 0.1 + 0.9 * (batch_idx_train / warm_step)
+            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    if params.use_transducer:
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            # logging.warning(
+            #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer/zipformer.py b/egs/librispeech/ASR/zipformer/zipformer.py
new file mode 100644
index 000000000..61ae378d8
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/zipformer.py
@@ -0,0 +1,2437 @@
+#!/usr/bin/env python3
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Daniel Povey,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import logging
+import math
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+)
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationDropoutAndLinear,
+    Balancer,
+    BiasNorm,
+    ChunkCausalDepthwiseConv1d,
+    Dropout2,
+    FloatLike,
+    ScheduledFloat,
+    Whiten,
+    convert_num_channels,
+    limit_param_value,
+    penalize_abs_values_gt,
+    softmax,
+)
+from torch import Tensor, nn
+
+
+class Zipformer2(EncoderInterface):
+    """
+    Args:
+
+    Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length
+    as downsampling_factor if they are single ints or one-element tuples.  The length of
+    downsampling_factor defines the number of stacks.
+
+        output_downsampling_factor (int): how much to downsample at the output.  Note:
+            we also downsample by a factor of 2 in the Conv2dSubsampling encoder.
+            You should probably leave this at 2.
+        downsampling_factor (Tuple[int]): downsampling factor for each encoder stack.
+           Note: this is in addition to the downsampling factor of 2 that is applied in
+           the frontend (self.encoder_embed).
+        encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per
+           encoder stack.
+        num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack
+        encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of
+            the encoder stacks for purposes of per-frame dropout (recommend 256 for
+            now).
+        query_head_dim (int or Tuple[int]): dimension of query and key per attention
+           head: per stack, if a tuple..
+        pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per
+           attention head
+        value_head_dim (int or Tuple[int]): dimension of value in each attention head
+        num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism.
+              Must be at least 4.
+        feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules
+        cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module
+
+        pos_dim (int): the dimension of each positional-encoding vector prior to projection,
+            e.g. 128.
+
+        dropout (float): dropout rate
+        warmup_batches (float): number of batches to warm up over; this controls
+          dropout of encoder layers.
+        causal (bool): if True, support chunkwise causal convolution.  This should
+          not hurt WER as no modeling power is lost, but the convolution modules will be
+          slightly slower and use more memory.  Enables use of the chunk_size and
+          left_context_chunks options in forward(), which simulates streaming
+          decoding.
+        chunk_size: (list of int): only set this to other than [-1] if causal;
+           the chunk size will be randomly chosen from this list.  -1 means no chunking.
+        left_context_frames: (list of int): determines the number of left-
+           context chunks for causal training; will be rounded to a number of
+           chunks.  Must not be less than cnn_module_kernel (after factoring in
+           rounding and downsampling); an error will be thrown if this is violated.
+    """
+
+    def __init__(
+        self,
+        output_downsampling_factor: int = 2,
+        downsampling_factor: Tuple[int] = (2, 4),
+        encoder_dim: Union[int, Tuple[int]] = 384,
+        num_encoder_layers: Union[int, Tuple[int]] = 4,
+        encoder_unmasked_dim: Union[int, Tuple[int]] = 256,
+        query_head_dim: Union[int, Tuple[int]] = 24,
+        pos_head_dim: Union[int, Tuple[int]] = 4,
+        value_head_dim: Union[int, Tuple[int]] = 12,
+        num_heads: Union[int, Tuple[int]] = 8,
+        feedforward_dim: Union[int, Tuple[int]] = 1536,
+        cnn_module_kernel: Union[int, Tuple[int]] = 31,
+        pos_dim: int = 192,
+        dropout: FloatLike = None,  # see code below for default
+        warmup_batches: float = 4000.0,
+        causal: bool = False,
+        chunk_size: Tuple[int] = [-1],
+        left_context_frames: Tuple[int] = [-1],
+    ) -> None:
+        super(Zipformer2, self).__init__()
+
+        if dropout is None:
+            dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1))
+
+        def _to_tuple(x):
+            """Converts a single int or a 1-tuple of an int to a tuple with the same length
+            as downsampling_factor"""
+            if isinstance(x, int):
+                x = (x,)
+            if len(x) == 1:
+                x = x * len(downsampling_factor)
+            else:
+                assert len(x) == len(downsampling_factor) and isinstance(x[0], int)
+            return x
+
+        self.output_downsampling_factor = output_downsampling_factor  # int
+        self.downsampling_factor = downsampling_factor  # tuple
+        self.encoder_dim = encoder_dim = _to_tuple(encoder_dim)  # tuple
+        self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple(
+            encoder_unmasked_dim
+        )  # tuple
+        num_encoder_layers = _to_tuple(num_encoder_layers)
+        self.num_encoder_layers = num_encoder_layers
+        self.query_head_dim = query_head_dim = _to_tuple(query_head_dim)
+        self.value_head_dim = value_head_dim = _to_tuple(value_head_dim)
+        pos_head_dim = _to_tuple(pos_head_dim)
+        self.num_heads = num_heads = _to_tuple(num_heads)
+        feedforward_dim = _to_tuple(feedforward_dim)
+        self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel)
+
+        self.causal = causal
+        self.chunk_size = chunk_size
+        self.left_context_frames = left_context_frames
+
+        for u, d in zip(encoder_unmasked_dim, encoder_dim):
+            assert u <= d
+
+        # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder
+        encoders = []
+
+        num_encoders = len(downsampling_factor)
+        for i in range(num_encoders):
+            encoder_layer = Zipformer2EncoderLayer(
+                embed_dim=encoder_dim[i],
+                pos_dim=pos_dim,
+                num_heads=num_heads[i],
+                query_head_dim=query_head_dim[i],
+                pos_head_dim=pos_head_dim[i],
+                value_head_dim=value_head_dim[i],
+                feedforward_dim=feedforward_dim[i],
+                dropout=dropout,
+                cnn_module_kernel=cnn_module_kernel[i],
+                causal=causal,
+            )
+
+            # For the segment of the warmup period, we let the Conv2dSubsampling
+            # layer learn something.  Then we start to warm up the other encoders.
+            encoder = Zipformer2Encoder(
+                encoder_layer,
+                num_encoder_layers[i],
+                pos_dim=pos_dim,
+                dropout=dropout,
+                warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1),
+                warmup_end=warmup_batches * (i + 2) / (num_encoders + 1),
+                final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5),
+            )
+
+            if downsampling_factor[i] != 1:
+                encoder = DownsampledZipformer2Encoder(
+                    encoder,
+                    dim=encoder_dim[i],
+                    downsample=downsampling_factor[i],
+                    dropout=dropout,
+                )
+
+            encoders.append(encoder)
+
+        self.encoders = nn.ModuleList(encoders)
+
+        self.downsample_output = SimpleDownsample(
+            max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout
+        )
+
+    def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]:
+        """
+        In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of
+        randomized feature masks, one per encoder.
+        On e.g. 15% of frames, these masks will zero out all enocder dims larger than
+        some supplied number, e.g. >256, so in effect on those frames we are using
+        a smaller encoer dim.
+
+        We generate the random masks at this level because we want the 2 masks to 'agree'
+        all the way up the encoder stack. This will mean that the 1st mask will have
+        mask values repeated self.zipformer_subsampling_factor times.
+
+        Args:
+           x: the embeddings (needed for the shape and dtype and device), of shape
+             (1, batch_size, encoder_dims0)
+        """
+        num_encoders = len(self.encoder_dim)
+        if not self.training:
+            return [1.0] * num_encoders
+
+        (num_frames0, batch_size, _encoder_dims0) = x.shape
+
+        assert self.encoder_dim[0] == _encoder_dims0, (
+            self.encoder_dim[0],
+            _encoder_dims0,
+        )
+
+        feature_mask_dropout_prob = 0.125
+
+        # mask1 shape: (1, batch_size, 1)
+        mask1 = (
+            torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob
+        ).to(x.dtype)
+
+        # mask2 has additional sequences masked, about twice the number.
+        mask2 = torch.logical_and(
+            mask1,
+            (
+                torch.rand(1, batch_size, 1, device=x.device)
+                > feature_mask_dropout_prob
+            ).to(x.dtype),
+        )
+
+        # dim: (1, batch_size, 2)
+        mask = torch.cat((mask1, mask2), dim=-1)
+
+        feature_masks = []
+        for i in range(num_encoders):
+            channels = self.encoder_dim[i]
+            feature_mask = torch.ones(
+                1, batch_size, channels, dtype=x.dtype, device=x.device
+            )
+            u1 = self.encoder_unmasked_dim[i]
+            u2 = u1 + (channels - u1) // 2
+
+            feature_mask[:, :, u1:u2] *= mask[..., 0:1]
+            feature_mask[:, :, u2:] *= mask[..., 1:2]
+
+            feature_masks.append(feature_mask)
+
+        return feature_masks
+
+    def get_chunk_info(self) -> Tuple[int, int]:
+        """
+        Returns chunk_size and left_context_chunks.
+        """
+        if not self.causal:
+            return -1, -1
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            assert len(self.chunk_size) == 1, self.chunk_size
+            chunk_size = self.chunk_size[0]
+        else:
+            chunk_size = random.choice(self.chunk_size)
+
+        if chunk_size == -1:
+            left_context_chunks = -1
+        else:
+            if torch.jit.is_scripting() or torch.jit.is_tracing():
+                assert len(self.left_context_frames) == 1, self.left_context_frames
+                left_context_frames = self.left_context_frames[0]
+            else:
+                left_context_frames = random.choice(self.left_context_frames)
+            # Note: in Python, -1 // n == -1 for n > 0
+            left_context_chunks = left_context_frames // chunk_size
+            if left_context_chunks == 0:
+                left_context_chunks = 1
+
+        return chunk_size, left_context_chunks
+
+    def forward(
+        self,
+        x: Tensor,
+        x_lens: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (seq_len, batch_size, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          src_key_padding_mask:
+            The mask for padding, of shape (batch_size, seq_len); True means
+            masked position. May be None.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim))
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        outputs = []
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            feature_masks = [1.0] * len(self.encoder_dim)
+        else:
+            feature_masks = self.get_feature_masks(x)
+
+        chunk_size, left_context_chunks = self.get_chunk_info()
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            # Not support exporting a model for simulating streaming decoding
+            attn_mask = None
+        else:
+            attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks)
+
+        for i, module in enumerate(self.encoders):
+            ds = self.downsampling_factor[i]
+            x = convert_num_channels(x, self.encoder_dim[i])
+
+            x = module(
+                x,
+                chunk_size=chunk_size,
+                feature_mask=feature_masks[i],
+                src_key_padding_mask=(
+                    None
+                    if src_key_padding_mask is None
+                    else src_key_padding_mask[..., ::ds]
+                ),
+                attn_mask=attn_mask,
+            )
+            outputs.append(x)
+
+        # if the last output has the largest dimension, x will be unchanged,
+        # it will be the same as outputs[-1].  Otherwise it will be concatenated
+        # from different pieces of 'outputs', taking each dimension from the
+        # most recent output that has it present.
+        x = self._get_full_dim_output(outputs)
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            lengths = (x_lens + 1) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                lengths = (x_lens + 1) // 2
+
+        return x, lengths
+
+    def _get_attn_mask(
+        self, x: Tensor, chunk_size: int, left_context_chunks: int
+    ) -> Optional[Tensor]:
+        """
+        Return None if chunk_size == -1, else return attention mask of shape
+          (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len).  True
+           means a masked position.
+        Args:
+           x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim).
+          chunk_size: chunk size, must divide
+        """
+        if chunk_size <= 0:
+            return None
+        assert all(chunk_size % d == 0 for d in self.downsampling_factor)
+        if left_context_chunks >= 0:
+            num_encoders = len(self.encoder_dim)
+            assert all(
+                chunk_size * left_context_chunks
+                >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i]
+                for i in range(num_encoders)
+            )
+        else:
+            left_context_chunks = 1000000
+
+        seq_len = x.shape[0]
+
+        # t is frame index, shape (seq_len,)
+        t = torch.arange(seq_len, dtype=torch.int32, device=x.device)
+        # c is chunk index for each frame, shape (seq_len,)
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            c = t // chunk_size
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                c = t // chunk_size
+        src_c = c
+        tgt_c = c.unsqueeze(-1)
+
+        attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks)
+        if __name__ == "__main__":
+            logging.info(f"attn_mask = {attn_mask}")
+        return attn_mask
+
+    def _get_full_dim_output(self, outputs: List[Tensor]):
+        num_encoders = len(self.encoder_dim)
+        assert len(outputs) == num_encoders
+        output_dim = max(self.encoder_dim)
+        output_pieces = [outputs[-1]]
+        cur_dim = self.encoder_dim[-1]
+        for i in range(num_encoders - 2, -1, -1):
+            d = self.encoder_dim[i]
+            if d > cur_dim:
+                this_output = outputs[i]
+                output_pieces.append(this_output[..., cur_dim:d])
+                cur_dim = d
+        assert cur_dim == output_dim
+        return torch.cat(output_pieces, dim=-1)
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        x_lens: Tensor,
+        states: List[Tensor],
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (seq_len, batch_size, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states: list of cached tensors of all encoder layers. For layer-i,
+            states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+            cached_conv1, cached_conv2).
+          src_key_padding_mask:
+            The mask for padding, of shape (batch_size, seq_len); True means
+            masked position. May be None.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim))
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+            - updated states
+        """
+        outputs = []
+        new_states = []
+        layer_offset = 0
+
+        for i, module in enumerate(self.encoders):
+            num_layers = module.num_layers
+            ds = self.downsampling_factor[i]
+            x = convert_num_channels(x, self.encoder_dim[i])
+
+            x, new_layer_states = module.streaming_forward(
+                x,
+                states=states[layer_offset * 6 : (layer_offset + num_layers) * 6],
+                left_context_len=self.left_context_frames[0] // ds,
+                src_key_padding_mask=src_key_padding_mask[..., ::ds],
+            )
+            layer_offset += num_layers
+            outputs.append(x)
+            new_states += new_layer_states
+
+        # if the last output has the largest dimension, x will be unchanged,
+        # it will be the same as outputs[-1].  Otherwise it will be concatenated
+        # from different pieces of 'outputs', taking each dimension from the
+        # most recent output that has it present.
+        x = self._get_full_dim_output(outputs)
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            lengths = (x_lens + 1) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                lengths = (x_lens + 1) // 2
+
+        return x, lengths, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[Tensor]:
+        """Get initial states.
+
+        A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        """
+        states = []
+        for i, module in enumerate(self.encoders):
+            num_layers = module.num_layers
+            embed_dim = self.encoder_dim[i]
+            ds = self.downsampling_factor[i]
+            num_heads = self.num_heads[i]
+            key_dim = self.query_head_dim[i] * num_heads
+            value_dim = self.value_head_dim[i] * num_heads
+            downsample_left = self.left_context_frames[0] // ds
+            nonlin_attn_head_dim = 3 * embed_dim // 4
+            conv_left_pad = self.cnn_module_kernel[i] // 2
+            for layer in range(num_layers):
+                cached_key = torch.zeros(downsample_left, batch_size, key_dim).to(
+                    device
+                )
+                cached_nonlin_attn = torch.zeros(
+                    1, batch_size, downsample_left, nonlin_attn_head_dim
+                ).to(device)
+                cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to(
+                    device
+                )
+                cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to(
+                    device
+                )
+                cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(
+                    device
+                )
+                cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(
+                    device
+                )
+                states += [
+                    cached_key,
+                    cached_nonlin_attn,
+                    cached_val1,
+                    cached_val2,
+                    cached_conv1,
+                    cached_conv2,
+                ]
+
+        return states
+
+
+def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat:
+    return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x)
+
+
+def _balancer_schedule(min_prob: float):
+    return ScheduledFloat((0.0, 0.4), (8000.0, min_prob))
+
+
+class Zipformer2EncoderLayer(nn.Module):
+    """
+    Args:
+        embed_dim: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        feedforward_dim: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        pos_dim: int,
+        num_heads: int,
+        query_head_dim: int,
+        pos_head_dim: int,
+        value_head_dim: int,
+        feedforward_dim: int,
+        dropout: FloatLike = 0.1,
+        cnn_module_kernel: int = 31,
+        causal: bool = False,
+        attention_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0
+        ),
+        conv_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0
+        ),
+        const_attention_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.25), (4000.0, 0.025), default=0
+        ),
+        ff2_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)
+        ),
+        ff3_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)
+        ),
+        bypass_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.5), (4000.0, 0.02), default=0
+        ),
+    ) -> None:
+        super(Zipformer2EncoderLayer, self).__init__()
+        self.embed_dim = embed_dim
+
+        # self.bypass implements layer skipping as well as bypass; see its default values.
+        self.bypass = BypassModule(
+            embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0
+        )
+        # bypass_mid is bypass used in the middle of the layer.
+        self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0)
+
+        # skip probability for dynamic modules (meaning: anything but feedforward).
+        self.attention_skip_rate = copy.deepcopy(attention_skip_rate)
+        # an additional skip probability that applies to ConvModule to stop it from
+        # contributing too much early on.
+        self.conv_skip_rate = copy.deepcopy(conv_skip_rate)
+
+        # ff2_skip_rate is to prevent the ff2 module from having output that's too big
+        # compared to its residual.
+        self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate)
+        self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate)
+
+        self.const_attention_rate = copy.deepcopy(const_attention_rate)
+
+        self.self_attn_weights = RelPositionMultiheadAttentionWeights(
+            embed_dim,
+            pos_dim=pos_dim,
+            num_heads=num_heads,
+            query_head_dim=query_head_dim,
+            pos_head_dim=pos_head_dim,
+            dropout=0.0,
+        )
+
+        self.self_attn1 = SelfAttention(embed_dim, num_heads, value_head_dim)
+
+        self.self_attn2 = SelfAttention(embed_dim, num_heads, value_head_dim)
+
+        self.feed_forward1 = FeedforwardModule(
+            embed_dim, (feedforward_dim * 3) // 4, dropout
+        )
+
+        self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout)
+
+        self.feed_forward3 = FeedforwardModule(
+            embed_dim, (feedforward_dim * 5) // 4, dropout
+        )
+
+        self.nonlin_attention = NonlinAttention(
+            embed_dim, hidden_channels=3 * embed_dim // 4
+        )
+
+        self.conv_module1 = ConvolutionModule(
+            embed_dim, cnn_module_kernel, causal=causal
+        )
+
+        self.conv_module2 = ConvolutionModule(
+            embed_dim, cnn_module_kernel, causal=causal
+        )
+
+        # TODO: remove it
+        self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5))
+
+        self.norm = BiasNorm(embed_dim)
+
+        self.balancer1 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            min_abs=0.2,
+            max_abs=4.0,
+        )
+
+        # balancer for output of NonlinAttentionModule
+        self.balancer_na = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)),
+            prob=0.05,  # out of concern for memory usage
+        )
+
+        # balancer for output of feedforward2, prevent it from staying too
+        # small.  give this a very small probability, even at the start of
+        # training, it's to fix a rare problem and it's OK to fix it slowly.
+        self.balancer_ff2 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0),
+            max_abs=2.0,
+            prob=0.05,
+        )
+
+        self.balancer_ff3 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0),
+            max_abs=4.0,
+            prob=0.05,
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(4.0, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.balancer2 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            min_abs=0.1,
+            max_abs=4.0,
+        )
+
+    def get_sequence_dropout_mask(
+        self, x: Tensor, dropout_rate: float
+    ) -> Optional[Tensor]:
+        if (
+            dropout_rate == 0.0
+            or not self.training
+            or torch.jit.is_scripting()
+            or torch.jit.is_tracing()
+        ):
+            return None
+        batch_size = x.shape[1]
+        mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype)
+        return mask
+
+    def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor:
+        """
+        Apply sequence-level dropout to x.
+        x shape: (seq_len, batch_size, embed_dim)
+        """
+        dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate)
+        if dropout_mask is None:
+            return x
+        else:
+            return x * dropout_mask
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        chunk_size: int = -1,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+            Pass the input through the encoder layer.
+            Args:
+                src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+             pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim)
+             chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking.
+           feature_mask: something that broadcasts with src, that we'll multiply `src`
+                  by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+             attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                    interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                   True means masked position. May be None.
+        src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+
+            Returns:
+               A tensor which has the same shape as src
+        """
+        src_orig = src
+
+        # dropout rate for non-feedforward submodules
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            attention_skip_rate = 0.0
+        else:
+            attention_skip_rate = (
+                float(self.attention_skip_rate) if self.training else 0.0
+            )
+
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        attn_weights = self.self_attn_weights(
+            src,
+            pos_emb=pos_emb,
+            attn_mask=attn_mask,
+            key_padding_mask=src_key_padding_mask,
+        )
+
+        src = src + self.feed_forward1(src)
+
+        self_attn_dropout_mask = self.get_sequence_dropout_mask(
+            src, attention_skip_rate
+        )
+
+        selected_attn_weights = attn_weights[0:1]
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            pass
+        elif not self.training and random.random() < float(self.const_attention_rate):
+            # Make attention weights constant.  The intention is to
+            # encourage these modules to do something similar to an
+            # averaging-over-time operation.
+            # only need the mask, can just use the 1st one and expand later
+            selected_attn_weights = selected_attn_weights[0:1]
+            selected_attn_weights = (selected_attn_weights > 0.0).to(
+                selected_attn_weights.dtype
+            )
+            selected_attn_weights = selected_attn_weights * (
+                1.0 / selected_attn_weights.sum(dim=-1, keepdim=True)
+            )
+
+        na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights))
+
+        src = src + (
+            na if self_attn_dropout_mask is None else na * self_attn_dropout_mask
+        )
+
+        self_attn = self.self_attn1(src, attn_weights)
+
+        src = src + (
+            self_attn
+            if self_attn_dropout_mask is None
+            else self_attn * self_attn_dropout_mask
+        )
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            conv_skip_rate = 0.0
+        else:
+            conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0
+        src = src + self.sequence_dropout(
+            self.conv_module1(
+                src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask
+            ),
+            conv_skip_rate,
+        )
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            ff2_skip_rate = 0.0
+        else:
+            ff2_skip_rate = float(self.ff2_skip_rate) if self.training else 0.0
+        src = src + self.sequence_dropout(
+            self.balancer_ff2(self.feed_forward2(src)), ff2_skip_rate
+        )
+
+        # bypass in the middle of the layer.
+        src = self.bypass_mid(src_orig, src)
+
+        self_attn = self.self_attn2(src, attn_weights)
+
+        src = src + (
+            self_attn
+            if self_attn_dropout_mask is None
+            else self_attn * self_attn_dropout_mask
+        )
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            conv_skip_rate = 0.0
+        else:
+            conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0
+        src = src + self.sequence_dropout(
+            self.conv_module2(
+                src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask
+            ),
+            conv_skip_rate,
+        )
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            ff3_skip_rate = 0.0
+        else:
+            ff3_skip_rate = float(self.ff3_skip_rate) if self.training else 0.0
+        src = src + self.sequence_dropout(
+            self.balancer_ff3(self.feed_forward3(src)), ff3_skip_rate
+        )
+
+        src = self.balancer1(src)
+        src = self.norm(src)
+
+        src = self.bypass(src_orig, src)
+
+        src = self.balancer2(src)
+        src = self.whiten(src)
+
+        return src
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        cached_key: Tensor,
+        cached_nonlin_attn: Tensor,
+        cached_val1: Tensor,
+        cached_val2: Tensor,
+        cached_conv1: Tensor,
+        cached_conv2: Tensor,
+        left_context_len: int,
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+        """Pass the input through the encoder layer in streaming forward mode.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            pos_emb: (1, left_context_len+2*seq_len-1, pos_emb_dim) or
+              (batch_size, left_context_len+2*seq_len-1, pos_emb_dim)
+            cached_key: cached attention key tensor of left context,
+              of shape (left_context_len, batch_size, key_dim)
+            cached_nonlin_attn: left context for nonlin_attention module, a Tensor of shape
+              (num_heads, batch_size, left_context_len, head_dim)
+            cached_val1: cached left context for the first attention module,
+              of shape (left_context_len, batch_size, value_dim)
+            cached_val2: cached left context for the second attention module,
+              of shape (left_context_len, batch_size, value_dim)
+            cached_conv1: cached left context for the first convolution module,
+              of shape (batch_size, channels, left_pad)
+            cached_conv2: cached left context for the second convolution module,
+              of shape (batch_size, channels, left_pad)
+            left_context_len: number of left context frames.
+            src_key_padding_mask:  the mask for padding, of shape
+              (batch_size, left_context_len + seq_len); True means masked position.
+              May be None.
+
+        Returns:
+            - x, with the same shape as src
+            - updated cached_key
+            - updated cached_nonlin_attn
+            - updated cached_val1
+            - updated cached_val2
+            - updated cached_conv1
+            - updated cached_conv2
+        """
+        src_orig = src
+
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        attn_weights, cached_key = self.self_attn_weights.streaming_forward(
+            src,
+            pos_emb=pos_emb,
+            cached_key=cached_key,
+            left_context_len=left_context_len,
+            key_padding_mask=src_key_padding_mask,
+        )
+
+        src = src + self.feed_forward1(src)
+
+        na, cached_nonlin_attn = self.nonlin_attention.streaming_forward(
+            src,
+            attn_weights[0:1],
+            cached_x=cached_nonlin_attn,
+            left_context_len=left_context_len,
+        )
+        src = src + na
+
+        self_attn, cached_val1 = self.self_attn1.streaming_forward(
+            src,
+            attn_weights=attn_weights,
+            cached_val=cached_val1,
+            left_context_len=left_context_len,
+        )
+        src = src + self_attn
+
+        src_conv, cached_conv1 = self.conv_module1.streaming_forward(
+            src,
+            cache=cached_conv1,
+            src_key_padding_mask=src_key_padding_mask[:, left_context_len:],
+        )
+        src = src + src_conv
+
+        src = src + self.feed_forward2(src)
+
+        # bypass in the middle of the layer.
+        src = self.bypass_mid(src_orig, src)
+
+        self_attn, cached_val2 = self.self_attn2.streaming_forward(
+            src,
+            attn_weights=attn_weights,
+            cached_val=cached_val2,
+            left_context_len=left_context_len,
+        )
+        src = src + self_attn
+
+        src_conv, cached_conv2 = self.conv_module2.streaming_forward(
+            src,
+            cache=cached_conv2,
+            src_key_padding_mask=src_key_padding_mask[:, left_context_len:],
+        )
+        src = src + src_conv
+
+        src = src + self.feed_forward3(src)
+
+        src = self.norm(src)
+
+        src = self.bypass(src_orig, src)
+
+        return (
+            src,
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        )
+
+
+class Zipformer2Encoder(nn.Module):
+    r"""Zipformer2Encoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the Zipformer2EncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+       pos_dim: the dimension for the relative positional encoding
+
+    Examples::
+        >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8)
+        >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = zipformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        pos_dim: int,
+        dropout: float,
+        warmup_begin: float,
+        warmup_end: float,
+        initial_layerdrop_rate: float = 0.5,
+        final_layerdrop_rate: float = 0.05,
+    ) -> None:
+        super().__init__()
+        self.encoder_pos = CompactRelPositionalEncoding(
+            pos_dim, dropout_rate=0.15, length_factor=1.0
+        )
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert 0 <= warmup_begin <= warmup_end
+
+        delta = (1.0 / num_layers) * (warmup_end - warmup_begin)
+        cur_begin = warmup_begin  # interpreted as a training batch index
+        for i in range(num_layers):
+            cur_end = cur_begin + delta
+            self.layers[i].bypass.skip_rate = ScheduledFloat(
+                (cur_begin, initial_layerdrop_rate),
+                (cur_end, final_layerdrop_rate),
+                default=0.0,
+            )
+            cur_begin = cur_end
+
+    def forward(
+        self,
+        src: Tensor,
+        chunk_size: int = -1,
+        feature_mask: Union[Tensor, float] = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking.
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+            attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                 interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                 True means masked position. May be None.
+            src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+
+        Returns: a Tensor with the same shape as src.
+        """
+        pos_emb = self.encoder_pos(src)
+        output = src
+
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            output = output * feature_mask
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                chunk_size=chunk_size,
+                attn_mask=attn_mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+            if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+                output = output * feature_mask
+
+        return output
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        states: List[Tensor],
+        left_context_len: int,
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is
+              (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+            left_context_len: Number of left context frames.
+            src_key_padding_mask:  the mask for padding, of shape
+              (batch_size, left_context_len + seq_len); True means masked position.
+              May be None.
+
+        Returns:
+          - output, a Tensor with the same shape as src.
+          - updated states
+        """
+        pos_emb = self.encoder_pos(src, left_context_len)
+        output = src
+
+        new_states = []
+        for i, mod in enumerate(self.layers):
+            (
+                cached_key,
+                cached_nonlin_attn,
+                cached_val1,
+                cached_val2,
+                cached_conv1,
+                cached_conv2,
+            ) = states[i * 6 : (i + 1) * 6]
+            (
+                output,
+                new_cached_key,
+                new_cached_nonlin_attn,
+                new_cached_val1,
+                new_cached_val2,
+                new_cached_conv1,
+                new_cached_conv2,
+            ) = mod.streaming_forward(
+                output,
+                pos_emb,
+                cached_key=cached_key,
+                cached_nonlin_attn=cached_nonlin_attn,
+                cached_val1=cached_val1,
+                cached_val2=cached_val2,
+                cached_conv1=cached_conv1,
+                cached_conv2=cached_conv2,
+                left_context_len=left_context_len,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+            new_states += [
+                new_cached_key,
+                new_cached_nonlin_attn,
+                new_cached_val1,
+                new_cached_val2,
+                new_cached_conv1,
+                new_cached_conv2,
+            ]
+
+        return output, new_states
+
+
+class BypassModule(nn.Module):
+    """
+    An nn.Module that implements a learnable bypass scale, and also randomized per-sequence
+    layer-skipping.  The bypass is limited during early stages of training to be close to
+    "straight-through", i.e. to not do the bypass operation much initially, in order to
+    force all the modules to learn something.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        skip_rate: FloatLike = 0.0,
+        straight_through_rate: FloatLike = 0.0,
+        scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0),
+        scale_max: FloatLike = 1.0,
+    ):
+        super().__init__()
+        self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5))
+        self.skip_rate = copy.deepcopy(skip_rate)
+        self.straight_through_rate = copy.deepcopy(straight_through_rate)
+        self.scale_min = copy.deepcopy(scale_min)
+        self.scale_max = copy.deepcopy(scale_max)
+
+    def _get_bypass_scale(self, batch_size: int):
+        # returns bypass-scale of shape (num_channels,),
+        # or (batch_size, num_channels,).  This is actually the
+        # scale on the non-residual term, so 0 correponds to bypassing
+        # this module.
+        if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training:
+            return self.bypass_scale
+        else:
+            ans = limit_param_value(
+                self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max)
+            )
+            skip_rate = float(self.skip_rate)
+            if skip_rate != 0.0:
+                mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate
+                ans = ans * mask
+                # now ans is of shape (batch_size, num_channels), and is zero for sequences
+                # on which we have randomly chosen to do layer-skipping.
+            straight_through_rate = float(self.straight_through_rate)
+            if straight_through_rate != 0.0:
+                mask = (
+                    torch.rand((batch_size, 1), device=ans.device)
+                    < straight_through_rate
+                )
+                ans = torch.maximum(ans, mask.to(ans.dtype))
+            return ans
+
+    def forward(self, src_orig: Tensor, src: Tensor):
+        """
+        Args: src_orig and src are both of shape (seq_len, batch_size, num_channels)
+        Returns: something with the same shape as src and src_orig
+        """
+        bypass_scale = self._get_bypass_scale(src.shape[1])
+        return src_orig + (src - src_orig) * bypass_scale
+
+
+class DownsampledZipformer2Encoder(nn.Module):
+    r"""
+    DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate,
+    after convolutional downsampling, and then upsampled again at the output, and combined
+    with the origin input, so that the output has the same shape as the input.
+    """
+
+    def __init__(
+        self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike
+    ):
+        super(DownsampledZipformer2Encoder, self).__init__()
+        self.downsample_factor = downsample
+        self.downsample = SimpleDownsample(dim, downsample, dropout)
+        self.num_layers = encoder.num_layers
+        self.encoder = encoder
+        self.upsample = SimpleUpsample(dim, downsample)
+        self.out_combiner = BypassModule(dim, straight_through_rate=0)
+
+    def forward(
+        self,
+        src: Tensor,
+        chunk_size: int = -1,
+        feature_mask: Union[Tensor, float] = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+            attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                 interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                 True means masked position. May be None.
+            src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+
+        Returns: a Tensor with the same shape as src.
+        """
+        src_orig = src
+        src = self.downsample(src)
+        ds = self.downsample_factor
+        if attn_mask is not None:
+            attn_mask = attn_mask[::ds, ::ds]
+
+        src = self.encoder(
+            src,
+            chunk_size=chunk_size // ds,
+            feature_mask=feature_mask,
+            attn_mask=attn_mask,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src)
+
+    def streaming_forward(
+        self,
+        src: Tensor,
+        states: List[Tensor],
+        left_context_len: int,
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        r"""Downsample, go through encoder, upsample, in streaming forward mode.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is
+              (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+            left_context_len: Number of left context frames.
+            src_key_padding_mask: the mask for padding, of shape (batch_size, left_context_len+seq_len);
+              True means masked position. May be None.
+
+        Returns:
+            - output, a Tensor with the same shape as src.
+            - updated states
+        """
+        src_orig = src
+        src = self.downsample(src)
+
+        src, new_states = self.encoder.streaming_forward(
+            src,
+            states=states,
+            left_context_len=left_context_len,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src), new_states
+
+
+class SimpleDownsample(torch.nn.Module):
+    """
+    Does downsampling with attention, by weighted sum, and a projection..
+    """
+
+    def __init__(self, channels: int, downsample: int, dropout: FloatLike):
+        super(SimpleDownsample, self).__init__()
+
+        self.bias = nn.Parameter(torch.zeros(downsample))
+
+        self.name = None  # will be set from training code
+        self.dropout = copy.deepcopy(dropout)
+
+        self.downsample = downsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, in_channels)
+        Returns a tensor of shape
+           ( (seq_len+downsample-1)//downsample, batch_size, channels)
+        """
+        (seq_len, batch_size, in_channels) = src.shape
+        ds = self.downsample
+        d_seq_len = (seq_len + ds - 1) // ds
+
+        # Pad to an exact multiple of self.downsample
+        # right-pad src, repeating the last element.
+        pad = d_seq_len * ds - seq_len
+        src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2])
+        src = torch.cat((src, src_extra), dim=0)
+        assert src.shape[0] == d_seq_len * ds
+
+        src = src.reshape(d_seq_len, ds, batch_size, in_channels)
+
+        weights = self.bias.softmax(dim=0)
+        # weights: (downsample, 1, 1)
+        weights = weights.unsqueeze(-1).unsqueeze(-1)
+
+        # ans1 is the first `in_channels` channels of the output
+        ans = (src * weights).sum(dim=1)
+
+        return ans
+
+
+class SimpleUpsample(torch.nn.Module):
+    """
+    A very simple form of upsampling that mostly just repeats the input, but
+    also adds a position-specific bias.
+    """
+
+    def __init__(self, num_channels: int, upsample: int):
+        super(SimpleUpsample, self).__init__()
+        self.upsample = upsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, num_channels)
+        Returns a tensor of shape
+           ( (seq_len*upsample), batch_size, num_channels)
+        """
+        upsample = self.upsample
+        (seq_len, batch_size, num_channels) = src.shape
+        src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels)
+        src = src.reshape(seq_len * upsample, batch_size, num_channels)
+        return src
+
+
+class CompactRelPositionalEncoding(torch.nn.Module):
+    """
+    Relative positional encoding module.  This version is "compact" meaning it is able to encode
+    the important information about the relative position in a relatively small number of dimensions.
+    The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001)
+    make very little difference to the embedding.   Such differences were potentially important
+    when encoding absolute position, but not important when encoding relative position because there
+    is now no need to compare two large offsets with each other.
+
+    Our embedding works done by projecting the interval [-infinity,infinity] to a finite interval
+    using the atan() function, before doing the fourier transform of that fixed interval.  The
+    atan() function would compress the "long tails" too small,
+    making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic
+    function to compress large offsets to a smaller range before applying atan().
+    Scalings are chosen in such a way that the embedding can clearly distinguish invidual offsets as long
+    as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim)
+
+
+    Args:
+        embed_dim: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length: just a heuristic for initialization.
+        length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives
+           less weight to small differences of offset near the origin.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        dropout_rate: FloatLike,
+        max_len: int = 1000,
+        length_factor: float = 1.0,
+    ) -> None:
+        """Construct a CompactRelPositionalEncoding object."""
+        super(CompactRelPositionalEncoding, self).__init__()
+        self.embed_dim = embed_dim
+        assert embed_dim % 2 == 0
+        self.dropout = Dropout2(dropout_rate)
+        self.pe = None
+        assert length_factor >= 1.0
+        self.length_factor = length_factor
+        self.extend_pe(torch.tensor(0.0).expand(max_len))
+
+    def extend_pe(self, x: Tensor, left_context_len: int = 0) -> None:
+        """Reset the positional encodings."""
+        T = x.size(0) + left_context_len
+
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(0) >= T * 2 - 1:
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+
+        # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ]
+        x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1)
+
+        freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device)
+
+        # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution
+        # for small time offsets but less resolution for large time offsets.
+        compression_length = self.embed_dim**0.5
+        # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity;
+        # but it does so more slowly than T for large absolute values of T.
+        # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which
+        # is important.
+        x_compressed = (
+            compression_length
+            * x.sign()
+            * ((x.abs() + compression_length).log() - math.log(compression_length))
+        )
+
+        # if self.length_factor == 1.0, then length_scale is chosen so that the
+        # FFT can exactly separate points close to the origin (T == 0).  So this
+        # part of the formulation is not really heuristic.
+        # But empirically, for ASR at least, length_factor > 1.0 seems to work better.
+        length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi)
+
+        # note for machine implementations: if atan is not available, we can use:
+        #   x.sign() * ((1 / (x.abs() + 1)) - 1)  * (-math.pi/2)
+        #  check on wolframalpha.com: plot(sign(x) *  (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x))
+        x_atan = (x_compressed / length_scale).atan()  # results between -pi and pi
+
+        cosines = (x_atan * freqs).cos()
+        sines = (x_atan * freqs).sin()
+
+        pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device)
+        pe[:, 0::2] = cosines
+        pe[:, 1::2] = sines
+        pe[:, -1] = 1.0  # for bias.
+
+        self.pe = pe.to(dtype=x.dtype)
+
+    def forward(self, x: Tensor, left_context_len: int = 0) -> Tensor:
+        """Create positional encoding.
+
+        Args:
+            x (Tensor): Input tensor (time, batch, `*`).
+            left_context_len: (int): Length of cached left context.
+
+        Returns:
+            positional embedding, of shape (batch, left_context_len + 2*time-1, `*`).
+        """
+        self.extend_pe(x, left_context_len)
+        x_size_left = x.size(0) + left_context_len
+        # length of positive side: x.size(0) + left_context_len
+        # length of negative side: x.size(0)
+        pos_emb = self.pe[
+            self.pe.size(0) // 2
+            - x_size_left
+            + 1 : self.pe.size(0) // 2  # noqa E203
+            + x.size(0),
+            :,
+        ]
+        pos_emb = pos_emb.unsqueeze(0)
+        return self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttentionWeights(nn.Module):
+    r"""Module that computes multi-head attention weights with relative position encoding.
+    Various other modules consume the resulting attention weights: see, for example, the
+    SimpleAttention module which allows you to compute conventional attention.
+
+    This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
+    we have to write up the differences.
+
+
+    Args:
+           embed_dim: number of channels at the input to this module, e.g. 256
+             pos_dim: dimension of the positional encoding vectors, e.g. 128.
+           num_heads:  number of heads to compute weights for, e.g. 8
+     query_head_dim: dimension of the query (and key), per head.  e.g. 24.
+       pos_head_dim: dimension of the projected positional encoding per head, e.g. 4.
+            dropout: dropout probability for attn_output_weights. Default: 0.0.
+       pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on
+                     any given call to forward(), in training time.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        pos_dim: int,
+        num_heads: int,
+        query_head_dim: int,
+        pos_head_dim: int,
+        dropout: float = 0.0,
+        pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)),
+    ) -> None:
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.query_head_dim = query_head_dim
+        self.pos_head_dim = pos_head_dim
+        self.dropout = dropout
+        self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate)
+        self.name = None  # will be overwritten in training code; for diagnostics.
+
+        key_head_dim = query_head_dim
+        in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5 that has been used in previous forms of attention,
+        # dividing it between the query and key.   Note: this module is intended
+        # to be used with the ScaledAdam optimizer; with most other optimizers,
+        # it would be necessary to apply the scaling factor in the forward function.
+        self.in_proj = ScaledLinear(
+            embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25
+        )
+
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=_whitening_schedule(3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+        # add a balancer for the keys that runs with very small probability, and
+        # tries to enforce that all dimensions have mean around zero.  The
+        # weights produced by this module are invariant to adding a constant to
+        # the keys, so the derivative of the bias is mathematically zero; but
+        # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero
+        # bias because the small numerical roundoff tends to have a non-random
+        # sign.  This module is intended to prevent that.  Use a very small
+        # probability; that should be suffixient to fix the problem.
+        self.balance_keys = Balancer(
+            key_head_dim * num_heads,
+            channel_dim=-1,
+            min_positive=0.4,
+            max_positive=0.6,
+            min_abs=0.0,
+            max_abs=100.0,
+            prob=0.025,
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(
+            pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05
+        )
+
+        # the following are for diagnosics only, see --print-diagnostics option
+        self.copy_pos_query = Identity()
+        self.copy_query = Identity()
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""
+        Args:
+            x: input of shape (seq_len, batch_size, embed_dim)
+            pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 1, pos_dim)
+            key_padding_mask: a bool tensor of shape (batch_size, seq_len).  Positions that
+               are True in this mask will be ignored as sources in the attention weighting.
+            attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len),
+               interpreted as ([batch_size,] tgt_seq_len, src_seq_len)
+               saying which positions are allowed to attend to which other positions.
+        Returns:
+           a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len)
+           interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len).
+        """
+        x = self.in_proj(x)
+        query_head_dim = self.query_head_dim
+        pos_head_dim = self.pos_head_dim
+        num_heads = self.num_heads
+
+        seq_len, batch_size, _ = x.shape
+
+        query_dim = query_head_dim * num_heads
+
+        # self-attention
+        q = x[..., 0:query_dim]
+        k = x[..., query_dim : 2 * query_dim]
+        # p is the position-encoding query
+        p = x[..., 2 * query_dim :]
+        assert p.shape[-1] == num_heads * pos_head_dim
+
+        q = self.copy_query(q)  # for diagnostics only, does nothing.
+        k = self.whiten_keys(self.balance_keys(k))  # does nothing in the forward pass.
+        p = self.copy_pos_query(p)  # for diagnostics only, does nothing.
+
+        q = q.reshape(seq_len, batch_size, num_heads, query_head_dim)
+        p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim)
+        k = k.reshape(seq_len, batch_size, num_heads, query_head_dim)
+
+        # time1 refers to target, time2 refers to source.
+        q = q.permute(2, 1, 0, 3)  # (head, batch, time1, query_head_dim)
+        p = p.permute(2, 1, 0, 3)  # (head, batch, time1, pos_head_dim)
+        k = k.permute(2, 1, 3, 0)  # (head, batch, d_k, time2)
+
+        attn_scores = torch.matmul(q, k)
+
+        use_pos_scores = False
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            # We can't put random.random() in the same line
+            use_pos_scores = True
+        elif not self.training or random.random() >= float(self.pos_emb_skip_rate):
+            use_pos_scores = True
+
+        if use_pos_scores:
+            pos_emb = self.linear_pos(pos_emb)
+            seq_len2 = 2 * seq_len - 1
+            pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute(
+                2, 0, 3, 1
+            )
+            # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2)
+
+            # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2)
+            #  [where seq_len2 represents relative position.]
+            pos_scores = torch.matmul(p, pos_emb)
+            # the following .as_strided() expression converts the last axis of pos_scores from relative
+            # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+            # not, but let this code define which way round it is supposed to be.
+            if torch.jit.is_tracing():
+                (num_heads, batch_size, time1, n) = pos_scores.shape
+                rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+                cols = torch.arange(seq_len)
+                rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+                indexes = rows + cols
+                pos_scores = pos_scores.reshape(-1, n)
+                pos_scores = torch.gather(pos_scores, dim=1, index=indexes)
+                pos_scores = pos_scores.reshape(num_heads, batch_size, time1, seq_len)
+            else:
+                pos_scores = pos_scores.as_strided(
+                    (num_heads, batch_size, seq_len, seq_len),
+                    (
+                        pos_scores.stride(0),
+                        pos_scores.stride(1),
+                        pos_scores.stride(2) - pos_scores.stride(3),
+                        pos_scores.stride(3),
+                    ),
+                    storage_offset=pos_scores.stride(3) * (seq_len - 1),
+                )
+
+            attn_scores = attn_scores + pos_scores
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            pass
+        elif self.training and random.random() < 0.1:
+            # This is a harder way of limiting the attention scores to not be
+            # too large.  It incurs a penalty if any of them has an absolute
+            # value greater than 50.0.  this should be outside the normal range
+            # of the attention scores.  We use this mechanism instead of, say,
+            # something added to the loss function involving the entropy,
+            # because once the entropy gets very small gradients through the
+            # softmax can become very small, and we'd get zero derivatives.  The
+            # choices of 1.0e-04 as the scale on the penalty makes this
+            # mechanism vulnerable to the absolute scale of the loss function,
+            # but we view this as a failsafe to avoid "implausible" parameter
+            # values rather than a regularization method that should be active
+            # under normal circumstances.
+            attn_scores = penalize_abs_values_gt(
+                attn_scores, limit=25.0, penalty=1.0e-04, name=self.name
+            )
+
+        assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        if attn_mask is not None:
+            assert attn_mask.dtype == torch.bool
+            # use -1000 to avoid nan's where attn_mask and key_padding_mask make
+            # all scores zero.  It's important that this be large enough that exp(-1000)
+            # is exactly zero, for reasons related to const_attention_rate, it
+            # compares the final weights with zero.
+            attn_scores = attn_scores.masked_fill(attn_mask, -1000)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (
+                batch_size,
+                seq_len,
+            ), key_padding_mask.shape
+            attn_scores = attn_scores.masked_fill(
+                key_padding_mask.unsqueeze(1),
+                -1000,
+            )
+
+        # We use our own version of softmax, defined in scaling.py, which should
+        # save a little of the memory used in backprop by, if we are in
+        # automatic mixed precision mode (amp / autocast), by only storing the
+        # half-precision output for backprop purposes.
+        attn_weights = softmax(attn_scores, dim=-1)
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            pass
+        elif random.random() < 0.001 and not self.training:
+            self._print_attn_entropy(attn_weights)
+
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )
+
+        return attn_weights
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        cached_key: Tensor,
+        left_context_len: int,
+        key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""
+        Args:
+            x: input of shape (seq_len, batch_size, embed_dim)
+            pos_emb: Positional embedding tensor, of shape (1, left_context_len+2*seq_len-1, pos_dim)
+            cached_key: cached attention key tensor of left context,
+              of shape (left_context_len, batch_size, key_dim)
+            left_context_len: number of left context frames.
+            key_padding_mask: a bool tensor of shape (batch_size, seq_len).  Positions that
+              are True in this mask will be ignored as sources in the attention weighting.
+
+        Returns:
+           - attention weights, of shape (hum_heads, batch_size, seq_len, seq_len2),
+             interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len).
+           - updated cached attention key tensor of left context.
+        """
+        x = self.in_proj(x)
+        query_head_dim = self.query_head_dim
+        pos_head_dim = self.pos_head_dim
+        num_heads = self.num_heads
+
+        seq_len, batch_size, _ = x.shape
+
+        query_dim = query_head_dim * num_heads
+
+        # self-attention
+        q = x[..., 0:query_dim]
+        k = x[..., query_dim : 2 * query_dim]
+        # p is the position-encoding query
+        p = x[..., 2 * query_dim :]
+        assert p.shape[-1] == num_heads * pos_head_dim
+
+        # Pad cached left contexts
+        assert cached_key.shape[0] == left_context_len, (
+            cached_key.shape[0],
+            left_context_len,
+        )
+        k = torch.cat([cached_key, k], dim=0)
+        # Update cached left contexts
+        cached_key = k[-left_context_len:, ...]
+
+        # The length of key
+        k_len = k.shape[0]
+
+        q = q.reshape(seq_len, batch_size, num_heads, query_head_dim)
+        p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim)
+        k = k.reshape(k_len, batch_size, num_heads, query_head_dim)
+
+        # time1 refers to target, time2 refers to source.
+        q = q.permute(2, 1, 0, 3)  # (head, batch, time1, query_head_dim)
+        p = p.permute(2, 1, 0, 3)  # (head, batch, time1, pos_head_dim)
+        k = k.permute(2, 1, 3, 0)  # (head, batch, d_k, time2)
+
+        attn_scores = torch.matmul(q, k)
+
+        pos_emb = self.linear_pos(pos_emb)
+        seq_len2 = 2 * seq_len - 1 + left_context_len
+        pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute(
+            2, 0, 3, 1
+        )
+        # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2)
+
+        # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2)
+        #  [where seq_len2 represents relative position.]
+        pos_scores = torch.matmul(p, pos_emb)
+
+        if torch.jit.is_tracing():
+            (num_heads, batch_size, time1, n) = pos_scores.shape
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(k_len)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+            pos_scores = pos_scores.reshape(-1, n)
+            pos_scores = torch.gather(pos_scores, dim=1, index=indexes)
+            pos_scores = pos_scores.reshape(num_heads, batch_size, time1, k_len)
+        # the following .as_strided() expression converts the last axis of pos_scores from relative
+        # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+        # not, but let this code define which way round it is supposed to be.
+        else:
+            pos_scores = pos_scores.as_strided(
+                (num_heads, batch_size, seq_len, k_len),
+                (
+                    pos_scores.stride(0),
+                    pos_scores.stride(1),
+                    pos_scores.stride(2) - pos_scores.stride(3),
+                    pos_scores.stride(3),
+                ),
+                storage_offset=pos_scores.stride(3) * (seq_len - 1),
+            )
+
+        attn_scores = attn_scores + pos_scores
+
+        assert attn_scores.shape == (
+            num_heads,
+            batch_size,
+            seq_len,
+            k_len,
+        ), attn_scores.shape
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (batch_size, k_len), key_padding_mask.shape
+            attn_scores = attn_scores.masked_fill(
+                key_padding_mask.unsqueeze(1),
+                -1000,
+            )
+
+        attn_weights = attn_scores.softmax(dim=-1)
+
+        return attn_weights, cached_key
+
+    def _print_attn_entropy(self, attn_weights: Tensor):
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .mean(dim=(1, 2))
+                )
+                logging.info(
+                    f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}"
+                )
+
+
+class SelfAttention(nn.Module):
+    """
+    The simplest possible attention module.  This one works with already-computed attention
+    weights, e.g. as computed by RelPositionMultiheadAttentionWeights.
+
+    Args:
+          embed_dim: the input and output embedding dimension
+          num_heads: the number of attention heads
+          value_head_dim: the value dimension per head
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        value_head_dim: int,
+    ) -> None:
+        super().__init__()
+        self.in_proj = nn.Linear(embed_dim, num_heads * value_head_dim, bias=True)
+
+        self.out_proj = ScaledLinear(
+            num_heads * value_head_dim, embed_dim, bias=True, initial_scale=0.05
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """
+        Args:
+          x: input tensor, of shape (seq_len, batch_size, embed_dim)
+         attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len),
+          with seq_len being interpreted as (tgt_seq_len, src_seq_len).  Expect
+          attn_weights.sum(dim=-1) == 1.
+        Returns:
+           a tensor with the same shape as x.
+        """
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        x = self.in_proj(x)  # (seq_len, batch_size, num_heads * value_head_dim)
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, value_head_dim)
+        value_head_dim = x.shape[-1]
+
+        # todo: see whether there is benefit in overriding matmul
+        x = torch.matmul(attn_weights, x)
+        # v: (num_heads, batch_size, seq_len, value_head_dim)
+
+        x = (
+            x.permute(2, 1, 0, 3)
+            .contiguous()
+            .view(seq_len, batch_size, num_heads * value_head_dim)
+        )
+
+        # returned value is of shape (seq_len, batch_size, embed_dim), like the input.
+        x = self.out_proj(x)
+        x = self.whiten(x)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_val: Tensor,
+        left_context_len: int,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            x: input tensor, of shape (seq_len, batch_size, embed_dim)
+            attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len),
+              with seq_len being interpreted as (tgt_seq_len, src_seq_len).  Expect
+              attn_weights.sum(dim=-1) == 1.
+            cached_val: cached attention value tensor of left context,
+              of shape (left_context_len, batch_size, value_dim)
+            left_context_len: number of left context frames.
+
+        Returns:
+           - attention weighted output, a tensor with the same shape as x.
+           - updated cached attention value tensor of left context.
+        """
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        seq_len2 = seq_len + left_context_len
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2)
+
+        x = self.in_proj(x)  # (seq_len, batch_size, num_heads * value_head_dim)
+
+        # Pad cached left contexts
+        assert cached_val.shape[0] == left_context_len, (
+            cached_val.shape[0],
+            left_context_len,
+        )
+        x = torch.cat([cached_val, x], dim=0)
+        # Update cached left contexts
+        cached_val = x[-left_context_len:, ...]
+
+        x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, value_head_dim)
+        value_head_dim = x.shape[-1]
+
+        # todo: see whether there is benefit in overriding matmul
+        x = torch.matmul(attn_weights, x)
+        # v: (num_heads, batch_size, seq_len, value_head_dim)
+
+        x = (
+            x.permute(2, 1, 0, 3)
+            .contiguous()
+            .view(seq_len, batch_size, num_heads * value_head_dim)
+        )
+
+        # returned value is of shape (seq_len, batch_size, embed_dim), like the input.
+        x = self.out_proj(x)
+
+        return x, cached_val
+
+
+class FeedforwardModule(nn.Module):
+    """Feedforward module in Zipformer2 model."""
+
+    def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike):
+        super(FeedforwardModule, self).__init__()
+        self.in_proj = nn.Linear(embed_dim, feedforward_dim)
+
+        self.hidden_balancer = Balancer(
+            feedforward_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=1.0,
+            min_abs=0.75,
+            max_abs=5.0,
+        )
+
+        # shared_dim=0 means we share the dropout mask along the time axis
+        self.out_proj = ActivationDropoutAndLinear(
+            feedforward_dim,
+            embed_dim,
+            activation="SwooshL",
+            dropout_p=dropout,
+            dropout_shared_dim=0,
+            bias=True,
+            initial_scale=0.1,
+        )
+
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(self, x: Tensor):
+        x = self.in_proj(x)
+        x = self.hidden_balancer(x)
+        # out_proj contains SwooshL activation, then dropout, then linear.
+        x = self.out_proj(x)
+        x = self.out_whiten(x)
+        return x
+
+
+class NonlinAttention(nn.Module):
+    """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed
+       from the attention module) in place of actual convolution.  We also took out the second nonlinearity, the
+       one after the attention mechanism.
+
+    Args:
+        channels (int): The number of channels of conv layers.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        hidden_channels: int,
+    ) -> None:
+        super().__init__()
+
+        self.hidden_channels = hidden_channels
+
+        self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True)
+
+        # balancer that goes before the sigmoid.  Have quite a large min_abs value, at 2.0,
+        # because we noticed that well-trained instances of this module have abs-value before the sigmoid
+        # starting from about 3, and poorly-trained instances of the module have smaller abs values
+        # before the sigmoid.
+        self.balancer = Balancer(
+            hidden_channels,
+            channel_dim=-1,
+            min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)),
+            max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)),
+            min_abs=0.5,
+            max_abs=5.0,
+        )
+        self.tanh = nn.Tanh()
+
+        self.identity1 = Identity()  # for diagnostics.
+        self.identity2 = Identity()  # for diagnostics.
+        self.identity3 = Identity()  # for diagnostics.
+
+        self.out_proj = ScaledLinear(
+            hidden_channels, channels, bias=True, initial_scale=0.05
+        )
+
+        self.whiten1 = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(5.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.whiten2 = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(5.0, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """.
+                Args:
+                   x: a Tensor of shape (seq_len, batch_size, num_channels)
+        attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
+                Returns:
+                   a Tensor with the same shape as x
+        """
+        x = self.in_proj(x)
+
+        (seq_len, batch_size, _) = x.shape
+        hidden_channels = self.hidden_channels
+
+        s, x, y = x.chunk(3, dim=2)
+
+        # s will go through tanh.
+
+        s = self.balancer(s)
+        s = self.tanh(s)
+
+        s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels)
+        x = self.whiten1(x)
+        x = x * s
+        x = self.identity1(x)  # diagnostics only, it's the identity.
+
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = torch.matmul(attn_weights, x)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1)
+
+        y = self.identity2(y)
+        x = x * y
+        x = self.identity3(x)
+
+        x = self.out_proj(x)
+        x = self.whiten2(x)
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_x: Tensor,
+        left_context_len: int,
+    ) -> Tuple[Tensor, Tensor]:
+        """.
+        Args:
+            x: a Tensor of shape (seq_len, batch_size, num_channels)
+            attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
+            cached_x: left context, a Tensor of shape
+              (num_heads, batch_size, left_context_len, head_dim)
+            left_context_len: number of left context frames.
+        Returns:
+            - a Tensor with the same shape as x
+            - updated left context with same shape as cached_x
+        """
+        x = self.in_proj(x)
+
+        (seq_len, batch_size, _) = x.shape
+        hidden_channels = self.hidden_channels
+
+        s, x, y = x.chunk(3, dim=2)
+
+        # s will go through tanh.
+        s = self.tanh(s)
+
+        s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels)
+        x = x * s
+
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (
+            num_heads,
+            batch_size,
+            seq_len,
+            left_context_len + seq_len,
+        )
+
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+
+        # Pad cached tensor
+        assert cached_x.shape[2] == left_context_len, (
+            cached_x.shape[2],
+            left_context_len,
+        )
+        x_pad = torch.cat([cached_x, x], dim=2)
+        # Update cached tensor
+        cached_x = x_pad[:, :, -left_context_len:, :]
+
+        x = torch.matmul(attn_weights, x_pad)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1)
+
+        x = x * y
+
+        x = self.out_proj(x)
+        return x, cached_x
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Zipformer2 model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        causal: bool,
+    ) -> None:
+        """Construct a ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        bottleneck_dim = channels
+        self.causal = causal
+
+        self.in_proj = nn.Linear(
+            channels,
+            2 * bottleneck_dim,
+        )
+        # the gradients on in_proj are a little noisy, likely to do with the
+        # sigmoid in glu.
+
+        # after in_proj we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.balancer1 = Balancer(
+            bottleneck_dim,
+            channel_dim=-1,
+            min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)),
+            max_positive=1.0,
+            min_abs=1.5,
+            max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0),
+        )
+
+        self.activation1 = Identity()  # for diagnostics
+
+        self.sigmoid = nn.Sigmoid()
+
+        self.activation2 = Identity()  # for diagnostics
+
+        assert kernel_size % 2 == 1
+
+        self.depthwise_conv = (
+            ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size)
+            if causal
+            else nn.Conv1d(
+                in_channels=bottleneck_dim,
+                out_channels=bottleneck_dim,
+                groups=bottleneck_dim,
+                kernel_size=kernel_size,
+                padding=kernel_size // 2,
+            )
+        )
+
+        self.balancer2 = Balancer(
+            bottleneck_dim,
+            channel_dim=1,
+            min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)),
+            max_positive=1.0,
+            min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)),
+            max_abs=10.0,
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.out_proj = ActivationDropoutAndLinear(
+            bottleneck_dim,
+            channels,
+            activation="SwooshR",
+            dropout_p=0.0,
+            initial_scale=0.05,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+        chunk_size: int = -1,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+           src_key_padding_mask: the mask for the src keys per batch (optional):
+               (batch, #time), contains True in masked positions.
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+
+        x = self.in_proj(x)  # (time, batch, 2*channels)
+
+        x, s = x.chunk(2, dim=2)
+        s = self.balancer1(s)
+        s = self.sigmoid(s)
+        x = self.activation1(x)  # identity.
+        x = x * s
+        x = self.activation2(x)  # identity
+
+        # (time, batch, channels)
+
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        if (
+            not torch.jit.is_scripting()
+            and not torch.jit.is_tracing()
+            and chunk_size >= 0
+        ):
+            # Not support exporting a model for simulated streaming decoding
+            assert (
+                self.causal
+            ), "Must initialize model with causal=True if you use chunk_size"
+            x = self.depthwise_conv(x, chunk_size=chunk_size)
+        else:
+            x = self.depthwise_conv(x)
+
+        x = self.balancer2(x)
+        x = x.permute(2, 0, 1)  # (time, batch, channels)
+
+        x = self.whiten(x)  # (time, batch, channels)
+        x = self.out_proj(x)  # (time, batch, channels)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cache: Tensor,
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module in streaming forward mode.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: cached left context for depthwise_conv of shape
+              (#batch, channels, left_pad)
+            src_key_padding_mask: the mask for the src keys per batch (optional):
+              (batch, #time), contains True in masked positions.
+
+        Returns:
+            - Output tensor (#time, batch, channels).
+            - Updated cache (#batch, channels, left_pad)
+        """
+
+        x = self.in_proj(x)  # (time, batch, 2*channels)
+
+        x, s = x.chunk(2, dim=2)
+        s = self.sigmoid(s)
+        x = x * s
+        # (time, batch, channels)
+
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        x, cache = self.depthwise_conv.streaming_forward(x, cache=cache)
+
+        x = x.permute(2, 0, 1)  # (time, batch, channels)
+
+        x = self.out_proj(x)  # (time, batch, channels)
+
+        return x, cache
+
+
+class ScalarMultiply(nn.Module):
+    def __init__(self, scale: float):
+        super().__init__()
+        self.scale = scale
+
+    def forward(self, x):
+        return x * self.scale
+
+
+def _test_zipformer_main(causal: bool = False):
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+
+    c = Zipformer2(
+        encoder_dim=(64, 96),
+        encoder_unmasked_dim=(48, 64),
+        num_heads=(4, 4),
+        causal=causal,
+        chunk_size=(4,) if causal else (-1,),
+        left_context_frames=(64,),
+    )
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(seq_len, batch_size, 64),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f[0].sum().backward()
+    c.eval()
+    f = c(
+        torch.randn(seq_len, batch_size, 64),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_zipformer_main(False)
+    _test_zipformer_main(True)
diff --git a/egs/librispeech/ASR/zipformer_ctc/__init__.py b/egs/librispeech/ASR/zipformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/zipformer_ctc/asr_datamodule.py b/egs/librispeech/ASR/zipformer_ctc/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/decode.py b/egs/librispeech/ASR/zipformer_ctc/decode.py
new file mode 100755
index 000000000..7f605e2c8
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/decode.py
@@ -0,0 +1,886 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_ctc_model, get_params
+from transformer import encoder_padding_mask
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=77,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    nnet_output, _ = model.encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(nnet_output)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+
+    nnet_output = nnet_output.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    mask = encoder_padding_mask(nnet_output.size(0), supervisions)
+    mask = mask.to(nnet_output.device) if mask is not None else None
+    mmodel = model.decoder.module if hasattr(model.decoder, "module") else model.decoder
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=mmodel,
+            memory=nnet_output,
+            memory_key_padding_mask=mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=mmodel,
+            memory=nnet_output,
+            memory_key_padding_mask=mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[int], List[int]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        if params.rnn_lm_avg == 1:
+            load_checkpoint(
+                f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                rnn_lm_model,
+            )
+            rnn_lm_model.to(device)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer_ctc/decoder.py b/egs/librispeech/ASR/zipformer_ctc/decoder.py
new file mode 100644
index 000000000..8dec048a1
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/decoder.py
@@ -0,0 +1,298 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from label_smoothing import LabelSmoothingLoss
+from torch.nn.utils.rnn import pad_sequence
+from transformer import PositionalEncoding, TransformerDecoderLayer
+
+
+class Decoder(nn.Module):
+    """This class implements Transformer based decoder for an attention-based encoder-decoder
+    model.
+    """
+
+    def __init__(
+        self,
+        num_layers: int,
+        num_classes: int,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+    ):
+        """
+        Args:
+          num_layers:
+            Number of layers.
+          num_classes:
+            Number of tokens of the modeling unit including blank.
+          d_model:
+            Dimension of the input embedding, and of the decoder output.
+        """
+        super().__init__()
+
+        if num_layers > 0:
+            self.decoder_num_class = num_classes  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    @torch.jit.export
+    def forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+    For instance, if sz is 3, it returns::
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+    Args:
+      sz: mask size
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/librispeech/ASR/zipformer_ctc/encoder_interface.py b/egs/librispeech/ASR/zipformer_ctc/encoder_interface.py
new file mode 120000
index 000000000..b8529e0b7
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/export.py b/egs/librispeech/ASR/zipformer_ctc/export.py
new file mode 100755
index 000000000..4c46aea2c
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/export.py
@@ -0,0 +1,237 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_ctc/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = get_ctc_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        convert_scaled_to_non_scaled(model, inplace=True)
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer_ctc/label_smoothing.py b/egs/librispeech/ASR/zipformer_ctc/label_smoothing.py
new file mode 120000
index 000000000..08734abd7
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/label_smoothing.py
@@ -0,0 +1 @@
+../conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/model.py b/egs/librispeech/ASR/zipformer_ctc/model.py
new file mode 100644
index 000000000..2aeb8a072
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/model.py
@@ -0,0 +1,158 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from transformer import encoder_padding_mask
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.utils import encode_supervisions
+
+
+class CTCModel(nn.Module):
+    """It implements a CTC model with an auxiliary attention head."""
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        encoder_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            An instance of `EncoderInterface`. The shared encoder for the CTC and attention
+            branches
+          decoder:
+            An instance of `nn.Module`. This is the decoder for the attention branch.
+          encoder_dim:
+            Dimension of the encoder output.
+          decoder_dim:
+            Dimension of the decoder output.
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder = encoder
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+        self.decoder = decoder
+
+    @torch.jit.ignore
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        supervisions: torch.Tensor,
+        graph_compiler: BpeCtcTrainingGraphCompiler,
+        subsampling_factor: int = 1,
+        beam_size: int = 10,
+        reduction: str = "sum",
+        use_double_scores: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Tensor of dimension (N, T, C) where N is the batch size,
+            T is the number of frames, and C is the feature dimension.
+          x_lens:
+            Tensor of dimension (N,) where N is the batch size.
+          supervisions:
+            Supervisions are used in training.
+          graph_compiler:
+            It is used to compile a decoding graph from texts.
+          subsampling_factor:
+            It is used to compute the `supervisions` for the encoder.
+          beam_size:
+            Beam size used in `k2.ctc_loss`.
+          reduction:
+            Reduction method used in `k2.ctc_loss`.
+          use_double_scores:
+            If True, use double precision in `k2.ctc_loss`.
+        Returns:
+          Return the CTC loss, attention loss, and the total number of frames.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+
+        nnet_output, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(nnet_output)
+
+        # NOTE: We need `encode_supervisions` to sort sequences with
+        # different duration in decreasing order, required by
+        # `k2.intersect_dense` called in `k2.ctc_loss`
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=subsampling_factor
+        )
+        num_frames = supervision_segments[:, 2].sum().item()
+
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            ctc_output,
+            supervision_segments.cpu(),
+            allow_truncate=subsampling_factor - 1,
+        )
+
+        ctc_loss = k2.ctc_loss(
+            decoding_graph=decoding_graph,
+            dense_fsa_vec=dense_fsa_vec,
+            output_beam=beam_size,
+            reduction=reduction,
+            use_double_scores=use_double_scores,
+        )
+
+        if self.decoder is not None:
+            nnet_output = nnet_output.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+            mmodel = (
+                self.decoder.module if hasattr(self.decoder, "module") else self.decoder
+            )
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            mask = encoder_padding_mask(nnet_output.size(0), supervisions)
+            mask = mask.to(nnet_output.device) if mask is not None else None
+            att_loss = mmodel.forward(
+                nnet_output,
+                mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        else:
+            att_loss = torch.tensor([0])
+
+        return ctc_loss, att_loss, num_frames
diff --git a/egs/librispeech/ASR/zipformer_ctc/optim.py b/egs/librispeech/ASR/zipformer_ctc/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/scaling.py b/egs/librispeech/ASR/zipformer_ctc/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/scaling_converter.py b/egs/librispeech/ASR/zipformer_ctc/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/subsampling.py b/egs/librispeech/ASR/zipformer_ctc/subsampling.py
new file mode 120000
index 000000000..6fee09e58
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/subsampling.py
@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/train.py b/egs/librispeech/ASR/zipformer_ctc/train.py
new file mode 100755
index 000000000..60990456d
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/train.py
@@ -0,0 +1,1135 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./zipformer_ctc/train.py \
+     --exp-dir ./zipformer_ctc/exp \
+     --world-size 4 \
+     --full-libri 1 \
+     --max-duration 500 \
+     --num-epochs 30
+"""
+
+import argparse
+import copy
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decoder import Decoder
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import CTCModel
+from optim import Eden, LRScheduler, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Normalization for the input features, can be a
+                              boolean indicating whether to do batch
+                              normalization, or a float which means just scaling
+                              the input features with this float value.
+                              If given a float value, we will remove batchnorm
+                              layer in `ConvolutionModule` as well.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        num_layers=params.num_decoder_layers,
+        num_classes=params.vocab_size,
+        d_model=int(params.encoder_dims.split(",")[-1]),
+    )
+    return decoder
+
+
+def get_ctc_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+
+    model = CTCModel(
+        encoder=encoder,
+        decoder=decoder,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+    with torch.set_grad_enabled(is_training):
+        ctc_loss, att_loss, tot_frames = model(
+            feature,
+            feature_lens,
+            supervisions,
+            graph_compiler,
+            subsampling_factor=params.subsampling_factor,
+            beam_size=params.beam_size,
+            reduction=params.reduction,
+            use_double_scores=params.use_double_scores,
+        )
+
+    info = MetricsTracker()
+    info["frames"] = tot_frames
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    assert loss.requires_grad == is_training, f"{loss.requires_grad} != {is_training}"
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    batch=batch,
+                    graph_compiler=graph_compiler,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    logging.info("About to create model")
+
+    model = get_ctc_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 25.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer_ctc/transformer.py b/egs/librispeech/ASR/zipformer_ctc/transformer.py
new file mode 120000
index 000000000..4c890cf29
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/transformer.py
@@ -0,0 +1 @@
+../conformer_ctc/transformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_ctc/zipformer.py b/egs/librispeech/ASR/zipformer_ctc/zipformer.py
new file mode 120000
index 000000000..79b076556
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_ctc/zipformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_mmi/README.md b/egs/librispeech/ASR/zipformer_mmi/README.md
new file mode 100644
index 000000000..e9a37a52a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/README.md
@@ -0,0 +1,26 @@
+This recipe implements Zipformer-MMI model.
+
+See https://k2-fsa.github.io/icefall/recipes/Non-streaming-ASR/librispeech/zipformer_mmi.html for detailed tutorials.
+
+It uses **CTC loss for warm-up** and then switches to MMI loss during training.
+
+For decoding, it uses HP (H is ctc_topo, P is token-level bi-gram) as decoding graph. Supported decoding methods are:
+- **1best**. Extract the best path from the decoding lattice as the decoding result.
+- **nbest**. Extract n paths from the decoding lattice; the path with the highest score is the decoding result.
+- **nbest-rescoring-LG**. Extract n paths from the decoding lattice, rescore them with an word-level 3-gram LM, the path with the highest score is the decoding result.
+- **nbest-rescoring-3-gram**. Extract n paths from the decoding lattice, rescore them with an token-level 3-gram LM, the path with the highest score is the decoding result.
+- **nbest-rescoring-4-gram**. Extract n paths from the decoding lattice, rescore them with an token-level 4-gram LM, the path with the highest score is the decoding result.
+
+Experimental results training on train-clean-100 (epoch-30-avg-10):
+- 1best. 6.43 & 17.44
+- nbest, nbest-scale=1.2, 6.43 & 17.45
+- nbest-rescoring-LG, nbest-scale=1.2, 5.87 & 16.35
+- nbest-rescoring-3-gram,  nbest-scale=1.2, 6.19 & 16.57
+- nbest-rescoring-4-gram,  nbest-scale=1.2, 5.87 & 16.07
+
+Experimental results training on full librispeech (epoch-30-avg-10):
+- 1best. 2.54 & 5.65
+- nbest, nbest-scale=1.2, 2.54 & 5.66
+- nbest-rescoring-LG, nbest-scale=1.2, 2.49 & 5.42
+- nbest-rescoring-3-gram,  nbest-scale=1.2, 2.52 & 5.62
+- nbest-rescoring-4-gram,  nbest-scale=1.2, 2.5 & 5.51
diff --git a/egs/librispeech/ASR/zipformer_mmi/__init__.py b/egs/librispeech/ASR/zipformer_mmi/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/librispeech/ASR/zipformer_mmi/asr_datamodule.py b/egs/librispeech/ASR/zipformer_mmi/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_mmi/decode.py b/egs/librispeech/ASR/zipformer_mmi/decode.py
new file mode 100755
index 000000000..33c0bf199
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/decode.py
@@ -0,0 +1,730 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) 1best
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --decoding-method 1best
+(2) nbest
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest
+(3) nbest-rescoring-LG
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest-rescoring-LG
+(4) nbest-rescoring-3-gram
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest-rescoring-3-gram
+(5) nbest-rescoring-4-gram
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest-rescoring-4-gram
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_rescore_with_LM,
+    one_best_decoding,
+)
+from icefall.lexicon import Lexicon
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_mmi/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="1best",
+        help="""Decoding method. Use HP as decoding graph, where H is
+        ctc_topo and P is token-level bi-gram lm.
+        Supported values are:
+        - (1) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (2) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring-LG. Extract n paths from the decoding lattice,
+          rescore them with an word-level 3-gram LM, the path with the
+          highest score is the decoding result.
+        - (5) nbest-rescoring-3-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 3-gram LM, the path with
+          the highest score is the decoding result.
+        - (6) nbest-rescoring-4-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 4-gram LM, the path with
+          the highest score is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hp-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `ctc_topo_P.scores`.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HP: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    G: Optional[k2.Fsa] = None,
+    LG: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring-LG", it uses nbest rescoring with word-level 3-gram LM.
+        - params.decoding_method is "nbest-rescoring-3-gram", it uses nbest rescoring with token-level 3-gram LM.
+        - params.decoding_method is "nbest-rescoring-4-gram", it uses nbest rescoring with token-level 4-gram LM.
+
+      model:
+        The neural model.
+      HP:
+        The decoding graph. H is ctc_topo, P is token-level bi-gram LM.
+      bpe_model:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      LG:
+        An LM. L is the lexicon, G is a word-level 3-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-LG".
+      G:
+        An LM. L is the lexicon, G is a token-level 3-gram or 4-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-3-gram"
+        or "nbest-rescoring-4-gram".
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    device = HP.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3, feature.shape
+    feature = feature.to(device)
+
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    nnet_output, encoder_out_lens = model(x=feature, x_lens=feature_lens)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HP,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    method = params.decoding_method
+
+    if method in ["1best", "nbest"]:
+        if method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using HP, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        return {key: hyps}
+
+    assert method in [
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if method == "nbest-rescoring-LG":
+        assert LG is not None
+        LM = LG
+    else:
+        assert G is not None
+        LM = G
+    best_path_dict = nbest_rescore_with_LM(
+        lattice=lattice,
+        LM=LM,
+        num_paths=params.num_paths,
+        lm_scale_list=lm_scale_list,
+        nbest_scale=params.nbest_scale,
+    )
+
+    ans = dict()
+    suffix = f"-nbest-scale-{params.nbest_scale}-{params.num_paths}"
+    for lm_scale_str, best_path in best_path_dict.items():
+        token_ids = get_texts(best_path)
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        ans[lm_scale_str + suffix] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HP: k2.Fsa,
+    bpe_model: spm.SentencePieceProcessor,
+    G: Optional[k2.Fsa] = None,
+    LG: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HP:
+        The decoding graph. H is ctc_topo, P is token-level bi-gram LM.
+      bpe_model:
+        The BPE model.
+      LG:
+        An LM. L is the lexicon, G is a word-level 3-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-LG".
+      G:
+        An LM. L is the lexicon, G is a token-level 3-gram or 4-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-3-gram"
+        or "nbest-rescoring-4-gram".
+
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HP=HP,
+            bpe_model=bpe_model,
+            batch=batch,
+            G=G,
+            LG=LG,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "1best",
+        "nbest",
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    ), params.decoding_method
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    HP = mmi_graph_compiler.ctc_topo_P
+    HP.scores *= params.hp_scale
+    if not hasattr(HP, "lm_scores"):
+        HP.lm_scores = HP.scores.clone()
+
+    LG = None
+    G = None
+
+    if params.decoding_method == "nbest-rescoring-LG":
+        lg_filename = params.lang_dir / "LG.pt"
+        logging.info(f"Loading {lg_filename}")
+        LG = k2.Fsa.from_dict(torch.load(lg_filename, map_location=device))
+        LG = k2.Fsa.from_fsas([LG]).to(device)
+        LG.lm_scores = LG.scores.clone()
+
+    elif params.decoding_method in ["nbest-rescoring-3-gram", "nbest-rescoring-4-gram"]:
+        order = params.decoding_method[-6]
+        assert order in ("3", "4"), (params.decoding_method, order)
+        order = int(order)
+        if not (params.lang_dir / f"{order}gram.pt").is_file():
+            logging.info(f"Loading {order}gram.fst.txt")
+            logging.warning("It may take a few minutes.")
+            with open(params.lang_dir / f"{order}gram.fst.txt") as f:
+                first_token_disambig_id = lexicon.token_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_token_disambig_id] = 0
+                G = k2.Fsa.from_fsas([G]).to(device)
+                # G = k2.remove_epsilon(G)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lang_dir / f"{order}gram.pt")
+        else:
+            logging.info(f"Loading pre-compiled {order}gram.pt")
+            d = torch.load(params.lang_dir / f"{order}gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        G.lm_scores = G.scores.clone()
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HP=HP,
+            bpe_model=bpe_model,
+            G=G,
+            LG=LG,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer_mmi/encoder_interface.py b/egs/librispeech/ASR/zipformer_mmi/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_mmi/export.py b/egs/librispeech/ASR/zipformer_mmi/export.py
new file mode 100755
index 000000000..1aec56420
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/export.py
@@ -0,0 +1,309 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `zipformer_mmi/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./zipformer_mmi/decode.py \
+        --exp-dir ./zipformer_mmi/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+    # You will find the pre-trained model in icefall-asr-librispeech-zipformer-mmi-2022-12-08/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_mmi/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt.",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer_mmi/jit_pretrained.py b/egs/librispeech/ASR/zipformer_mmi/jit_pretrained.py
new file mode 100755
index 000000000..c9ef16ffa
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/jit_pretrained.py
@@ -0,0 +1,391 @@
+#!/usr/bin/env python3
+# Copyright      2021-2022  Xiaomi Corp.   (authors: Fangjun Kuang,
+#                                                    Zengwei)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+(1) 1best
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(2) nbest
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(3) nbest-rescoring-LG
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-LG \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(4) nbest-rescoring-3-gram
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-3-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(5) nbest-rescoring-4-gram
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-4-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_rescore_with_LM,
+    one_best_decoding,
+)
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method. Use HP as decoding graph, where H is
+        ctc_topo and P is token-level bi-gram lm.
+        Supported values are:
+        - (1) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (2) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring-LG. Extract n paths from the decoding lattice,
+          rescore them with an word-level 3-gram LM, the path with the
+          highest score is the decoding result.
+        - (5) nbest-rescoring-3-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 3-gram LM, the path with
+          the highest score is the decoding result.
+        - (6) nbest-rescoring-4-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 4-gram LM, the path with
+          the highest score is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.2,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.1,
+        help="""
+        Used when method is nbest-rescoring-LG, nbest-rescoring-3-gram,
+        and nbest-rescoring-4-gram.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--hp-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `ctc_topo_P.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(params.nn_model_filename)
+    model.eval()
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    HP = mmi_graph_compiler.ctc_topo_P
+    HP.scores *= params.hp_scale
+    if not hasattr(HP, "lm_scores"):
+        HP.lm_scores = HP.scores.clone()
+
+    method = params.method
+    assert method in (
+        "1best",
+        "nbest",
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    )
+    # loading language model for rescoring
+    LM = None
+    if method == "nbest-rescoring-LG":
+        lg_filename = params.lang_dir / "LG.pt"
+        logging.info(f"Loading {lg_filename}")
+        LG = k2.Fsa.from_dict(torch.load(lg_filename, map_location=device))
+        LG = k2.Fsa.from_fsas([LG]).to(device)
+        LG.lm_scores = LG.scores.clone()
+        LM = LG
+    elif method in ["nbest-rescoring-3-gram", "nbest-rescoring-4-gram"]:
+        order = method[-6]
+        assert order in ("3", "4")
+        order = int(order)
+        logging.info(f"Loading pre-compiled {order}gram.pt")
+        d = torch.load(params.lang_dir / f"{order}gram.pt", map_location=device)
+        G = k2.Fsa.from_dict(d)
+        G.lm_scores = G.scores.clone()
+        LM = G
+
+    # Encoder forward
+    nnet_output, encoder_out_lens = model(x=features, x_lens=feature_lengths)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HP,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if method in ["1best", "nbest"]:
+        if method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+    else:
+        best_path_dict = nbest_rescore_with_LM(
+            lattice=lattice,
+            LM=LM,
+            num_paths=params.num_paths,
+            lm_scale_list=[params.ngram_lm_scale],
+            nbest_scale=params.nbest_scale,
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    # Note: `best_path.aux_labels` contains token IDs, not word IDs
+    # since we are using HP, not HLG here.
+    #
+    # token_ids is a lit-of-list of IDs
+    token_ids = get_texts(best_path)
+    # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+    hyps = bpe_model.decode(token_ids)
+    # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+    hyps = [s.split() for s in hyps]
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer_mmi/model.py b/egs/librispeech/ASR/zipformer_mmi/model.py
new file mode 100644
index 000000000..4045c8b64
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/model.py
@@ -0,0 +1,75 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+
+class CTCModel(nn.Module):
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        encoder_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder = encoder
+
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+        Returns:
+          Return the ctc outputs and encoder output lengths.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(encoder_out)
+
+        return ctc_output, x_lens
diff --git a/egs/librispeech/ASR/zipformer_mmi/optim.py b/egs/librispeech/ASR/zipformer_mmi/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_mmi/pretrained.py b/egs/librispeech/ASR/zipformer_mmi/pretrained.py
new file mode 100755
index 000000000..3ba4da5dd
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/pretrained.py
@@ -0,0 +1,411 @@
+#!/usr/bin/env python3
+# Copyright      2021-2022  Xiaomi Corp.   (authors: Fangjun Kuang,
+#                                                    Zengwei)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) 1best
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(2) nbest
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --nbest-scale 1.2 \
+    --method nbest \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(3) nbest-rescoring-LG
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-LG \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(4) nbest-rescoring-3-gram
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-3-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(5) nbest-rescoring-4-gram
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --tokens data/lang_bpe_500/tokens.txt \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-4-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+You can also use `./zipformer_mmi/exp/epoch-xx.pt`.
+
+Note: ./zipformer_mmi/exp/pretrained.pt is generated by
+./zipformer_mmi/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_rescore_with_LM,
+    one_best_decoding,
+)
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import get_texts, num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method. Use HP as decoding graph, where H is
+        ctc_topo and P is token-level bi-gram lm.
+        Supported values are:
+        - (1) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (2) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring-LG. Extract n paths from the decoding lattice,
+          rescore them with an word-level 3-gram LM, the path with the
+          highest score is the decoding result.
+        - (5) nbest-rescoring-3-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 3-gram LM, the path with
+          the highest score is the decoding result.
+        - (6) nbest-rescoring-4-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 4-gram LM, the path with
+          the highest score is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.2,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.1,
+        help="""
+        Used when method is nbest-rescoring-LG, nbest-rescoring-3-gram,
+        and nbest-rescoring-4-gram.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--hp-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `ctc_topo_P.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    # Load id of the <blk> token and the vocab size
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    HP = mmi_graph_compiler.ctc_topo_P
+    HP.scores *= params.hp_scale
+    if not hasattr(HP, "lm_scores"):
+        HP.lm_scores = HP.scores.clone()
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    method = params.method
+    assert method in (
+        "1best",
+        "nbest",
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    )
+    # loading language model for rescoring
+    LM = None
+    if method == "nbest-rescoring-LG":
+        lg_filename = params.lang_dir / "LG.pt"
+        logging.info(f"Loading {lg_filename}")
+        LG = k2.Fsa.from_dict(torch.load(lg_filename, map_location=device))
+        LG = k2.Fsa.from_fsas([LG]).to(device)
+        LG.lm_scores = LG.scores.clone()
+        LM = LG
+    elif method in ["nbest-rescoring-3-gram", "nbest-rescoring-4-gram"]:
+        order = method[-6]
+        assert order in ("3", "4")
+        order = int(order)
+        logging.info(f"Loading pre-compiled {order}gram.pt")
+        d = torch.load(params.lang_dir / f"{order}gram.pt", map_location=device)
+        G = k2.Fsa.from_dict(d)
+        G.lm_scores = G.scores.clone()
+        LM = G
+
+    # Encoder forward
+    nnet_output, encoder_out_lens = model(x=features, x_lens=feature_lengths)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HP,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if method in ["1best", "nbest"]:
+        if method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+    else:
+        best_path_dict = nbest_rescore_with_LM(
+            lattice=lattice,
+            LM=LM,
+            num_paths=params.num_paths,
+            lm_scale_list=[params.ngram_lm_scale],
+            nbest_scale=params.nbest_scale,
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    # Note: `best_path.aux_labels` contains token IDs, not word IDs
+    # since we are using HP, not HLG here.
+    #
+    # token_ids is a lit-of-list of IDs
+    token_ids = get_texts(best_path)
+    hyps = [token_ids_to_words(ids) for ids in token_ids]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer_mmi/scaling.py b/egs/librispeech/ASR/zipformer_mmi/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_mmi/scaling_converter.py b/egs/librispeech/ASR/zipformer_mmi/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/zipformer_mmi/test_model.py b/egs/librispeech/ASR/zipformer_mmi/test_model.py
new file mode 100755
index 000000000..7782845f4
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/test_model.py
@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./zipformer_mmi/test_model.py
+"""
+
+import torch
+from train import get_ctc_model, get_params
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.num_encoder_layers = "2,4,3,2,4"
+    #  params.feedforward_dims = "1024,1024,1536,1536,1024"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    features = torch.randn(2, 100, 80)
+    feature_lengths = torch.full((2,), 100)
+    model(x=features, x_lens=feature_lengths)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer_mmi/train.py b/egs/librispeech/ASR/zipformer_mmi/train.py
new file mode 100755
index 000000000..4b50acdde
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/train.py
@@ -0,0 +1,1187 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_mmi/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir zipformer_mmi/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./zipformer_mmi/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_mmi/exp \
+  --full-libri 1 \
+  --max-duration 500
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import CTCModel
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon, UniqLexicon
+from icefall.mmi import LFMMILoss
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_mmi/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--use-pruned-intersect",
+        type=str2bool,
+        default=False,
+        help="""Whether to use `intersect_dense_pruned` to get denominator
+        lattice.""",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            # parameters for mmi loss
+            "mmi_beam_size": 6,
+            "den_scale": 1.0,
+            # parameters for mmi loss
+            "ctc_beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_ctc_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+
+    model = CTCModel(
+        encoder=encoder,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    ctc_graph_compiler: BpeCtcTrainingGraphCompiler,
+    mmi_graph_compiler: MmiTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute ctc loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_out_lens = model(x=feature, x_lens=feature_lens)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `LFMMILoss.forward()`
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    info = MetricsTracker()
+    if batch_idx_train < warm_step:
+        # Training with ctc loss
+        # Works with a BPE model
+        token_ids = ctc_graph_compiler.texts_to_ids(texts)
+        decoding_graph = ctc_graph_compiler.compile(token_ids)
+        loss = k2.ctc_loss(
+            decoding_graph=decoding_graph,
+            dense_fsa_vec=dense_fsa_vec,
+            output_beam=params.ctc_beam_size,
+            reduction=params.reduction,
+            use_double_scores=params.use_double_scores,
+        )
+        info["ctc_loss"] = loss.detach().cpu().item()
+        info["mmi_loss"] = 0
+    else:
+        # Training with mmi loss
+        loss_fn = LFMMILoss(
+            graph_compiler=mmi_graph_compiler,
+            use_pruned_intersect=params.use_pruned_intersect,
+            den_scale=params.den_scale,
+            beam_size=params.mmi_beam_size,
+        )
+        loss = loss_fn(dense_fsa_vec=dense_fsa_vec, texts=texts)
+        info["ctc_loss"] = 0
+        info["mmi_loss"] = loss.detach().cpu().item()
+
+    assert loss.requires_grad == is_training
+
+    info["frames"] = encoder_out_lens.sum().cpu().item()
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    ctc_graph_compiler: BpeCtcTrainingGraphCompiler,
+    mmi_graph_compiler: MmiTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            ctc_graph_compiler=ctc_graph_compiler,
+            mmi_graph_compiler=mmi_graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    ctc_graph_compiler: BpeCtcTrainingGraphCompiler,
+    mmi_graph_compiler: MmiTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    ctc_graph_compiler=ctc_graph_compiler,
+                    mmi_graph_compiler=mmi_graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(
+                batch, params=params, graph_compiler=mmi_graph_compiler
+            )
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                ctc_graph_compiler=ctc_graph_compiler,
+                mmi_graph_compiler=mmi_graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    assert "lang_bpe" in str(params.lang_dir)
+    ctc_graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    # train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        # train_cuts += librispeech.train_clean_360_cuts()
+        # train_cuts += librispeech.train_other_500_cuts()
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            ctc_graph_compiler=ctc_graph_compiler,
+            mmi_graph_compiler=mmi_graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            ctc_graph_compiler=ctc_graph_compiler,
+            mmi_graph_compiler=mmi_graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: MmiTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    ctc_graph_compiler: BpeCtcTrainingGraphCompiler,
+    mmi_graph_compiler: MmiTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    ctc_graph_compiler=ctc_graph_compiler,
+                    mmi_graph_compiler=mmi_graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(
+                batch, params=params, graph_compiler=mmi_graph_compiler
+            )
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/ASR/zipformer_mmi/zipformer.py b/egs/librispeech/ASR/zipformer_mmi/zipformer.py
new file mode 120000
index 000000000..79b076556
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer_mmi/zipformer.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/README.md b/egs/librispeech/WSASR/README.md
new file mode 100644
index 000000000..3b8822fd2
--- /dev/null
+++ b/egs/librispeech/WSASR/README.md
@@ -0,0 +1,224 @@
+# Introduction
+
+This is a weakly supervised ASR recipe for the LibriSpeech (clean 100 hours) dataset. We train a
+conformer model using [Bypass Temporal Classification](https://arxiv.org/pdf/2306.01031.pdf) (BTC)/[Omni-temporal Classification](https://arxiv.org/pdf/2309.15796.pdf) (OTC) with transcripts with synthetic errors. In this README, we will describe
+the task and the BTC/OTC training process.
+
+Note that OTC is an extension of BTC and supports all BTC functions. Therefore, in the following, we only describe OTC.
+## Task
+We propose BTC/OTC to directly train an ASR system leveraging weak supervision, i.e., speech with non-verbatim transcripts. This is achieved by using a special token $\star$ to model uncertainties (i.e., substitution errors, insertion errors, and deletion errors) 
+within the WFST framework during training.
+
+
+<div style="display: flex;flex; justify-content: space-between">
+  <figure style="flex: 2; text-align: center; margin: 5px;">
+    <img src="figures/sub.png" alt="Image 1" width="25%" />
+
+  </figure>
+  <figure style="flex: 2; text-align: center; margin: 5px;">
+    <img src="figures/ins.png" alt="Image 2" width="25%" />
+
+  </figure>
+  <figure style="flex: 2; text-align: center;margin: 5px;">
+    <img src="figures/del.png" alt="Image 3" width="25%" />
+
+  </figure>
+</div>
+<figcaption> Examples of errors (substitution, insertion, and deletion) in the transcript. The grey box is the verbatim transcript and the red box is the inaccurate transcript. Inaccurate words are marked in bold.</figcaption> <br><br>
+
+
+We modify $G(\mathbf{y})$ by adding self-loop arcs into each state and bypass arcs into each arc. 
+  <p align="center">
+    <img src="figures/otc_g.png" alt="Image Alt Text" width="50%" />
+
+  </p>
+
+We incorporate the penalty strategy and apply different configurations for the self-loop arc and bypass arc. The penalties are set as
+
+$$\lambda_{1_{i}} = \beta_{1} * \tau_{1}^{i},\quad \lambda_{2_{i}} = \beta_{2} * \tau_{2}^{i}$$
+
+for the $i$-th training epoch. $\beta$ is the initial penalty that encourages the model to rely more on the given transcript at the start of training. 
+It decays exponentially by a factor of $\tau \in (0, 1)$, gradually encouraging the model to align speech with $\star$ when getting confused. 
+
+After composing the modified WFST $G_{\text{otc}}(\mathbf{y})$ with $L$ and $T$, the OTC training graph is shown in this figure:
+<figure style="text-align: center">
+  <img src="figures/otc_training_graph.drawio.png" alt="Image Alt Text" />
+  <figcaption>OTC training graph. The self-loop arcs and bypass arcs are highlighted in green and blue, respectively.</figcaption>
+</figure>
+
+The $\star$ is represented as the average probability of all non-blank tokens.
+  <p align="center">
+    <img src="figures/otc_emission.drawio.png" width="50%" />
+  </p>
+
+The weight of $\star$ is the log average probability of "a" and "b": $\log \frac{e^{-1.2} + e^{-2.3}}{2} = -1.6$ and $\log \frac{e^{-1.9} + e^{-0.5}}{2} = -1.0$ for 2 frames.
+
+## Description of the recipe
+### Preparation
+```
+# feature_type can be ssl or fbank
+feature_type=ssl
+feature_dir="data/${feature_type}"
+manifest_dir="${feature_dir}"
+lang_dir="data/lang"
+lm_dir="data/lm"
+exp_dir="conformer_ctc2/exp"
+otc_token="<star>"
+
+./prepare.sh \
+  --feature-type "${feature_type}" \
+  --feature-dir "${feature_dir}" \
+  --lang-dir "${lang_dir}" \
+  --lm-dir "${lm_dir}" \
+  --otc-token "${otc_token}" 
+```
+This script adds the 'otc_token' ('\<star\>') and its corresponding sentence-piece ('▁\<star\>') to 'words.txt' and 'tokens.txt,' respectively. Additionally, it computes SSL features using the 'wav2vec2-base' model. (You can use GPU to accelerate feature extraction).
+
+### Making synthetic errors to the transcript (train-clean-100) [optional]
+```
+sub_er=0.17
+ins_er=0.17
+del_er=0.17
+synthetic_train_manifest="librispeech_cuts_train-clean-100_${sub_er}_${ins_er}_${del_er}.jsonl.gz"
+
+./local/make_error_cutset.py \
+  --input-cutset "${manifest_dir}/librispeech_cuts_train-clean-100.jsonl.gz" \
+  --words-file "${lang_dir}/words.txt" \
+  --sub-error-rate "${sub_er}" \
+  --ins-error-rate "${ins_er}" \
+  --del-error-rate "${del_er}" \
+  --output-cutset "${manifest_dir}/${synthetic_train_manifest}"
+```
+This script generates synthetic substitution, insertion, and deletion errors in the transcript with ratios 'sub_er', 'ins_er', and 'del_er', respectively. The original transcript is saved as 'verbatim transcript' in the cutset, along with information on how the transcript is corrupted:
+
+  - '[hello]' indicates the original word 'hello' is substituted by another word
+  - '[]' indicates an extra word is inserted into the transcript
+  - '-hello-' indicates the word 'hello' is deleted from the transcript
+  
+So if the original transcript is "have a nice day" and the synthetic one is "a very good day", the 'verbatim transcript' would be:
+```
+original:  have  a      nice  day
+synthetic:       a very good  day
+verbatim: -have- a  [] [nice] day
+```
+
+### Training
+The training uses synthetic data based on the train-clean-100 subset.
+```
+otc_lang_dir=data/lang_bpe_200
+
+allow_bypass_arc=true
+allow_self_loop_arc=true
+initial_bypass_weight=-19
+initial_self_loop_weight=3.75
+bypass_weight_decay=0.975
+self_loop_weight_decay=0.999
+
+show_alignment=true
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./conformer_ctc2/train.py \
+  --world-size 4 \
+  --manifest-dir "${manifest_dir}" \
+  --train-manifest "${synthetic_train_manifest}" \
+  --exp-dir "${exp_dir}" \
+  --lang-dir "${otc_lang_dir}" \
+  --otc-token "${otc_token}" \
+  --allow-bypass-arc "${allow_bypass_arc}" \
+  --allow-self-loop-arc "${allow_self_loop_arc}" \
+  --initial-bypass-weight "${initial_bypass_weight}" \
+  --initial-self-loop-weight "${initial_self_loop_weight}" \
+  --bypass-weight-decay "${bypass_weight_decay}" \
+  --self-loop-weight-decay "${self_loop_weight_decay}" \
+  --show-alignment "${show_alignment}"
+```
+The bypass arc deals with substitution and insertion errors, while the self-loop arc deals with deletion errors. Using "--show-alignment" would print the best alignment during training, which is very helpful for tuning hyperparameters and debugging.
+
+### Decoding
+```
+export CUDA_VISIBLE_DEVICES="0"
+./conformer_ctc2/decode.py \
+  --manifest-dir "${manifest_dir}" \
+  --exp-dir "${exp_dir}" \
+  --lang-dir "${otc_lang_dir}" \
+  --lm-dir "${lm_dir}" \
+  --otc-token "${otc_token}"
+```
+
+### Results (ctc-greedy-search)
+<table>
+  <tr>
+    <td rowspan=2>Training Criterion</td>
+    <td colspan=2>ssl</td>
+    <td colspan=2>fbank</td>
+  </tr>
+  <tr>
+    <td>test-clean</td>
+    <td>test-other</td>
+    <td>test-clean</td>
+    <td>test-other</td>
+  </tr>
+  <tr>
+    <td>CTC</td>
+    <td>100.0</td>
+    <td>100.0</td>
+    <td>99.89</td>
+    <td>99.98</td>
+  </tr>
+  <tr>
+    <td>OTC</td>
+    <td>11.89</td>
+    <td>25.46</td>
+    <td>20.14</td>
+    <td>44.24</td>
+  </tr>
+</table>
+
+### Results (1best, blank_bias=-4)
+<table>
+  <tr>
+    <td rowspan=2>Training Criterion</td>
+    <td colspan=2>ssl</td>
+    <td colspan=2>fbank</td>
+  </tr>
+  <tr>
+    <td>test-clean</td>
+    <td>test-other</td>
+    <td>test-clean</td>
+    <td>test-other</td>
+  </tr>
+  <tr>
+    <td>CTC</td>
+    <td>98.40</td>
+    <td>98.68</td>
+    <td>99.79</td>
+    <td>99.86</td>
+  </tr>
+  <tr>
+    <td>OTC</td>
+    <td>6.59</td>
+    <td>15.98</td>
+    <td>11.78</td>
+    <td>32.38</td>
+  </tr>
+</table>
+
+## Pre-trained Model
+Pre-trained model: <https://huggingface.co/dgao/icefall-otc-librispeech-conformer-ctc2>
+
+## Citations
+```
+@inproceedings{gao2023bypass,
+  title={Bypass Temporal Classification: Weakly Supervised Automatic Speech Recognition with Imperfect Transcripts},
+  author={Gao, Dongji and Wiesner, Matthew and Xu, Hainan and Garcia, Leibny Paola and Povey, Daniel and Khudanpur, Sanjeev},
+  booktitle={INTERSPEECH},
+  year={2023}
+}
+
+@inproceedings{gao2023learning,
+  title={Learning from Flawed Data: Weakly Supervised Automatic Speech Recognition},
+  author={Gao, Dongji and Xu, Hainan and Raj, Desh and Garcia, Leibny Paola and Povey, Daniel and Khudanpur, Sanjeev},
+  booktitle={IEEE ASRU},
+  year={2023}
+}
+```
diff --git a/egs/librispeech/WSASR/conformer_ctc2/__init__.py b/egs/librispeech/WSASR/conformer_ctc2/__init__.py
new file mode 120000
index 000000000..43a85af20
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/__init__.py
@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py b/egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py
new file mode 100644
index 000000000..2f8e658c5
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py
@@ -0,0 +1,369 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#                2023  John Hopkins University  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import AudioSamples  # noqa F401 For AudioSamples
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--full-libri",
+            type=str2bool,
+            default=False,
+            help="""Used only when --mini-libri is False.When enabled,
+            use 960h LibriSpeech. Otherwise, use 100h subset.""",
+        )
+        group.add_argument(
+            "--mini-libri",
+            type=str2bool,
+            default=False,
+            help="True for mini librispeech",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/ssl"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+        group.add_argument(
+            "--train-manifest",
+            type=str,
+            default="librispeech_cuts_train-clean-100.jsonl.gz",
+            help="Train manifest file.",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+
+        validate = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_clean_5_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get train-clean-5 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-5.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / self.args.train_manifest)
+
+    @lru_cache()
+    def train_all_shuf_cuts(self) -> CutSet:
+        logging.info(
+            "About to get the shuffled train-clean-100, \
+            train-clean-360 and train-other-500 cuts"
+        )
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_2_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get dev-clean-2 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean-2.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
diff --git a/egs/librispeech/WSASR/conformer_ctc2/attention.py b/egs/librispeech/WSASR/conformer_ctc2/attention.py
new file mode 120000
index 000000000..e808a6f20
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/attention.py
@@ -0,0 +1 @@
+../../ASR/conformer_ctc2/attention.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/conformer.py b/egs/librispeech/WSASR/conformer_ctc2/conformer.py
new file mode 100644
index 000000000..db4821d37
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/conformer.py
@@ -0,0 +1,949 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corp.                              (author: Quandong Wang)
+#                2023  Johns Hopkins University                  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling, Conv2dSubsampling2
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 2,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.2,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4 and subsampling_factor != 2:
+            raise NotImplementedError("Support only 'subsampling_factor=4 or 2'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if self.subsampling_factor == 4:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+        elif self.subsampling_factor == 2:
+            self.encoder_embed = Conv2dSubsampling2(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), self.subsampling_factor, supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        x = self.encoder(
+            x, pos_emb, src_key_padding_mask=mask, warmup=warmup
+        )  # (T, N, C)
+
+        # x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        # return x, lengths
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src = src + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/librispeech/WSASR/conformer_ctc2/decode.py b/egs/librispeech/WSASR/conformer_ctc2/decode.py
new file mode 100755
index 000000000..3fa045533
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/decode.py
@@ -0,0 +1,718 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Quandong Wang)
+#           2023 Johns Hopkins University (Author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.otc_graph_compiler import OtcTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        default="<star>",
+        help="OTC token",
+    )
+
+    parser.add_argument(
+        "--blank-bias",
+        type=float,
+        default=0,
+        help="bias (log-prob) added to blank token during decoding",
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="ctc-greedy-search",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) ctc-greedy-search. It only use CTC output and a sentence piece
+              model for decoding. It produces the same results with ctc-decoding.
+            - (2) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+       """,
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=0,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_200",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 2,
+            "feature_dim": 768,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "encoder_dim": 512,
+            "num_encoder_layers": 12,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def ctc_greedy_search(
+    nnet_output: torch.Tensor,
+    memory: torch.Tensor,
+    memory_key_padding_mask: torch.Tensor,
+) -> List[List[int]]:
+    """Apply CTC greedy search
+
+     Args:
+         speech (torch.Tensor): (batch, max_len, feat_dim)
+         speech_length (torch.Tensor): (batch, )
+    Returns:
+         List[List[int]]: best path result
+    """
+    batch_size = memory.shape[1]
+    # Let's assume B = batch_size
+    encoder_out = memory
+    encoder_mask = memory_key_padding_mask
+    maxlen = encoder_out.size(0)
+
+    ctc_probs = nnet_output  # (B, maxlen, vocab_size)
+    topk_prob, topk_index = ctc_probs.topk(1, dim=2)  # (B, maxlen, 1)
+    topk_index = topk_index.view(batch_size, maxlen)  # (B, maxlen)
+    topk_index = topk_index.masked_fill_(encoder_mask, 0)  # (B, maxlen)
+    hyps = [hyp.tolist() for hyp in topk_index]
+    scores = topk_prob.max(1)
+    hyps = [remove_duplicates_and_blank(hyp) for hyp in hyps]
+    return hyps, scores
+
+
+def remove_duplicates_and_blank(hyp: List[int]) -> List[int]:
+    # from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/common.py
+    new_hyp: List[int] = []
+    cur = 0
+    while cur < len(hyp):
+        if hyp[cur] != 0:
+            new_hyp.append(hyp[cur])
+        prev = cur
+        while cur < len(hyp) and hyp[cur] == hyp[prev]:
+            cur += 1
+    return new_hyp
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+    nnet_output[:, :, 0] += params.blank_bias
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor + 2,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "ctc-greedy-search":
+        hyps, _ = ctc_greedy_search(
+            nnet_output,
+            memory,
+            memory_key_padding_mask,
+        )
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(hyps)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-greedy-search"
+        return {key: hyps}
+
+    if params.method in ["1best"]:
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        key = "no_rescore"
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+
+        return {key: hyps}
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+    assert "▁" not in args.otc_token
+    args.otc_token = f"▁{args.otc_token}"
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    # remove otc_token from decoding units
+    max_token_id = max(lexicon.tokens) - 1
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = OtcTrainingGraphCompiler(
+        params.lang_dir,
+        params.otc_token,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding" or params.method == "ctc-greedy-search":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/conformer_ctc2/export.py b/egs/librispeech/WSASR/conformer_ctc2/export.py
new file mode 120000
index 000000000..5f484e391
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/export.py
@@ -0,0 +1 @@
+../../ASR/conformer_ctc2/export.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/label_smoothing.py b/egs/librispeech/WSASR/conformer_ctc2/label_smoothing.py
new file mode 120000
index 000000000..c050ea637
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/label_smoothing.py
@@ -0,0 +1 @@
+../../ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/optim.py b/egs/librispeech/WSASR/conformer_ctc2/optim.py
new file mode 120000
index 000000000..db836b5e0
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/optim.py
@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/scaling.py b/egs/librispeech/WSASR/conformer_ctc2/scaling.py
new file mode 120000
index 000000000..bd0abfeee
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/scaling.py
@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/subsampling.py b/egs/librispeech/WSASR/conformer_ctc2/subsampling.py
new file mode 100644
index 000000000..2ba802866
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/subsampling.py
@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corporation                        (author: Quandong Wang)
+#                2023  Johns Hopkins University                  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+
+class Conv2dSubsampling(torch.nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = torch.nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+class Conv2dSubsampling2(torch.nn.Module):
+    """Convolutional 2D subsampling (to 1/2 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim) where
+    T' = (T - 1) // 2 - 2, which approximates T' == T // 2
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = torch.nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * ((in_channels - 1) // 2 - 2), out_channels
+        )
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.unsqueeze(1)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
diff --git a/egs/librispeech/WSASR/conformer_ctc2/train.py b/egs/librispeech/WSASR/conformer_ctc2/train.py
new file mode 100755
index 000000000..fe6c5af91
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/train.py
@@ -0,0 +1,1115 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,
+#                                                  Zengwei Yao,
+#                                                  Quandong Wang)
+#              2023 Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc2/train.py \
+  --world-size 4 \
+  --manifest-dir data/ssl \
+  --train-manifest librispeech_cuts_train-clean-100_0.17_0.17_0.17.jsonl.gz \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_200 \
+  --otc-token "<star>" \
+  --allow-bypass-arc true \
+  --allow-self-loop-arc true \
+  --initial-bypass-weight -19 \
+  --initial-self-loop-weight 3.75 \
+  --bypass-weight-decay 0.975 \
+  --self-loop-weight-decay 0.999 \
+  --show-alignment true
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.decode import one_best_decoding
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.otc_graph_compiler import OtcTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions_otc,
+    get_texts,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_200",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.0,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=0,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=10,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        default="_<star>",
+        help="OTC token",
+    )
+
+    parser.add_argument(
+        "--allow-bypass-arc",
+        type=str2bool,
+        default=True,
+        help="""Whether to add bypass arc to training graph for substitution
+        and insertion errors (wrong or extra words in the transcript).""",
+    )
+
+    parser.add_argument(
+        "--allow-self-loop-arc",
+        type=str2bool,
+        default=True,
+        help="""Whether to self-loop bypass arc to training graph for deletion errors
+        (missing words in the transcript).""",
+    )
+
+    parser.add_argument(
+        "--initial-bypass-weight",
+        type=float,
+        default=0.0,
+        help="Initial weight associated with bypass arc",
+    )
+
+    parser.add_argument(
+        "--initial-self-loop-weight",
+        type=float,
+        default=0.0,
+        help="Initial weight associated with self-loop arc",
+    )
+
+    parser.add_argument(
+        "--bypass-weight-decay",
+        type=float,
+        default=1.0,
+        help="""Weight decay factor of bypass arc weight:
+        bypass_arc_weight = intial_bypass_weight * bypass_weight_decay ^ ith-epoch""",
+    )
+
+    parser.add_argument(
+        "--self-loop-weight-decay",
+        type=float,
+        default=1.0,
+        help="""Weight decay factor of self-loop arc weight:
+        self_loop_arc_weight = intial_self_loop_weight * self_loop_weight_decay ^ ith-epoch""",
+    )
+
+    parser.add_argument(
+        "--show-alignment",
+        type=str2bool,
+        default=True,
+        help="Whether to print OTC alignment during training",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 800,  # For the 100h subset, use 800
+            "alignment_interval": 25,
+            # parameters for conformer
+            "feature_dim": 768,
+            "subsampling_factor": 2,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for ctc loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    batch: dict,
+    graph_compiler: OtcTrainingGraphCompiler,
+    is_training: bool,
+    warmup: float = 2.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute OTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(
+            feature, supervisions, warmup=warmup
+        )
+        # Set the probability of OTC token as the average of non-blank tokens
+        # under the assumption that blank is the first and
+        # OTC token is the last token in tokens.txt
+        _, _, V = nnet_output.shape
+
+        otc_token_log_prob = torch.logsumexp(
+            nnet_output[:, :, 1:], dim=-1, keepdim=True
+        ) - torch.log(torch.tensor([V - 1])).to(device)
+
+        nnet_output = torch.cat([nnet_output, otc_token_log_prob], dim=-1)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts, utt_ids, verbatim_texts = encode_supervisions_otc(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    bypass_weight = graph_compiler.initial_bypass_weight * (
+        graph_compiler.bypass_weight_decay ** (params.cur_epoch - 1)
+    )
+    self_loop_weight = graph_compiler.initial_self_loop_weight * (
+        graph_compiler.self_loop_weight_decay ** (params.cur_epoch - 1)
+    )
+
+    decoding_graph = graph_compiler.compile(
+        texts=texts,
+        allow_bypass_arc=params.allow_bypass_arc,
+        allow_self_loop_arc=params.allow_self_loop_arc,
+        bypass_weight=bypass_weight,
+        self_loop_weight=self_loop_weight,
+    )
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=3,
+    )
+
+    otc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert params.att_rate == 0.0
+    loss = otc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+    info["otc_loss"] = otc_loss.detach().cpu().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    if params.show_alignment:
+        if params.batch_idx_train % params.alignment_interval == 0:
+            for index, utt_id in enumerate(utt_ids):
+                verbatim_text = verbatim_texts[index]
+                utt_id = utt_ids[index]
+
+                lattice = k2.intersect_dense(
+                    decoding_graph,
+                    dense_fsa_vec,
+                    params.beam_size,
+                )
+                best_path = one_best_decoding(
+                    lattice=lattice,
+                    use_double_scores=params.use_double_scores,
+                )
+                hyp_ids = get_texts(best_path)[index]
+                hyp_text_list = [graph_compiler.token_table[i] for i in hyp_ids]
+                hyp_text = "".join(hyp_text_list).replace("▁", " ")
+
+                logging.info(f"[utterance id]: {utt_id}")
+                logging.info(f"[verbatim text]: {verbatim_text}")
+                logging.info(f"[best alignment]: {hyp_text}")
+                logging.info(bypass_weight)
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: OtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: OtcTrainingGraphCompiler,
+    scheduler: LRSchedulerType,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        # scaler.scale(loss).backward()
+
+        try:
+            # loss.backward()
+            scaler.scale(loss).backward()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(f"failing batch size:{batch_size} ")
+            raise
+
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+            if loss_info["otc_loss"] == float("inf"):
+                logging.error("Your loss contains inf, something goes wrong")
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = OtcTrainingGraphCompiler(
+        params.lang_dir,
+        otc_token=params.otc_token,
+        device=device,
+        initial_bypass_weight=params.initial_bypass_weight,
+        initial_self_loop_weight=params.initial_self_loop_weight,
+        bypass_weight_decay=params.bypass_weight_decay,
+        self_loop_weight_decay=params.self_loop_weight_decay,
+    )
+
+    # remove OTC token as it is the average of all non-blank tokens
+    max_token_id = graph_compiler.get_max_token_id() - 1
+    # add blank
+    num_classes = max_token_id + 1
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    print(model)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            scheduler=scheduler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: OtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    batch=batch,
+                    graph_compiler=graph_compiler,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    assert "▁" not in args.otc_token
+    args.otc_token = f"▁{args.otc_token}"
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/conformer_ctc2/transformer.py b/egs/librispeech/WSASR/conformer_ctc2/transformer.py
new file mode 100644
index 000000000..41e6cd357
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/transformer.py
@@ -0,0 +1,1055 @@
+# Copyright    2021  University of Chinese Academy of Sciences (author: Han Zhu)
+# Copyright    2022  Xiaomi Corp.                              (author: Quandong Wang)
+#              2023  Johns Hopkins University                  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from attention import MultiheadAttention
+from label_smoothing import LabelSmoothingLoss
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledEmbedding,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          layer_dropout (float): layer-dropout rate.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4 and subsampling_factor != 2:
+            raise NotImplementedError("Support only 'subsampling_factor=4 or 2'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.encoder = TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), ScaledLinear(d_model, num_classes, bias=True)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = ScaledEmbedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+            )
+
+            self.decoder = TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+            )
+
+            self.decoder_output_layer = ScaledLinear(
+                d_model, self.decoder_num_class, bias=True
+            )
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        supervision: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+
+        encoder_memory, memory_key_padding_mask = self.run_encoder(
+            x, supervision, warmup
+        )
+
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask, warmup=warmup)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.self_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # src_att = self.self_attn(src, src, src, src_mask)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.layer_dropout = layer_dropout
+        self.self_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+          warmup: controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+
+
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        tgt_orig = tgt
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # tgt_att = self.self_attn(tgt, tgt, tgt, tgt_mask)
+        tgt_att = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(tgt_att)
+
+        # src_att = self.src_attn(tgt, memory, memory, memory_mask)
+        src_att = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(src_att)
+
+        tgt = tgt + self.dropout(self.feed_forward(tgt))
+
+        tgt = self.norm_final(self.balancer(tgt))
+
+        if alpha != 1.0:
+            tgt = alpha * tgt + (1 - alpha) * tgt_orig
+
+        return tgt
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class TransformerDecoder(nn.Module):
+    r"""TransformerDecoder is a stack of N decoder layers
+
+    Args:
+        decoder_layer: an instance of the TransformerDecoderLayer() class (required).
+        num_layers: the number of sub-decoder-layers in the decoder (required).
+
+    Examples::
+        >>> decoder_layer = TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> transformer_decoder = TransformerDecoder(decoder_layer, num_layers=6)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(10, 32, 512)
+        >>> out = transformer_decoder(tgt, memory)
+    """
+
+    def __init__(self, decoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(decoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        r"""Pass the input through the decoder layers in turn.
+
+        Args:
+            tgt: the sequence to the decoder (required).
+            memory: the sequence from the last layer of the encoder (required).
+            tgt_mask: the mask for the tgt sequence (optional).
+            memory_mask: the mask for the memory sequence (optional).
+            tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
+            memory_key_padding_mask: the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (S, N, E).
+            tgt_mask: (S, S).
+            tgt_key_padding_mask: (N, S).
+
+        """
+        output = tgt
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                memory,
+                tgt_mask=tgt_mask,
+                memory_mask=memory_mask,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                memory_key_padding_mask=memory_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int,
+    subsampling_factor: Optional[int] = 4,
+    supervisions: Optional[Supervisions] = None,
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4 or 2. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    if subsampling_factor == 4:
+        lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    elif subsampling_factor == 2:
+        lengths = [(i - 1) // 2 - 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/librispeech/WSASR/figures/del.png b/egs/librispeech/WSASR/figures/del.png
new file mode 100644
index 000000000..38973980b
Binary files /dev/null and b/egs/librispeech/WSASR/figures/del.png differ
diff --git a/egs/librispeech/WSASR/figures/ins.png b/egs/librispeech/WSASR/figures/ins.png
new file mode 100644
index 000000000..2d0e807a9
Binary files /dev/null and b/egs/librispeech/WSASR/figures/ins.png differ
diff --git a/egs/librispeech/WSASR/figures/otc_emission.drawio.png b/egs/librispeech/WSASR/figures/otc_emission.drawio.png
new file mode 100644
index 000000000..6cea5531d
Binary files /dev/null and b/egs/librispeech/WSASR/figures/otc_emission.drawio.png differ
diff --git a/egs/librispeech/WSASR/figures/otc_g.png b/egs/librispeech/WSASR/figures/otc_g.png
new file mode 100644
index 000000000..ebad49180
Binary files /dev/null and b/egs/librispeech/WSASR/figures/otc_g.png differ
diff --git a/egs/librispeech/WSASR/figures/otc_training_graph.drawio.png b/egs/librispeech/WSASR/figures/otc_training_graph.drawio.png
new file mode 100644
index 000000000..8978158d8
Binary files /dev/null and b/egs/librispeech/WSASR/figures/otc_training_graph.drawio.png differ
diff --git a/egs/librispeech/WSASR/figures/sub.png b/egs/librispeech/WSASR/figures/sub.png
new file mode 100644
index 000000000..5674e9feb
Binary files /dev/null and b/egs/librispeech/WSASR/figures/sub.png differ
diff --git a/egs/librispeech/WSASR/local/compile_hlg.py b/egs/librispeech/WSASR/local/compile_hlg.py
new file mode 100755
index 000000000..63791f4cc
--- /dev/null
+++ b/egs/librispeech/WSASR/local/compile_hlg.py
@@ -0,0 +1,173 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_n_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lm",
+        type=str,
+        default="G_3_gram",
+        help="""Stem name for LM used in HLG compiling.
+        """,
+    )
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        help="""LM directory.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lm_dir: str, lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+      lm:
+        The language stem base name.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path(f"{lm_dir}/{lm}.pt").is_file():
+        logging.info(f"Loading pre-compiled {lm}")
+        d = torch.load(f"{lm_dir}/{lm}.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info(f"Loading {lm}.fst.txt")
+        with open(f"{lm_dir}/{lm}.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), f"{lm_dir}/{lm}.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # See https://github.com/k2-fsa/k2/issues/874
+    # for why we need to set LG.properties to None
+    LG.__dict__["_properties"] = None
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lm_dir = Path(args.lm_dir)
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lm_dir, lang_dir, args.lm)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/WSASR/local/compute_fbank_librispeech.py b/egs/librispeech/WSASR/local/compute_fbank_librispeech.py
new file mode 100755
index 000000000..a387d54c9
--- /dev/null
+++ b/egs/librispeech/WSASR/local/compute_fbank_librispeech.py
@@ -0,0 +1,162 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=True,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_librispeech(
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = (
+            "dev-clean",
+            "dev-other",
+            "test-clean",
+            "test-other",
+            "train-clean-100",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = "librispeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+
+            if "train" in partition:
+                if bpe_model:
+                    cut_set = filter_cuts(cut_set, sp)
+                if perturb_speed:
+                    logging.info(f"Doing speed perturb")
+                    cut_set = (
+                        cut_set
+                        + cut_set.perturb_speed(0.9)
+                        + cut_set.perturb_speed(1.1)
+                    )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_librispeech(
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/librispeech/WSASR/local/compute_ssl_librispeech.py b/egs/librispeech/WSASR/local/compute_ssl_librispeech.py
new file mode 100755
index 000000000..f405c468c
--- /dev/null
+++ b/egs/librispeech/WSASR/local/compute_ssl_librispeech.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import S3PRLSSL, CutSet, NumpyFilesWriter, S3PRLSSLConfig
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_ssl_librispeech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/ssl")
+    num_jobs = 1
+
+    dataset_parts = (
+        "dev-clean",
+        "dev-other",
+        "test-clean",
+        "test-other",
+        "train-clean-100",
+    )
+    prefix = "librispeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = S3PRLSSL(S3PRLSSLConfig(ssl_model="wav2vec2", device="cuda"))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                storage_type=NumpyFilesWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_ssl_librispeech()
diff --git a/egs/librispeech/WSASR/local/filter_cuts.py b/egs/librispeech/WSASR/local/filter_cuts.py
new file mode 100644
index 000000000..fbcc9e24a
--- /dev/null
+++ b/egs/librispeech/WSASR/local/filter_cuts.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script removes short and long utterances from a cutset.
+
+Caution:
+  You may need to tune the thresholds for your own dataset.
+
+Usage example:
+
+  python3 ./local/filter_cuts.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --in-cuts data/fbank/librispeech_cuts_test-clean.jsonl.gz \
+    --out-cuts data/fbank-filtered/librispeech_cuts_test-clean.jsonl.gz
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=Path,
+        help="Path to the bpe.model",
+    )
+
+    parser.add_argument(
+        "--in-cuts",
+        type=Path,
+        help="Path to the input cutset",
+    )
+
+    parser.add_argument(
+        "--out-cuts",
+        type=Path,
+        help="Path to the output cutset",
+    )
+
+    return parser.parse_args()
+
+
+def filter_cuts(cut_set: CutSet, sp: spm.SentencePieceProcessor):
+    total = 0  # number of total utterances before removal
+    removed = 0  # number of removed utterances
+
+    def remove_short_and_long_utterances(c: Cut):
+        """Return False to exclude the input cut"""
+        nonlocal removed, total
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ./display_manifest_statistics.py
+        #
+        # You should use ./display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        total += 1
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            removed += 1
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./pruned_transducer_stateless2/conformer.py, the
+        # conv module uses the following expression
+        # for subsampling
+        if c.num_frames is None:
+            num_frames = c.duration * 100  # approximate
+        else:
+            num_frames = c.num_frames
+
+        T = ((num_frames - 1) // 2 - 1) // 2
+        # Note: for ./lstm_transducer_stateless/lstm.py, the formula is
+        #  T = ((num_frames - 3) // 2 - 1) // 2
+
+        # Note: for ./pruned_transducer_stateless7/zipformer.py, the formula is
+        # T = ((num_frames - 7) // 2 + 1) // 2
+
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            removed += 1
+            return False
+
+        return True
+
+    # We use to_eager() here so that we can print out the value of total
+    # and removed below.
+    ans = cut_set.filter(remove_short_and_long_utterances).to_eager()
+    ratio = removed / total * 100
+    logging.info(
+        f"Removed {removed} cuts from {total} cuts. {ratio:.3f}% data is removed."
+    )
+    return ans
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    if args.out_cuts.is_file():
+        logging.info(f"{args.out_cuts} already exists - skipping")
+        return
+
+    assert args.in_cuts.is_file(), f"{args.in_cuts} does not exist"
+    assert args.bpe_model.is_file(), f"{args.bpe_model} does not exist"
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    cut_set = load_manifest_lazy(args.in_cuts)
+    assert isinstance(cut_set, CutSet)
+
+    cut_set = filter_cuts(cut_set, sp)
+    logging.info(f"Saving to {args.out_cuts}")
+    args.out_cuts.parent.mkdir(parents=True, exist_ok=True)
+    cut_set.to_file(args.out_cuts)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/WSASR/local/get_words_from_lexicon.py b/egs/librispeech/WSASR/local/get_words_from_lexicon.py
new file mode 100755
index 000000000..0cc740b36
--- /dev/null
+++ b/egs/librispeech/WSASR/local/get_words_from_lexicon.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+
+import argparse
+from pathlib import Path
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        help="OTC token to be added to words.txt",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    otc_token = args.otc_token
+
+    lexicon = read_lexicon(lang_dir / "lexicon.txt")
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    words = ["<eps>"] + sorted_ans + [otc_token] + ["#0", "<s>", "</s>"]
+
+    words_file = lang_dir / "words.txt"
+    with open(words_file, "w") as wf:
+        for i, word in enumerate(words):
+            wf.write(f"{word} {i}\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/make_error_cutset.py b/egs/librispeech/WSASR/local/make_error_cutset.py
new file mode 100755
index 000000000..8463a380e
--- /dev/null
+++ b/egs/librispeech/WSASR/local/make_error_cutset.py
@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+
+# Copyright 2023 Johns Hopkins University (author: Dongji Gao)
+
+import argparse
+import random
+from pathlib import Path
+from typing import List
+
+from lhotse import CutSet, load_manifest
+from lhotse.cut.base import Cut
+
+from icefall.utils import str2bool
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "--input-cutset",
+        type=str,
+        help="Supervision manifest that contains verbatim transcript",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="words.txt file",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        help="OTC token in words.txt",
+    )
+
+    parser.add_argument(
+        "--sub-error-rate",
+        type=float,
+        default=0.0,
+        help="Substitution error rate",
+    )
+
+    parser.add_argument(
+        "--ins-error-rate",
+        type=float,
+        default=0.0,
+        help="Insertion error rate",
+    )
+
+    parser.add_argument(
+        "--del-error-rate",
+        type=float,
+        default=0.0,
+        help="Deletion error rate",
+    )
+
+    parser.add_argument(
+        "--output-cutset",
+        type=str,
+        default="",
+        help="Supervision manifest that contains modified non-verbatim transcript",
+    )
+
+    parser.add_argument("--verbose", type=str2bool, help="show details of errors")
+    return parser.parse_args()
+
+
+def check_args(args):
+    total_error_rate = args.sub_error_rate + args.ins_error_rate + args.del_error_rate
+    assert args.sub_error_rate >= 0 and args.sub_error_rate <= 1.0
+    assert args.ins_error_rate >= 0 and args.sub_error_rate <= 1.0
+    assert args.del_error_rate >= 0 and args.sub_error_rate <= 1.0
+    assert total_error_rate <= 1.0
+
+
+def get_word_list(token_path: str) -> List:
+    word_list = []
+    with open(Path(token_path), "r") as tp:
+        for line in tp.readlines():
+            token = line.split()[0]
+            assert token not in word_list
+            word_list.append(token)
+    return word_list
+
+
+def modify_cut_text(
+    cut: Cut,
+    words_list: List,
+    non_words: List,
+    sub_ratio: float = 0.0,
+    ins_ratio: float = 0.0,
+    del_ratio: float = 0.0,
+):
+    text = cut.supervisions[0].text
+    text_list = text.split()
+
+    # We save the modified information of the original verbatim text for debugging
+    marked_verbatim_text_list = []
+    modified_text_list = []
+
+    del_index_set = set()
+    sub_index_set = set()
+    ins_index_set = set()
+
+    # We follow the order: deletion -> substitution -> insertion
+    for token in text_list:
+        marked_token = token
+        modified_token = token
+
+        prob = random.random()
+
+        if prob <= del_ratio:
+            marked_token = f"-{token}-"
+            modified_token = ""
+        elif prob <= del_ratio + sub_ratio + ins_ratio:
+            if prob <= del_ratio + sub_ratio:
+                marked_token = f"[{token}]"
+            else:
+                marked_verbatim_text_list.append(marked_token)
+                modified_text_list.append(modified_token)
+                marked_token = "[]"
+
+            # get new_token
+            while (
+                modified_token == token
+                or modified_token in non_words
+                or modified_token.startswith("#")
+            ):
+                modified_token = random.choice(words_list)
+
+        marked_verbatim_text_list.append(marked_token)
+        modified_text_list.append(modified_token)
+
+    marked_text = " ".join(marked_verbatim_text_list)
+    modified_text = " ".join(modified_text_list)
+
+    if not hasattr(cut.supervisions[0], "verbatim_text"):
+        cut.supervisions[0].verbatim_text = marked_text
+    cut.supervisions[0].text = modified_text
+
+    return cut
+
+
+def main():
+    args = get_args()
+    check_args(args)
+
+    otc_token = args.otc_token
+    non_words = set(("sil", "<UNK>", "<eps>"))
+    non_words.add(otc_token)
+
+    words_list = get_word_list(args.words_file)
+    cutset = load_manifest(Path(args.input_cutset))
+
+    cuts = []
+
+    for cut in cutset:
+        modified_cut = modify_cut_text(
+            cut=cut,
+            words_list=words_list,
+            non_words=non_words,
+            sub_ratio=args.sub_error_rate,
+            ins_ratio=args.ins_error_rate,
+            del_ratio=args.del_error_rate,
+        )
+        cuts.append(modified_cut)
+
+    output_cutset = CutSet.from_cuts(cuts)
+    output_cutset.to_file(args.output_cutset)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/prepare_lang.py b/egs/librispeech/WSASR/local/prepare_lang.py
new file mode 100755
index 000000000..d913756a1
--- /dev/null
+++ b/egs/librispeech/WSASR/local/prepare_lang.py
@@ -0,0 +1,413 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+from icefall.utils import str2bool
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    lexicon_filename = lang_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(lang_dir / "tokens.txt", token2id)
+    write_mapping(lang_dir / "words.txt", word2id)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/prepare_otc_lang_bpe.py b/egs/librispeech/WSASR/local/prepare_otc_lang_bpe.py
new file mode 100755
index 000000000..415bdff6f
--- /dev/null
+++ b/egs/librispeech/WSASR/local/prepare_otc_lang_bpe.py
@@ -0,0 +1,295 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.utils import str2bool
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_otc_lexicon(
+    model_file: str,
+    words: List[str],
+    oov: str,
+    otc_token: str,
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
+      otc_token:
+        The OTC token in lexicon.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    words_pieces_ids: List[List[int]] = sp.encode(words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
+
+    # Add OTC token to the last.
+    lexicon.append((otc_token, [f"▁{otc_token}"]))
+    otc_token_index = len(token2id)
+    token2id[f"▁{otc_token}"] = otc_token_index
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        default="<star>",
+        help="The OTC token in lexicon.",
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bpe.model"
+    otc_token = args.otc_token
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = [
+        "<eps>",
+        "!SIL",
+        "<SPOKEN_NOISE>",
+        args.oov,
+        otc_token,
+        "#0",
+        "<s>",
+        "</s>",
+    ]
+
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_otc_lexicon(
+        model_file, words, args.oov, otc_token
+    )
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/train_bpe_model.py b/egs/librispeech/WSASR/local/train_bpe_model.py
new file mode 100755
index 000000000..43142aee4
--- /dev/null
+++ b/egs/librispeech/WSASR/local/train_bpe_model.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/validate_bpe_lexicon.py b/egs/librispeech/WSASR/local/validate_bpe_lexicon.py
new file mode 100755
index 000000000..16a489c11
--- /dev/null
+++ b/egs/librispeech/WSASR/local/validate_bpe_lexicon.py
@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks that there are no OOV tokens in the BPE-based lexicon.
+
+Usage example:
+
+    python3 ./local/validate_bpe_lexicon.py \
+            --lexicon /path/to/lexicon.txt \
+            --bpe-model /path/to/bpe.model
+"""
+
+import argparse
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+
+from icefall.lexicon import read_lexicon
+
+# Map word to word pieces
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--lexicon",
+        required=True,
+        type=Path,
+        help="Path to lexicon.txt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        required=True,
+        type=Path,
+        help="Path to bpe.model",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        required=True,
+        type=str,
+        help="OTC token",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    assert args.lexicon.is_file(), args.lexicon
+    assert args.bpe_model.is_file(), args.bpe_model
+
+    lexicon = read_lexicon(args.lexicon)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    word_pieces = set(sp.id_to_piece(list(range(sp.vocab_size()))))
+    word_pieces.add(f"▁{args.otc_token}")
+    for word, pieces in lexicon:
+        for p in pieces:
+            if p not in word_pieces:
+                raise ValueError(f"The word {word} contains an OOV token {p}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/validate_manifest.py b/egs/librispeech/WSASR/local/validate_manifest.py
new file mode 100755
index 000000000..f620b91ea
--- /dev/null
+++ b/egs/librispeech/WSASR/local/validate_manifest.py
@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+- Supervision time bounds are within cut time bounds
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def validate_one_supervision_per_cut(c: Cut):
+    if len(c.supervisions) != 1:
+        raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
+
+
+def validate_supervision_and_cut_time_bounds(c: Cut):
+    s = c.supervisions[0]
+    if s.start < c.start:
+        raise ValueError(
+            f"{c.id}: Supervision start time {s.start} is less "
+            f"than cut start time {c.start}"
+        )
+
+    if s.end > c.end:
+        raise ValueError(
+            f"{c.id}: Supervision end time {s.end} is larger "
+            f"than cut end time {c.end}"
+        )
+
+
+def main():
+    args = get_args()
+
+    manifest = args.manifest
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    for c in cut_set:
+        validate_one_supervision_per_cut(c)
+        validate_supervision_and_cut_time_bounds(c)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/WSASR/prepare.sh b/egs/librispeech/WSASR/prepare.sh
new file mode 100755
index 000000000..0d2a67259
--- /dev/null
+++ b/egs/librispeech/WSASR/prepare.sh
@@ -0,0 +1,233 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/LibriSpeech
+#      You can find BOOKS.TXT, test-clean, train-clean-360, etc, inside it.
+#      You can download them from https://www.openslr.org/12
+#
+#  - $dl_dir/lm
+#      This directory contains the following files downloaded from
+#       http://www.openslr.org/resources/11
+#
+#        - 3-gram.pruned.1e-7.arpa.gz
+#        - 3-gram.pruned.1e-7.arpa
+#        - 4-gram.arpa.gz
+#        - 4-gram.arpa
+#        - librispeech-vocab.txt
+#        - librispeech-lexicon.txt
+#        - librispeech-lm-norm.txt.gz
+#
+otc_token="<star>"
+feature_type="ssl"
+
+dl_dir=$PWD/download
+manifests_dir="data/manifests"
+feature_dir="data/${feature_type}"
+lang_dir="data/lang"
+lm_dir="data/lm"
+
+perturb_speed=false
+
+# ssl or fbank
+
+. ./cmd.sh
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  200
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: ${dl_dir}"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download LM"
+  mkdir -p ${dl_dir}/lm
+  if [ ! -e ${dl_dir}/lm/.done ]; then
+    ./local/download_lm.py --out-dir=${dl_dir}/lm
+    touch ${dl_dir}/lm/.done
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/LibriSpeech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/LibriSpeech $dl_dir/LibriSpeech
+  #
+  if [ ! -d $dl_dir/LibriSpeech/train-clean-100 ]; then
+    lhotse download librispeech --full ${dl_dir}
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LibriSpeech manifest"
+  # We assume that you have downloaded the LibriSpeech corpus
+  # to $dl_dir/LibriSpeech
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.librispeech.done ]; then
+    lhotse prepare librispeech -j ${nj} \
+      -p dev-clean \
+      -p dev-other \
+      -p test-clean \
+      -p test-other \
+      -p train-clean-100 "${dl_dir}/LibriSpeech" "${manifests_dir}"
+    touch data/manifests/.librispeech.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute ${feature_type} feature for librispeech (train-clean-100)"
+  mkdir -p "${feature_dir}"  
+  if [ ! -e "${feature_dir}/.librispeech.done" ]; then
+    if [ "${feature_type}" = ssl ]; then
+      ./local/compute_ssl_librispeech.py
+    elif [ "${feature_type}" = fbank ]; then
+      ./local/compute_fbank_librispeech.py --perturb-speed ${perturb_speed}
+    else
+      log "Error: not supported --feature-type '${feature_type}'" 
+      exit 2
+    fi
+
+    touch "${feature_dir}.librispeech.done"
+  fi
+
+  if [ ! -e "${feature_dir}/.librispeech-validated.done" ]; then
+    log "Validating data/ssl for LibriSpeech"
+    parts=(
+      train-clean-100
+      test-clean
+      test-other
+      dev-clean
+      dev-other
+    )
+    for part in ${parts[@]}; do
+      python3 ./local/validate_manifest.py \
+        "${feature_dir}/librispeech_cuts_${part}.jsonl.gz"
+    done
+    touch "${feature_dir}/.librispeech-validated.done"
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare words.txt"
+  mkdir -p ${lang_dir}
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/lm/librispeech-lexicon.txt |
+    sort | uniq > ${lang_dir}/lexicon.txt
+
+  local/get_words_from_lexicon.py \
+    --lang-dir ${lang_dir} \
+    --otc-token ${otc_token}
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    bpe_lang_dir="data/lang_bpe_${vocab_size}"
+    mkdir -p "${bpe_lang_dir}"
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp "${lang_dir}/words.txt" "${bpe_lang_dir}"
+
+    if [ ! -f "${bpe_lang_dir}/transcript_words.txt" ]; then
+      log "Generate data for BPE training"
+      files=$(
+        find "$dl_dir/LibriSpeech/train-clean-100" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/train-clean-360" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/train-other-500" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > "${bpe_lang_dir}/transcript_words.txt"
+    fi
+
+    if [ ! -f ${bpe_lang_dir}/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir ${bpe_lang_dir} \
+        --vocab-size ${vocab_size} \
+        --transcript ${bpe_lang_dir}/transcript_words.txt 
+    fi
+
+    if [ ! -f ${bpe_lang_dir}/L_disambig.pt ]; then
+      ./local/prepare_otc_lang_bpe.py \
+        --lang-dir "${bpe_lang_dir}" \
+        --otc-token "${otc_token}"
+
+      log "Validating ${bpe_lang_dir}/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon ${bpe_lang_dir}/lexicon.txt \
+        --bpe-model ${bpe_lang_dir}/bpe.model \
+        --otc-token "${otc_token}"
+    fi
+  done
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p "${lm_dir}"
+  if [ ! -f ${lm_dir}/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="${lang_dir}/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      ${dl_dir}/lm/3-gram.pruned.1e-7.arpa > ${lm_dir}/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f ${lm_dir}/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="${lang_dir}/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      ${dl_dir}/lm/4-gram.arpa > ${lm_dir}/G_4_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compile HLG"
+  # Note If ./local/compile_hlg.py throws OOM,
+  # please switch to the following command
+  #
+  # ./local/compile_hlg_using_openfst.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    bpe_lang_dir="data/lang_bpe_${vocab_size}"
+    echo "LM DIR: ${lm_dir}"
+    ./local/compile_hlg.py \
+        --lm-dir "${lm_dir}" \
+        --lang-dir "${bpe_lang_dir}"
+  done
+fi
diff --git a/egs/ljspeech/TTS/local/compute_spectrogram_ljspeech.py b/egs/ljspeech/TTS/local/compute_spectrogram_ljspeech.py
new file mode 100755
index 000000000..97c9008fc
--- /dev/null
+++ b/egs/ljspeech/TTS/local/compute_spectrogram_ljspeech.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LJSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated spectrogram features are saved in data/spectrogram.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    LilcomChunkyWriter,
+    Spectrogram,
+    SpectrogramConfig,
+    load_manifest,
+)
+from lhotse.audio import RecordingSet
+from lhotse.supervision import SupervisionSet
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_spectrogram_ljspeech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/spectrogram")
+    num_jobs = min(4, os.cpu_count())
+
+    sampling_rate = 22050
+    frame_length = 1024 / sampling_rate  # (in second)
+    frame_shift = 256 / sampling_rate  # (in second)
+    use_fft_mag = True
+
+    prefix = "ljspeech"
+    suffix = "jsonl.gz"
+    partition = "all"
+
+    recordings = load_manifest(
+        src_dir / f"{prefix}_recordings_{partition}.{suffix}", RecordingSet
+    )
+    supervisions = load_manifest(
+        src_dir / f"{prefix}_supervisions_{partition}.{suffix}", SupervisionSet
+    )
+
+    config = SpectrogramConfig(
+        sampling_rate=sampling_rate,
+        frame_length=frame_length,
+        frame_shift=frame_shift,
+        use_fft_mag=use_fft_mag,
+    )
+    extractor = Spectrogram(config)
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{cuts_filename} already exists - skipping.")
+            return
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=recordings, supervisions=supervisions
+        )
+
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+        cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_spectrogram_ljspeech()
diff --git a/egs/ljspeech/TTS/local/display_manifest_statistics.py b/egs/ljspeech/TTS/local/display_manifest_statistics.py
new file mode 100755
index 000000000..93f0044f0
--- /dev/null
+++ b/egs/ljspeech/TTS/local/display_manifest_statistics.py
@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in vits/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    path = "./data/spectrogram/ljspeech_cuts_all.jsonl.gz"
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Cut statistics:
+    ╒═══════════════════════════╤══════════╕
+    │ Cuts count:               │ 13100    │
+    ├───────────────────────────┼──────────┤
+    │ Total duration (hh:mm:ss) │ 23:55:18 │
+    ├───────────────────────────┼──────────┤
+    │ mean                      │ 6.6      │
+    ├───────────────────────────┼──────────┤
+    │ std                       │ 2.2      │
+    ├───────────────────────────┼──────────┤
+    │ min                       │ 1.1      │
+    ├───────────────────────────┼──────────┤
+    │ 25%                       │ 5.0      │
+    ├───────────────────────────┼──────────┤
+    │ 50%                       │ 6.8      │
+    ├───────────────────────────┼──────────┤
+    │ 75%                       │ 8.4      │
+    ├───────────────────────────┼──────────┤
+    │ 99%                       │ 10.0     │
+    ├───────────────────────────┼──────────┤
+    │ 99.5%                     │ 10.1     │
+    ├───────────────────────────┼──────────┤
+    │ 99.9%                     │ 10.1     │
+    ├───────────────────────────┼──────────┤
+    │ max                       │ 10.1     │
+    ├───────────────────────────┼──────────┤
+    │ Recordings available:     │ 13100    │
+    ├───────────────────────────┼──────────┤
+    │ Features available:       │ 13100    │
+    ├───────────────────────────┼──────────┤
+    │ Supervisions available:   │ 13100    │
+    ╘═══════════════════════════╧══════════╛
+"""
diff --git a/egs/ljspeech/TTS/local/prepare_token_file.py b/egs/ljspeech/TTS/local/prepare_token_file.py
new file mode 100755
index 000000000..df976804a
--- /dev/null
+++ b/egs/ljspeech/TTS/local/prepare_token_file.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file reads the texts in given manifest and generates the file that maps tokens to IDs.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict
+
+from lhotse import load_manifest
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest-file",
+        type=Path,
+        default=Path("data/spectrogram/ljspeech_cuts_train.jsonl.gz"),
+        help="Path to the manifest file",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=Path,
+        default=Path("data/tokens.txt"),
+        help="Path to the tokens",
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_token2id(manifest_file: Path) -> Dict[str, int]:
+    """Return a dict that maps token to IDs."""
+    extra_tokens = [
+        "<blk>",  # 0 for blank
+        "<sos/eos>",  # 1 for sos and eos symbols.
+        "<unk>",  # 2 for OOV
+    ]
+    all_tokens = set()
+
+    cut_set = load_manifest(manifest_file)
+
+    for cut in cut_set:
+        # Each cut only contain one supervision
+        assert len(cut.supervisions) == 1, len(cut.supervisions)
+        for t in cut.tokens:
+            all_tokens.add(t)
+
+    all_tokens = extra_tokens + list(all_tokens)
+
+    token2id: Dict[str, int] = {token: i for i, token in enumerate(all_tokens)}
+    return token2id
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    manifest_file = Path(args.manifest_file)
+    out_file = Path(args.tokens)
+
+    token2id = get_token2id(manifest_file)
+    write_mapping(out_file, token2id)
diff --git a/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py b/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py
new file mode 100755
index 000000000..fcd0137a0
--- /dev/null
+++ b/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file reads the texts in given manifest and save the new cuts with phoneme tokens.
+"""
+
+import logging
+from pathlib import Path
+
+import g2p_en
+import tacotron_cleaner.cleaners
+from lhotse import CutSet, load_manifest
+
+
+def prepare_tokens_ljspeech():
+    output_dir = Path("data/spectrogram")
+    prefix = "ljspeech"
+    suffix = "jsonl.gz"
+    partition = "all"
+
+    cut_set = load_manifest(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+    g2p = g2p_en.G2p()
+
+    new_cuts = []
+    for cut in cut_set:
+        # Each cut only contains one supervision
+        assert len(cut.supervisions) == 1, len(cut.supervisions)
+        text = cut.supervisions[0].normalized_text
+        # Text normalization
+        text = tacotron_cleaner.cleaners.custom_english_cleaners(text)
+        # Convert to phonemes
+        cut.tokens = g2p(text)
+        new_cuts.append(cut)
+
+    new_cut_set = CutSet.from_cuts(new_cuts)
+    new_cut_set.to_file(output_dir / f"{prefix}_cuts_with_tokens_{partition}.{suffix}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    prepare_tokens_ljspeech()
diff --git a/egs/ljspeech/TTS/local/validate_manifest.py b/egs/ljspeech/TTS/local/validate_manifest.py
new file mode 100755
index 000000000..68159ae03
--- /dev/null
+++ b/egs/ljspeech/TTS/local/validate_manifest.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/spectrogram/ljspeech_cuts_all.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.dataset.speech_synthesis import validate_for_tts
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    manifest = args.manifest
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet), type(cut_set)
+
+    validate_for_tts(cut_set)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/ljspeech/TTS/prepare.sh b/egs/ljspeech/TTS/prepare.sh
new file mode 100755
index 000000000..ed0a07f5e
--- /dev/null
+++ b/egs/ljspeech/TTS/prepare.sh
@@ -0,0 +1,125 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=0
+stop_stage=100
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: build monotonic_align lib"
+  if [ ! -d vits/monotonic_align/build ]; then
+    cd vits/monotonic_align
+    python setup.py build_ext --inplace
+    cd ../../
+  else 
+    log "monotonic_align lib already built"
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # The directory $dl_dir/LJSpeech-1.1 will contain:
+  #   - wavs, which contains the audio files
+  #   - metadata.csv, which provides the transcript text for each audio clip
+
+  # If you have pre-downloaded it to /path/to/LJSpeech-1.1, you can create a symlink
+  #
+  #   ln -sfv /path/to/LJSpeech-1.1 $dl_dir/LJSpeech-1.1
+  #
+  if [ ! -d $dl_dir/LJSpeech-1.1 ]; then
+    lhotse download ljspeech $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LJSpeech manifest"
+  # We assume that you have downloaded the LJSpeech corpus
+  # to $dl_dir/LJSpeech
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.ljspeech.done ]; then
+    lhotse prepare ljspeech $dl_dir/LJSpeech-1.1 data/manifests
+    touch data/manifests/.ljspeech.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute spectrogram for LJSpeech"
+  mkdir -p data/spectrogram
+  if [ ! -e data/spectrogram/.ljspeech.done ]; then
+    ./local/compute_spectrogram_ljspeech.py
+    touch data/spectrogram/.ljspeech.done
+  fi
+
+  if [ ! -e data/spectrogram/.ljspeech-validated.done ]; then
+    log "Validating data/spectrogram for LJSpeech"
+    python3 ./local/validate_manifest.py \
+      data/spectrogram/ljspeech_cuts_all.jsonl.gz
+    touch data/spectrogram/.ljspeech-validated.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare phoneme tokens for LJSpeech"
+  if [ ! -e data/spectrogram/.ljspeech_with_token.done ]; then
+    ./local/prepare_tokens_ljspeech.py
+    mv data/spectrogram/ljspeech_cuts_with_tokens_all.jsonl.gz \
+      data/spectrogram/ljspeech_cuts_all.jsonl.gz
+    touch data/spectrogram/.ljspeech_with_token.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Split the LJSpeech cuts into train, valid and test sets"
+  if [ ! -e data/spectrogram/.ljspeech_split.done ]; then
+    lhotse subset --last 600 \
+      data/spectrogram/ljspeech_cuts_all.jsonl.gz \
+      data/spectrogram/ljspeech_cuts_validtest.jsonl.gz
+    lhotse subset --first 100 \
+      data/spectrogram/ljspeech_cuts_validtest.jsonl.gz \
+      data/spectrogram/ljspeech_cuts_valid.jsonl.gz
+    lhotse subset --last 500 \
+      data/spectrogram/ljspeech_cuts_validtest.jsonl.gz \
+      data/spectrogram/ljspeech_cuts_test.jsonl.gz
+
+    rm data/spectrogram/ljspeech_cuts_validtest.jsonl.gz
+
+    n=$(( $(gunzip -c data/spectrogram/ljspeech_cuts_all.jsonl.gz | wc -l) - 600 ))
+    lhotse subset --first $n  \
+      data/spectrogram/ljspeech_cuts_all.jsonl.gz \
+      data/spectrogram/ljspeech_cuts_train.jsonl.gz
+      touch data/spectrogram/.ljspeech_split.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Generate token file"
+  # We assume you have installed g2p_en and espnet_tts_frontend.
+  # If not, please install them with:
+  #   - g2p_en: `pip install g2p_en`, refer to https://github.com/Kyubyong/g2p
+  #   - espnet_tts_frontend, `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/
+  if [ ! -e data/tokens.txt ]; then
+    ./local/prepare_token_file.py \
+      --manifest-file data/spectrogram/ljspeech_cuts_train.jsonl.gz \
+      --tokens data/tokens.txt
+  fi
+fi
diff --git a/egs/ljspeech/TTS/shared/parse_options.sh b/egs/ljspeech/TTS/shared/parse_options.sh
new file mode 120000
index 000000000..e4665e7de
--- /dev/null
+++ b/egs/ljspeech/TTS/shared/parse_options.sh
@@ -0,0 +1 @@
+../../../librispeech/ASR/shared/parse_options.sh
\ No newline at end of file
diff --git a/egs/ljspeech/TTS/vits/README.md b/egs/ljspeech/TTS/vits/README.md
new file mode 100644
index 000000000..1141326b9
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/README.md
@@ -0,0 +1,3 @@
+See https://k2-fsa.github.io/icefall/recipes/TTS/ljspeech/vits.html for detailed tutorials.
+
+Training logs, Tensorboard logs, and checkpoints are uploaded to https://huggingface.co/Zengwei/icefall-tts-ljspeech-vits-2023-11-29.
diff --git a/egs/ljspeech/TTS/vits/duration_predictor.py b/egs/ljspeech/TTS/vits/duration_predictor.py
new file mode 100644
index 000000000..1a8190014
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/duration_predictor.py
@@ -0,0 +1,193 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/duration_predictor.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Stochastic duration predictor modules in VITS.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+import math
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from flow import (
+    ConvFlow,
+    DilatedDepthSeparableConv,
+    ElementwiseAffineFlow,
+    FlipFlow,
+    LogFlow,
+)
+
+
+class StochasticDurationPredictor(torch.nn.Module):
+    """Stochastic duration predictor module.
+
+    This is a module of stochastic duration predictor described in `Conditional
+    Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech`_.
+
+    .. _`Conditional Variational Autoencoder with Adversarial Learning for End-to-End
+        Text-to-Speech`: https://arxiv.org/abs/2006.04558
+
+    """
+
+    def __init__(
+        self,
+        channels: int = 192,
+        kernel_size: int = 3,
+        dropout_rate: float = 0.5,
+        flows: int = 4,
+        dds_conv_layers: int = 3,
+        global_channels: int = -1,
+    ):
+        """Initialize StochasticDurationPredictor module.
+
+        Args:
+            channels (int): Number of channels.
+            kernel_size (int): Kernel size.
+            dropout_rate (float): Dropout rate.
+            flows (int): Number of flows.
+            dds_conv_layers (int): Number of conv layers in DDS conv.
+            global_channels (int): Number of global conditioning channels.
+
+        """
+        super().__init__()
+
+        self.pre = torch.nn.Conv1d(channels, channels, 1)
+        self.dds = DilatedDepthSeparableConv(
+            channels,
+            kernel_size,
+            layers=dds_conv_layers,
+            dropout_rate=dropout_rate,
+        )
+        self.proj = torch.nn.Conv1d(channels, channels, 1)
+
+        self.log_flow = LogFlow()
+        self.flows = torch.nn.ModuleList()
+        self.flows += [ElementwiseAffineFlow(2)]
+        for i in range(flows):
+            self.flows += [
+                ConvFlow(
+                    2,
+                    channels,
+                    kernel_size,
+                    layers=dds_conv_layers,
+                )
+            ]
+            self.flows += [FlipFlow()]
+
+        self.post_pre = torch.nn.Conv1d(1, channels, 1)
+        self.post_dds = DilatedDepthSeparableConv(
+            channels,
+            kernel_size,
+            layers=dds_conv_layers,
+            dropout_rate=dropout_rate,
+        )
+        self.post_proj = torch.nn.Conv1d(channels, channels, 1)
+        self.post_flows = torch.nn.ModuleList()
+        self.post_flows += [ElementwiseAffineFlow(2)]
+        for i in range(flows):
+            self.post_flows += [
+                ConvFlow(
+                    2,
+                    channels,
+                    kernel_size,
+                    layers=dds_conv_layers,
+                )
+            ]
+            self.post_flows += [FlipFlow()]
+
+        if global_channels > 0:
+            self.global_conv = torch.nn.Conv1d(global_channels, channels, 1)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        w: Optional[torch.Tensor] = None,
+        g: Optional[torch.Tensor] = None,
+        inverse: bool = False,
+        noise_scale: float = 1.0,
+    ) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, channels, T_text).
+            x_mask (Tensor): Mask tensor (B, 1, T_text).
+            w (Optional[Tensor]): Duration tensor (B, 1, T_text).
+            g (Optional[Tensor]): Global conditioning tensor (B, channels, 1)
+            inverse (bool): Whether to inverse the flow.
+            noise_scale (float): Noise scale value.
+
+        Returns:
+            Tensor: If not inverse, negative log-likelihood (NLL) tensor (B,).
+                If inverse, log-duration tensor (B, 1, T_text).
+
+        """
+        x = x.detach()  # stop gradient
+        x = self.pre(x)
+        if g is not None:
+            x = x + self.global_conv(g.detach())  # stop gradient
+        x = self.dds(x, x_mask)
+        x = self.proj(x) * x_mask
+
+        if not inverse:
+            assert w is not None, "w must be provided."
+            h_w = self.post_pre(w)
+            h_w = self.post_dds(h_w, x_mask)
+            h_w = self.post_proj(h_w) * x_mask
+            e_q = (
+                torch.randn(
+                    w.size(0),
+                    2,
+                    w.size(2),
+                ).to(device=x.device, dtype=x.dtype)
+                * x_mask
+            )
+            z_q = e_q
+            logdet_tot_q = 0.0
+            for flow in self.post_flows:
+                z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
+                logdet_tot_q += logdet_q
+            z_u, z1 = torch.split(z_q, [1, 1], 1)
+            u = torch.sigmoid(z_u) * x_mask
+            z0 = (w - u) * x_mask
+            logdet_tot_q += torch.sum(
+                (F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2]
+            )
+            logq = (
+                torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask, [1, 2])
+                - logdet_tot_q
+            )
+
+            logdet_tot = 0
+            z0, logdet = self.log_flow(z0, x_mask)
+            logdet_tot += logdet
+            z = torch.cat([z0, z1], 1)
+            for flow in self.flows:
+                z, logdet = flow(z, x_mask, g=x, inverse=inverse)
+                logdet_tot = logdet_tot + logdet
+            nll = (
+                torch.sum(0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask, [1, 2])
+                - logdet_tot
+            )
+            return nll + logq  # (B,)
+        else:
+            flows = list(reversed(self.flows))
+            flows = flows[:-2] + [flows[-1]]  # remove a useless vflow
+            z = (
+                torch.randn(
+                    x.size(0),
+                    2,
+                    x.size(2),
+                ).to(device=x.device, dtype=x.dtype)
+                * noise_scale
+            )
+            for flow in flows:
+                z = flow(z, x_mask, g=x, inverse=inverse)
+            z0, z1 = z.split(1, 1)
+            logw = z0
+            return logw
diff --git a/egs/ljspeech/TTS/vits/export-onnx.py b/egs/ljspeech/TTS/vits/export-onnx.py
new file mode 100755
index 000000000..36a9de27f
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/export-onnx.py
@@ -0,0 +1,267 @@
+#!/usr/bin/env python3
+#
+# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script exports a VITS model from PyTorch to ONNX.
+
+Export the model to ONNX:
+./vits/export-onnx.py \
+  --epoch 1000 \
+  --exp-dir vits/exp \
+  --tokens data/tokens.txt
+
+It will generate two files inside vits/exp:
+  - vits-epoch-1000.onnx
+  - vits-epoch-1000.int8.onnx (quantizated model)
+
+See ./test_onnx.py for how to use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import torch
+import torch.nn as nn
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from tokenizer import Tokenizer
+from train import get_model, get_params
+
+from icefall.checkpoint import load_checkpoint
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=1000,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="vits/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxModel(nn.Module):
+    """A wrapper for VITS generator."""
+
+    def __init__(self, model: nn.Module):
+        """
+        Args:
+          model:
+            A VITS generator.
+          frame_shift:
+            The frame shift in samples.
+        """
+        super().__init__()
+        self.model = model
+
+    def forward(
+        self,
+        tokens: torch.Tensor,
+        tokens_lens: torch.Tensor,
+        noise_scale: float = 0.667,
+        alpha: float = 1.0,
+        noise_scale_dur: float = 0.8,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of VITS.inference_batch
+
+        Args:
+          tokens:
+            Input text token indexes (1, T_text)
+          tokens_lens:
+            Number of tokens of shape (1,)
+          noise_scale (float):
+            Noise scale parameter for flow.
+          noise_scale_dur (float):
+            Noise scale parameter for duration predictor.
+          alpha (float):
+            Alpha parameter to control the speed of generated speech.
+
+        Returns:
+          Return a tuple containing:
+            - audio, generated wavform tensor, (B, T_wav)
+        """
+        audio, _, _ = self.model.inference(
+            text=tokens,
+            text_lengths=tokens_lens,
+            noise_scale=noise_scale,
+            noise_scale_dur=noise_scale_dur,
+            alpha=alpha,
+        )
+        return audio
+
+
+def export_model_onnx(
+    model: nn.Module,
+    model_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given generator model to ONNX format.
+    The exported model has one input:
+
+        - tokens, a tensor of shape (1, T_text); dtype is torch.int64
+
+    and it has one output:
+
+        - audio, a tensor of shape (1, T'); dtype is torch.float32
+
+    Args:
+      model:
+        The VITS generator.
+      model_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    tokens = torch.randint(low=0, high=79, size=(1, 13), dtype=torch.int64)
+    tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64)
+    noise_scale = torch.tensor([1], dtype=torch.float32)
+    noise_scale_dur = torch.tensor([1], dtype=torch.float32)
+    alpha = torch.tensor([1], dtype=torch.float32)
+
+    torch.onnx.export(
+        model,
+        (tokens, tokens_lens, noise_scale, alpha, noise_scale_dur),
+        model_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "tokens",
+            "tokens_lens",
+            "noise_scale",
+            "alpha",
+            "noise_scale_dur",
+        ],
+        output_names=["audio"],
+        dynamic_axes={
+            "tokens": {0: "N", 1: "T"},
+            "tokens_lens": {0: "N"},
+            "audio": {0: "N", 1: "T"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "VITS",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "VITS generator",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=model_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.blank_id
+    params.oov_id = tokenizer.oov_id
+    params.vocab_size = tokenizer.vocab_size
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+
+    model = model.generator
+    model.to("cpu")
+    model.eval()
+
+    model = OnnxModel(model=model)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"generator parameters: {num_param}")
+
+    suffix = f"epoch-{params.epoch}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    model_filename = params.exp_dir / f"vits-{suffix}.onnx"
+    export_model_onnx(
+        model,
+        model_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported generator to {model_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    model_filename_int8 = params.exp_dir / f"vits-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=model_filename,
+        model_output=model_filename_int8,
+        weight_type=QuantType.QUInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/ljspeech/TTS/vits/flow.py b/egs/ljspeech/TTS/vits/flow.py
new file mode 100644
index 000000000..2b84f6434
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/flow.py
@@ -0,0 +1,311 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/flow.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Basic Flow modules used in VITS.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+from transform import piecewise_rational_quadratic_transform
+
+
+class FlipFlow(torch.nn.Module):
+    """Flip flow module."""
+
+    def forward(
+        self, x: torch.Tensor, *args, inverse: bool = False, **kwargs
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, channels, T).
+            inverse (bool): Whether to inverse the flow.
+
+        Returns:
+            Tensor: Flipped tensor (B, channels, T).
+            Tensor: Log-determinant tensor for NLL (B,) if not inverse.
+
+        """
+        x = torch.flip(x, [1])
+        if not inverse:
+            logdet = x.new_zeros(x.size(0))
+            return x, logdet
+        else:
+            return x
+
+
+class LogFlow(torch.nn.Module):
+    """Log flow module."""
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        inverse: bool = False,
+        eps: float = 1e-5,
+        **kwargs
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, channels, T).
+            x_mask (Tensor): Mask tensor (B, 1, T).
+            inverse (bool): Whether to inverse the flow.
+            eps (float): Epsilon for log.
+
+        Returns:
+            Tensor: Output tensor (B, channels, T).
+            Tensor: Log-determinant tensor for NLL (B,) if not inverse.
+
+        """
+        if not inverse:
+            y = torch.log(torch.clamp_min(x, eps)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class ElementwiseAffineFlow(torch.nn.Module):
+    """Elementwise affine flow module."""
+
+    def __init__(self, channels: int):
+        """Initialize ElementwiseAffineFlow module.
+
+        Args:
+            channels (int): Number of channels.
+
+        """
+        super().__init__()
+        self.channels = channels
+        self.register_parameter("m", torch.nn.Parameter(torch.zeros(channels, 1)))
+        self.register_parameter("logs", torch.nn.Parameter(torch.zeros(channels, 1)))
+
+    def forward(
+        self, x: torch.Tensor, x_mask: torch.Tensor, inverse: bool = False, **kwargs
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, channels, T).
+            x_lengths (Tensor): Length tensor (B,).
+            inverse (bool): Whether to inverse the flow.
+
+        Returns:
+            Tensor: Output tensor (B, channels, T).
+            Tensor: Log-determinant tensor for NLL (B,) if not inverse.
+
+        """
+        if not inverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class Transpose(torch.nn.Module):
+    """Transpose module for torch.nn.Sequential()."""
+
+    def __init__(self, dim1: int, dim2: int):
+        """Initialize Transpose module."""
+        super().__init__()
+        self.dim1 = dim1
+        self.dim2 = dim2
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Transpose."""
+        return x.transpose(self.dim1, self.dim2)
+
+
+class DilatedDepthSeparableConv(torch.nn.Module):
+    """Dilated depth-separable conv module."""
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        layers: int,
+        dropout_rate: float = 0.0,
+        eps: float = 1e-5,
+    ):
+        """Initialize DilatedDepthSeparableConv module.
+
+        Args:
+            channels (int): Number of channels.
+            kernel_size (int): Kernel size.
+            layers (int): Number of layers.
+            dropout_rate (float): Dropout rate.
+            eps (float): Epsilon for layer norm.
+
+        """
+        super().__init__()
+
+        self.convs = torch.nn.ModuleList()
+        for i in range(layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs += [
+                torch.nn.Sequential(
+                    torch.nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        groups=channels,
+                        dilation=dilation,
+                        padding=padding,
+                    ),
+                    Transpose(1, 2),
+                    torch.nn.LayerNorm(
+                        channels,
+                        eps=eps,
+                        elementwise_affine=True,
+                    ),
+                    Transpose(1, 2),
+                    torch.nn.GELU(),
+                    torch.nn.Conv1d(
+                        channels,
+                        channels,
+                        1,
+                    ),
+                    Transpose(1, 2),
+                    torch.nn.LayerNorm(
+                        channels,
+                        eps=eps,
+                        elementwise_affine=True,
+                    ),
+                    Transpose(1, 2),
+                    torch.nn.GELU(),
+                    torch.nn.Dropout(dropout_rate),
+                )
+            ]
+
+    def forward(
+        self, x: torch.Tensor, x_mask: torch.Tensor, g: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, in_channels, T).
+            x_mask (Tensor): Mask tensor (B, 1, T).
+            g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1).
+
+        Returns:
+            Tensor: Output tensor (B, channels, T).
+
+        """
+        if g is not None:
+            x = x + g
+        for f in self.convs:
+            y = f(x * x_mask)
+            x = x + y
+        return x * x_mask
+
+
+class ConvFlow(torch.nn.Module):
+    """Convolutional flow module."""
+
+    def __init__(
+        self,
+        in_channels: int,
+        hidden_channels: int,
+        kernel_size: int,
+        layers: int,
+        bins: int = 10,
+        tail_bound: float = 5.0,
+    ):
+        """Initialize ConvFlow module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            hidden_channels (int): Number of hidden channels.
+            kernel_size (int): Kernel size.
+            layers (int): Number of layers.
+            bins (int): Number of bins.
+            tail_bound (float): Tail bound value.
+
+        """
+        super().__init__()
+        self.half_channels = in_channels // 2
+        self.hidden_channels = hidden_channels
+        self.bins = bins
+        self.tail_bound = tail_bound
+
+        self.input_conv = torch.nn.Conv1d(
+            self.half_channels,
+            hidden_channels,
+            1,
+        )
+        self.dds_conv = DilatedDepthSeparableConv(
+            hidden_channels,
+            kernel_size,
+            layers,
+            dropout_rate=0.0,
+        )
+        self.proj = torch.nn.Conv1d(
+            hidden_channels,
+            self.half_channels * (bins * 3 - 1),
+            1,
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        inverse: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, channels, T).
+            x_mask (Tensor): Mask tensor (B,).
+            g (Optional[Tensor]): Global conditioning tensor (B, channels, 1).
+            inverse (bool): Whether to inverse the flow.
+
+        Returns:
+            Tensor: Output tensor (B, channels, T).
+            Tensor: Log-determinant tensor for NLL (B,) if not inverse.
+
+        """
+        xa, xb = x.split(x.size(1) // 2, 1)
+        h = self.input_conv(xa)
+        h = self.dds_conv(h, x_mask, g=g)
+        h = self.proj(h) * x_mask  # (B, half_channels * (bins * 3 - 1), T)
+
+        b, c, t = xa.shape
+        # (B, half_channels, bins * 3 - 1, T) -> (B, half_channels, T, bins * 3 - 1)
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)
+
+        # TODO(kan-bayashi): Understand this calculation
+        denom = math.sqrt(self.hidden_channels)
+        unnorm_widths = h[..., : self.bins] / denom
+        unnorm_heights = h[..., self.bins : 2 * self.bins] / denom
+        unnorm_derivatives = h[..., 2 * self.bins :]
+        xb, logdet_abs = piecewise_rational_quadratic_transform(
+            xb,
+            unnorm_widths,
+            unnorm_heights,
+            unnorm_derivatives,
+            inverse=inverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+        x = torch.cat([xa, xb], 1) * x_mask
+        logdet = torch.sum(logdet_abs * x_mask, [1, 2])
+        if not inverse:
+            return x, logdet
+        else:
+            return x
diff --git a/egs/ljspeech/TTS/vits/generator.py b/egs/ljspeech/TTS/vits/generator.py
new file mode 100644
index 000000000..66c8cedb1
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/generator.py
@@ -0,0 +1,530 @@
+# based on https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/generator.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Generator module in VITS.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+
+import math
+from typing import List, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from duration_predictor import StochasticDurationPredictor
+from hifigan import HiFiGANGenerator
+from posterior_encoder import PosteriorEncoder
+from residual_coupling import ResidualAffineCouplingBlock
+from text_encoder import TextEncoder
+from utils import get_random_segments
+
+from icefall.utils import make_pad_mask
+
+
+class VITSGenerator(torch.nn.Module):
+    """Generator module in VITS, `Conditional Variational Autoencoder
+    with Adversarial Learning for End-to-End Text-to-Speech`.
+    """
+
+    def __init__(
+        self,
+        vocabs: int,
+        aux_channels: int = 513,
+        hidden_channels: int = 192,
+        spks: Optional[int] = None,
+        langs: Optional[int] = None,
+        spk_embed_dim: Optional[int] = None,
+        global_channels: int = -1,
+        segment_size: int = 32,
+        text_encoder_attention_heads: int = 2,
+        text_encoder_ffn_expand: int = 4,
+        text_encoder_cnn_module_kernel: int = 5,
+        text_encoder_blocks: int = 6,
+        text_encoder_dropout_rate: float = 0.1,
+        decoder_kernel_size: int = 7,
+        decoder_channels: int = 512,
+        decoder_upsample_scales: List[int] = [8, 8, 2, 2],
+        decoder_upsample_kernel_sizes: List[int] = [16, 16, 4, 4],
+        decoder_resblock_kernel_sizes: List[int] = [3, 7, 11],
+        decoder_resblock_dilations: List[List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+        use_weight_norm_in_decoder: bool = True,
+        posterior_encoder_kernel_size: int = 5,
+        posterior_encoder_layers: int = 16,
+        posterior_encoder_stacks: int = 1,
+        posterior_encoder_base_dilation: int = 1,
+        posterior_encoder_dropout_rate: float = 0.0,
+        use_weight_norm_in_posterior_encoder: bool = True,
+        flow_flows: int = 4,
+        flow_kernel_size: int = 5,
+        flow_base_dilation: int = 1,
+        flow_layers: int = 4,
+        flow_dropout_rate: float = 0.0,
+        use_weight_norm_in_flow: bool = True,
+        use_only_mean_in_flow: bool = True,
+        stochastic_duration_predictor_kernel_size: int = 3,
+        stochastic_duration_predictor_dropout_rate: float = 0.5,
+        stochastic_duration_predictor_flows: int = 4,
+        stochastic_duration_predictor_dds_conv_layers: int = 3,
+    ):
+        """Initialize VITS generator module.
+
+        Args:
+            vocabs (int): Input vocabulary size.
+            aux_channels (int): Number of acoustic feature channels.
+            hidden_channels (int): Number of hidden channels.
+            spks (Optional[int]): Number of speakers. If set to > 1, assume that the
+                sids will be provided as the input and use sid embedding layer.
+            langs (Optional[int]): Number of languages. If set to > 1, assume that the
+                lids will be provided as the input and use sid embedding layer.
+            spk_embed_dim (Optional[int]): Speaker embedding dimension. If set to > 0,
+                assume that spembs will be provided as the input.
+            global_channels (int): Number of global conditioning channels.
+            segment_size (int): Segment size for decoder.
+            text_encoder_attention_heads (int): Number of heads in conformer block
+                of text encoder.
+            text_encoder_ffn_expand (int): Expansion ratio of FFN in conformer block
+                of text encoder.
+            text_encoder_cnn_module_kernel (int): Convolution kernel size in text encoder.
+            text_encoder_blocks (int): Number of conformer blocks in text encoder.
+            text_encoder_dropout_rate (float): Dropout rate in conformer block of
+                text encoder.
+            decoder_kernel_size (int): Decoder kernel size.
+            decoder_channels (int): Number of decoder initial channels.
+            decoder_upsample_scales (List[int]): List of upsampling scales in decoder.
+            decoder_upsample_kernel_sizes (List[int]): List of kernel size for
+                upsampling layers in decoder.
+            decoder_resblock_kernel_sizes (List[int]): List of kernel size for resblocks
+                in decoder.
+            decoder_resblock_dilations (List[List[int]]): List of list of dilations for
+                resblocks in decoder.
+            use_weight_norm_in_decoder (bool): Whether to apply weight normalization in
+                decoder.
+            posterior_encoder_kernel_size (int): Posterior encoder kernel size.
+            posterior_encoder_layers (int): Number of layers of posterior encoder.
+            posterior_encoder_stacks (int): Number of stacks of posterior encoder.
+            posterior_encoder_base_dilation (int): Base dilation of posterior encoder.
+            posterior_encoder_dropout_rate (float): Dropout rate for posterior encoder.
+            use_weight_norm_in_posterior_encoder (bool): Whether to apply weight
+                normalization in posterior encoder.
+            flow_flows (int): Number of flows in flow.
+            flow_kernel_size (int): Kernel size in flow.
+            flow_base_dilation (int): Base dilation in flow.
+            flow_layers (int): Number of layers in flow.
+            flow_dropout_rate (float): Dropout rate in flow
+            use_weight_norm_in_flow (bool): Whether to apply weight normalization in
+                flow.
+            use_only_mean_in_flow (bool): Whether to use only mean in flow.
+            stochastic_duration_predictor_kernel_size (int): Kernel size in stochastic
+                duration predictor.
+            stochastic_duration_predictor_dropout_rate (float): Dropout rate in
+                stochastic duration predictor.
+            stochastic_duration_predictor_flows (int): Number of flows in stochastic
+                duration predictor.
+            stochastic_duration_predictor_dds_conv_layers (int): Number of DDS conv
+                layers in stochastic duration predictor.
+
+        """
+        super().__init__()
+        self.segment_size = segment_size
+        self.text_encoder = TextEncoder(
+            vocabs=vocabs,
+            d_model=hidden_channels,
+            num_heads=text_encoder_attention_heads,
+            dim_feedforward=hidden_channels * text_encoder_ffn_expand,
+            cnn_module_kernel=text_encoder_cnn_module_kernel,
+            num_layers=text_encoder_blocks,
+            dropout=text_encoder_dropout_rate,
+        )
+        self.decoder = HiFiGANGenerator(
+            in_channels=hidden_channels,
+            out_channels=1,
+            channels=decoder_channels,
+            global_channels=global_channels,
+            kernel_size=decoder_kernel_size,
+            upsample_scales=decoder_upsample_scales,
+            upsample_kernel_sizes=decoder_upsample_kernel_sizes,
+            resblock_kernel_sizes=decoder_resblock_kernel_sizes,
+            resblock_dilations=decoder_resblock_dilations,
+            use_weight_norm=use_weight_norm_in_decoder,
+        )
+        self.posterior_encoder = PosteriorEncoder(
+            in_channels=aux_channels,
+            out_channels=hidden_channels,
+            hidden_channels=hidden_channels,
+            kernel_size=posterior_encoder_kernel_size,
+            layers=posterior_encoder_layers,
+            stacks=posterior_encoder_stacks,
+            base_dilation=posterior_encoder_base_dilation,
+            global_channels=global_channels,
+            dropout_rate=posterior_encoder_dropout_rate,
+            use_weight_norm=use_weight_norm_in_posterior_encoder,
+        )
+        self.flow = ResidualAffineCouplingBlock(
+            in_channels=hidden_channels,
+            hidden_channels=hidden_channels,
+            flows=flow_flows,
+            kernel_size=flow_kernel_size,
+            base_dilation=flow_base_dilation,
+            layers=flow_layers,
+            global_channels=global_channels,
+            dropout_rate=flow_dropout_rate,
+            use_weight_norm=use_weight_norm_in_flow,
+            use_only_mean=use_only_mean_in_flow,
+        )
+        # TODO(kan-bayashi): Add deterministic version as an option
+        self.duration_predictor = StochasticDurationPredictor(
+            channels=hidden_channels,
+            kernel_size=stochastic_duration_predictor_kernel_size,
+            dropout_rate=stochastic_duration_predictor_dropout_rate,
+            flows=stochastic_duration_predictor_flows,
+            dds_conv_layers=stochastic_duration_predictor_dds_conv_layers,
+            global_channels=global_channels,
+        )
+
+        self.upsample_factor = int(np.prod(decoder_upsample_scales))
+        self.spks = None
+        if spks is not None and spks > 1:
+            assert global_channels > 0
+            self.spks = spks
+            self.global_emb = torch.nn.Embedding(spks, global_channels)
+        self.spk_embed_dim = None
+        if spk_embed_dim is not None and spk_embed_dim > 0:
+            assert global_channels > 0
+            self.spk_embed_dim = spk_embed_dim
+            self.spemb_proj = torch.nn.Linear(spk_embed_dim, global_channels)
+        self.langs = None
+        if langs is not None and langs > 1:
+            assert global_channels > 0
+            self.langs = langs
+            self.lang_emb = torch.nn.Embedding(langs, global_channels)
+
+        # delayed import
+        from monotonic_align import maximum_path
+
+        self.maximum_path = maximum_path
+
+    def forward(
+        self,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        feats: torch.Tensor,
+        feats_lengths: torch.Tensor,
+        sids: Optional[torch.Tensor] = None,
+        spembs: Optional[torch.Tensor] = None,
+        lids: Optional[torch.Tensor] = None,
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        Tuple[
+            torch.Tensor,
+            torch.Tensor,
+            torch.Tensor,
+            torch.Tensor,
+            torch.Tensor,
+            torch.Tensor,
+        ],
+    ]:
+        """Calculate forward propagation.
+
+        Args:
+            text (Tensor): Text index tensor (B, T_text).
+            text_lengths (Tensor): Text length tensor (B,).
+            feats (Tensor): Feature tensor (B, aux_channels, T_feats).
+            feats_lengths (Tensor): Feature length tensor (B,).
+            sids (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]): Language index tensor (B,) or (B, 1).
+
+        Returns:
+            Tensor: Waveform tensor (B, 1, segment_size * upsample_factor).
+            Tensor: Duration negative log-likelihood (NLL) tensor (B,).
+            Tensor: Monotonic attention weight tensor (B, 1, T_feats, T_text).
+            Tensor: Segments start index tensor (B,).
+            Tensor: Text mask tensor (B, 1, T_text).
+            Tensor: Feature mask tensor (B, 1, T_feats).
+            tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+                - Tensor: Posterior encoder hidden representation (B, H, T_feats).
+                - Tensor: Flow hidden representation (B, H, T_feats).
+                - Tensor: Expanded text encoder projected mean (B, H, T_feats).
+                - Tensor: Expanded text encoder projected scale (B, H, T_feats).
+                - Tensor: Posterior encoder projected mean (B, H, T_feats).
+                - Tensor: Posterior encoder projected scale (B, H, T_feats).
+
+        """
+        # forward text encoder
+        x, m_p, logs_p, x_mask = self.text_encoder(text, text_lengths)
+
+        # calculate global conditioning
+        g = None
+        if self.spks is not None:
+            # speaker one-hot vector embedding: (B, global_channels, 1)
+            g = self.global_emb(sids.view(-1)).unsqueeze(-1)
+        if self.spk_embed_dim is not None:
+            # pretreined speaker embedding, e.g., X-vector (B, global_channels, 1)
+            g_ = self.spemb_proj(F.normalize(spembs)).unsqueeze(-1)
+            if g is None:
+                g = g_
+            else:
+                g = g + g_
+        if self.langs is not None:
+            # language one-hot vector embedding: (B, global_channels, 1)
+            g_ = self.lang_emb(lids.view(-1)).unsqueeze(-1)
+            if g is None:
+                g = g_
+            else:
+                g = g + g_
+
+        # forward posterior encoder
+        z, m_q, logs_q, y_mask = self.posterior_encoder(feats, feats_lengths, g=g)
+
+        # forward flow
+        z_p = self.flow(z, y_mask, g=g)  # (B, H, T_feats)
+
+        # monotonic alignment search
+        with torch.no_grad():
+            # negative cross-entropy
+            s_p_sq_r = torch.exp(-2 * logs_p)  # (B, H, T_text)
+            # (B, 1, T_text)
+            neg_x_ent_1 = torch.sum(
+                -0.5 * math.log(2 * math.pi) - logs_p,
+                [1],
+                keepdim=True,
+            )
+            # (B, T_feats, H) x (B, H, T_text) = (B, T_feats, T_text)
+            neg_x_ent_2 = torch.matmul(
+                -0.5 * (z_p**2).transpose(1, 2),
+                s_p_sq_r,
+            )
+            # (B, T_feats, H) x (B, H, T_text) = (B, T_feats, T_text)
+            neg_x_ent_3 = torch.matmul(
+                z_p.transpose(1, 2),
+                (m_p * s_p_sq_r),
+            )
+            # (B, 1, T_text)
+            neg_x_ent_4 = torch.sum(
+                -0.5 * (m_p**2) * s_p_sq_r,
+                [1],
+                keepdim=True,
+            )
+            # (B, T_feats, T_text)
+            neg_x_ent = neg_x_ent_1 + neg_x_ent_2 + neg_x_ent_3 + neg_x_ent_4
+            # (B, 1, T_feats, T_text)
+            attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+            # monotonic attention weight: (B, 1, T_feats, T_text)
+            attn = (
+                self.maximum_path(
+                    neg_x_ent,
+                    attn_mask.squeeze(1),
+                )
+                .unsqueeze(1)
+                .detach()
+            )
+
+        # forward duration predictor
+        w = attn.sum(2)  # (B, 1, T_text)
+        dur_nll = self.duration_predictor(x, x_mask, w=w, g=g)
+        dur_nll = dur_nll / torch.sum(x_mask)
+
+        # expand the length to match with the feature sequence
+        # (B, T_feats, T_text) x (B, T_text, H) -> (B, H, T_feats)
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
+        # (B, T_feats, T_text) x (B, T_text, H) -> (B, H, T_feats)
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
+
+        # get random segments
+        z_segments, z_start_idxs = get_random_segments(
+            z,
+            feats_lengths,
+            self.segment_size,
+        )
+
+        # forward decoder with random segments
+        wav = self.decoder(z_segments, g=g)
+
+        return (
+            wav,
+            dur_nll,
+            attn,
+            z_start_idxs,
+            x_mask,
+            y_mask,
+            (z, z_p, m_p, logs_p, m_q, logs_q),
+        )
+
+    def inference(
+        self,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        feats: Optional[torch.Tensor] = None,
+        feats_lengths: Optional[torch.Tensor] = None,
+        sids: Optional[torch.Tensor] = None,
+        spembs: Optional[torch.Tensor] = None,
+        lids: Optional[torch.Tensor] = None,
+        dur: Optional[torch.Tensor] = None,
+        noise_scale: float = 0.667,
+        noise_scale_dur: float = 0.8,
+        alpha: float = 1.0,
+        max_len: Optional[int] = None,
+        use_teacher_forcing: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Run inference.
+
+        Args:
+            text (Tensor): Input text index tensor (B, T_text,).
+            text_lengths (Tensor): Text length tensor (B,).
+            feats (Tensor): Feature tensor (B, aux_channels, T_feats,).
+            feats_lengths (Tensor): Feature length tensor (B,).
+            sids (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]): Language index tensor (B,) or (B, 1).
+            dur (Optional[Tensor]): Ground-truth duration (B, T_text,). If provided,
+                skip the prediction of durations (i.e., teacher forcing).
+            noise_scale (float): Noise scale parameter for flow.
+            noise_scale_dur (float): Noise scale parameter for duration predictor.
+            alpha (float): Alpha parameter to control the speed of generated speech.
+            max_len (Optional[int]): Maximum length of acoustic feature sequence.
+            use_teacher_forcing (bool): Whether to use teacher forcing.
+
+        Returns:
+            Tensor: Generated waveform tensor (B, T_wav).
+            Tensor: Monotonic attention weight tensor (B, T_feats, T_text).
+            Tensor: Duration tensor (B, T_text).
+
+        """
+        # encoder
+        x, m_p, logs_p, x_mask = self.text_encoder(text, text_lengths)
+        x_mask = x_mask.to(x.dtype)
+        g = None
+        if self.spks is not None:
+            # (B, global_channels, 1)
+            g = self.global_emb(sids.view(-1)).unsqueeze(-1)
+        if self.spk_embed_dim is not None:
+            # (B, global_channels, 1)
+            g_ = self.spemb_proj(F.normalize(spembs.unsqueeze(0))).unsqueeze(-1)
+            if g is None:
+                g = g_
+            else:
+                g = g + g_
+        if self.langs is not None:
+            # (B, global_channels, 1)
+            g_ = self.lang_emb(lids.view(-1)).unsqueeze(-1)
+            if g is None:
+                g = g_
+            else:
+                g = g + g_
+
+        if use_teacher_forcing:
+            # forward posterior encoder
+            z, m_q, logs_q, y_mask = self.posterior_encoder(feats, feats_lengths, g=g)
+
+            # forward flow
+            z_p = self.flow(z, y_mask, g=g)  # (B, H, T_feats)
+
+            # monotonic alignment search
+            s_p_sq_r = torch.exp(-2 * logs_p)  # (B, H, T_text)
+            # (B, 1, T_text)
+            neg_x_ent_1 = torch.sum(
+                -0.5 * math.log(2 * math.pi) - logs_p,
+                [1],
+                keepdim=True,
+            )
+            # (B, T_feats, H) x (B, H, T_text) = (B, T_feats, T_text)
+            neg_x_ent_2 = torch.matmul(
+                -0.5 * (z_p**2).transpose(1, 2),
+                s_p_sq_r,
+            )
+            # (B, T_feats, H) x (B, H, T_text) = (B, T_feats, T_text)
+            neg_x_ent_3 = torch.matmul(
+                z_p.transpose(1, 2),
+                (m_p * s_p_sq_r),
+            )
+            # (B, 1, T_text)
+            neg_x_ent_4 = torch.sum(
+                -0.5 * (m_p**2) * s_p_sq_r,
+                [1],
+                keepdim=True,
+            )
+            # (B, T_feats, T_text)
+            neg_x_ent = neg_x_ent_1 + neg_x_ent_2 + neg_x_ent_3 + neg_x_ent_4
+            # (B, 1, T_feats, T_text)
+            attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+            # monotonic attention weight: (B, 1, T_feats, T_text)
+            attn = self.maximum_path(
+                neg_x_ent,
+                attn_mask.squeeze(1),
+            ).unsqueeze(1)
+            dur = attn.sum(2)  # (B, 1, T_text)
+
+            # forward decoder with random segments
+            wav = self.decoder(z * y_mask, g=g)
+        else:
+            # duration
+            if dur is None:
+                logw = self.duration_predictor(
+                    x,
+                    x_mask,
+                    g=g,
+                    inverse=True,
+                    noise_scale=noise_scale_dur,
+                )
+                w = torch.exp(logw) * x_mask * alpha
+                dur = torch.ceil(w)
+            y_lengths = torch.clamp_min(torch.sum(dur, [1, 2]), 1).long()
+            y_mask = (~make_pad_mask(y_lengths)).unsqueeze(1).to(text.device)
+            y_mask = y_mask.to(x.dtype)
+            attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+            attn = self._generate_path(dur, attn_mask)
+
+            # expand the length to match with the feature sequence
+            # (B, T_feats, T_text) x (B, T_text, H) -> (B, H, T_feats)
+            m_p = torch.matmul(
+                attn.squeeze(1),
+                m_p.transpose(1, 2),
+            ).transpose(1, 2)
+            # (B, T_feats, T_text) x (B, T_text, H) -> (B, H, T_feats)
+            logs_p = torch.matmul(
+                attn.squeeze(1),
+                logs_p.transpose(1, 2),
+            ).transpose(1, 2)
+
+            # decoder
+            z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
+            z = self.flow(z_p, y_mask, g=g, inverse=True)
+            wav = self.decoder((z * y_mask)[:, :, :max_len], g=g)
+
+        return wav.squeeze(1), attn.squeeze(1), dur.squeeze(1)
+
+    def _generate_path(self, dur: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
+        """Generate path a.k.a. monotonic attention.
+
+        Args:
+            dur (Tensor): Duration tensor (B, 1, T_text).
+            mask (Tensor): Attention mask tensor (B, 1, T_feats, T_text).
+
+        Returns:
+            Tensor: Path tensor (B, 1, T_feats, T_text).
+
+        """
+        b, _, t_y, t_x = mask.shape
+        cum_dur = torch.cumsum(dur, -1)
+        cum_dur_flat = cum_dur.view(b * t_x)
+        path = torch.arange(t_y, dtype=dur.dtype, device=dur.device)
+        path = path.unsqueeze(0) < cum_dur_flat.unsqueeze(1)
+        # path = path.view(b, t_x, t_y).to(dtype=mask.dtype)
+        path = path.view(b, t_x, t_y).to(dtype=torch.float)
+        # path will be like (t_x = 3, t_y = 5):
+        # [[[1., 1., 0., 0., 0.],      [[[1., 1., 0., 0., 0.],
+        #   [1., 1., 1., 1., 0.],  -->   [0., 0., 1., 1., 0.],
+        #   [1., 1., 1., 1., 1.]]]       [0., 0., 0., 0., 1.]]]
+        path = path - F.pad(path, [0, 0, 1, 0, 0, 0])[:, :-1]
+        # path = path.to(dtype=mask.dtype)
+        return path.unsqueeze(1).transpose(2, 3) * mask
diff --git a/egs/ljspeech/TTS/vits/hifigan.py b/egs/ljspeech/TTS/vits/hifigan.py
new file mode 100644
index 000000000..589ac30f6
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/hifigan.py
@@ -0,0 +1,933 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/hifigan/hifigan.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""HiFi-GAN Modules.
+
+This code is modified from https://github.com/kan-bayashi/ParallelWaveGAN.
+
+"""
+
+import copy
+import logging
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+class HiFiGANGenerator(torch.nn.Module):
+    """HiFiGAN generator module."""
+
+    def __init__(
+        self,
+        in_channels: int = 80,
+        out_channels: int = 1,
+        channels: int = 512,
+        global_channels: int = -1,
+        kernel_size: int = 7,
+        upsample_scales: List[int] = [8, 8, 2, 2],
+        upsample_kernel_sizes: List[int] = [16, 16, 4, 4],
+        resblock_kernel_sizes: List[int] = [3, 7, 11],
+        resblock_dilations: List[List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+        use_additional_convs: bool = True,
+        bias: bool = True,
+        nonlinear_activation: str = "LeakyReLU",
+        nonlinear_activation_params: Dict[str, Any] = {"negative_slope": 0.1},
+        use_weight_norm: bool = True,
+    ):
+        """Initialize HiFiGANGenerator module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            out_channels (int): Number of output channels.
+            channels (int): Number of hidden representation channels.
+            global_channels (int): Number of global conditioning channels.
+            kernel_size (int): Kernel size of initial and final conv layer.
+            upsample_scales (List[int]): List of upsampling scales.
+            upsample_kernel_sizes (List[int]): List of kernel sizes for upsample layers.
+            resblock_kernel_sizes (List[int]): List of kernel sizes for residual blocks.
+            resblock_dilations (List[List[int]]): List of list of dilations for residual
+                blocks.
+            use_additional_convs (bool): Whether to use additional conv layers in
+                residual blocks.
+            bias (bool): Whether to add bias parameter in convolution layers.
+            nonlinear_activation (str): Activation function module name.
+            nonlinear_activation_params (Dict[str, Any]): Hyperparameters for activation
+                function.
+            use_weight_norm (bool): Whether to use weight norm. If set to true, it will
+                be applied to all of the conv layers.
+
+        """
+        super().__init__()
+
+        # check hyperparameters are valid
+        assert kernel_size % 2 == 1, "Kernel size must be odd number."
+        assert len(upsample_scales) == len(upsample_kernel_sizes)
+        assert len(resblock_dilations) == len(resblock_kernel_sizes)
+
+        # define modules
+        self.upsample_factor = int(np.prod(upsample_scales) * out_channels)
+        self.num_upsamples = len(upsample_kernel_sizes)
+        self.num_blocks = len(resblock_kernel_sizes)
+        self.input_conv = torch.nn.Conv1d(
+            in_channels,
+            channels,
+            kernel_size,
+            1,
+            padding=(kernel_size - 1) // 2,
+        )
+        self.upsamples = torch.nn.ModuleList()
+        self.blocks = torch.nn.ModuleList()
+        for i in range(len(upsample_kernel_sizes)):
+            assert upsample_kernel_sizes[i] == 2 * upsample_scales[i]
+            self.upsamples += [
+                torch.nn.Sequential(
+                    getattr(torch.nn, nonlinear_activation)(
+                        **nonlinear_activation_params
+                    ),
+                    torch.nn.ConvTranspose1d(
+                        channels // (2**i),
+                        channels // (2 ** (i + 1)),
+                        upsample_kernel_sizes[i],
+                        upsample_scales[i],
+                        padding=upsample_scales[i] // 2 + upsample_scales[i] % 2,
+                        output_padding=upsample_scales[i] % 2,
+                    ),
+                )
+            ]
+            for j in range(len(resblock_kernel_sizes)):
+                self.blocks += [
+                    ResidualBlock(
+                        kernel_size=resblock_kernel_sizes[j],
+                        channels=channels // (2 ** (i + 1)),
+                        dilations=resblock_dilations[j],
+                        bias=bias,
+                        use_additional_convs=use_additional_convs,
+                        nonlinear_activation=nonlinear_activation,
+                        nonlinear_activation_params=nonlinear_activation_params,
+                    )
+                ]
+        self.output_conv = torch.nn.Sequential(
+            # NOTE(kan-bayashi): follow official implementation but why
+            #   using different slope parameter here? (0.1 vs. 0.01)
+            torch.nn.LeakyReLU(),
+            torch.nn.Conv1d(
+                channels // (2 ** (i + 1)),
+                out_channels,
+                kernel_size,
+                1,
+                padding=(kernel_size - 1) // 2,
+            ),
+            torch.nn.Tanh(),
+        )
+        if global_channels > 0:
+            self.global_conv = torch.nn.Conv1d(global_channels, channels, 1)
+
+        # apply weight norm
+        if use_weight_norm:
+            self.apply_weight_norm()
+
+        # reset parameters
+        self.reset_parameters()
+
+    def forward(
+        self, c: torch.Tensor, g: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            c (Tensor): Input tensor (B, in_channels, T).
+            g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1).
+
+        Returns:
+            Tensor: Output tensor (B, out_channels, T).
+
+        """
+        c = self.input_conv(c)
+        if g is not None:
+            c = c + self.global_conv(g)
+        for i in range(self.num_upsamples):
+            c = self.upsamples[i](c)
+            cs = 0.0  # initialize
+            for j in range(self.num_blocks):
+                cs += self.blocks[i * self.num_blocks + j](c)
+            c = cs / self.num_blocks
+        c = self.output_conv(c)
+
+        return c
+
+    def reset_parameters(self):
+        """Reset parameters.
+
+        This initialization follows the official implementation manner.
+        https://github.com/jik876/hifi-gan/blob/master/models.py
+
+        """
+
+        def _reset_parameters(m: torch.nn.Module):
+            if isinstance(m, (torch.nn.Conv1d, torch.nn.ConvTranspose1d)):
+                m.weight.data.normal_(0.0, 0.01)
+                logging.debug(f"Reset parameters in {m}.")
+
+        self.apply(_reset_parameters)
+
+    def remove_weight_norm(self):
+        """Remove weight normalization module from all of the layers."""
+
+        def _remove_weight_norm(m: torch.nn.Module):
+            try:
+                logging.debug(f"Weight norm is removed from {m}.")
+                torch.nn.utils.remove_weight_norm(m)
+            except ValueError:  # this module didn't have weight norm
+                return
+
+        self.apply(_remove_weight_norm)
+
+    def apply_weight_norm(self):
+        """Apply weight normalization module from all of the layers."""
+
+        def _apply_weight_norm(m: torch.nn.Module):
+            if isinstance(m, torch.nn.Conv1d) or isinstance(
+                m, torch.nn.ConvTranspose1d
+            ):
+                torch.nn.utils.weight_norm(m)
+                logging.debug(f"Weight norm is applied to {m}.")
+
+        self.apply(_apply_weight_norm)
+
+    def inference(
+        self, c: torch.Tensor, g: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        """Perform inference.
+
+        Args:
+            c (torch.Tensor): Input tensor (T, in_channels).
+            g (Optional[Tensor]): Global conditioning tensor (global_channels, 1).
+
+        Returns:
+            Tensor: Output tensor (T ** upsample_factor, out_channels).
+
+        """
+        if g is not None:
+            g = g.unsqueeze(0)
+        c = self.forward(c.transpose(1, 0).unsqueeze(0), g=g)
+        return c.squeeze(0).transpose(1, 0)
+
+
+class ResidualBlock(torch.nn.Module):
+    """Residual block module in HiFiGAN."""
+
+    def __init__(
+        self,
+        kernel_size: int = 3,
+        channels: int = 512,
+        dilations: List[int] = [1, 3, 5],
+        bias: bool = True,
+        use_additional_convs: bool = True,
+        nonlinear_activation: str = "LeakyReLU",
+        nonlinear_activation_params: Dict[str, Any] = {"negative_slope": 0.1},
+    ):
+        """Initialize ResidualBlock module.
+
+        Args:
+            kernel_size (int): Kernel size of dilation convolution layer.
+            channels (int): Number of channels for convolution layer.
+            dilations (List[int]): List of dilation factors.
+            use_additional_convs (bool): Whether to use additional convolution layers.
+            bias (bool): Whether to add bias parameter in convolution layers.
+            nonlinear_activation (str): Activation function module name.
+            nonlinear_activation_params (Dict[str, Any]): Hyperparameters for activation
+                function.
+
+        """
+        super().__init__()
+        self.use_additional_convs = use_additional_convs
+        self.convs1 = torch.nn.ModuleList()
+        if use_additional_convs:
+            self.convs2 = torch.nn.ModuleList()
+        assert kernel_size % 2 == 1, "Kernel size must be odd number."
+        for dilation in dilations:
+            self.convs1 += [
+                torch.nn.Sequential(
+                    getattr(torch.nn, nonlinear_activation)(
+                        **nonlinear_activation_params
+                    ),
+                    torch.nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation,
+                        bias=bias,
+                        padding=(kernel_size - 1) // 2 * dilation,
+                    ),
+                )
+            ]
+            if use_additional_convs:
+                self.convs2 += [
+                    torch.nn.Sequential(
+                        getattr(torch.nn, nonlinear_activation)(
+                            **nonlinear_activation_params
+                        ),
+                        torch.nn.Conv1d(
+                            channels,
+                            channels,
+                            kernel_size,
+                            1,
+                            dilation=1,
+                            bias=bias,
+                            padding=(kernel_size - 1) // 2,
+                        ),
+                    )
+                ]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, channels, T).
+
+        Returns:
+            Tensor: Output tensor (B, channels, T).
+
+        """
+        for idx in range(len(self.convs1)):
+            xt = self.convs1[idx](x)
+            if self.use_additional_convs:
+                xt = self.convs2[idx](xt)
+            x = xt + x
+        return x
+
+
+class HiFiGANPeriodDiscriminator(torch.nn.Module):
+    """HiFiGAN period discriminator module."""
+
+    def __init__(
+        self,
+        in_channels: int = 1,
+        out_channels: int = 1,
+        period: int = 3,
+        kernel_sizes: List[int] = [5, 3],
+        channels: int = 32,
+        downsample_scales: List[int] = [3, 3, 3, 3, 1],
+        max_downsample_channels: int = 1024,
+        bias: bool = True,
+        nonlinear_activation: str = "LeakyReLU",
+        nonlinear_activation_params: Dict[str, Any] = {"negative_slope": 0.1},
+        use_weight_norm: bool = True,
+        use_spectral_norm: bool = False,
+    ):
+        """Initialize HiFiGANPeriodDiscriminator module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            out_channels (int): Number of output channels.
+            period (int): Period.
+            kernel_sizes (list): Kernel sizes of initial conv layers and the final conv
+                layer.
+            channels (int): Number of initial channels.
+            downsample_scales (List[int]): List of downsampling scales.
+            max_downsample_channels (int): Number of maximum downsampling channels.
+            use_additional_convs (bool): Whether to use additional conv layers in
+                residual blocks.
+            bias (bool): Whether to add bias parameter in convolution layers.
+            nonlinear_activation (str): Activation function module name.
+            nonlinear_activation_params (Dict[str, Any]): Hyperparameters for activation
+                function.
+            use_weight_norm (bool): Whether to use weight norm.
+                If set to true, it will be applied to all of the conv layers.
+            use_spectral_norm (bool): Whether to use spectral norm.
+                If set to true, it will be applied to all of the conv layers.
+
+        """
+        super().__init__()
+        assert len(kernel_sizes) == 2
+        assert kernel_sizes[0] % 2 == 1, "Kernel size must be odd number."
+        assert kernel_sizes[1] % 2 == 1, "Kernel size must be odd number."
+
+        self.period = period
+        self.convs = torch.nn.ModuleList()
+        in_chs = in_channels
+        out_chs = channels
+        for downsample_scale in downsample_scales:
+            self.convs += [
+                torch.nn.Sequential(
+                    torch.nn.Conv2d(
+                        in_chs,
+                        out_chs,
+                        (kernel_sizes[0], 1),
+                        (downsample_scale, 1),
+                        padding=((kernel_sizes[0] - 1) // 2, 0),
+                    ),
+                    getattr(torch.nn, nonlinear_activation)(
+                        **nonlinear_activation_params
+                    ),
+                )
+            ]
+            in_chs = out_chs
+            # NOTE(kan-bayashi): Use downsample_scale + 1?
+            out_chs = min(out_chs * 4, max_downsample_channels)
+        self.output_conv = torch.nn.Conv2d(
+            out_chs,
+            out_channels,
+            (kernel_sizes[1] - 1, 1),
+            1,
+            padding=((kernel_sizes[1] - 1) // 2, 0),
+        )
+
+        if use_weight_norm and use_spectral_norm:
+            raise ValueError("Either use use_weight_norm or use_spectral_norm.")
+
+        # apply weight norm
+        if use_weight_norm:
+            self.apply_weight_norm()
+
+        # apply spectral norm
+        if use_spectral_norm:
+            self.apply_spectral_norm()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            c (Tensor): Input tensor (B, in_channels, T).
+
+        Returns:
+            list: List of each layer's tensors.
+
+        """
+        # transform 1d to 2d -> (B, C, T/P, P)
+        b, c, t = x.shape
+        if t % self.period != 0:
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t += n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        # forward conv
+        outs = []
+        for layer in self.convs:
+            x = layer(x)
+            outs += [x]
+        x = self.output_conv(x)
+        x = torch.flatten(x, 1, -1)
+        outs += [x]
+
+        return outs
+
+    def apply_weight_norm(self):
+        """Apply weight normalization module from all of the layers."""
+
+        def _apply_weight_norm(m: torch.nn.Module):
+            if isinstance(m, torch.nn.Conv2d):
+                torch.nn.utils.weight_norm(m)
+                logging.debug(f"Weight norm is applied to {m}.")
+
+        self.apply(_apply_weight_norm)
+
+    def apply_spectral_norm(self):
+        """Apply spectral normalization module from all of the layers."""
+
+        def _apply_spectral_norm(m: torch.nn.Module):
+            if isinstance(m, torch.nn.Conv2d):
+                torch.nn.utils.spectral_norm(m)
+                logging.debug(f"Spectral norm is applied to {m}.")
+
+        self.apply(_apply_spectral_norm)
+
+
+class HiFiGANMultiPeriodDiscriminator(torch.nn.Module):
+    """HiFiGAN multi-period discriminator module."""
+
+    def __init__(
+        self,
+        periods: List[int] = [2, 3, 5, 7, 11],
+        discriminator_params: Dict[str, Any] = {
+            "in_channels": 1,
+            "out_channels": 1,
+            "kernel_sizes": [5, 3],
+            "channels": 32,
+            "downsample_scales": [3, 3, 3, 3, 1],
+            "max_downsample_channels": 1024,
+            "bias": True,
+            "nonlinear_activation": "LeakyReLU",
+            "nonlinear_activation_params": {"negative_slope": 0.1},
+            "use_weight_norm": True,
+            "use_spectral_norm": False,
+        },
+    ):
+        """Initialize HiFiGANMultiPeriodDiscriminator module.
+
+        Args:
+            periods (List[int]): List of periods.
+            discriminator_params (Dict[str, Any]): Parameters for hifi-gan period
+                discriminator module. The period parameter will be overwritten.
+
+        """
+        super().__init__()
+        self.discriminators = torch.nn.ModuleList()
+        for period in periods:
+            params = copy.deepcopy(discriminator_params)
+            params["period"] = period
+            self.discriminators += [HiFiGANPeriodDiscriminator(**params)]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input noise signal (B, 1, T).
+
+        Returns:
+            List: List of list of each discriminator outputs, which consists of each
+                layer output tensors.
+
+        """
+        outs = []
+        for f in self.discriminators:
+            outs += [f(x)]
+
+        return outs
+
+
+class HiFiGANScaleDiscriminator(torch.nn.Module):
+    """HiFi-GAN scale discriminator module."""
+
+    def __init__(
+        self,
+        in_channels: int = 1,
+        out_channels: int = 1,
+        kernel_sizes: List[int] = [15, 41, 5, 3],
+        channels: int = 128,
+        max_downsample_channels: int = 1024,
+        max_groups: int = 16,
+        bias: int = True,
+        downsample_scales: List[int] = [2, 2, 4, 4, 1],
+        nonlinear_activation: str = "LeakyReLU",
+        nonlinear_activation_params: Dict[str, Any] = {"negative_slope": 0.1},
+        use_weight_norm: bool = True,
+        use_spectral_norm: bool = False,
+    ):
+        """Initilize HiFiGAN scale discriminator module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            out_channels (int): Number of output channels.
+            kernel_sizes (List[int]): List of four kernel sizes. The first will be used
+                for the first conv layer, and the second is for downsampling part, and
+                the remaining two are for the last two output layers.
+            channels (int): Initial number of channels for conv layer.
+            max_downsample_channels (int): Maximum number of channels for downsampling
+                layers.
+            bias (bool): Whether to add bias parameter in convolution layers.
+            downsample_scales (List[int]): List of downsampling scales.
+            nonlinear_activation (str): Activation function module name.
+            nonlinear_activation_params (Dict[str, Any]): Hyperparameters for activation
+                function.
+            use_weight_norm (bool): Whether to use weight norm. If set to true, it will
+                be applied to all of the conv layers.
+            use_spectral_norm (bool): Whether to use spectral norm. If set to true, it
+                will be applied to all of the conv layers.
+
+        """
+        super().__init__()
+        self.layers = torch.nn.ModuleList()
+
+        # check kernel size is valid
+        assert len(kernel_sizes) == 4
+        for ks in kernel_sizes:
+            assert ks % 2 == 1
+
+        # add first layer
+        self.layers += [
+            torch.nn.Sequential(
+                torch.nn.Conv1d(
+                    in_channels,
+                    channels,
+                    # NOTE(kan-bayashi): Use always the same kernel size
+                    kernel_sizes[0],
+                    bias=bias,
+                    padding=(kernel_sizes[0] - 1) // 2,
+                ),
+                getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params),
+            )
+        ]
+
+        # add downsample layers
+        in_chs = channels
+        out_chs = channels
+        # NOTE(kan-bayashi): Remove hard coding?
+        groups = 4
+        for downsample_scale in downsample_scales:
+            self.layers += [
+                torch.nn.Sequential(
+                    torch.nn.Conv1d(
+                        in_chs,
+                        out_chs,
+                        kernel_size=kernel_sizes[1],
+                        stride=downsample_scale,
+                        padding=(kernel_sizes[1] - 1) // 2,
+                        groups=groups,
+                        bias=bias,
+                    ),
+                    getattr(torch.nn, nonlinear_activation)(
+                        **nonlinear_activation_params
+                    ),
+                )
+            ]
+            in_chs = out_chs
+            # NOTE(kan-bayashi): Remove hard coding?
+            out_chs = min(in_chs * 2, max_downsample_channels)
+            # NOTE(kan-bayashi): Remove hard coding?
+            groups = min(groups * 4, max_groups)
+
+        # add final layers
+        out_chs = min(in_chs * 2, max_downsample_channels)
+        self.layers += [
+            torch.nn.Sequential(
+                torch.nn.Conv1d(
+                    in_chs,
+                    out_chs,
+                    kernel_size=kernel_sizes[2],
+                    stride=1,
+                    padding=(kernel_sizes[2] - 1) // 2,
+                    bias=bias,
+                ),
+                getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params),
+            )
+        ]
+        self.layers += [
+            torch.nn.Conv1d(
+                out_chs,
+                out_channels,
+                kernel_size=kernel_sizes[3],
+                stride=1,
+                padding=(kernel_sizes[3] - 1) // 2,
+                bias=bias,
+            ),
+        ]
+
+        if use_weight_norm and use_spectral_norm:
+            raise ValueError("Either use use_weight_norm or use_spectral_norm.")
+
+        # apply weight norm
+        self.use_weight_norm = use_weight_norm
+        if use_weight_norm:
+            self.apply_weight_norm()
+
+        # apply spectral norm
+        self.use_spectral_norm = use_spectral_norm
+        if use_spectral_norm:
+            self.apply_spectral_norm()
+
+        # backward compatibility
+        self._register_load_state_dict_pre_hook(self._load_state_dict_pre_hook)
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input noise signal (B, 1, T).
+
+        Returns:
+            List[Tensor]: List of output tensors of each layer.
+
+        """
+        outs = []
+        for f in self.layers:
+            x = f(x)
+            outs += [x]
+
+        return outs
+
+    def apply_weight_norm(self):
+        """Apply weight normalization module from all of the layers."""
+
+        def _apply_weight_norm(m: torch.nn.Module):
+            if isinstance(m, torch.nn.Conv1d):
+                torch.nn.utils.weight_norm(m)
+                logging.debug(f"Weight norm is applied to {m}.")
+
+        self.apply(_apply_weight_norm)
+
+    def apply_spectral_norm(self):
+        """Apply spectral normalization module from all of the layers."""
+
+        def _apply_spectral_norm(m: torch.nn.Module):
+            if isinstance(m, torch.nn.Conv1d):
+                torch.nn.utils.spectral_norm(m)
+                logging.debug(f"Spectral norm is applied to {m}.")
+
+        self.apply(_apply_spectral_norm)
+
+    def remove_weight_norm(self):
+        """Remove weight normalization module from all of the layers."""
+
+        def _remove_weight_norm(m):
+            try:
+                logging.debug(f"Weight norm is removed from {m}.")
+                torch.nn.utils.remove_weight_norm(m)
+            except ValueError:  # this module didn't have weight norm
+                return
+
+        self.apply(_remove_weight_norm)
+
+    def remove_spectral_norm(self):
+        """Remove spectral normalization module from all of the layers."""
+
+        def _remove_spectral_norm(m):
+            try:
+                logging.debug(f"Spectral norm is removed from {m}.")
+                torch.nn.utils.remove_spectral_norm(m)
+            except ValueError:  # this module didn't have weight norm
+                return
+
+        self.apply(_remove_spectral_norm)
+
+    def _load_state_dict_pre_hook(
+        self,
+        state_dict,
+        prefix,
+        local_metadata,
+        strict,
+        missing_keys,
+        unexpected_keys,
+        error_msgs,
+    ):
+        """Fix the compatibility of weight / spectral normalization issue.
+
+        Some pretrained models are trained with configs that use weight / spectral
+        normalization, but actually, the norm is not applied. This causes the mismatch
+        of the parameters with configs. To solve this issue, when parameter mismatch
+        happens in loading pretrained model, we remove the norm from the current model.
+
+        See also:
+            - https://github.com/espnet/espnet/pull/5240
+            - https://github.com/espnet/espnet/pull/5249
+            - https://github.com/kan-bayashi/ParallelWaveGAN/pull/409
+
+        """
+        current_module_keys = [x for x in state_dict.keys() if x.startswith(prefix)]
+        if self.use_weight_norm and any(
+            [k.endswith("weight") for k in current_module_keys]
+        ):
+            logging.warning(
+                "It seems weight norm is not applied in the pretrained model but the"
+                " current model uses it. To keep the compatibility, we remove the norm"
+                " from the current model. This may cause unexpected behavior due to the"
+                " parameter mismatch in finetuning. To avoid this issue, please change"
+                " the following parameters in config to false:\n"
+                " - discriminator_params.follow_official_norm\n"
+                " - discriminator_params.scale_discriminator_params.use_weight_norm\n"
+                " - discriminator_params.scale_discriminator_params.use_spectral_norm\n"
+                "\n"
+                "See also:\n"
+                " - https://github.com/espnet/espnet/pull/5240\n"
+                " - https://github.com/espnet/espnet/pull/5249"
+            )
+            self.remove_weight_norm()
+            self.use_weight_norm = False
+            for k in current_module_keys:
+                if k.endswith("weight_g") or k.endswith("weight_v"):
+                    del state_dict[k]
+
+        if self.use_spectral_norm and any(
+            [k.endswith("weight") for k in current_module_keys]
+        ):
+            logging.warning(
+                "It seems spectral norm is not applied in the pretrained model but the"
+                " current model uses it. To keep the compatibility, we remove the norm"
+                " from the current model. This may cause unexpected behavior due to the"
+                " parameter mismatch in finetuning. To avoid this issue, please change"
+                " the following parameters in config to false:\n"
+                " - discriminator_params.follow_official_norm\n"
+                " - discriminator_params.scale_discriminator_params.use_weight_norm\n"
+                " - discriminator_params.scale_discriminator_params.use_spectral_norm\n"
+                "\n"
+                "See also:\n"
+                " - https://github.com/espnet/espnet/pull/5240\n"
+                " - https://github.com/espnet/espnet/pull/5249"
+            )
+            self.remove_spectral_norm()
+            self.use_spectral_norm = False
+            for k in current_module_keys:
+                if (
+                    k.endswith("weight_u")
+                    or k.endswith("weight_v")
+                    or k.endswith("weight_orig")
+                ):
+                    del state_dict[k]
+
+
+class HiFiGANMultiScaleDiscriminator(torch.nn.Module):
+    """HiFi-GAN multi-scale discriminator module."""
+
+    def __init__(
+        self,
+        scales: int = 3,
+        downsample_pooling: str = "AvgPool1d",
+        # follow the official implementation setting
+        downsample_pooling_params: Dict[str, Any] = {
+            "kernel_size": 4,
+            "stride": 2,
+            "padding": 2,
+        },
+        discriminator_params: Dict[str, Any] = {
+            "in_channels": 1,
+            "out_channels": 1,
+            "kernel_sizes": [15, 41, 5, 3],
+            "channels": 128,
+            "max_downsample_channels": 1024,
+            "max_groups": 16,
+            "bias": True,
+            "downsample_scales": [2, 2, 4, 4, 1],
+            "nonlinear_activation": "LeakyReLU",
+            "nonlinear_activation_params": {"negative_slope": 0.1},
+        },
+        follow_official_norm: bool = False,
+    ):
+        """Initilize HiFiGAN multi-scale discriminator module.
+
+        Args:
+            scales (int): Number of multi-scales.
+            downsample_pooling (str): Pooling module name for downsampling of the
+                inputs.
+            downsample_pooling_params (Dict[str, Any]): Parameters for the above pooling
+                module.
+            discriminator_params (Dict[str, Any]): Parameters for hifi-gan scale
+                discriminator module.
+            follow_official_norm (bool): Whether to follow the norm setting of the
+                official implementaion. The first discriminator uses spectral norm
+                and the other discriminators use weight norm.
+
+        """
+        super().__init__()
+        self.discriminators = torch.nn.ModuleList()
+
+        # add discriminators
+        for i in range(scales):
+            params = copy.deepcopy(discriminator_params)
+            if follow_official_norm:
+                if i == 0:
+                    params["use_weight_norm"] = False
+                    params["use_spectral_norm"] = True
+                else:
+                    params["use_weight_norm"] = True
+                    params["use_spectral_norm"] = False
+            self.discriminators += [HiFiGANScaleDiscriminator(**params)]
+        self.pooling = None
+        if scales > 1:
+            self.pooling = getattr(torch.nn, downsample_pooling)(
+                **downsample_pooling_params
+            )
+
+    def forward(self, x: torch.Tensor) -> List[List[torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input noise signal (B, 1, T).
+
+        Returns:
+            List[List[torch.Tensor]]: List of list of each discriminator outputs,
+                which consists of eachlayer output tensors.
+
+        """
+        outs = []
+        for f in self.discriminators:
+            outs += [f(x)]
+            if self.pooling is not None:
+                x = self.pooling(x)
+
+        return outs
+
+
+class HiFiGANMultiScaleMultiPeriodDiscriminator(torch.nn.Module):
+    """HiFi-GAN multi-scale + multi-period discriminator module."""
+
+    def __init__(
+        self,
+        # Multi-scale discriminator related
+        scales: int = 3,
+        scale_downsample_pooling: str = "AvgPool1d",
+        scale_downsample_pooling_params: Dict[str, Any] = {
+            "kernel_size": 4,
+            "stride": 2,
+            "padding": 2,
+        },
+        scale_discriminator_params: Dict[str, Any] = {
+            "in_channels": 1,
+            "out_channels": 1,
+            "kernel_sizes": [15, 41, 5, 3],
+            "channels": 128,
+            "max_downsample_channels": 1024,
+            "max_groups": 16,
+            "bias": True,
+            "downsample_scales": [2, 2, 4, 4, 1],
+            "nonlinear_activation": "LeakyReLU",
+            "nonlinear_activation_params": {"negative_slope": 0.1},
+        },
+        follow_official_norm: bool = True,
+        # Multi-period discriminator related
+        periods: List[int] = [2, 3, 5, 7, 11],
+        period_discriminator_params: Dict[str, Any] = {
+            "in_channels": 1,
+            "out_channels": 1,
+            "kernel_sizes": [5, 3],
+            "channels": 32,
+            "downsample_scales": [3, 3, 3, 3, 1],
+            "max_downsample_channels": 1024,
+            "bias": True,
+            "nonlinear_activation": "LeakyReLU",
+            "nonlinear_activation_params": {"negative_slope": 0.1},
+            "use_weight_norm": True,
+            "use_spectral_norm": False,
+        },
+    ):
+        """Initilize HiFiGAN multi-scale + multi-period discriminator module.
+
+        Args:
+            scales (int): Number of multi-scales.
+            scale_downsample_pooling (str): Pooling module name for downsampling of the
+                inputs.
+            scale_downsample_pooling_params (dict): Parameters for the above pooling
+                module.
+            scale_discriminator_params (dict): Parameters for hifi-gan scale
+                discriminator module.
+            follow_official_norm (bool): Whether to follow the norm setting of the
+                official implementaion. The first discriminator uses spectral norm and
+                the other discriminators use weight norm.
+            periods (list): List of periods.
+            period_discriminator_params (dict): Parameters for hifi-gan period
+                discriminator module. The period parameter will be overwritten.
+
+        """
+        super().__init__()
+        self.msd = HiFiGANMultiScaleDiscriminator(
+            scales=scales,
+            downsample_pooling=scale_downsample_pooling,
+            downsample_pooling_params=scale_downsample_pooling_params,
+            discriminator_params=scale_discriminator_params,
+            follow_official_norm=follow_official_norm,
+        )
+        self.mpd = HiFiGANMultiPeriodDiscriminator(
+            periods=periods,
+            discriminator_params=period_discriminator_params,
+        )
+
+    def forward(self, x: torch.Tensor) -> List[List[torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input noise signal (B, 1, T).
+
+        Returns:
+            List[List[Tensor]]: List of list of each discriminator outputs,
+                which consists of each layer output tensors. Multi scale and
+                multi period ones are concatenated.
+
+        """
+        msd_outs = self.msd(x)
+        mpd_outs = self.mpd(x)
+        return msd_outs + mpd_outs
diff --git a/egs/ljspeech/TTS/vits/infer.py b/egs/ljspeech/TTS/vits/infer.py
new file mode 100755
index 000000000..cf0d20ae2
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/infer.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+#
+# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script performs model inference on test set.
+
+Usage:
+./vits/infer.py \
+    --epoch 1000 \
+    --exp-dir ./vits/exp \
+    --max-duration 500
+"""
+
+
+import argparse
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+import torchaudio
+from tokenizer import Tokenizer
+from train import get_model, get_params
+from tts_datamodule import LJSpeechTtsDataModule
+
+from icefall.checkpoint import load_checkpoint
+from icefall.utils import AttributeDict, setup_logger
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=1000,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="vits/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    return parser
+
+
+def infer_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    tokenizer: Tokenizer,
+) -> None:
+    """Decode dataset.
+    The ground-truth and generated audio pairs will be saved to `params.save_wav_dir`.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      tokenizer:
+        Used to convert text to phonemes.
+    """
+    #  Background worker save audios to disk.
+    def _save_worker(
+        batch_size: int,
+        cut_ids: List[str],
+        audio: torch.Tensor,
+        audio_pred: torch.Tensor,
+        audio_lens: List[int],
+        audio_lens_pred: List[int],
+    ):
+        for i in range(batch_size):
+            torchaudio.save(
+                str(params.save_wav_dir / f"{cut_ids[i]}_gt.wav"),
+                audio[i : i + 1, : audio_lens[i]],
+                sample_rate=params.sampling_rate,
+            )
+            torchaudio.save(
+                str(params.save_wav_dir / f"{cut_ids[i]}_pred.wav"),
+                audio_pred[i : i + 1, : audio_lens_pred[i]],
+                sample_rate=params.sampling_rate,
+            )
+
+    device = next(model.parameters()).device
+    num_cuts = 0
+    log_interval = 5
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    futures = []
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        for batch_idx, batch in enumerate(dl):
+            batch_size = len(batch["tokens"])
+
+            tokens = batch["tokens"]
+            tokens = tokenizer.tokens_to_token_ids(tokens)
+            tokens = k2.RaggedTensor(tokens)
+            row_splits = tokens.shape.row_splits(1)
+            tokens_lens = row_splits[1:] - row_splits[:-1]
+            tokens = tokens.to(device)
+            tokens_lens = tokens_lens.to(device)
+            # tensor of shape (B, T)
+            tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id)
+
+            audio = batch["audio"]
+            audio_lens = batch["audio_lens"].tolist()
+            cut_ids = [cut.id for cut in batch["cut"]]
+
+            audio_pred, _, durations = model.inference_batch(
+                text=tokens, text_lengths=tokens_lens
+            )
+            audio_pred = audio_pred.detach().cpu()
+            # convert to samples
+            audio_lens_pred = (
+                (durations.sum(1) * params.frame_shift).to(dtype=torch.int64).tolist()
+            )
+
+            futures.append(
+                executor.submit(
+                    _save_worker,
+                    batch_size,
+                    cut_ids,
+                    audio,
+                    audio_pred,
+                    audio_lens,
+                    audio_lens_pred,
+                )
+            )
+
+            num_cuts += batch_size
+
+            if batch_idx % log_interval == 0:
+                batch_str = f"{batch_idx}/{num_batches}"
+
+                logging.info(
+                    f"batch {batch_str}, cuts processed until now is {num_cuts}"
+                )
+        # return results
+        for f in futures:
+            f.result()
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LJSpeechTtsDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.suffix = f"epoch-{params.epoch}"
+
+    params.res_dir = params.exp_dir / "infer" / params.suffix
+    params.save_wav_dir = params.res_dir / "wav"
+    params.save_wav_dir.mkdir(parents=True, exist_ok=True)
+
+    setup_logger(f"{params.res_dir}/log-infer-{params.suffix}")
+    logging.info("Infer started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.blank_id
+    params.oov_id = tokenizer.oov_id
+    params.vocab_size = tokenizer.vocab_size
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+
+    model.to(device)
+    model.eval()
+
+    num_param_g = sum([p.numel() for p in model.generator.parameters()])
+    logging.info(f"Number of parameters in generator: {num_param_g}")
+    num_param_d = sum([p.numel() for p in model.discriminator.parameters()])
+    logging.info(f"Number of parameters in discriminator: {num_param_d}")
+    logging.info(f"Total number of parameters: {num_param_g + num_param_d}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    ljspeech = LJSpeechTtsDataModule(args)
+
+    test_cuts = ljspeech.test_cuts()
+    test_dl = ljspeech.test_dataloaders(test_cuts)
+
+    infer_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+        tokenizer=tokenizer,
+    )
+
+    logging.info(f"Wav files are saved to {params.save_wav_dir}")
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ljspeech/TTS/vits/loss.py b/egs/ljspeech/TTS/vits/loss.py
new file mode 100644
index 000000000..2f4dc9bc0
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/loss.py
@@ -0,0 +1,335 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/hifigan/loss.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""HiFiGAN-related loss modules.
+
+This code is modified from https://github.com/kan-bayashi/ParallelWaveGAN.
+
+"""
+
+from typing import List, Tuple, Union
+
+import torch
+import torch.distributions as D
+import torch.nn.functional as F
+from lhotse.features.kaldi import Wav2LogFilterBank
+
+
+class GeneratorAdversarialLoss(torch.nn.Module):
+    """Generator adversarial loss module."""
+
+    def __init__(
+        self,
+        average_by_discriminators: bool = True,
+        loss_type: str = "mse",
+    ):
+        """Initialize GeneratorAversarialLoss module.
+
+        Args:
+            average_by_discriminators (bool): Whether to average the loss by
+                the number of discriminators.
+            loss_type (str): Loss type, "mse" or "hinge".
+
+        """
+        super().__init__()
+        self.average_by_discriminators = average_by_discriminators
+        assert loss_type in ["mse", "hinge"], f"{loss_type} is not supported."
+        if loss_type == "mse":
+            self.criterion = self._mse_loss
+        else:
+            self.criterion = self._hinge_loss
+
+    def forward(
+        self,
+        outputs: Union[List[List[torch.Tensor]], List[torch.Tensor], torch.Tensor],
+    ) -> torch.Tensor:
+        """Calcualate generator adversarial loss.
+
+        Args:
+            outputs (Union[List[List[Tensor]], List[Tensor], Tensor]): Discriminator
+                outputs, list of discriminator outputs, or list of list of discriminator
+                outputs..
+
+        Returns:
+            Tensor: Generator adversarial loss value.
+
+        """
+        if isinstance(outputs, (tuple, list)):
+            adv_loss = 0.0
+            for i, outputs_ in enumerate(outputs):
+                if isinstance(outputs_, (tuple, list)):
+                    # NOTE(kan-bayashi): case including feature maps
+                    outputs_ = outputs_[-1]
+                adv_loss += self.criterion(outputs_)
+            if self.average_by_discriminators:
+                adv_loss /= i + 1
+        else:
+            adv_loss = self.criterion(outputs)
+
+        return adv_loss
+
+    def _mse_loss(self, x):
+        return F.mse_loss(x, x.new_ones(x.size()))
+
+    def _hinge_loss(self, x):
+        return -x.mean()
+
+
+class DiscriminatorAdversarialLoss(torch.nn.Module):
+    """Discriminator adversarial loss module."""
+
+    def __init__(
+        self,
+        average_by_discriminators: bool = True,
+        loss_type: str = "mse",
+    ):
+        """Initialize DiscriminatorAversarialLoss module.
+
+        Args:
+            average_by_discriminators (bool): Whether to average the loss by
+                the number of discriminators.
+            loss_type (str): Loss type, "mse" or "hinge".
+
+        """
+        super().__init__()
+        self.average_by_discriminators = average_by_discriminators
+        assert loss_type in ["mse", "hinge"], f"{loss_type} is not supported."
+        if loss_type == "mse":
+            self.fake_criterion = self._mse_fake_loss
+            self.real_criterion = self._mse_real_loss
+        else:
+            self.fake_criterion = self._hinge_fake_loss
+            self.real_criterion = self._hinge_real_loss
+
+    def forward(
+        self,
+        outputs_hat: Union[List[List[torch.Tensor]], List[torch.Tensor], torch.Tensor],
+        outputs: Union[List[List[torch.Tensor]], List[torch.Tensor], torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Calcualate discriminator adversarial loss.
+
+        Args:
+            outputs_hat (Union[List[List[Tensor]], List[Tensor], Tensor]): Discriminator
+                outputs, list of discriminator outputs, or list of list of discriminator
+                outputs calculated from generator.
+            outputs (Union[List[List[Tensor]], List[Tensor], Tensor]): Discriminator
+                outputs, list of discriminator outputs, or list of list of discriminator
+                outputs calculated from groundtruth.
+
+        Returns:
+            Tensor: Discriminator real loss value.
+            Tensor: Discriminator fake loss value.
+
+        """
+        if isinstance(outputs, (tuple, list)):
+            real_loss = 0.0
+            fake_loss = 0.0
+            for i, (outputs_hat_, outputs_) in enumerate(zip(outputs_hat, outputs)):
+                if isinstance(outputs_hat_, (tuple, list)):
+                    # NOTE(kan-bayashi): case including feature maps
+                    outputs_hat_ = outputs_hat_[-1]
+                    outputs_ = outputs_[-1]
+                real_loss += self.real_criterion(outputs_)
+                fake_loss += self.fake_criterion(outputs_hat_)
+            if self.average_by_discriminators:
+                fake_loss /= i + 1
+                real_loss /= i + 1
+        else:
+            real_loss = self.real_criterion(outputs)
+            fake_loss = self.fake_criterion(outputs_hat)
+
+        return real_loss, fake_loss
+
+    def _mse_real_loss(self, x: torch.Tensor) -> torch.Tensor:
+        return F.mse_loss(x, x.new_ones(x.size()))
+
+    def _mse_fake_loss(self, x: torch.Tensor) -> torch.Tensor:
+        return F.mse_loss(x, x.new_zeros(x.size()))
+
+    def _hinge_real_loss(self, x: torch.Tensor) -> torch.Tensor:
+        return -torch.mean(torch.min(x - 1, x.new_zeros(x.size())))
+
+    def _hinge_fake_loss(self, x: torch.Tensor) -> torch.Tensor:
+        return -torch.mean(torch.min(-x - 1, x.new_zeros(x.size())))
+
+
+class FeatureMatchLoss(torch.nn.Module):
+    """Feature matching loss module."""
+
+    def __init__(
+        self,
+        average_by_layers: bool = True,
+        average_by_discriminators: bool = True,
+        include_final_outputs: bool = False,
+    ):
+        """Initialize FeatureMatchLoss module.
+
+        Args:
+            average_by_layers (bool): Whether to average the loss by the number
+                of layers.
+            average_by_discriminators (bool): Whether to average the loss by
+                the number of discriminators.
+            include_final_outputs (bool): Whether to include the final output of
+                each discriminator for loss calculation.
+
+        """
+        super().__init__()
+        self.average_by_layers = average_by_layers
+        self.average_by_discriminators = average_by_discriminators
+        self.include_final_outputs = include_final_outputs
+
+    def forward(
+        self,
+        feats_hat: Union[List[List[torch.Tensor]], List[torch.Tensor]],
+        feats: Union[List[List[torch.Tensor]], List[torch.Tensor]],
+    ) -> torch.Tensor:
+        """Calculate feature matching loss.
+
+        Args:
+            feats_hat (Union[List[List[Tensor]], List[Tensor]]): List of list of
+                discriminator outputs or list of discriminator outputs calcuated
+                from generator's outputs.
+            feats (Union[List[List[Tensor]], List[Tensor]]): List of list of
+                discriminator outputs or list of discriminator outputs calcuated
+                from groundtruth..
+
+        Returns:
+            Tensor: Feature matching loss value.
+
+        """
+        feat_match_loss = 0.0
+        for i, (feats_hat_, feats_) in enumerate(zip(feats_hat, feats)):
+            feat_match_loss_ = 0.0
+            if not self.include_final_outputs:
+                feats_hat_ = feats_hat_[:-1]
+                feats_ = feats_[:-1]
+            for j, (feat_hat_, feat_) in enumerate(zip(feats_hat_, feats_)):
+                feat_match_loss_ += F.l1_loss(feat_hat_, feat_.detach())
+            if self.average_by_layers:
+                feat_match_loss_ /= j + 1
+            feat_match_loss += feat_match_loss_
+        if self.average_by_discriminators:
+            feat_match_loss /= i + 1
+
+        return feat_match_loss
+
+
+class MelSpectrogramLoss(torch.nn.Module):
+    """Mel-spectrogram loss."""
+
+    def __init__(
+        self,
+        sampling_rate: int = 22050,
+        frame_length: int = 1024,  # in samples
+        frame_shift: int = 256,  # in samples
+        n_mels: int = 80,
+        use_fft_mag: bool = True,
+    ):
+        super().__init__()
+        self.wav_to_mel = Wav2LogFilterBank(
+            sampling_rate=sampling_rate,
+            frame_length=frame_length / sampling_rate,  # in second
+            frame_shift=frame_shift / sampling_rate,  # in second
+            use_fft_mag=use_fft_mag,
+            num_filters=n_mels,
+        )
+
+    def forward(
+        self,
+        y_hat: torch.Tensor,
+        y: torch.Tensor,
+        return_mel: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]]:
+        """Calculate Mel-spectrogram loss.
+
+        Args:
+            y_hat (Tensor): Generated waveform tensor (B, 1, T).
+            y (Tensor): Groundtruth waveform tensor (B, 1, T).
+            spec (Optional[Tensor]): Groundtruth linear amplitude spectrum tensor
+                (B, T, n_fft // 2 + 1).  if provided, use it instead of groundtruth
+                waveform.
+
+        Returns:
+            Tensor: Mel-spectrogram loss value.
+
+        """
+        mel_hat = self.wav_to_mel(y_hat.squeeze(1))
+        mel = self.wav_to_mel(y.squeeze(1))
+        mel_loss = F.l1_loss(mel_hat, mel)
+
+        if return_mel:
+            return mel_loss, (mel_hat, mel)
+
+        return mel_loss
+
+
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/loss.py
+
+"""VITS-related loss modules.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+
+class KLDivergenceLoss(torch.nn.Module):
+    """KL divergence loss."""
+
+    def forward(
+        self,
+        z_p: torch.Tensor,
+        logs_q: torch.Tensor,
+        m_p: torch.Tensor,
+        logs_p: torch.Tensor,
+        z_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        """Calculate KL divergence loss.
+
+        Args:
+            z_p (Tensor): Flow hidden representation (B, H, T_feats).
+            logs_q (Tensor): Posterior encoder projected scale (B, H, T_feats).
+            m_p (Tensor): Expanded text encoder projected mean (B, H, T_feats).
+            logs_p (Tensor): Expanded text encoder projected scale (B, H, T_feats).
+            z_mask (Tensor): Mask tensor (B, 1, T_feats).
+
+        Returns:
+            Tensor: KL divergence loss.
+
+        """
+        z_p = z_p.float()
+        logs_q = logs_q.float()
+        m_p = m_p.float()
+        logs_p = logs_p.float()
+        z_mask = z_mask.float()
+        kl = logs_p - logs_q - 0.5
+        kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
+        kl = torch.sum(kl * z_mask)
+        loss = kl / torch.sum(z_mask)
+
+        return loss
+
+
+class KLDivergenceLossWithoutFlow(torch.nn.Module):
+    """KL divergence loss without flow."""
+
+    def forward(
+        self,
+        m_q: torch.Tensor,
+        logs_q: torch.Tensor,
+        m_p: torch.Tensor,
+        logs_p: torch.Tensor,
+    ) -> torch.Tensor:
+        """Calculate KL divergence loss without flow.
+
+        Args:
+            m_q (Tensor): Posterior encoder projected mean (B, H, T_feats).
+            logs_q (Tensor): Posterior encoder projected scale (B, H, T_feats).
+            m_p (Tensor): Expanded text encoder projected mean (B, H, T_feats).
+            logs_p (Tensor): Expanded text encoder projected scale (B, H, T_feats).
+        """
+        posterior_norm = D.Normal(m_q, torch.exp(logs_q))
+        prior_norm = D.Normal(m_p, torch.exp(logs_p))
+        loss = D.kl_divergence(posterior_norm, prior_norm).mean()
+        return loss
diff --git a/egs/ljspeech/TTS/vits/monotonic_align/__init__.py b/egs/ljspeech/TTS/vits/monotonic_align/__init__.py
new file mode 100644
index 000000000..2b35654f5
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/monotonic_align/__init__.py
@@ -0,0 +1,81 @@
+# https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/monotonic_align/__init__.py
+
+"""Maximum path calculation module.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+import warnings
+
+import numpy as np
+import torch
+from numba import njit, prange
+
+try:
+    from .core import maximum_path_c
+
+    is_cython_avalable = True
+except ImportError:
+    is_cython_avalable = False
+    warnings.warn(
+        "Cython version is not available. Fallback to 'EXPERIMETAL' numba version. "
+        "If you want to use the cython version, please build it as follows: "
+        "`cd espnet2/gan_tts/vits/monotonic_align; python setup.py build_ext --inplace`"
+    )
+
+
+def maximum_path(neg_x_ent: torch.Tensor, attn_mask: torch.Tensor) -> torch.Tensor:
+    """Calculate maximum path.
+
+    Args:
+        neg_x_ent (Tensor): Negative X entropy tensor (B, T_feats, T_text).
+        attn_mask (Tensor): Attention mask (B, T_feats, T_text).
+
+    Returns:
+        Tensor: Maximum path tensor (B, T_feats, T_text).
+
+    """
+    device, dtype = neg_x_ent.device, neg_x_ent.dtype
+    neg_x_ent = neg_x_ent.cpu().numpy().astype(np.float32)
+    path = np.zeros(neg_x_ent.shape, dtype=np.int32)
+    t_t_max = attn_mask.sum(1)[:, 0].cpu().numpy().astype(np.int32)
+    t_s_max = attn_mask.sum(2)[:, 0].cpu().numpy().astype(np.int32)
+    if is_cython_avalable:
+        maximum_path_c(path, neg_x_ent, t_t_max, t_s_max)
+    else:
+        maximum_path_numba(path, neg_x_ent, t_t_max, t_s_max)
+
+    return torch.from_numpy(path).to(device=device, dtype=dtype)
+
+
+@njit
+def maximum_path_each_numba(path, value, t_y, t_x, max_neg_val=-np.inf):
+    """Calculate a single maximum path with numba."""
+    index = t_x - 1
+    for y in range(t_y):
+        for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+            if x == y:
+                v_cur = max_neg_val
+            else:
+                v_cur = value[y - 1, x]
+            if x == 0:
+                if y == 0:
+                    v_prev = 0.0
+                else:
+                    v_prev = max_neg_val
+            else:
+                v_prev = value[y - 1, x - 1]
+            value[y, x] += max(v_prev, v_cur)
+
+    for y in range(t_y - 1, -1, -1):
+        path[y, index] = 1
+        if index != 0 and (index == y or value[y - 1, index] < value[y - 1, index - 1]):
+            index = index - 1
+
+
+@njit(parallel=True)
+def maximum_path_numba(paths, values, t_ys, t_xs):
+    """Calculate batch maximum path with numba."""
+    for i in prange(paths.shape[0]):
+        maximum_path_each_numba(paths[i], values[i], t_ys[i], t_xs[i])
diff --git a/egs/ljspeech/TTS/vits/monotonic_align/core.pyx b/egs/ljspeech/TTS/vits/monotonic_align/core.pyx
new file mode 100644
index 000000000..c02c2d02e
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/monotonic_align/core.pyx
@@ -0,0 +1,51 @@
+# https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/monotonic_align/core.pyx
+
+"""Maximum path calculation module with cython optimization.
+
+This code is copied from https://github.com/jaywalnut310/vits and modifed code format.
+
+"""
+
+cimport cython
+
+from cython.parallel import prange
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cdef void maximum_path_each(int[:, ::1] path, float[:, ::1] value, int t_y, int t_x, float max_neg_val=-1e9) nogil:
+    cdef int x
+    cdef int y
+    cdef float v_prev
+    cdef float v_cur
+    cdef float tmp
+    cdef int index = t_x - 1
+
+    for y in range(t_y):
+        for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+            if x == y:
+                v_cur = max_neg_val
+            else:
+                v_cur = value[y - 1, x]
+            if x == 0:
+                if y == 0:
+                    v_prev = 0.0
+                else:
+                    v_prev = max_neg_val
+            else:
+                v_prev = value[y - 1, x - 1]
+            value[y, x] += max(v_prev, v_cur)
+
+    for y in range(t_y - 1, -1, -1):
+        path[y, index] = 1
+        if index != 0 and (index == y or value[y - 1, index] < value[y - 1, index - 1]):
+            index = index - 1
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cpdef void maximum_path_c(int[:, :, ::1] paths, float[:, :, ::1] values, int[::1] t_ys, int[::1] t_xs) nogil:
+    cdef int b = paths.shape[0]
+    cdef int i
+    for i in prange(b, nogil=True):
+        maximum_path_each(paths[i], values[i], t_ys[i], t_xs[i])
diff --git a/egs/ljspeech/TTS/vits/monotonic_align/setup.py b/egs/ljspeech/TTS/vits/monotonic_align/setup.py
new file mode 100644
index 000000000..33d75e176
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/monotonic_align/setup.py
@@ -0,0 +1,31 @@
+# https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/monotonic_align/setup.py
+"""Setup cython code."""
+
+from Cython.Build import cythonize
+from setuptools import Extension, setup
+from setuptools.command.build_ext import build_ext as _build_ext
+
+
+class build_ext(_build_ext):
+    """Overwrite build_ext."""
+
+    def finalize_options(self):
+        """Prevent numpy from thinking it is still in its setup process."""
+        _build_ext.finalize_options(self)
+        __builtins__.__NUMPY_SETUP__ = False
+        import numpy
+
+        self.include_dirs.append(numpy.get_include())
+
+
+exts = [
+    Extension(
+        name="core",
+        sources=["core.pyx"],
+    )
+]
+setup(
+    name="monotonic_align",
+    ext_modules=cythonize(exts, language_level=3),
+    cmdclass={"build_ext": build_ext},
+)
diff --git a/egs/ljspeech/TTS/vits/posterior_encoder.py b/egs/ljspeech/TTS/vits/posterior_encoder.py
new file mode 100644
index 000000000..1104fb864
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/posterior_encoder.py
@@ -0,0 +1,117 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/posterior_encoder.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Posterior encoder module in VITS.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+from typing import Optional, Tuple
+
+import torch
+from wavenet import Conv1d, WaveNet
+
+from icefall.utils import make_pad_mask
+
+
+class PosteriorEncoder(torch.nn.Module):
+    """Posterior encoder module in VITS.
+
+    This is a module of posterior encoder described in `Conditional Variational
+    Autoencoder with Adversarial Learning for End-to-End Text-to-Speech`_.
+
+    .. _`Conditional Variational Autoencoder with Adversarial Learning for End-to-End
+        Text-to-Speech`: https://arxiv.org/abs/2006.04558
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 513,
+        out_channels: int = 192,
+        hidden_channels: int = 192,
+        kernel_size: int = 5,
+        layers: int = 16,
+        stacks: int = 1,
+        base_dilation: int = 1,
+        global_channels: int = -1,
+        dropout_rate: float = 0.0,
+        bias: bool = True,
+        use_weight_norm: bool = True,
+    ):
+        """Initilialize PosteriorEncoder module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            out_channels (int): Number of output channels.
+            hidden_channels (int): Number of hidden channels.
+            kernel_size (int): Kernel size in WaveNet.
+            layers (int): Number of layers of WaveNet.
+            stacks (int): Number of repeat stacking of WaveNet.
+            base_dilation (int): Base dilation factor.
+            global_channels (int): Number of global conditioning channels.
+            dropout_rate (float): Dropout rate.
+            bias (bool): Whether to use bias parameters in conv.
+            use_weight_norm (bool): Whether to apply weight norm.
+
+        """
+        super().__init__()
+
+        # define modules
+        self.input_conv = Conv1d(in_channels, hidden_channels, 1)
+        self.encoder = WaveNet(
+            in_channels=-1,
+            out_channels=-1,
+            kernel_size=kernel_size,
+            layers=layers,
+            stacks=stacks,
+            base_dilation=base_dilation,
+            residual_channels=hidden_channels,
+            aux_channels=-1,
+            gate_channels=hidden_channels * 2,
+            skip_channels=hidden_channels,
+            global_channels=global_channels,
+            dropout_rate=dropout_rate,
+            bias=bias,
+            use_weight_norm=use_weight_norm,
+            use_first_conv=False,
+            use_last_conv=False,
+            scale_residual=False,
+            scale_skip_connect=True,
+        )
+        self.proj = Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(
+        self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, in_channels, T_feats).
+            x_lengths (Tensor): Length tensor (B,).
+            g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1).
+
+        Returns:
+            Tensor: Encoded hidden representation tensor (B, out_channels, T_feats).
+            Tensor: Projected mean tensor (B, out_channels, T_feats).
+            Tensor: Projected scale tensor (B, out_channels, T_feats).
+            Tensor: Mask tensor for input tensor (B, 1, T_feats).
+
+        """
+        x_mask = (
+            (~make_pad_mask(x_lengths))
+            .unsqueeze(1)
+            .to(
+                dtype=x.dtype,
+                device=x.device,
+            )
+        )
+        x = self.input_conv(x) * x_mask
+        x = self.encoder(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = stats.split(stats.size(1) // 2, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+
+        return z, m, logs, x_mask
diff --git a/egs/ljspeech/TTS/vits/residual_coupling.py b/egs/ljspeech/TTS/vits/residual_coupling.py
new file mode 100644
index 000000000..f9a2a3786
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/residual_coupling.py
@@ -0,0 +1,228 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/residual_coupling.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Residual affine coupling modules in VITS.
+
+This code is based on https://github.com/jaywalnut310/vits.
+
+"""
+
+from typing import Optional, Tuple, Union
+
+import torch
+from flow import FlipFlow
+from wavenet import WaveNet
+
+
+class ResidualAffineCouplingBlock(torch.nn.Module):
+    """Residual affine coupling block module.
+
+    This is a module of residual affine coupling block, which used as "Flow" in
+    `Conditional Variational Autoencoder with Adversarial Learning for End-to-End
+    Text-to-Speech`_.
+
+    .. _`Conditional Variational Autoencoder with Adversarial Learning for End-to-End
+        Text-to-Speech`: https://arxiv.org/abs/2006.04558
+
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 192,
+        hidden_channels: int = 192,
+        flows: int = 4,
+        kernel_size: int = 5,
+        base_dilation: int = 1,
+        layers: int = 4,
+        global_channels: int = -1,
+        dropout_rate: float = 0.0,
+        use_weight_norm: bool = True,
+        bias: bool = True,
+        use_only_mean: bool = True,
+    ):
+        """Initilize ResidualAffineCouplingBlock module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            hidden_channels (int): Number of hidden channels.
+            flows (int): Number of flows.
+            kernel_size (int): Kernel size for WaveNet.
+            base_dilation (int): Base dilation factor for WaveNet.
+            layers (int): Number of layers of WaveNet.
+            stacks (int): Number of stacks of WaveNet.
+            global_channels (int): Number of global channels.
+            dropout_rate (float): Dropout rate.
+            use_weight_norm (bool): Whether to use weight normalization in WaveNet.
+            bias (bool): Whether to use bias paramters in WaveNet.
+            use_only_mean (bool): Whether to estimate only mean.
+
+        """
+        super().__init__()
+
+        self.flows = torch.nn.ModuleList()
+        for i in range(flows):
+            self.flows += [
+                ResidualAffineCouplingLayer(
+                    in_channels=in_channels,
+                    hidden_channels=hidden_channels,
+                    kernel_size=kernel_size,
+                    base_dilation=base_dilation,
+                    layers=layers,
+                    stacks=1,
+                    global_channels=global_channels,
+                    dropout_rate=dropout_rate,
+                    use_weight_norm=use_weight_norm,
+                    bias=bias,
+                    use_only_mean=use_only_mean,
+                )
+            ]
+            self.flows += [FlipFlow()]
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        inverse: bool = False,
+    ) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, in_channels, T).
+            x_lengths (Tensor): Length tensor (B,).
+            g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1).
+            inverse (bool): Whether to inverse the flow.
+
+        Returns:
+            Tensor: Output tensor (B, in_channels, T).
+
+        """
+        if not inverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, inverse=inverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, inverse=inverse)
+        return x
+
+
+class ResidualAffineCouplingLayer(torch.nn.Module):
+    """Residual affine coupling layer."""
+
+    def __init__(
+        self,
+        in_channels: int = 192,
+        hidden_channels: int = 192,
+        kernel_size: int = 5,
+        base_dilation: int = 1,
+        layers: int = 5,
+        stacks: int = 1,
+        global_channels: int = -1,
+        dropout_rate: float = 0.0,
+        use_weight_norm: bool = True,
+        bias: bool = True,
+        use_only_mean: bool = True,
+    ):
+        """Initialzie ResidualAffineCouplingLayer module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            hidden_channels (int): Number of hidden channels.
+            kernel_size (int): Kernel size for WaveNet.
+            base_dilation (int): Base dilation factor for WaveNet.
+            layers (int): Number of layers of WaveNet.
+            stacks (int): Number of stacks of WaveNet.
+            global_channels (int): Number of global channels.
+            dropout_rate (float): Dropout rate.
+            use_weight_norm (bool): Whether to use weight normalization in WaveNet.
+            bias (bool): Whether to use bias paramters in WaveNet.
+            use_only_mean (bool): Whether to estimate only mean.
+
+        """
+        assert in_channels % 2 == 0, "in_channels should be divisible by 2"
+        super().__init__()
+        self.half_channels = in_channels // 2
+        self.use_only_mean = use_only_mean
+
+        # define modules
+        self.input_conv = torch.nn.Conv1d(
+            self.half_channels,
+            hidden_channels,
+            1,
+        )
+        self.encoder = WaveNet(
+            in_channels=-1,
+            out_channels=-1,
+            kernel_size=kernel_size,
+            layers=layers,
+            stacks=stacks,
+            base_dilation=base_dilation,
+            residual_channels=hidden_channels,
+            aux_channels=-1,
+            gate_channels=hidden_channels * 2,
+            skip_channels=hidden_channels,
+            global_channels=global_channels,
+            dropout_rate=dropout_rate,
+            bias=bias,
+            use_weight_norm=use_weight_norm,
+            use_first_conv=False,
+            use_last_conv=False,
+            scale_residual=False,
+            scale_skip_connect=True,
+        )
+        if use_only_mean:
+            self.proj = torch.nn.Conv1d(
+                hidden_channels,
+                self.half_channels,
+                1,
+            )
+        else:
+            self.proj = torch.nn.Conv1d(
+                hidden_channels,
+                self.half_channels * 2,
+                1,
+            )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        inverse: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, in_channels, T).
+            x_lengths (Tensor): Length tensor (B,).
+            g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1).
+            inverse (bool): Whether to inverse the flow.
+
+        Returns:
+            Tensor: Output tensor (B, in_channels, T).
+            Tensor: Log-determinant tensor for NLL (B,) if not inverse.
+
+        """
+        xa, xb = x.split(x.size(1) // 2, dim=1)
+        h = self.input_conv(xa) * x_mask
+        h = self.encoder(h, x_mask, g=g)
+        stats = self.proj(h) * x_mask
+        if not self.use_only_mean:
+            m, logs = stats.split(stats.size(1) // 2, dim=1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not inverse:
+            xb = m + xb * torch.exp(logs) * x_mask
+            x = torch.cat([xa, xb], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            xb = (xb - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([xa, xb], 1)
+            return x
diff --git a/egs/ljspeech/TTS/vits/test_onnx.py b/egs/ljspeech/TTS/vits/test_onnx.py
new file mode 100755
index 000000000..fcbc1d663
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/test_onnx.py
@@ -0,0 +1,123 @@
+#!/usr/bin/env python3
+#
+# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script is used to test the exported onnx model by vits/export-onnx.py
+
+Use the onnx model to generate a wav:
+./vits/test_onnx.py \
+  --model-filename vits/exp/vits-epoch-1000.onnx \
+  --tokens data/tokens.txt
+"""
+
+
+import argparse
+import logging
+
+import onnxruntime as ort
+import torch
+import torchaudio
+from tokenizer import Tokenizer
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the onnx model.",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(self, model_filename: str):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 4
+
+        self.session_opts = session_opts
+
+        self.model = ort.InferenceSession(
+            model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        logging.info(f"{self.model.get_modelmeta().custom_metadata_map}")
+
+    def __call__(self, tokens: torch.Tensor, tokens_lens: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          tokens:
+            A 1-D tensor of shape (1, T)
+        Returns:
+            A tensor of shape (1, T')
+        """
+        noise_scale = torch.tensor([0.667], dtype=torch.float32)
+        noise_scale_dur = torch.tensor([0.8], dtype=torch.float32)
+        alpha = torch.tensor([1.0], dtype=torch.float32)
+
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: tokens.numpy(),
+                self.model.get_inputs()[1].name: tokens_lens.numpy(),
+                self.model.get_inputs()[2].name: noise_scale.numpy(),
+                self.model.get_inputs()[3].name: alpha.numpy(),
+                self.model.get_inputs()[4].name: noise_scale_dur.numpy(),
+            },
+        )[0]
+        return torch.from_numpy(out)
+
+
+def main():
+    args = get_parser().parse_args()
+
+    tokenizer = Tokenizer(args.tokens)
+
+    logging.info("About to create onnx model")
+    model = OnnxModel(args.model_filename)
+
+    text = "I went there to see the land, the people and how their system works, end quote."
+    tokens = tokenizer.texts_to_token_ids([text])
+    tokens = torch.tensor(tokens)  # (1, T)
+    tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64)  # (1, T)
+    audio = model(tokens, tokens_lens)  # (1, T')
+
+    torchaudio.save(str("test_onnx.wav"), audio, sample_rate=22050)
+    logging.info("Saved to test_onnx.wav")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/ljspeech/TTS/vits/text_encoder.py b/egs/ljspeech/TTS/vits/text_encoder.py
new file mode 100644
index 000000000..fcbae7103
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/text_encoder.py
@@ -0,0 +1,684 @@
+#!/usr/bin/env python3
+# Copyright           2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Text encoder module in VITS.
+
+This code is based on
+  - https://github.com/jaywalnut310/vits
+  - https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/text_encoder.py
+  - https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/transducer_stateless/conformer.py
+"""
+
+import copy
+import math
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+
+from icefall.utils import is_jit_tracing, make_pad_mask
+
+
+class TextEncoder(torch.nn.Module):
+    """Text encoder module in VITS.
+
+    This is a module of text encoder described in `Conditional Variational Autoencoder
+    with Adversarial Learning for End-to-End Text-to-Speech`.
+    """
+
+    def __init__(
+        self,
+        vocabs: int,
+        d_model: int = 192,
+        num_heads: int = 2,
+        dim_feedforward: int = 768,
+        cnn_module_kernel: int = 5,
+        num_layers: int = 6,
+        dropout: float = 0.1,
+    ):
+        """Initialize TextEncoder module.
+
+        Args:
+            vocabs (int): Vocabulary size.
+            d_model (int): attention dimension
+            num_heads (int): number of attention heads
+            dim_feedforward (int): feedforward dimention
+            cnn_module_kernel (int): convolution kernel size
+            num_layers (int): number of encoder layers
+            dropout (float): dropout rate
+        """
+        super().__init__()
+        self.d_model = d_model
+
+        # define modules
+        self.emb = torch.nn.Embedding(vocabs, d_model)
+        torch.nn.init.normal_(self.emb.weight, 0.0, d_model**-0.5)
+
+        # We use conformer as text encoder
+        self.encoder = Transformer(
+            d_model=d_model,
+            num_heads=num_heads,
+            dim_feedforward=dim_feedforward,
+            cnn_module_kernel=cnn_module_kernel,
+            num_layers=num_layers,
+            dropout=dropout,
+        )
+
+        self.proj = torch.nn.Conv1d(d_model, d_model * 2, 1)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lengths: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input index tensor (B, T_text).
+            x_lengths (Tensor): Length tensor (B,).
+
+        Returns:
+            Tensor: Encoded hidden representation (B, attention_dim, T_text).
+            Tensor: Projected mean tensor (B, attention_dim, T_text).
+            Tensor: Projected scale tensor (B, attention_dim, T_text).
+            Tensor: Mask tensor for input tensor (B, 1, T_text).
+
+        """
+        # (B, T_text, embed_dim)
+        x = self.emb(x) * math.sqrt(self.d_model)
+
+        assert x.size(1) == x_lengths.max().item()
+
+        # (B, T_text)
+        pad_mask = make_pad_mask(x_lengths)
+
+        # encoder assume the channel last (B, T_text, embed_dim)
+        x = self.encoder(x, key_padding_mask=pad_mask)
+
+        # convert the channel first (B, embed_dim, T_text)
+        x = x.transpose(1, 2)
+        non_pad_mask = (~pad_mask).unsqueeze(1)
+        stats = self.proj(x) * non_pad_mask
+        m, logs = stats.split(stats.size(1) // 2, dim=1)
+
+        return x, m, logs, non_pad_mask
+
+
+class Transformer(nn.Module):
+    """
+    Args:
+        d_model (int): attention dimension
+        num_heads (int): number of attention heads
+        dim_feedforward (int): feedforward dimention
+        cnn_module_kernel (int): convolution kernel size
+        num_layers (int): number of encoder layers
+        dropout (float): dropout rate
+    """
+
+    def __init__(
+        self,
+        d_model: int = 192,
+        num_heads: int = 2,
+        dim_feedforward: int = 768,
+        cnn_module_kernel: int = 5,
+        num_layers: int = 6,
+        dropout: float = 0.1,
+    ) -> None:
+        super().__init__()
+
+        self.num_layers = num_layers
+        self.d_model = d_model
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            num_heads=num_heads,
+            dim_feedforward=dim_feedforward,
+            cnn_module_kernel=cnn_module_kernel,
+            dropout=dropout,
+        )
+        self.encoder = TransformerEncoder(encoder_layer, num_layers)
+        self.after_norm = nn.LayerNorm(d_model)
+
+    def forward(
+        self, x: Tensor, key_padding_mask: Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          lengths:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        """
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        x = self.encoder(x, pos_emb, key_padding_mask=key_padding_mask)  # (T, N, C)
+
+        x = self.after_norm(x)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        return x
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    TransformerEncoderLayer is made up of self-attn and feedforward.
+
+    Args:
+        d_model: the number of expected features in the input.
+        num_heads: the number of heads in the multi-head attention models.
+        dim_feedforward: the dimension of the feed-forward network model.
+        dropout: the dropout value (default=0.1).
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        num_heads: int,
+        dim_feedforward: int,
+        cnn_module_kernel: int,
+        dropout: float = 0.1,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.self_attn = RelPositionMultiheadAttention(
+            d_model, num_heads, dropout=dropout
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.norm_ff_macaron = nn.LayerNorm(d_model)  # for the macaron style FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(d_model)  # for the final output of the block
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+
+        self.ff_scale = 0.5
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the transformer encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer, of shape (seq_len, batch_size, embed_dim).
+            pos_emb: Positional embedding tensor, of shape (1, seq_len*2-1, pos_dim).
+            key_padding_mask: the mask for the src keys per batch, of shape (batch_size, seq_len)
+        """
+        # macaron style feed-forward module
+        src = src + self.ff_scale * self.dropout(
+            self.feed_forward_macaron(self.norm_ff_macaron(src))
+        )
+
+        # multi-head self-attention module
+        src_attn = self.self_attn(
+            self.norm_mha(src),
+            pos_emb=pos_emb,
+            key_padding_mask=key_padding_mask,
+        )
+        src = src + self.dropout(src_attn)
+
+        # convolution module
+        src = src + self.dropout(self.conv_module(self.norm_conv(src)))
+
+        # feed-forward module
+        src = src + self.dropout(self.feed_forward(self.norm_ff(src)))
+
+        src = self.norm_final(src)
+
+        return src
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer class.
+        num_layers: the number of sub-encoder-layers in the encoder.
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder layer, of shape (seq_len, batch_size, embed_dim).
+            pos_emb: Positional embedding tensor, of shape (1, seq_len*2-1, pos_dim).
+            key_padding_mask: the mask for the src keys per batch, of shape (batch_size, seq_len)
+        """
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                key_padding_mask=key_padding_mask,
+            )
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        x_size = x.size(1)
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size, self.d_model)
+        pe_negative = torch.zeros(x_size, self.d_model)
+        position = torch.arange(0, x_size, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, seq_len, 2*seq_len-1).
+
+        Returns:
+            Tensor: tensor of shape (batch, head, seq_len, seq_len)
+        """
+        (batch_size, num_heads, seq_len, n) = x.shape
+
+        if not is_jit_tracing():
+            assert n == 2 * seq_len - 1, f"{n} == 2 * {seq_len} - 1"
+
+        if is_jit_tracing():
+            rows = torch.arange(start=seq_len - 1, end=-1, step=-1)
+            cols = torch.arange(seq_len)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, seq_len, seq_len)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, seq_len, seq_len),
+                (batch_stride, head_stride, time_stride - n_stride, n_stride),
+                storage_offset=n_stride * (seq_len - 1),
+            )
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Args:
+            x: Input tensor of shape (seq_len, batch_size, embed_dim)
+            pos_emb: Positional embedding tensor, (1, 2*seq_len-1, pos_dim)
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+                Its shape is (batch_size, seq_len).
+
+        Outputs:
+            A tensor of shape (seq_len, batch_size, embed_dim).
+        """
+        seq_len, batch_size, _ = x.shape
+        scaling = float(self.head_dim) ** -0.5
+
+        q, k, v = self.in_proj(x).chunk(3, dim=-1)
+
+        q = q.contiguous().view(seq_len, batch_size, self.num_heads, self.head_dim)
+        k = k.contiguous().view(seq_len, batch_size, self.num_heads, self.head_dim)
+        v = (
+            v.contiguous()
+            .view(seq_len, batch_size * self.num_heads, self.head_dim)
+            .transpose(0, 1)
+        )
+
+        q = q.transpose(0, 1)  # (batch_size, seq_len, num_head, head_dim)
+
+        p = self.linear_pos(pos_emb).view(
+            pos_emb.size(0), -1, self.num_heads, self.head_dim
+        )
+        # (1, 2*seq_len, num_head, head_dim) -> (1, num_head, head_dim, 2*seq_len-1)
+        p = p.permute(0, 2, 3, 1)
+
+        # (batch_size, num_head, seq_len, head_dim)
+        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
+        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch_size, num_head, head_dim, seq_len)
+        matrix_ac = torch.matmul(
+            q_with_bias_u, k
+        )  # (batch_size, num_head, seq_len, seq_len)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p
+        )  # (batch_size, num_head, seq_len, 2*seq_len-1)
+        matrix_bd = self.rel_shift(
+            matrix_bd
+        )  # (batch_size, num_head, seq_len, seq_len)
+
+        # (batch_size, num_head, seq_len, seq_len)
+        attn_output_weights = (matrix_ac + matrix_bd) * scaling
+        attn_output_weights = attn_output_weights.view(
+            batch_size * self.num_heads, seq_len, seq_len
+        )
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (batch_size, seq_len)
+            attn_output_weights = attn_output_weights.view(
+                batch_size, self.num_heads, seq_len, seq_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                batch_size * self.num_heads, seq_len, seq_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=self.dropout, training=self.training
+        )
+
+        # (batch_size * num_head, seq_len, head_dim)
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert attn_output.shape == (
+            batch_size * self.num_heads,
+            seq_len,
+            self.head_dim,
+        )
+
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(seq_len, batch_size, self.embed_dim)
+        )
+        # (seq_len, batch_size, embed_dim)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        padding = (kernel_size - 1) // 2
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.LayerNorm(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+        # x is (batch, channels, time)
+        x = x.permute(0, 2, 1)
+        x = self.norm(x)
+        x = x.permute(0, 2, 1)
+
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def _test_text_encoder():
+    vocabs = 500
+    d_model = 192
+    batch_size = 5
+    seq_len = 100
+
+    m = TextEncoder(vocabs=vocabs, d_model=d_model)
+    x, m, logs, mask = m(
+        x=torch.randint(low=0, high=vocabs, size=(batch_size, seq_len)),
+        x_lengths=torch.full((batch_size,), seq_len),
+    )
+    print(x.shape, m.shape, logs.shape, mask.shape)
+
+
+if __name__ == "__main__":
+    _test_text_encoder()
diff --git a/egs/ljspeech/TTS/vits/tokenizer.py b/egs/ljspeech/TTS/vits/tokenizer.py
new file mode 100644
index 000000000..70f1240b4
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/tokenizer.py
@@ -0,0 +1,108 @@
+# Copyright      2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict, List
+
+import g2p_en
+import tacotron_cleaner.cleaners
+from utils import intersperse
+
+
+class Tokenizer(object):
+    def __init__(self, tokens: str):
+        """
+        Args:
+            tokens: the file that maps tokens to ids
+        """
+        # Parse token file
+        self.token2id: Dict[str, int] = {}
+        with open(tokens, "r", encoding="utf-8") as f:
+            for line in f.readlines():
+                info = line.rstrip().split()
+                if len(info) == 1:
+                    # case of space
+                    token = " "
+                    id = int(info[0])
+                else:
+                    token, id = info[0], int(info[1])
+                self.token2id[token] = id
+
+        self.blank_id = self.token2id["<blk>"]
+        self.oov_id = self.token2id["<unk>"]
+        self.vocab_size = len(self.token2id)
+
+        self.g2p = g2p_en.G2p()
+
+    def texts_to_token_ids(self, texts: List[str], intersperse_blank: bool = True):
+        """
+        Args:
+          texts:
+            A list of transcripts.
+          intersperse_blank:
+            Whether to intersperse blanks in the token sequence.
+
+        Returns:
+          Return a list of token id list [utterance][token_id]
+        """
+        token_ids_list = []
+
+        for text in texts:
+            # Text normalization
+            text = tacotron_cleaner.cleaners.custom_english_cleaners(text)
+            # Convert to phonemes
+            tokens = self.g2p(text)
+            token_ids = []
+            for t in tokens:
+                if t in self.token2id:
+                    token_ids.append(self.token2id[t])
+                else:
+                    token_ids.append(self.oov_id)
+
+            if intersperse_blank:
+                token_ids = intersperse(token_ids, self.blank_id)
+
+                token_ids_list.append(token_ids)
+
+        return token_ids_list
+
+    def tokens_to_token_ids(
+        self, tokens_list: List[str], intersperse_blank: bool = True
+    ):
+        """
+        Args:
+          tokens_list:
+            A list of token list, each corresponding to one utterance.
+          intersperse_blank:
+            Whether to intersperse blanks in the token sequence.
+
+        Returns:
+          Return a list of token id list [utterance][token_id]
+        """
+        token_ids_list = []
+
+        for tokens in tokens_list:
+            token_ids = []
+            for t in tokens:
+                if t in self.token2id:
+                    token_ids.append(self.token2id[t])
+                else:
+                    token_ids.append(self.oov_id)
+
+            if intersperse_blank:
+                token_ids = intersperse(token_ids, self.blank_id)
+                token_ids_list.append(token_ids)
+
+        return token_ids_list
diff --git a/egs/ljspeech/TTS/vits/train.py b/egs/ljspeech/TTS/vits/train.py
new file mode 100755
index 000000000..71c4224fa
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/train.py
@@ -0,0 +1,926 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import numpy as np
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from tokenizer import Tokenizer
+from torch.cuda.amp import GradScaler, autocast
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import Optimizer
+from torch.utils.tensorboard import SummaryWriter
+from tts_datamodule import LJSpeechTtsDataModule
+from utils import MetricsTracker, plot_feature, save_checkpoint
+from vits import VITS
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+LRSchedulerType = torch.optim.lr_scheduler._LRScheduler
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=1000,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="vits/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    parser.add_argument(
+        "--lr", type=float, default=2.0e-4, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=20,
+        help="""Save checkpoint after processing this number of epochs"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.cur_epoch % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/epoch-{params.cur_epoch}.pt'.
+        Since it will take around 1000 epochs, we suggest using a large
+        save_every_n to save disk space.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            # training params
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": -1,  # 0
+            "log_interval": 50,
+            "valid_interval": 200,
+            "env_info": get_env_info(),
+            "sampling_rate": 22050,
+            "frame_shift": 256,
+            "frame_length": 1024,
+            "feature_dim": 513,  # 1024 // 2 + 1, 1024 is fft_length
+            "n_mels": 80,
+            "lambda_adv": 1.0,  # loss scaling coefficient for adversarial loss
+            "lambda_mel": 45.0,  # loss scaling coefficient for Mel loss
+            "lambda_feat_match": 2.0,  # loss scaling coefficient for feat match loss
+            "lambda_dur": 1.0,  # loss scaling coefficient for duration loss
+            "lambda_kl": 1.0,  # loss scaling coefficient for KL divergence loss
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict, model: nn.Module
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(filename, model=model)
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    mel_loss_params = {
+        "n_mels": params.n_mels,
+        "frame_length": params.frame_length,
+        "frame_shift": params.frame_shift,
+    }
+    model = VITS(
+        vocab_size=params.vocab_size,
+        feature_dim=params.feature_dim,
+        sampling_rate=params.sampling_rate,
+        mel_loss_params=mel_loss_params,
+        lambda_adv=params.lambda_adv,
+        lambda_mel=params.lambda_mel,
+        lambda_feat_match=params.lambda_feat_match,
+        lambda_dur=params.lambda_dur,
+        lambda_kl=params.lambda_kl,
+    )
+    return model
+
+
+def prepare_input(batch: dict, tokenizer: Tokenizer, device: torch.device):
+    """Parse batch data"""
+    audio = batch["audio"].to(device)
+    features = batch["features"].to(device)
+    audio_lens = batch["audio_lens"].to(device)
+    features_lens = batch["features_lens"].to(device)
+    tokens = batch["tokens"]
+
+    tokens = tokenizer.tokens_to_token_ids(tokens)
+    tokens = k2.RaggedTensor(tokens)
+    row_splits = tokens.shape.row_splits(1)
+    tokens_lens = row_splits[1:] - row_splits[:-1]
+    tokens = tokens.to(device)
+    tokens_lens = tokens_lens.to(device)
+    # a tensor of shape (B, T)
+    tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id)
+
+    return audio, audio_lens, features, features_lens, tokens, tokens_lens
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    tokenizer: Tokenizer,
+    optimizer_g: Optimizer,
+    optimizer_d: Optimizer,
+    scheduler_g: LRSchedulerType,
+    scheduler_d: LRSchedulerType,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      tokenizer:
+        Used to convert text to phonemes.
+      optimizer_g:
+        The optimizer for generator.
+      optimizer_d:
+        The optimizer for discriminator.
+      scheduler_g:
+        The learning rate scheduler for generator, we call step() every epoch.
+      scheduler_d:
+        The learning rate scheduler for discriminator, we call step() every epoch.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+
+    # used to summary the stats over iterations in one epoch
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            params=params,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            scheduler_g=scheduler_g,
+            scheduler_d=scheduler_d,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+
+        batch_size = len(batch["tokens"])
+        audio, audio_lens, features, features_lens, tokens, tokens_lens = prepare_input(
+            batch, tokenizer, device
+        )
+
+        loss_info = MetricsTracker()
+        loss_info["samples"] = batch_size
+
+        try:
+            with autocast(enabled=params.use_fp16):
+                # forward discriminator
+                loss_d, stats_d = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    forward_generator=False,
+                )
+            for k, v in stats_d.items():
+                loss_info[k] = v * batch_size
+            # update discriminator
+            optimizer_d.zero_grad()
+            scaler.scale(loss_d).backward()
+            scaler.step(optimizer_d)
+
+            with autocast(enabled=params.use_fp16):
+                # forward generator
+                loss_g, stats_g = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    forward_generator=True,
+                    return_sample=params.batch_idx_train % params.log_interval == 0,
+                )
+            for k, v in stats_g.items():
+                if "returned_sample" not in k:
+                    loss_info[k] = v * batch_size
+            # update generator
+            optimizer_g.zero_grad()
+            scaler.scale(loss_g).backward()
+            scaler.step(optimizer_g)
+            scaler.update()
+
+            # summary stats
+            tot_loss = tot_loss + loss_info
+        except:  # noqa
+            save_bad_model()
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if params.batch_idx_train % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (
+                cur_grad_scale < 32.0 and params.batch_idx_train % 400 == 0
+            ):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if params.batch_idx_train % params.log_interval == 0:
+            cur_lr_g = max(scheduler_g.get_last_lr())
+            cur_lr_d = max(scheduler_d.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"global_batch_idx: {params.batch_idx_train}, batch size: {batch_size}, "
+                f"loss[{loss_info}], tot_loss[{tot_loss}], "
+                f"cur_lr_g: {cur_lr_g:.2e}, cur_lr_d: {cur_lr_d:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate_g", cur_lr_g, params.batch_idx_train
+                )
+                tb_writer.add_scalar(
+                    "train/learning_rate_d", cur_lr_d, params.batch_idx_train
+                )
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+                if "returned_sample" in stats_g:
+                    speech_hat_, speech_, mel_hat_, mel_ = stats_g["returned_sample"]
+                    tb_writer.add_audio(
+                        "train/speech_hat_",
+                        speech_hat_,
+                        params.batch_idx_train,
+                        params.sampling_rate,
+                    )
+                    tb_writer.add_audio(
+                        "train/speech_",
+                        speech_,
+                        params.batch_idx_train,
+                        params.sampling_rate,
+                    )
+                    tb_writer.add_image(
+                        "train/mel_hat_",
+                        plot_feature(mel_hat_),
+                        params.batch_idx_train,
+                        dataformats="HWC",
+                    )
+                    tb_writer.add_image(
+                        "train/mel_",
+                        plot_feature(mel_),
+                        params.batch_idx_train,
+                        dataformats="HWC",
+                    )
+
+        if (
+            params.batch_idx_train % params.valid_interval == 0
+            and not params.print_diagnostics
+        ):
+            logging.info("Computing validation loss")
+            valid_info, (speech_hat, speech) = compute_validation_loss(
+                params=params,
+                model=model,
+                tokenizer=tokenizer,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+                tb_writer.add_audio(
+                    "train/valdi_speech_hat",
+                    speech_hat,
+                    params.batch_idx_train,
+                    params.sampling_rate,
+                )
+                tb_writer.add_audio(
+                    "train/valdi_speech",
+                    speech,
+                    params.batch_idx_train,
+                    params.sampling_rate,
+                )
+
+    loss_value = tot_loss["generator_loss"] / tot_loss["samples"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    tokenizer: Tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+    rank: int = 0,
+) -> Tuple[MetricsTracker, Tuple[np.ndarray, np.ndarray]]:
+    """Run the validation process."""
+    model.eval()
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+
+    # used to summary the stats over iterations
+    tot_loss = MetricsTracker()
+    returned_sample = None
+
+    with torch.no_grad():
+        for batch_idx, batch in enumerate(valid_dl):
+            batch_size = len(batch["tokens"])
+            (
+                audio,
+                audio_lens,
+                features,
+                features_lens,
+                tokens,
+                tokens_lens,
+            ) = prepare_input(batch, tokenizer, device)
+
+            loss_info = MetricsTracker()
+            loss_info["samples"] = batch_size
+
+            # forward discriminator
+            loss_d, stats_d = model(
+                text=tokens,
+                text_lengths=tokens_lens,
+                feats=features,
+                feats_lengths=features_lens,
+                speech=audio,
+                speech_lengths=audio_lens,
+                forward_generator=False,
+            )
+            assert loss_d.requires_grad is False
+            for k, v in stats_d.items():
+                loss_info[k] = v * batch_size
+
+            # forward generator
+            loss_g, stats_g = model(
+                text=tokens,
+                text_lengths=tokens_lens,
+                feats=features,
+                feats_lengths=features_lens,
+                speech=audio,
+                speech_lengths=audio_lens,
+                forward_generator=True,
+            )
+            assert loss_g.requires_grad is False
+            for k, v in stats_g.items():
+                loss_info[k] = v * batch_size
+
+            # summary stats
+            tot_loss = tot_loss + loss_info
+
+            # infer for first batch:
+            if batch_idx == 0 and rank == 0:
+                inner_model = model.module if isinstance(model, DDP) else model
+                audio_pred, _, duration = inner_model.inference(
+                    text=tokens[0, : tokens_lens[0].item()]
+                )
+                audio_pred = audio_pred.data.cpu().numpy()
+                audio_len_pred = (
+                    (duration.sum(0) * params.frame_shift).to(dtype=torch.int64).item()
+                )
+                assert audio_len_pred == len(audio_pred), (
+                    audio_len_pred,
+                    len(audio_pred),
+                )
+                audio_gt = audio[0, : audio_lens[0].item()].data.cpu().numpy()
+                returned_sample = (audio_pred, audio_gt)
+
+    if world_size > 1:
+        tot_loss.reduce(device)
+
+    loss_value = tot_loss["generator_loss"] / tot_loss["samples"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss, returned_sample
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    tokenizer: Tokenizer,
+    optimizer_g: torch.optim.Optimizer,
+    optimizer_d: torch.optim.Optimizer,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        audio, audio_lens, features, features_lens, tokens, tokens_lens = prepare_input(
+            batch, tokenizer, device
+        )
+        try:
+            # for discriminator
+            with autocast(enabled=params.use_fp16):
+                loss_d, stats_d = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    forward_generator=False,
+                )
+            optimizer_d.zero_grad()
+            loss_d.backward()
+            # for generator
+            with autocast(enabled=params.use_fp16):
+                loss_g, stats_g = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    forward_generator=True,
+                )
+            optimizer_g.zero_grad()
+            loss_g.backward()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.blank_id
+    params.oov_id = tokenizer.oov_id
+    params.vocab_size = tokenizer.vocab_size
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+    generator = model.generator
+    discriminator = model.discriminator
+
+    num_param_g = sum([p.numel() for p in generator.parameters()])
+    logging.info(f"Number of parameters in generator: {num_param_g}")
+    num_param_d = sum([p.numel() for p in discriminator.parameters()])
+    logging.info(f"Number of parameters in discriminator: {num_param_d}")
+    logging.info(f"Total number of parameters: {num_param_g + num_param_d}")
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer_g = torch.optim.AdamW(
+        generator.parameters(), lr=params.lr, betas=(0.8, 0.99), eps=1e-9
+    )
+    optimizer_d = torch.optim.AdamW(
+        discriminator.parameters(), lr=params.lr, betas=(0.8, 0.99), eps=1e-9
+    )
+
+    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optimizer_g, gamma=0.999875)
+    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optimizer_d, gamma=0.999875)
+
+    if checkpoints is not None:
+        # load state_dict for optimizers
+        if "optimizer_g" in checkpoints:
+            logging.info("Loading optimizer_g state dict")
+            optimizer_g.load_state_dict(checkpoints["optimizer_g"])
+        if "optimizer_d" in checkpoints:
+            logging.info("Loading optimizer_d state dict")
+            optimizer_d.load_state_dict(checkpoints["optimizer_d"])
+
+        # load state_dict for schedulers
+        if "scheduler_g" in checkpoints:
+            logging.info("Loading scheduler_g state dict")
+            scheduler_g.load_state_dict(checkpoints["scheduler_g"])
+        if "scheduler_d" in checkpoints:
+            logging.info("Loading scheduler_d state dict")
+            scheduler_d.load_state_dict(checkpoints["scheduler_d"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    ljspeech = LJSpeechTtsDataModule(args)
+
+    train_cuts = ljspeech.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            # logging.warning(
+            #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_dl = ljspeech.train_dataloaders(train_cuts)
+
+    valid_cuts = ljspeech.valid_cuts()
+    valid_dl = ljspeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            tokenizer=tokenizer,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        logging.info(f"Start epoch {epoch}")
+
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        params.cur_epoch = epoch
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            tokenizer=tokenizer,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            scheduler_g=scheduler_g,
+            scheduler_d=scheduler_d,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        if epoch % params.save_every_n == 0 or epoch == params.num_epochs:
+            filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+            save_checkpoint(
+                filename=filename,
+                params=params,
+                model=model,
+                optimizer_g=optimizer_g,
+                optimizer_d=optimizer_d,
+                scheduler_g=scheduler_g,
+                scheduler_d=scheduler_d,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            if rank == 0:
+                if params.best_train_epoch == params.cur_epoch:
+                    best_train_filename = params.exp_dir / "best-train-loss.pt"
+                    copyfile(src=filename, dst=best_train_filename)
+
+                if params.best_valid_epoch == params.cur_epoch:
+                    best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+                    copyfile(src=filename, dst=best_valid_filename)
+
+        # step per epoch
+        scheduler_g.step()
+        scheduler_d.step()
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    LJSpeechTtsDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/ljspeech/TTS/vits/transform.py b/egs/ljspeech/TTS/vits/transform.py
new file mode 100644
index 000000000..c20d13130
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/transform.py
@@ -0,0 +1,218 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/transform.py
+
+"""Flow-related transformation.
+
+This code is derived from https://github.com/bayesiains/nflows.
+
+"""
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+# TODO(kan-bayashi): Documentation and type hint
+def piecewise_rational_quadratic_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = rational_quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unconstrained_rational_quadratic_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+# TODO(kan-bayashi): Documentation and type hint
+def unconstrained_rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = rational_quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+# TODO(kan-bayashi): Documentation and type hint
+def rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = _searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = _searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet
+
+
+def _searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
diff --git a/egs/ljspeech/TTS/vits/tts_datamodule.py b/egs/ljspeech/TTS/vits/tts_datamodule.py
new file mode 100644
index 000000000..81bb9ed13
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/tts_datamodule.py
@@ -0,0 +1,325 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022-2023  Xiaomi Corporation     (Authors: Mingshuang Luo,
+#                                                            Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Spectrogram, SpectrogramConfig, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+    SpeechSynthesisDataset,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LJSpeechTtsDataModule:
+    """
+    DataModule for tts experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="TTS data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/spectrogram"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, each batch will have the "
+            "field: batch['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to create train dataset")
+        train = SpeechSynthesisDataset(
+            return_text=False,
+            return_tokens=True,
+            feature_input_strategy=eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            sampling_rate = 22050
+            config = SpectrogramConfig(
+                sampling_rate=sampling_rate,
+                frame_length=1024 / sampling_rate,  # (in second),
+                frame_shift=256 / sampling_rate,  # (in second)
+                use_fft_mag=True,
+            )
+            train = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                feature_input_strategy=OnTheFlyFeatures(Spectrogram(config)),
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            sampling_rate = 22050
+            config = SpectrogramConfig(
+                sampling_rate=sampling_rate,
+                frame_length=1024 / sampling_rate,  # (in second),
+                frame_shift=256 / sampling_rate,  # (in second)
+                use_fft_mag=True,
+            )
+            validate = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                feature_input_strategy=OnTheFlyFeatures(Spectrogram(config)),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                feature_input_strategy=eval(self.args.input_strategy)(),
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create valid dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.info("About to create test dataset")
+        if self.args.on_the_fly_feats:
+            sampling_rate = 22050
+            config = SpectrogramConfig(
+                sampling_rate=sampling_rate,
+                frame_length=1024 / sampling_rate,  # (in second),
+                frame_shift=256 / sampling_rate,  # (in second)
+                use_fft_mag=True,
+            )
+            test = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                feature_input_strategy=OnTheFlyFeatures(Spectrogram(config)),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            test = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                feature_input_strategy=eval(self.args.input_strategy)(),
+                return_cuts=self.args.return_cuts,
+            )
+        test_sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=test_sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "ljspeech_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get validation cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "ljspeech_cuts_valid.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "ljspeech_cuts_test.jsonl.gz"
+        )
diff --git a/egs/ljspeech/TTS/vits/utils.py b/egs/ljspeech/TTS/vits/utils.py
new file mode 100644
index 000000000..6a067f596
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/utils.py
@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import collections
+import logging
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from lhotse.dataset.sampling.base import CutSampler
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import Optimizer
+from torch.utils.tensorboard import SummaryWriter
+
+
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/utils/get_random_segments.py
+def get_random_segments(
+    x: torch.Tensor,
+    x_lengths: torch.Tensor,
+    segment_size: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Get random segments.
+
+    Args:
+        x (Tensor): Input tensor (B, C, T).
+        x_lengths (Tensor): Length tensor (B,).
+        segment_size (int): Segment size.
+
+    Returns:
+        Tensor: Segmented tensor (B, C, segment_size).
+        Tensor: Start index tensor (B,).
+
+    """
+    b, c, t = x.size()
+    max_start_idx = x_lengths - segment_size
+    max_start_idx[max_start_idx < 0] = 0
+    start_idxs = (torch.rand([b]).to(x.device) * max_start_idx).to(
+        dtype=torch.long,
+    )
+    segments = get_segments(x, start_idxs, segment_size)
+
+    return segments, start_idxs
+
+
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/utils/get_random_segments.py
+def get_segments(
+    x: torch.Tensor,
+    start_idxs: torch.Tensor,
+    segment_size: int,
+) -> torch.Tensor:
+    """Get segments.
+
+    Args:
+        x (Tensor): Input tensor (B, C, T).
+        start_idxs (Tensor): Start index tensor (B,).
+        segment_size (int): Segment size.
+
+    Returns:
+        Tensor: Segmented tensor (B, C, segment_size).
+
+    """
+    b, c, t = x.size()
+    segments = x.new_zeros(b, c, segment_size)
+    for i, start_idx in enumerate(start_idxs):
+        segments[i] = x[i, :, start_idx : start_idx + segment_size]
+    return segments
+
+
+# from https://github.com/jaywalnut310/vit://github.com/jaywalnut310/vits/blob/main/commons.py
+def intersperse(sequence, item=0):
+    result = [item] * (len(sequence) * 2 + 1)
+    result[1::2] = sequence
+    return result
+
+
+# from https://github.com/jaywalnut310/vits/blob/main/utils.py
+MATPLOTLIB_FLAG = False
+
+
+def plot_feature(spectrogram):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+
+class MetricsTracker(collections.defaultdict):
+    def __init__(self):
+        # Passing the type 'int' to the base-class constructor
+        # makes undefined items default to int() which is zero.
+        # This class will play a role as metrics tracker.
+        # It can record many metrics, including but not limited to loss.
+        super(MetricsTracker, self).__init__(int)
+
+    def __add__(self, other: "MetricsTracker") -> "MetricsTracker":
+        ans = MetricsTracker()
+        for k, v in self.items():
+            ans[k] = v
+        for k, v in other.items():
+            ans[k] = ans[k] + v
+        return ans
+
+    def __mul__(self, alpha: float) -> "MetricsTracker":
+        ans = MetricsTracker()
+        for k, v in self.items():
+            ans[k] = v * alpha
+        return ans
+
+    def __str__(self) -> str:
+        ans = ""
+        for k, v in self.norm_items():
+            norm_value = "%.4g" % v
+            ans += str(k) + "=" + str(norm_value) + ", "
+        samples = "%.2f" % self["samples"]
+        ans += "over " + str(samples) + " samples."
+        return ans
+
+    def norm_items(self) -> List[Tuple[str, float]]:
+        """
+        Returns a list of pairs, like:
+          [('loss_1', 0.1), ('loss_2', 0.07)]
+        """
+        samples = self["samples"] if "samples" in self else 1
+        ans = []
+        for k, v in self.items():
+            if k == "samples":
+                continue
+            norm_value = float(v) / samples
+            ans.append((k, norm_value))
+        return ans
+
+    def reduce(self, device):
+        """
+        Reduce using torch.distributed, which I believe ensures that
+        all processes get the total.
+        """
+        keys = sorted(self.keys())
+        s = torch.tensor([float(self[k]) for k in keys], device=device)
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        for k, v in zip(keys, s.cpu().tolist()):
+            self[k] = v
+
+    def write_summary(
+        self,
+        tb_writer: SummaryWriter,
+        prefix: str,
+        batch_idx: int,
+    ) -> None:
+        """Add logging information to a TensorBoard writer.
+
+        Args:
+            tb_writer: a TensorBoard writer
+            prefix: a prefix for the name of the loss, e.g. "train/valid_",
+                or "train/current_"
+            batch_idx: The current batch index, used as the x-axis of the plot.
+        """
+        for k, v in self.norm_items():
+            tb_writer.add_scalar(prefix + k, v, batch_idx)
+
+
+# checkpoint saving and loading
+LRSchedulerType = torch.optim.lr_scheduler._LRScheduler
+
+
+def save_checkpoint(
+    filename: Path,
+    model: Union[nn.Module, DDP],
+    params: Optional[Dict[str, Any]] = None,
+    optimizer_g: Optional[Optimizer] = None,
+    optimizer_d: Optional[Optimizer] = None,
+    scheduler_g: Optional[LRSchedulerType] = None,
+    scheduler_d: Optional[LRSchedulerType] = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+) -> None:
+    """Save training information to a file.
+
+    Args:
+      filename:
+        The checkpoint filename.
+      model:
+        The model to be saved. We only save its `state_dict()`.
+      model_avg:
+        The stored model averaged from the start of training.
+      params:
+        User defined parameters, e.g., epoch, loss.
+      optimizer_g:
+        The optimizer for generator used in the training.
+        Its `state_dict` will be saved.
+      optimizer_d:
+        The optimizer for discriminator used in the training.
+        Its `state_dict` will be saved.
+      scheduler_g:
+        The learning rate scheduler for generator used in the training.
+        Its `state_dict` will be saved.
+      scheduler_d:
+        The learning rate scheduler for discriminator used in the training.
+        Its `state_dict` will be saved.
+      scalar:
+        The GradScaler to be saved. We only save its `state_dict()`.
+      rank:
+        Used in DDP. We save checkpoint only for the node whose rank is 0.
+    Returns:
+      Return None.
+    """
+    if rank != 0:
+        return
+
+    logging.info(f"Saving checkpoint to {filename}")
+
+    if isinstance(model, DDP):
+        model = model.module
+
+    checkpoint = {
+        "model": model.state_dict(),
+        "optimizer_g": optimizer_g.state_dict() if optimizer_g is not None else None,
+        "optimizer_d": optimizer_d.state_dict() if optimizer_d is not None else None,
+        "scheduler_g": scheduler_g.state_dict() if scheduler_g is not None else None,
+        "scheduler_d": scheduler_d.state_dict() if scheduler_d is not None else None,
+        "grad_scaler": scaler.state_dict() if scaler is not None else None,
+        "sampler": sampler.state_dict() if sampler is not None else None,
+    }
+
+    if params:
+        for k, v in params.items():
+            assert k not in checkpoint
+            checkpoint[k] = v
+
+    torch.save(checkpoint, filename)
diff --git a/egs/ljspeech/TTS/vits/vits.py b/egs/ljspeech/TTS/vits/vits.py
new file mode 100644
index 000000000..b4f0c21e6
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/vits.py
@@ -0,0 +1,607 @@
+# based on https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/vits/vits.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""VITS module for GAN-TTS task."""
+
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from generator import VITSGenerator
+from hifigan import (
+    HiFiGANMultiPeriodDiscriminator,
+    HiFiGANMultiScaleDiscriminator,
+    HiFiGANMultiScaleMultiPeriodDiscriminator,
+    HiFiGANPeriodDiscriminator,
+    HiFiGANScaleDiscriminator,
+)
+from loss import (
+    DiscriminatorAdversarialLoss,
+    FeatureMatchLoss,
+    GeneratorAdversarialLoss,
+    KLDivergenceLoss,
+    MelSpectrogramLoss,
+)
+from torch.cuda.amp import autocast
+from utils import get_segments
+
+AVAILABLE_GENERATERS = {
+    "vits_generator": VITSGenerator,
+}
+AVAILABLE_DISCRIMINATORS = {
+    "hifigan_period_discriminator": HiFiGANPeriodDiscriminator,
+    "hifigan_scale_discriminator": HiFiGANScaleDiscriminator,
+    "hifigan_multi_period_discriminator": HiFiGANMultiPeriodDiscriminator,
+    "hifigan_multi_scale_discriminator": HiFiGANMultiScaleDiscriminator,
+    "hifigan_multi_scale_multi_period_discriminator": HiFiGANMultiScaleMultiPeriodDiscriminator,  # NOQA
+}
+
+
+class VITS(nn.Module):
+    """Implement VITS, `Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech`"""
+
+    def __init__(
+        self,
+        # generator related
+        vocab_size: int,
+        feature_dim: int = 513,
+        sampling_rate: int = 22050,
+        generator_type: str = "vits_generator",
+        generator_params: Dict[str, Any] = {
+            "hidden_channels": 192,
+            "spks": None,
+            "langs": None,
+            "spk_embed_dim": None,
+            "global_channels": -1,
+            "segment_size": 32,
+            "text_encoder_attention_heads": 2,
+            "text_encoder_ffn_expand": 4,
+            "text_encoder_cnn_module_kernel": 5,
+            "text_encoder_blocks": 6,
+            "text_encoder_dropout_rate": 0.1,
+            "decoder_kernel_size": 7,
+            "decoder_channels": 512,
+            "decoder_upsample_scales": [8, 8, 2, 2],
+            "decoder_upsample_kernel_sizes": [16, 16, 4, 4],
+            "decoder_resblock_kernel_sizes": [3, 7, 11],
+            "decoder_resblock_dilations": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            "use_weight_norm_in_decoder": True,
+            "posterior_encoder_kernel_size": 5,
+            "posterior_encoder_layers": 16,
+            "posterior_encoder_stacks": 1,
+            "posterior_encoder_base_dilation": 1,
+            "posterior_encoder_dropout_rate": 0.0,
+            "use_weight_norm_in_posterior_encoder": True,
+            "flow_flows": 4,
+            "flow_kernel_size": 5,
+            "flow_base_dilation": 1,
+            "flow_layers": 4,
+            "flow_dropout_rate": 0.0,
+            "use_weight_norm_in_flow": True,
+            "use_only_mean_in_flow": True,
+            "stochastic_duration_predictor_kernel_size": 3,
+            "stochastic_duration_predictor_dropout_rate": 0.5,
+            "stochastic_duration_predictor_flows": 4,
+            "stochastic_duration_predictor_dds_conv_layers": 3,
+        },
+        # discriminator related
+        discriminator_type: str = "hifigan_multi_scale_multi_period_discriminator",
+        discriminator_params: Dict[str, Any] = {
+            "scales": 1,
+            "scale_downsample_pooling": "AvgPool1d",
+            "scale_downsample_pooling_params": {
+                "kernel_size": 4,
+                "stride": 2,
+                "padding": 2,
+            },
+            "scale_discriminator_params": {
+                "in_channels": 1,
+                "out_channels": 1,
+                "kernel_sizes": [15, 41, 5, 3],
+                "channels": 128,
+                "max_downsample_channels": 1024,
+                "max_groups": 16,
+                "bias": True,
+                "downsample_scales": [2, 2, 4, 4, 1],
+                "nonlinear_activation": "LeakyReLU",
+                "nonlinear_activation_params": {"negative_slope": 0.1},
+                "use_weight_norm": True,
+                "use_spectral_norm": False,
+            },
+            "follow_official_norm": False,
+            "periods": [2, 3, 5, 7, 11],
+            "period_discriminator_params": {
+                "in_channels": 1,
+                "out_channels": 1,
+                "kernel_sizes": [5, 3],
+                "channels": 32,
+                "downsample_scales": [3, 3, 3, 3, 1],
+                "max_downsample_channels": 1024,
+                "bias": True,
+                "nonlinear_activation": "LeakyReLU",
+                "nonlinear_activation_params": {"negative_slope": 0.1},
+                "use_weight_norm": True,
+                "use_spectral_norm": False,
+            },
+        },
+        # loss related
+        generator_adv_loss_params: Dict[str, Any] = {
+            "average_by_discriminators": False,
+            "loss_type": "mse",
+        },
+        discriminator_adv_loss_params: Dict[str, Any] = {
+            "average_by_discriminators": False,
+            "loss_type": "mse",
+        },
+        feat_match_loss_params: Dict[str, Any] = {
+            "average_by_discriminators": False,
+            "average_by_layers": False,
+            "include_final_outputs": True,
+        },
+        mel_loss_params: Dict[str, Any] = {
+            "frame_shift": 256,
+            "frame_length": 1024,
+            "n_mels": 80,
+        },
+        lambda_adv: float = 1.0,
+        lambda_mel: float = 45.0,
+        lambda_feat_match: float = 2.0,
+        lambda_dur: float = 1.0,
+        lambda_kl: float = 1.0,
+        cache_generator_outputs: bool = True,
+    ):
+        """Initialize VITS module.
+
+        Args:
+            idim (int): Input vocabrary size.
+            odim (int): Acoustic feature dimension. The actual output channels will
+                be 1 since VITS is the end-to-end text-to-wave model but for the
+                compatibility odim is used to indicate the acoustic feature dimension.
+            sampling_rate (int): Sampling rate, not used for the training but it will
+                be referred in saving waveform during the inference.
+            generator_type (str): Generator type.
+            generator_params (Dict[str, Any]): Parameter dict for generator.
+            discriminator_type (str): Discriminator type.
+            discriminator_params (Dict[str, Any]): Parameter dict for discriminator.
+            generator_adv_loss_params (Dict[str, Any]): Parameter dict for generator
+                adversarial loss.
+            discriminator_adv_loss_params (Dict[str, Any]): Parameter dict for
+                discriminator adversarial loss.
+            feat_match_loss_params (Dict[str, Any]): Parameter dict for feat match loss.
+            mel_loss_params (Dict[str, Any]): Parameter dict for mel loss.
+            lambda_adv (float): Loss scaling coefficient for adversarial loss.
+            lambda_mel (float): Loss scaling coefficient for mel spectrogram loss.
+            lambda_feat_match (float): Loss scaling coefficient for feat match loss.
+            lambda_dur (float): Loss scaling coefficient for duration loss.
+            lambda_kl (float): Loss scaling coefficient for KL divergence loss.
+            cache_generator_outputs (bool): Whether to cache generator outputs.
+
+        """
+        super().__init__()
+
+        # define modules
+        generator_class = AVAILABLE_GENERATERS[generator_type]
+        if generator_type == "vits_generator":
+            # NOTE(kan-bayashi): Update parameters for the compatibility.
+            #   The idim and odim is automatically decided from input data,
+            #   where idim represents #vocabularies and odim represents
+            #   the input acoustic feature dimension.
+            generator_params.update(vocabs=vocab_size, aux_channels=feature_dim)
+        self.generator = generator_class(
+            **generator_params,
+        )
+        discriminator_class = AVAILABLE_DISCRIMINATORS[discriminator_type]
+        self.discriminator = discriminator_class(
+            **discriminator_params,
+        )
+        self.generator_adv_loss = GeneratorAdversarialLoss(
+            **generator_adv_loss_params,
+        )
+        self.discriminator_adv_loss = DiscriminatorAdversarialLoss(
+            **discriminator_adv_loss_params,
+        )
+        self.feat_match_loss = FeatureMatchLoss(
+            **feat_match_loss_params,
+        )
+        mel_loss_params.update(sampling_rate=sampling_rate)
+        self.mel_loss = MelSpectrogramLoss(
+            **mel_loss_params,
+        )
+        self.kl_loss = KLDivergenceLoss()
+
+        # coefficients
+        self.lambda_adv = lambda_adv
+        self.lambda_mel = lambda_mel
+        self.lambda_kl = lambda_kl
+        self.lambda_feat_match = lambda_feat_match
+        self.lambda_dur = lambda_dur
+
+        # cache
+        self.cache_generator_outputs = cache_generator_outputs
+        self._cache = None
+
+        # store sampling rate for saving wav file
+        # (not used for the training)
+        self.sampling_rate = sampling_rate
+
+        # store parameters for test compatibility
+        self.spks = self.generator.spks
+        self.langs = self.generator.langs
+        self.spk_embed_dim = self.generator.spk_embed_dim
+
+    def forward(
+        self,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        feats: torch.Tensor,
+        feats_lengths: torch.Tensor,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        return_sample: bool = False,
+        sids: Optional[torch.Tensor] = None,
+        spembs: Optional[torch.Tensor] = None,
+        lids: Optional[torch.Tensor] = None,
+        forward_generator: bool = True,
+    ) -> Tuple[torch.Tensor, Dict[str, Any]]:
+        """Perform generator forward.
+
+        Args:
+            text (Tensor): Text index tensor (B, T_text).
+            text_lengths (Tensor): Text length tensor (B,).
+            feats (Tensor): Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor): Feature length tensor (B,).
+            speech (Tensor): Speech waveform tensor (B, T_wav).
+            speech_lengths (Tensor): Speech length tensor (B,).
+            sids (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]): Language index tensor (B,) or (B, 1).
+            forward_generator (bool): Whether to forward generator.
+
+        Returns:
+            - loss (Tensor): Loss scalar tensor.
+            - stats (Dict[str, float]): Statistics to be monitored.
+        """
+        if forward_generator:
+            return self._forward_generator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                speech=speech,
+                speech_lengths=speech_lengths,
+                return_sample=return_sample,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+            )
+        else:
+            return self._forward_discrminator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                speech=speech,
+                speech_lengths=speech_lengths,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+            )
+
+    def _forward_generator(
+        self,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        feats: torch.Tensor,
+        feats_lengths: torch.Tensor,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        return_sample: bool = False,
+        sids: Optional[torch.Tensor] = None,
+        spembs: Optional[torch.Tensor] = None,
+        lids: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Dict[str, Any]]:
+        """Perform generator forward.
+
+        Args:
+            text (Tensor): Text index tensor (B, T_text).
+            text_lengths (Tensor): Text length tensor (B,).
+            feats (Tensor): Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor): Feature length tensor (B,).
+            speech (Tensor): Speech waveform tensor (B, T_wav).
+            speech_lengths (Tensor): Speech length tensor (B,).
+            sids (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]): Language index tensor (B,) or (B, 1).
+
+        Returns:
+            * loss (Tensor): Loss scalar tensor.
+            * stats (Dict[str, float]): Statistics to be monitored.
+        """
+        # setup
+        feats = feats.transpose(1, 2)
+        speech = speech.unsqueeze(1)
+
+        # calculate generator outputs
+        reuse_cache = True
+        if not self.cache_generator_outputs or self._cache is None:
+            reuse_cache = False
+            outs = self.generator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+            )
+        else:
+            outs = self._cache
+
+        # store cache
+        if self.training and self.cache_generator_outputs and not reuse_cache:
+            self._cache = outs
+
+        # parse outputs
+        speech_hat_, dur_nll, _, start_idxs, _, z_mask, outs_ = outs
+        _, z_p, m_p, logs_p, _, logs_q = outs_
+        speech_ = get_segments(
+            x=speech,
+            start_idxs=start_idxs * self.generator.upsample_factor,
+            segment_size=self.generator.segment_size * self.generator.upsample_factor,
+        )
+
+        # calculate discriminator outputs
+        p_hat = self.discriminator(speech_hat_)
+        with torch.no_grad():
+            # do not store discriminator gradient in generator turn
+            p = self.discriminator(speech_)
+
+        # calculate losses
+        with autocast(enabled=False):
+            if not return_sample:
+                mel_loss = self.mel_loss(speech_hat_, speech_)
+            else:
+                mel_loss, (mel_hat_, mel_) = self.mel_loss(
+                    speech_hat_, speech_, return_mel=True
+                )
+            kl_loss = self.kl_loss(z_p, logs_q, m_p, logs_p, z_mask)
+            dur_loss = torch.sum(dur_nll.float())
+            adv_loss = self.generator_adv_loss(p_hat)
+            feat_match_loss = self.feat_match_loss(p_hat, p)
+
+            mel_loss = mel_loss * self.lambda_mel
+            kl_loss = kl_loss * self.lambda_kl
+            dur_loss = dur_loss * self.lambda_dur
+            adv_loss = adv_loss * self.lambda_adv
+            feat_match_loss = feat_match_loss * self.lambda_feat_match
+            loss = mel_loss + kl_loss + dur_loss + adv_loss + feat_match_loss
+
+        stats = dict(
+            generator_loss=loss.item(),
+            generator_mel_loss=mel_loss.item(),
+            generator_kl_loss=kl_loss.item(),
+            generator_dur_loss=dur_loss.item(),
+            generator_adv_loss=adv_loss.item(),
+            generator_feat_match_loss=feat_match_loss.item(),
+        )
+
+        if return_sample:
+            stats["returned_sample"] = (
+                speech_hat_[0].data.cpu().numpy(),
+                speech_[0].data.cpu().numpy(),
+                mel_hat_[0].data.cpu().numpy(),
+                mel_[0].data.cpu().numpy(),
+            )
+
+        # reset cache
+        if reuse_cache or not self.training:
+            self._cache = None
+
+        return loss, stats
+
+    def _forward_discrminator(
+        self,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        feats: torch.Tensor,
+        feats_lengths: torch.Tensor,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        sids: Optional[torch.Tensor] = None,
+        spembs: Optional[torch.Tensor] = None,
+        lids: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Dict[str, Any]]:
+        """Perform discriminator forward.
+
+        Args:
+            text (Tensor): Text index tensor (B, T_text).
+            text_lengths (Tensor): Text length tensor (B,).
+            feats (Tensor): Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor): Feature length tensor (B,).
+            speech (Tensor): Speech waveform tensor (B, T_wav).
+            speech_lengths (Tensor): Speech length tensor (B,).
+            sids (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]): Language index tensor (B,) or (B, 1).
+
+        Returns:
+            * loss (Tensor): Loss scalar tensor.
+            * stats (Dict[str, float]): Statistics to be monitored.
+        """
+        # setup
+        feats = feats.transpose(1, 2)
+        speech = speech.unsqueeze(1)
+
+        # calculate generator outputs
+        reuse_cache = True
+        if not self.cache_generator_outputs or self._cache is None:
+            reuse_cache = False
+            outs = self.generator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+            )
+        else:
+            outs = self._cache
+
+        # store cache
+        if self.cache_generator_outputs and not reuse_cache:
+            self._cache = outs
+
+        # parse outputs
+        speech_hat_, _, _, start_idxs, *_ = outs
+        speech_ = get_segments(
+            x=speech,
+            start_idxs=start_idxs * self.generator.upsample_factor,
+            segment_size=self.generator.segment_size * self.generator.upsample_factor,
+        )
+
+        # calculate discriminator outputs
+        p_hat = self.discriminator(speech_hat_.detach())
+        p = self.discriminator(speech_)
+
+        # calculate losses
+        with autocast(enabled=False):
+            real_loss, fake_loss = self.discriminator_adv_loss(p_hat, p)
+            loss = real_loss + fake_loss
+
+        stats = dict(
+            discriminator_loss=loss.item(),
+            discriminator_real_loss=real_loss.item(),
+            discriminator_fake_loss=fake_loss.item(),
+        )
+
+        # reset cache
+        if reuse_cache or not self.training:
+            self._cache = None
+
+        return loss, stats
+
+    def inference(
+        self,
+        text: torch.Tensor,
+        feats: Optional[torch.Tensor] = None,
+        sids: Optional[torch.Tensor] = None,
+        spembs: Optional[torch.Tensor] = None,
+        lids: Optional[torch.Tensor] = None,
+        durations: Optional[torch.Tensor] = None,
+        noise_scale: float = 0.667,
+        noise_scale_dur: float = 0.8,
+        alpha: float = 1.0,
+        max_len: Optional[int] = None,
+        use_teacher_forcing: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Run inference for single sample.
+
+        Args:
+            text (Tensor): Input text index tensor (T_text,).
+            feats (Tensor): Feature tensor (T_feats, aux_channels).
+            sids (Tensor): Speaker index tensor (1,).
+            spembs (Optional[Tensor]): Speaker embedding tensor (spk_embed_dim,).
+            lids (Tensor): Language index tensor (1,).
+            durations (Tensor): Ground-truth duration tensor (T_text,).
+            noise_scale (float): Noise scale value for flow.
+            noise_scale_dur (float): Noise scale value for duration predictor.
+            alpha (float): Alpha parameter to control the speed of generated speech.
+            max_len (Optional[int]): Maximum length.
+            use_teacher_forcing (bool): Whether to use teacher forcing.
+
+        Returns:
+            * wav (Tensor): Generated waveform tensor (T_wav,).
+            * att_w (Tensor): Monotonic attention weight tensor (T_feats, T_text).
+            * duration (Tensor): Predicted duration tensor (T_text,).
+        """
+        # setup
+        text = text[None]
+        text_lengths = torch.tensor(
+            [text.size(1)],
+            dtype=torch.long,
+            device=text.device,
+        )
+        if sids is not None:
+            sids = sids.view(1)
+        if lids is not None:
+            lids = lids.view(1)
+        if durations is not None:
+            durations = durations.view(1, 1, -1)
+
+        # inference
+        if use_teacher_forcing:
+            assert feats is not None
+            feats = feats[None].transpose(1, 2)
+            feats_lengths = torch.tensor(
+                [feats.size(2)],
+                dtype=torch.long,
+                device=feats.device,
+            )
+            wav, att_w, dur = self.generator.inference(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+                max_len=max_len,
+                use_teacher_forcing=use_teacher_forcing,
+            )
+        else:
+            wav, att_w, dur = self.generator.inference(
+                text=text,
+                text_lengths=text_lengths,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+                dur=durations,
+                noise_scale=noise_scale,
+                noise_scale_dur=noise_scale_dur,
+                alpha=alpha,
+                max_len=max_len,
+            )
+        return wav.view(-1), att_w[0], dur[0]
+
+    def inference_batch(
+        self,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        sids: Optional[torch.Tensor] = None,
+        durations: Optional[torch.Tensor] = None,
+        noise_scale: float = 0.667,
+        noise_scale_dur: float = 0.8,
+        alpha: float = 1.0,
+        max_len: Optional[int] = None,
+        use_teacher_forcing: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Run inference for one batch.
+
+        Args:
+            text (Tensor): Input text index tensor (B, T_text).
+            text_lengths (Tensor): Input text index tensor (B,).
+            sids (Tensor): Speaker index tensor (B,).
+            noise_scale (float): Noise scale value for flow.
+            noise_scale_dur (float): Noise scale value for duration predictor.
+            alpha (float): Alpha parameter to control the speed of generated speech.
+            max_len (Optional[int]): Maximum length.
+
+        Returns:
+            * wav (Tensor): Generated waveform tensor (B, T_wav).
+            * att_w (Tensor): Monotonic attention weight tensor (B, T_feats, T_text).
+            * duration (Tensor): Predicted duration tensor (B, T_text).
+        """
+        # inference
+        wav, att_w, dur = self.generator.inference(
+            text=text,
+            text_lengths=text_lengths,
+            sids=sids,
+            noise_scale=noise_scale,
+            noise_scale_dur=noise_scale_dur,
+            alpha=alpha,
+            max_len=max_len,
+        )
+        return wav, att_w, dur
diff --git a/egs/ljspeech/TTS/vits/wavenet.py b/egs/ljspeech/TTS/vits/wavenet.py
new file mode 100644
index 000000000..5db461d5c
--- /dev/null
+++ b/egs/ljspeech/TTS/vits/wavenet.py
@@ -0,0 +1,348 @@
+# from https://github.com/espnet/espnet/blob/master/espnet2/gan_tts/wavenet/wavenet.py
+
+# Copyright 2021 Tomoki Hayashi
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""WaveNet modules.
+
+This code is modified from https://github.com/kan-bayashi/ParallelWaveGAN.
+
+"""
+
+import logging
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+
+
+class WaveNet(torch.nn.Module):
+    """WaveNet with global conditioning."""
+
+    def __init__(
+        self,
+        in_channels: int = 1,
+        out_channels: int = 1,
+        kernel_size: int = 3,
+        layers: int = 30,
+        stacks: int = 3,
+        base_dilation: int = 2,
+        residual_channels: int = 64,
+        aux_channels: int = -1,
+        gate_channels: int = 128,
+        skip_channels: int = 64,
+        global_channels: int = -1,
+        dropout_rate: float = 0.0,
+        bias: bool = True,
+        use_weight_norm: bool = True,
+        use_first_conv: bool = False,
+        use_last_conv: bool = False,
+        scale_residual: bool = False,
+        scale_skip_connect: bool = False,
+    ):
+        """Initialize WaveNet module.
+
+        Args:
+            in_channels (int): Number of input channels.
+            out_channels (int): Number of output channels.
+            kernel_size (int): Kernel size of dilated convolution.
+            layers (int): Number of residual block layers.
+            stacks (int): Number of stacks i.e., dilation cycles.
+            base_dilation (int): Base dilation factor.
+            residual_channels (int): Number of channels in residual conv.
+            gate_channels (int):  Number of channels in gated conv.
+            skip_channels (int): Number of channels in skip conv.
+            aux_channels (int): Number of channels for local conditioning feature.
+            global_channels (int): Number of channels for global conditioning feature.
+            dropout_rate (float): Dropout rate. 0.0 means no dropout applied.
+            bias (bool): Whether to use bias parameter in conv layer.
+            use_weight_norm (bool): Whether to use weight norm. If set to true, it will
+                be applied to all of the conv layers.
+            use_first_conv (bool): Whether to use the first conv layers.
+            use_last_conv (bool): Whether to use the last conv layers.
+            scale_residual (bool): Whether to scale the residual outputs.
+            scale_skip_connect (bool): Whether to scale the skip connection outputs.
+
+        """
+        super().__init__()
+        self.layers = layers
+        self.stacks = stacks
+        self.kernel_size = kernel_size
+        self.base_dilation = base_dilation
+        self.use_first_conv = use_first_conv
+        self.use_last_conv = use_last_conv
+        self.scale_skip_connect = scale_skip_connect
+
+        # check the number of layers and stacks
+        assert layers % stacks == 0
+        layers_per_stack = layers // stacks
+
+        # define first convolution
+        if self.use_first_conv:
+            self.first_conv = Conv1d1x1(in_channels, residual_channels, bias=True)
+
+        # define residual blocks
+        self.conv_layers = torch.nn.ModuleList()
+        for layer in range(layers):
+            dilation = base_dilation ** (layer % layers_per_stack)
+            conv = ResidualBlock(
+                kernel_size=kernel_size,
+                residual_channels=residual_channels,
+                gate_channels=gate_channels,
+                skip_channels=skip_channels,
+                aux_channels=aux_channels,
+                global_channels=global_channels,
+                dilation=dilation,
+                dropout_rate=dropout_rate,
+                bias=bias,
+                scale_residual=scale_residual,
+            )
+            self.conv_layers += [conv]
+
+        # define output layers
+        if self.use_last_conv:
+            self.last_conv = torch.nn.Sequential(
+                torch.nn.ReLU(inplace=True),
+                Conv1d1x1(skip_channels, skip_channels, bias=True),
+                torch.nn.ReLU(inplace=True),
+                Conv1d1x1(skip_channels, out_channels, bias=True),
+            )
+
+        # apply weight norm
+        if use_weight_norm:
+            self.apply_weight_norm()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: Optional[torch.Tensor] = None,
+        c: Optional[torch.Tensor] = None,
+        g: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input noise signal (B, 1, T) if use_first_conv else
+                (B, residual_channels, T).
+            x_mask (Optional[Tensor]): Mask tensor (B, 1, T).
+            c (Optional[Tensor]): Local conditioning features (B, aux_channels, T).
+            g (Optional[Tensor]): Global conditioning features (B, global_channels, 1).
+
+        Returns:
+            Tensor: Output tensor (B, out_channels, T) if use_last_conv else
+                (B, residual_channels, T).
+
+        """
+        # encode to hidden representation
+        if self.use_first_conv:
+            x = self.first_conv(x)
+
+        # residual block
+        skips = 0.0
+        for f in self.conv_layers:
+            x, h = f(x, x_mask=x_mask, c=c, g=g)
+            skips = skips + h
+        x = skips
+        if self.scale_skip_connect:
+            x = x * math.sqrt(1.0 / len(self.conv_layers))
+
+        # apply final layers
+        if self.use_last_conv:
+            x = self.last_conv(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        """Remove weight normalization module from all of the layers."""
+
+        def _remove_weight_norm(m: torch.nn.Module):
+            try:
+                logging.debug(f"Weight norm is removed from {m}.")
+                torch.nn.utils.remove_weight_norm(m)
+            except ValueError:  # this module didn't have weight norm
+                return
+
+        self.apply(_remove_weight_norm)
+
+    def apply_weight_norm(self):
+        """Apply weight normalization module from all of the layers."""
+
+        def _apply_weight_norm(m: torch.nn.Module):
+            if isinstance(m, torch.nn.Conv1d) or isinstance(m, torch.nn.Conv2d):
+                torch.nn.utils.weight_norm(m)
+                logging.debug(f"Weight norm is applied to {m}.")
+
+        self.apply(_apply_weight_norm)
+
+    @staticmethod
+    def _get_receptive_field_size(
+        layers: int,
+        stacks: int,
+        kernel_size: int,
+        base_dilation: int,
+    ) -> int:
+        assert layers % stacks == 0
+        layers_per_cycle = layers // stacks
+        dilations = [base_dilation ** (i % layers_per_cycle) for i in range(layers)]
+        return (kernel_size - 1) * sum(dilations) + 1
+
+    @property
+    def receptive_field_size(self) -> int:
+        """Return receptive field size."""
+        return self._get_receptive_field_size(
+            self.layers, self.stacks, self.kernel_size, self.base_dilation
+        )
+
+
+class Conv1d(torch.nn.Conv1d):
+    """Conv1d module with customized initialization."""
+
+    def __init__(self, *args, **kwargs):
+        """Initialize Conv1d module."""
+        super().__init__(*args, **kwargs)
+
+    def reset_parameters(self):
+        """Reset parameters."""
+        torch.nn.init.kaiming_normal_(self.weight, nonlinearity="relu")
+        if self.bias is not None:
+            torch.nn.init.constant_(self.bias, 0.0)
+
+
+class Conv1d1x1(Conv1d):
+    """1x1 Conv1d with customized initialization."""
+
+    def __init__(self, in_channels: int, out_channels: int, bias: bool):
+        """Initialize 1x1 Conv1d module."""
+        super().__init__(
+            in_channels, out_channels, kernel_size=1, padding=0, dilation=1, bias=bias
+        )
+
+
+class ResidualBlock(torch.nn.Module):
+    """Residual block module in WaveNet."""
+
+    def __init__(
+        self,
+        kernel_size: int = 3,
+        residual_channels: int = 64,
+        gate_channels: int = 128,
+        skip_channels: int = 64,
+        aux_channels: int = 80,
+        global_channels: int = -1,
+        dropout_rate: float = 0.0,
+        dilation: int = 1,
+        bias: bool = True,
+        scale_residual: bool = False,
+    ):
+        """Initialize ResidualBlock module.
+
+        Args:
+            kernel_size (int): Kernel size of dilation convolution layer.
+            residual_channels (int): Number of channels for residual connection.
+            skip_channels (int): Number of channels for skip connection.
+            aux_channels (int): Number of local conditioning channels.
+            dropout (float): Dropout probability.
+            dilation (int): Dilation factor.
+            bias (bool): Whether to add bias parameter in convolution layers.
+            scale_residual (bool): Whether to scale the residual outputs.
+
+        """
+        super().__init__()
+        self.dropout_rate = dropout_rate
+        self.residual_channels = residual_channels
+        self.skip_channels = skip_channels
+        self.scale_residual = scale_residual
+
+        # check
+        assert (kernel_size - 1) % 2 == 0, "Not support even number kernel size."
+        assert gate_channels % 2 == 0
+
+        # dilation conv
+        padding = (kernel_size - 1) // 2 * dilation
+        self.conv = Conv1d(
+            residual_channels,
+            gate_channels,
+            kernel_size,
+            padding=padding,
+            dilation=dilation,
+            bias=bias,
+        )
+
+        # local conditioning
+        if aux_channels > 0:
+            self.conv1x1_aux = Conv1d1x1(aux_channels, gate_channels, bias=False)
+        else:
+            self.conv1x1_aux = None
+
+        # global conditioning
+        if global_channels > 0:
+            self.conv1x1_glo = Conv1d1x1(global_channels, gate_channels, bias=False)
+        else:
+            self.conv1x1_glo = None
+
+        # conv output is split into two groups
+        gate_out_channels = gate_channels // 2
+
+        # NOTE(kan-bayashi): concat two convs into a single conv for the efficiency
+        #   (integrate res 1x1 + skip 1x1 convs)
+        self.conv1x1_out = Conv1d1x1(
+            gate_out_channels, residual_channels + skip_channels, bias=bias
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: Optional[torch.Tensor] = None,
+        c: Optional[torch.Tensor] = None,
+        g: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Calculate forward propagation.
+
+        Args:
+            x (Tensor): Input tensor (B, residual_channels, T).
+            x_mask Optional[torch.Tensor]: Mask tensor (B, 1, T).
+            c (Optional[Tensor]): Local conditioning tensor (B, aux_channels, T).
+            g (Optional[Tensor]): Global conditioning tensor (B, global_channels, 1).
+
+        Returns:
+            Tensor: Output tensor for residual connection (B, residual_channels, T).
+            Tensor: Output tensor for skip connection (B, skip_channels, T).
+
+        """
+        residual = x
+        x = F.dropout(x, p=self.dropout_rate, training=self.training)
+        x = self.conv(x)
+
+        # split into two part for gated activation
+        splitdim = 1
+        xa, xb = x.split(x.size(splitdim) // 2, dim=splitdim)
+
+        # local conditioning
+        if c is not None:
+            c = self.conv1x1_aux(c)
+            ca, cb = c.split(c.size(splitdim) // 2, dim=splitdim)
+            xa, xb = xa + ca, xb + cb
+
+        # global conditioning
+        if g is not None:
+            g = self.conv1x1_glo(g)
+            ga, gb = g.split(g.size(splitdim) // 2, dim=splitdim)
+            xa, xb = xa + ga, xb + gb
+
+        x = torch.tanh(xa) * torch.sigmoid(xb)
+
+        # residual + skip 1x1 conv
+        x = self.conv1x1_out(x)
+        if x_mask is not None:
+            x = x * x_mask
+
+        # split integrated conv results
+        x, s = x.split([self.residual_channels, self.skip_channels], dim=1)
+
+        # for residual connection
+        x = x + residual
+        if self.scale_residual:
+            x = x * math.sqrt(0.5)
+
+        return x, s
diff --git a/egs/mgb2/ASR/README.md b/egs/mgb2/ASR/README.md
new file mode 100644
index 000000000..2bc4b000b
--- /dev/null
+++ b/egs/mgb2/ASR/README.md
@@ -0,0 +1,43 @@
+# MGB2
+
+The Multi-Dialect Broadcast News Arabic Speech Recognition (MGB-2):
+The second edition of the Multi-Genre Broadcast (MGB-2) Challenge is
+an evaluation of speech recognition and lightly supervised alignment
+using TV recordings in Arabic. The speech data is broad and multi-genre,
+spanning the whole range of TV output, and represents a challenging task for
+speech technology. In 2016, the challenge featured two new Arabic tracks based
+on TV data from Aljazeera. It was an official challenge at the 2016 IEEE
+Workshop on Spoken Language Technology. The 1,200 hours MGB-2: from Aljazeera
+TV programs have been manually captioned with no timing information.
+QCRI Arabic ASR system has been used to recognize all programs. The ASR output
+was used to align the manual captioning and produce speech segments for
+training speech recognition. More than 20 hours from 2015 programs have been
+transcribed verbatim and manually segmented. This data is split into a
+development set of 10 hours, and a similar evaluation set of 10 hours.
+Both the development and evaluation data have been released in the 2016 MGB
+challenge
+
+Official reference:
+
+Ali, Ahmed, et al. "The MGB-2 challenge: Arabic multi-dialect broadcast media recognition." 
+2016 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2016.
+
+IEEE link: https://ieeexplore.ieee.org/abstract/document/7846277
+
+## Stateless Pruned Transducer Performance Record (after 30 epochs)
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 15.52      | 15.28      | --epoch 18, --avg 5, --max-duration 200  |
+| modified beam search               | 13.88      | 13.7       | --epoch 18, --avg 5, --max-duration 200  |
+| fast beam search                   | 14.62      | 14.36      | --epoch 18, --avg 5, --max-duration 200  |
+
+## Conformer-CTC Performance Record (after 40 epochs)
+
+| Decoding method           | dev WER    | test WER |
+|---------------------------|------------|---------|
+| attention-decoder         | 15.62      |  15.01  |
+| whole-lattice-rescoring   | 15.89      |  15.08  |
+
+
+See [RESULTS](/egs/mgb2/ASR/RESULTS.md) for details.
diff --git a/egs/mgb2/ASR/RESULTS.md b/egs/mgb2/ASR/RESULTS.md
new file mode 100644
index 000000000..2a7ea7664
--- /dev/null
+++ b/egs/mgb2/ASR/RESULTS.md
@@ -0,0 +1,236 @@
+# Results
+
+
+### MGB2 all data BPE training results (Stateless Pruned Transducer)
+
+#### 2022-09-07
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 15.52      | 15.28      | --epoch 18, --avg 5, --max-duration 200 |
+| modified beam search               | 13.88      | 13.7       | --epoch 18, --avg 5, --max-duration 200 |
+| fast beam search                   | 14.62      | 14.36      | --epoch 18, --avg 5, --max-duration 200|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+
+  
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --max-duration 300 \
+  --num-buckets 50
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/YyNv45pfQ0GqWzZ898WOlw/#scalars
+
+The decoding command is:
+```
+epoch=18
+avg=5
+for method in greedy_search modified_beam_search fast_beam_search; do
+  ./pruned_transducer_stateless5/decode.py \
+    --epoch $epoch \
+	--beam-size 10 \
+    --avg $avg \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 200 \
+    --decoding-method $method \
+    --max-sym-per-frame 1 \
+    --num-encoder-layers 12 \
+    --dim-feedforward 2048 \
+    --nhead 8 \
+    --encoder-dim 512 \
+    --decoder-dim 512 \
+    --joiner-dim 512 \
+    --use-averaged-model True
+done
+```
+
+### MGB2 all data BPE training results (Conformer-CTC) (after 40 epochs)
+
+#### 2022-06-04
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+https://huggingface.co/AmirHussein/icefall-asr-mgb2-conformer_ctc-2022-27-06
+
+The best WER, as of 2022-06-04, for the MGB2 test dataset is below
+
+Using whole lattice HLG decoding + n-gram LM rescoring 
+
+|     | dev        | test       |
+|-----|------------|------------|
+| WER | 15.62      |  15.01     |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+| 0.1            | -            |
+
+
+Using n-best (n=0.5) attention decoder rescoring
+
+|     | dev        | test       |
+|-----|------------|------------|
+| WER |    15.89   |  15.08     |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+| 0.01           | 0.5             |
+
+
+To reproduce the above result, use the following commands for training:
+
+# Note: the model was trained on V-100 32GB GPU
+
+```
+cd egs/mgb2/ASR
+. ./path.sh
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1"
+./conformer_ctc/train.py \
+  --lang-dir data/lang_bpe_5000 \
+  --att-rate 0.8 \
+  --lr-factor 10 \
+  --max-duration  \
+  --concatenate-cuts 0 \
+  --world-size 2 \
+  --bucketing-sampler 1 \
+  --max-duration 100 \ 
+  --start-epoch 0 \
+  --num-epochs 40
+  
+```
+
+and the following command for nbest decoding
+
+```
+./conformer_ctc/decode.py \
+  --lang-dir data/lang_bpe_5000 \
+  --max-duration 30 \
+  --concatenate-cuts 0 \
+  --bucketing-sampler 1 \
+  --num-paths 1000 \
+  --epoch 40 \
+  --avg 5 \
+  --method attention-decoder \
+  --nbest-scale 0.5
+```
+
+and the following command for whole-lattice decoding
+
+```
+./conformer_ctc/decode.py \
+  --epoch 40 \
+  --avg 5 \
+  --exp-dir conformer_ctc/exp_5000_att0.8 \
+  --lang-dir data/lang_bpe_5000 \
+  --max-duration 30 \
+  --concatenate-cuts 0 \
+  --bucketing-sampler 1 \
+  --num-paths 1000 \
+  --method  whole-lattice-rescoring
+```
+
+
+The tensorboard log for training is available at
+https://tensorboard.dev/experiment/QYNzOi52RwOX8yvtpl3hMw/#scalars
+
+
+### MGB2 100h BPE training results (Conformer-CTC) (after 33 epochs)
+
+#### 2022-06-04
+
+The best WER, as of 2022-06-04, for the MGB2 test dataset is below
+
+Using whole lattice HLG decoding + n-gram LM rescoring 
+
+|     | dev        | test       |
+|-----|------------|------------|
+| WER | 25.32      |  23.53     |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+| 0.1            | -            |
+
+
+Using n-best (n=0.5) HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|     | dev        | test       |
+|-----|------------|------------|
+| WER |    27.87   |  26.12     |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+| 0.01           | 0.3             |
+
+
+To reproduce the above result, use the following commands for training:
+
+# Note: the model was trained on V-100 32GB GPU
+
+```
+cd egs/mgb2/ASR
+. ./path.sh
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1"
+./conformer_ctc/train.py \
+  --lang-dir data/lang_bpe_5000 \
+  --att-rate 0.8 \
+  --lr-factor 10 \
+  --max-duration  \
+  --concatenate-cuts 0 \
+  --world-size 2 \
+  --bucketing-sampler 1 \
+  --max-duration 100 \ 
+  --start-epoch 0 \
+  --num-epochs 40
+  
+```
+
+and the following command for nbest decoding
+
+```
+./conformer_ctc/decode.py \
+  --lang-dir data/lang_bpe_5000 \
+  --max-duration 30 \
+  --concatenate-cuts 0 \
+  --bucketing-sampler 1 \
+  --num-paths 1000 \
+  --epoch 40 \
+  --avg 5 \
+  --method attention-decoder \
+  --nbest-scale 0.5
+```
+
+and the following command for whole-lattice decoding
+
+```
+./conformer_ctc/decode.py \
+  --lang-dir data/lang_bpe_5000 \
+  --max-duration 30 \
+  --concatenate-cuts 0 \
+  --bucketing-sampler 1 \
+  --num-paths 1000 \
+  --epoch 40 \
+  --avg 5 \
+  --method  whole-lattice-rescoring
+```
+
+The tensorboard log for training is available at
+<https://tensorboard.dev/experiment/zy6FnumCQlmiO7BPsdCmEg/#scalars>
+
+
+
+
diff --git a/egs/mgb2/ASR/conformer_ctc/__init__.py b/egs/mgb2/ASR/conformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/mgb2/ASR/conformer_ctc/ali.py b/egs/mgb2/ASR/conformer_ctc/ali.py
new file mode 100755
index 000000000..aea962dcd
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/ali.py
@@ -0,0 +1,395 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+    ./conformer_ctc/ali.py \
+            --exp-dir ./conformer_ctc/exp \
+            --lang-dir ./data/lang_bpe_500 \
+            --epoch 20 \
+            --avg 10 \
+            --max-duration 300 \
+            --dataset train-clean-100 \
+            --out-dir data/ali
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import numpy as np
+import torch
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse import CutSet
+from lhotse.features.io import FeaturesWriter, NumpyHdf5Writer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import one_best_decoding
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    encode_supervisions,
+    get_alignments,
+    setup_logger,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=34,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--out-dir",
+        type=str,
+        required=True,
+        help="""Output directory.
+        It contains 3 generated files:
+
+        - labels_xxx.h5
+        - aux_labels_xxx.h5
+        - cuts_xxx.json.gz
+
+        where xxx is the value of `--dataset`. For instance, if
+        `--dataset` is `train-clean-100`, it will contain 3 files:
+
+        - `labels_train-clean-100.h5`
+        - `aux_labels_train-clean-100.h5`
+        - `cuts_train-clean-100.json.gz`
+
+        Note: Both labels_xxx.h5 and aux_labels_xxx.h5 contain framewise
+        alignment. The difference is that labels_xxx.h5 contains repeats.
+        """,
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="""The name of the dataset to compute alignments for.
+        Possible values are:
+            - test-clean.
+            - test-other
+            - train-clean-100
+            - train-clean-360
+            - train-other-500
+            - dev-clean
+            - dev-other
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "lm_dir": Path("data/lm"),
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "subsampling_factor": 4,
+            # Set it to 0 since attention decoder
+            # is not used for computing alignments
+            "num_decoder_layers": 0,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "output_beam": 10,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def compute_alignments(
+    model: torch.nn.Module,
+    dl: torch.utils.data.DataLoader,
+    labels_writer: FeaturesWriter,
+    aux_labels_writer: FeaturesWriter,
+    params: AttributeDict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+) -> CutSet:
+    """Compute the framewise alignments of a dataset.
+
+    Args:
+      model:
+        The neural network model.
+      dl:
+        Dataloader containing the dataset.
+      params:
+        Parameters for computing alignments.
+      graph_compiler:
+        It converts token IDs to decoding graphs.
+    Returns:
+      Return a CutSet. Each cut has two custom fields: labels_alignment
+      and aux_labels_alignment, containing framewise alignments information.
+      Both are of type `lhotse.array.TemporalArray`. The difference between
+      the two alignments is that `labels_alignment` contain repeats.
+    """
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+    num_cuts = 0
+
+    device = graph_compiler.device
+    cuts = []
+    for batch_idx, batch in enumerate(dl):
+        feature = batch["inputs"]
+
+        # at entry, feature is [N, T, C]
+        assert feature.ndim == 3
+        feature = feature.to(device)
+
+        supervisions = batch["supervisions"]
+        cut_list = supervisions["cut"]
+
+        for cut in cut_list:
+            assert len(cut.supervisions) == 1, f"{len(cut.supervisions)}"
+
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is [N, T, C]
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+        # we need also to sort cut_ids as encode_supervisions()
+        # reorders "texts".
+        # In general, new2old is an identity map since lhotse sorts the returned
+        # cuts by duration in descending order
+        new2old = supervision_segments[:, 0].tolist()
+
+        cut_list = [cut_list[i] for i in new2old]
+
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            nnet_output,
+            supervision_segments,
+            allow_truncate=params.subsampling_factor - 1,
+        )
+
+        lattice = k2.intersect_dense(
+            decoding_graph,
+            dense_fsa_vec,
+            params.output_beam,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice,
+            use_double_scores=params.use_double_scores,
+        )
+
+        labels_ali = get_alignments(best_path, kind="labels")
+        aux_labels_ali = get_alignments(best_path, kind="aux_labels")
+        assert len(labels_ali) == len(aux_labels_ali) == len(cut_list)
+        for cut, labels, aux_labels in zip(cut_list, labels_ali, aux_labels_ali):
+            cut.labels_alignment = labels_writer.store_array(
+                key=cut.id,
+                value=np.asarray(labels, dtype=np.int32),
+                # frame shift is 0.01s, subsampling_factor is 4
+                frame_shift=0.04,
+                temporal_dim=0,
+                start=0,
+            )
+            cut.aux_labels_alignment = aux_labels_writer.store_array(
+                key=cut.id,
+                value=np.asarray(aux_labels, dtype=np.int32),
+                # frame shift is 0.01s, subsampling_factor is 4
+                frame_shift=0.04,
+                temporal_dim=0,
+                start=0,
+            )
+
+        cuts += cut_list
+
+        num_cuts += len(cut_list)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return CutSet.from_cuts(cuts)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    args.enable_spec_aug = False
+    args.enable_musan = False
+    args.return_cuts = True
+    args.concatenate_cuts = False
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-ali")
+
+    logging.info(f"Computing alignments for {params.dataset} - started")
+    logging.info(params)
+
+    out_dir = Path(params.out_dir)
+    out_dir.mkdir(exist_ok=True)
+
+    out_labels_ali_filename = out_dir / f"labels_{params.dataset}.h5"
+    out_aux_labels_ali_filename = out_dir / f"aux_labels_{params.dataset}.h5"
+    out_manifest_filename = out_dir / f"cuts_{params.dataset}.json.gz"
+
+    for f in (
+        out_labels_ali_filename,
+        out_aux_labels_ali_filename,
+        out_manifest_filename,
+    ):
+        if f.exists():
+            logging.info(f"{f} exists - skipping")
+            return
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(
+            f"{params.exp_dir}/epoch-{params.epoch}.pt", model, strict=False
+        )
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.eval()
+
+    librispeech = LibriSpeechAsrDataModule(args)
+    if params.dataset == "test-clean":
+        test_clean_cuts = librispeech.test_clean_cuts()
+        dl = librispeech.test_dataloaders(test_clean_cuts)
+    elif params.dataset == "test-other":
+        test_other_cuts = librispeech.test_other_cuts()
+        dl = librispeech.test_dataloaders(test_other_cuts)
+    elif params.dataset == "train-clean-100":
+        train_clean_100_cuts = librispeech.train_clean_100_cuts()
+        dl = librispeech.train_dataloaders(train_clean_100_cuts)
+    elif params.dataset == "train-clean-360":
+        train_clean_360_cuts = librispeech.train_clean_360_cuts()
+        dl = librispeech.train_dataloaders(train_clean_360_cuts)
+    elif params.dataset == "train-other-500":
+        train_other_500_cuts = librispeech.train_other_500_cuts()
+        dl = librispeech.train_dataloaders(train_other_500_cuts)
+    elif params.dataset == "dev-clean":
+        dev_clean_cuts = librispeech.dev_clean_cuts()
+        dl = librispeech.valid_dataloaders(dev_clean_cuts)
+    else:
+        assert params.dataset == "dev-other", f"{params.dataset}"
+        dev_other_cuts = librispeech.dev_other_cuts()
+        dl = librispeech.valid_dataloaders(dev_other_cuts)
+
+    logging.info(f"Processing {params.dataset}")
+    with NumpyHdf5Writer(out_labels_ali_filename) as labels_writer:
+        with NumpyHdf5Writer(out_aux_labels_ali_filename) as aux_labels_writer:
+            cut_set = compute_alignments(
+                model=model,
+                dl=dl,
+                labels_writer=labels_writer,
+                aux_labels_writer=aux_labels_writer,
+                params=params,
+                graph_compiler=graph_compiler,
+            )
+
+    cut_set.to_file(out_manifest_filename)
+
+    logging.info(
+        f"For dataset {params.dataset}, its alignments with repeats are "
+        f"saved to {out_labels_ali_filename}, the alignments without repeats "
+        f"are saved to {out_aux_labels_ali_filename}, and the cut manifest "
+        f"file is {out_manifest_filename}. Number of cuts: {len(cut_set)}"
+    )
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py b/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py
new file mode 100644
index 000000000..7753d1674
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py
@@ -0,0 +1,371 @@
+# Copyright 2022 Johns Hopkins University  (Amir Hussein)
+# Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class MGB2AsrDataModule:
+
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=1,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.jsonl.gz")
+
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts, max_duration=self.args.max_duration, shuffle=False
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_train_shuf.jsonl.gz")
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_test.jsonl.gz")
diff --git a/egs/mgb2/ASR/conformer_ctc/compile_hlg.py b/egs/mgb2/ASR/conformer_ctc/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/compute_fbank_musan.py b/egs/mgb2/ASR/conformer_ctc/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/conformer.py b/egs/mgb2/ASR/conformer_ctc/conformer.py
new file mode 120000
index 000000000..d1f4209d7
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/conformer.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/convert_transcript_words_to_tokens.py b/egs/mgb2/ASR/conformer_ctc/convert_transcript_words_to_tokens.py
new file mode 120000
index 000000000..2ce13fd69
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/convert_transcript_words_to_tokens.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/convert_transcript_words_to_tokens.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/decode.py b/egs/mgb2/ASR/conformer_ctc/decode.py
new file mode 100755
index 000000000..f771d7f1e
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/decode.py
@@ -0,0 +1,695 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import pdb
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import MGB2AsrDataModule
+from conformer import Conformer
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=50,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=20,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+        # TODO: pass `lattice` instead of `rescored_lattice` to
+        # `rescore_with_attention_decoder`
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        # pdb.set_trace()
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for hyp_words, ref_text in zip(hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    MGB2AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in ["whole-lattice-rescoring", "attention-decoder"]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    MGB2 = MGB2AsrDataModule(args)
+
+    test_cuts = MGB2.test_cuts()
+    dev_cuts = MGB2.dev_cuts()
+
+    test_dl = MGB2.test_dataloaders(test_cuts)
+    dev_dl = MGB2.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_all_dl = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_all_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/conformer_ctc/download_lm.py b/egs/mgb2/ASR/conformer_ctc/download_lm.py
new file mode 120000
index 000000000..c9668bd2d
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/download_lm.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/download_lm.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/export.py b/egs/mgb2/ASR/conformer_ctc/export.py
new file mode 120000
index 000000000..60e314d9d
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/export.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/generate_unique_lexicon.py b/egs/mgb2/ASR/conformer_ctc/generate_unique_lexicon.py
new file mode 120000
index 000000000..c0aea1403
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/generate_unique_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/generate_unique_lexicon.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/label_smoothing.py b/egs/mgb2/ASR/conformer_ctc/label_smoothing.py
new file mode 120000
index 000000000..e9d239fff
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/label_smoothing.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/pretrained.py b/egs/mgb2/ASR/conformer_ctc/pretrained.py
new file mode 100755
index 000000000..d30ca98d8
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/pretrained.py
@@ -0,0 +1,430 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from conformer import Conformer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_whole_lattice,
+)
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="""Path to words.txt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.pt.
+        Used only when method is not ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (0) ctc-decoding - Use CTC decoding. It uses a sentence
+            piece model, i.e., lang_dir/bpe.model, to convert
+            word pieces to words. It needs neither a lexicon
+            nor an n-gram LM.
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        (3) attention-decoder - Extract n paths from the rescored
+            lattice and use the transformer attention decoder for
+            rescoring.
+            We call it HLG decoding + n-gram LM rescoring + attention
+            decoder rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring or attention-decoder.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=1.3,
+        help="""
+        Used only when method is whole-lattice-rescoring and attention-decoder.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--attention-decoder-scale",
+        type=float,
+        default=1.2,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for attention decoder scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sos-id",
+        type=int,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the SOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-classes",
+        type=int,
+        default=500,
+        help="""
+        Vocab size in the BPE model.
+        """,
+    )
+
+    parser.add_argument(
+        "--eos-id",
+        type=int,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the EOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    if args.method != "attention-decoder":
+        # to save memory as the attention decoder
+        # will not be used
+        params.num_decoder_layers = 0
+
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    # Note: We don't use key padding mask for attention during decoding
+    with torch.no_grad():
+        nnet_output, memory, memory_key_padding_mask = model(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    if params.method == "ctc-decoding":
+        logging.info("Use CTC decoding")
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(params.bpe_model)
+        max_token_id = params.num_classes - 1
+
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=H,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        token_ids = get_texts(best_path)
+        hyps = bpe_model.decode(token_ids)
+        hyps = [s.split() for s in hyps]
+    elif params.method in [
+        "1best",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]:
+        logging.info(f"Loading HLG from {params.HLG}")
+        HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+        HLG = HLG.to(device)
+        if not hasattr(HLG, "lm_scores"):
+            # For whole-lattice-rescoring and attention-decoder
+            HLG.lm_scores = HLG.scores.clone()
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+        ]:
+            logging.info(f"Loading G from {params.G}")
+            G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = G.to(device)
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G.lm_scores = G.scores.clone()
+
+        lattice = get_lattice(
+            nnet_output=nnet_output,
+            decoding_graph=HLG,
+            supervision_segments=supervision_segments,
+            search_beam=params.search_beam,
+            output_beam=params.output_beam,
+            min_active_states=params.min_active_states,
+            max_active_states=params.max_active_states,
+            subsampling_factor=params.subsampling_factor,
+        )
+
+        if params.method == "1best":
+            logging.info("Use HLG decoding")
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        elif params.method == "whole-lattice-rescoring":
+            logging.info("Use HLG decoding + LM rescoring")
+            best_path_dict = rescore_with_whole_lattice(
+                lattice=lattice,
+                G_with_epsilon_loops=G,
+                lm_scale_list=[params.ngram_lm_scale],
+            )
+            best_path = next(iter(best_path_dict.values()))
+        elif params.method == "attention-decoder":
+            logging.info("Use HLG + LM rescoring + attention decoder rescoring")
+            rescored_lattice = rescore_with_whole_lattice(
+                lattice=lattice, G_with_epsilon_loops=G, lm_scale_list=None
+            )
+            best_path_dict = rescore_with_attention_decoder(
+                lattice=rescored_lattice,
+                num_paths=params.num_paths,
+                model=model,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                sos_id=params.sos_id,
+                eos_id=params.eos_id,
+                nbest_scale=params.nbest_scale,
+                ngram_lm_scale=params.ngram_lm_scale,
+                attention_scale=params.attention_decoder_scale,
+            )
+            best_path = next(iter(best_path_dict.values()))
+
+        hyps = get_texts(best_path)
+        word_sym_table = k2.SymbolTable.from_file(params.words_file)
+        hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/mgb2/ASR/conformer_ctc/subsampling.py b/egs/mgb2/ASR/conformer_ctc/subsampling.py
new file mode 120000
index 000000000..16354dc73
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/test_label_smoothing.py b/egs/mgb2/ASR/conformer_ctc/test_label_smoothing.py
new file mode 120000
index 000000000..04b959ecf
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/test_label_smoothing.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/test_label_smoothing.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/test_subsampling.py b/egs/mgb2/ASR/conformer_ctc/test_subsampling.py
new file mode 120000
index 000000000..98c3be3e6
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/test_subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/test_subsampling.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/test_transformer.py b/egs/mgb2/ASR/conformer_ctc/test_transformer.py
new file mode 120000
index 000000000..8b0990ec6
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/test_transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/test_transformer.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/conformer_ctc/train.py b/egs/mgb2/ASR/conformer_ctc/train.py
new file mode 100755
index 000000000..08ffee210
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/train.py
@@ -0,0 +1,766 @@
+#!/usr/bin/env python3
+# Copyright 2022 Johns Hopkins University  (Amir Hussein)
+# Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import MGB2AsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=50,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Normalization for the input features, can be a
+                              boolean indicating whether to do batch
+                              normalization, or a float which means just scaling
+                              the input features with this float value.
+                              If given a float value, we will remove batchnorm
+                              layer in `ConvolutionModule` as well.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            "num_decoder_layers": 6,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    token_ids = graph_compiler.texts_to_ids(texts)
+
+    decoding_graph = graph_compiler.compile(token_ids)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction="none",
+        use_double_scores=params.use_double_scores,
+    )
+    # filter inf from ctc_loss
+    ctc_loss = torch.sum(
+        torch.where(
+            ctc_loss != float("inf"),
+            ctc_loss,
+            torch.tensor(0, dtype=torch.float32).to(device),
+        )
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch["inputs"].shape[0] == len(batch["supervisions"]["text"]):
+            params.batch_idx_train += 1
+            batch_size = len(batch["supervisions"]["text"])
+
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+            # if tot_loss is None:
+            #     logging.warning("Batch mismatch. Skipping ...")
+            #     del batch
+            #     del tot_loss
+            #     continue;
+            # elif tot_loss.isinf() or tot_loss.isnan():
+            #     logging.warning("NaN or Inf loss. Skipping ...")
+            #     del batch
+            #     del tot_loss
+            #     continue;
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+
+            optimizer.zero_grad()
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+
+            if batch_idx % params.log_interval == 0:
+                logging.info(
+                    f"Epoch {params.cur_epoch}, "
+                    f"batch {batch_idx}, loss[{loss_info}], "
+                    f"tot_loss[{tot_loss}], batch size: {batch_size}"
+                )
+
+            if batch_idx % params.log_interval == 0:
+
+                if tb_writer is not None:
+                    loss_info.write_summary(
+                        tb_writer, "train/current_", params.batch_idx_train
+                    )
+                    tot_loss.write_summary(
+                        tb_writer, "train/tot_", params.batch_idx_train
+                    )
+
+            if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+                logging.info("Computing validation loss")
+                valid_info = compute_validation_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    valid_dl=valid_dl,
+                    world_size=world_size,
+                )
+                model.train()
+                logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+                if tb_writer is not None:
+                    valid_info.write_summary(
+                        tb_writer, "train/valid_", params.batch_idx_train
+                    )
+        else:
+            logging.warning(
+                f"Batch {batch_idx} mismatch in dimentions between the input and the output. Skipping ..."
+            )
+            continue
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(42)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    MGB2 = MGB2AsrDataModule(args)
+
+    train_cuts = MGB2.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 0.5 <= c.duration <= 30.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_dl = MGB2.train_dataloaders(train_cuts)
+
+    valid_cuts = MGB2.dev_cuts()
+    valid_dl = MGB2.test_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    MGB2AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/conformer_ctc/transformer.py b/egs/mgb2/ASR/conformer_ctc/transformer.py
new file mode 120000
index 000000000..1c3f43fcf
--- /dev/null
+++ b/egs/mgb2/ASR/conformer_ctc/transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/transformer.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/local/__init__.py b/egs/mgb2/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/mgb2/ASR/local/compile_hlg.py b/egs/mgb2/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/mgb2/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/local/compute_fbank_mgb2.py b/egs/mgb2/ASR/local/compute_fbank_mgb2.py
new file mode 100755
index 000000000..6cae69e41
--- /dev/null
+++ b/egs/mgb2/ASR/local/compute_fbank_mgb2.py
@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+# Copyright 2022 Johns Hopkins University  (Amir Hussein)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the MGB2 dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_mgb2():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "train",
+        "test",
+        "dev",
+    )
+    manifests = read_manifests_if_cached(
+        prefix="mgb2", dataset_parts=dataset_parts, output_dir=src_dir
+    )
+    assert manifests is not None
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"cuts_{partition}.json.gz").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            logging.info("About to split cuts into smaller chunks.")
+            cut_set = cut_set.trim_to_supervisions(
+                keep_overlapping=False, min_duration=None
+            )
+            cut_set.to_file(output_dir / f"cuts_{partition}.jsonl.gz")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_mgb2()
diff --git a/egs/mgb2/ASR/local/compute_fbank_musan.py b/egs/mgb2/ASR/local/compute_fbank_musan.py
new file mode 100755
index 000000000..5d0d69a13
--- /dev/null
+++ b/egs/mgb2/ASR/local/compute_fbank_musan.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import (
+    ChunkedLilcomHdf5Writer,
+    CutSet,
+    Fbank,
+    FbankConfig,
+    LilcomChunkyWriter,
+    combine,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        prefix=prefix,
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix=suffix,
+    )
+    assert manifests is not None
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+    )
+
+    musan_cuts_path = output_dir / "cuts_musan.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        # create chunks of Musan with duration 5 - 10 seconds
+        musan_cuts = (
+            CutSet.from_manifests(
+                recordings=combine(part["recordings"] for part in manifests.values())
+            )
+            .cut_into_windows(10.0)
+            .filter(lambda c: c.duration > 5)
+            .compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/feats_musan",
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        musan_cuts.to_file(musan_cuts_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()
diff --git a/egs/mgb2/ASR/local/convert_transcript_words_to_tokens.py b/egs/mgb2/ASR/local/convert_transcript_words_to_tokens.py
new file mode 100755
index 000000000..a8d5117c9
--- /dev/null
+++ b/egs/mgb2/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+"""
+Convert a transcript file containing words to a corpus file containing tokens
+for LM training with the help of a lexicon.
+
+If the lexicon contains phones, the resulting LM will be a phone LM; If the
+lexicon contains word pieces, the resulting LM will be a word piece LM.
+
+If a word has multiple pronunciations, the one that appears first in the lexicon
+is kept; others are removed.
+
+If the input transcript is:
+
+    hello zoo world hello
+    world zoo
+    foo zoo world hellO
+
+and if the lexicon is
+
+    <UNK> SPN
+    hello h e l l o 2
+    hello h e l l o
+    world w o r l d
+    zoo z o o
+
+Then the output is
+
+    h e l l o 2 z o o w o r l d h e l l o 2
+    w o r l d z o o
+    SPN z o o w o r l d SPN
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+from generate_unique_lexicon import filter_multiple_pronunications
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="The input transcript file."
+        "We assume that the transcript file consists of "
+        "lines. Each line consists of space separated words.",
+    )
+    parser.add_argument("--lexicon", type=str, help="The input lexicon file.")
+    parser.add_argument("--oov", type=str, default="<UNK>", help="The OOV word.")
+
+    return parser.parse_args()
+
+
+def process_line(lexicon: Dict[str, List[str]], line: str, oov_token: str) -> None:
+    """
+    Args:
+      lexicon:
+        A dict containing pronunciations. Its keys are words and values
+        are pronunciations (i.e., tokens).
+      line:
+        A line of transcript consisting of space(s) separated words.
+      oov_token:
+        The pronunciation of the oov word if a word in `line` is not present
+        in the lexicon.
+    Returns:
+      Return None.
+    """
+    s = ""
+    words = line.strip().split()
+    for i, w in enumerate(words):
+        tokens = lexicon.get(w, oov_token)
+        s += " ".join(tokens)
+        s += " "
+    print(s.strip())
+
+
+def main():
+    args = get_args()
+    assert Path(args.lexicon).is_file()
+    assert Path(args.transcript).is_file()
+    assert len(args.oov) > 0
+
+    # Only the first pronunciation of a word is kept
+    lexicon = filter_multiple_pronunications(read_lexicon(args.lexicon))
+
+    lexicon = dict(lexicon)
+
+    assert args.oov in lexicon
+
+    oov_token = lexicon[args.oov]
+
+    with open(args.transcript) as f:
+        for line in f:
+            process_line(lexicon=lexicon, line=line, oov_token=oov_token)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/local/display_manifest_statistics.py b/egs/mgb2/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..d3e224905
--- /dev/null
+++ b/egs/mgb2/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest
+
+
+def main():
+    # path = "./data/fbank/cuts_train.jsonl.gz"
+    path = "./data/fbank/cuts_dev.jsonl.gz"
+    # path = "./data/fbank/cuts_test.jsonl.gz"
+
+    cuts = load_manifest(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+# train
+
+Cuts count: 1125309
+Total duration (hours): 3403.9
+Speech duration (hours): 3403.9 (100.0%)
+***
+Duration statistics (seconds):
+mean    10.9
+std     10.1
+min     0.2
+25%     5.2
+50%     7.8
+75%     12.7
+99%     52.0
+99.5%   65.1
+99.9%   99.5
+max     228.9
+
+
+# test
+Cuts count: 5365
+Total duration (hours): 9.6
+Speech duration (hours): 9.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.4
+std     1.5
+min     1.6
+25%     5.3
+50%     6.5
+75%     7.6
+99%     9.5
+99.5%   9.7
+99.9%   10.3
+max     12.4
+
+# dev
+Cuts count: 5002
+Total duration (hours): 8.5
+Speech duration (hours): 8.5 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.1
+std     1.7
+min     1.5
+25%     4.8
+50%     6.2
+75%     7.4
+99%     9.5
+99.5%   9.7
+99.9%   10.1
+max     20.3
+
+"""
diff --git a/egs/mgb2/ASR/local/generate_unique_lexicon.py b/egs/mgb2/ASR/local/generate_unique_lexicon.py
new file mode 120000
index 000000000..c0aea1403
--- /dev/null
+++ b/egs/mgb2/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/generate_unique_lexicon.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/local/prep_mgb2_lexicon.sh b/egs/mgb2/ASR/local/prep_mgb2_lexicon.sh
new file mode 100755
index 000000000..3b673db6f
--- /dev/null
+++ b/egs/mgb2/ASR/local/prep_mgb2_lexicon.sh
@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+
+# Copyright 2022 QCRI (author: Amir Hussein)
+# Apache 2.0
+# This script prepares the graphemic lexicon.
+
+dir=data/local/dict
+lexicon_url1="https://arabicspeech.org/arabicspeech-portal-resources/lexicon/ar-ar_grapheme_lexicon_20160209.bz2";
+lexicon_url2="https://arabicspeech.org/arabicspeech-portal-resources/lexicon/ar-ar_phoneme_lexicon_20140317.bz2";
+stage=0
+lang_dir=download/lm
+mkdir -p $lang_dir
+
+if [ $stage -le 0 ]; then
+  echo "$0: Downloading text for lexicon... $(date)."
+  wget --no-check-certificate -P $lang_dir $lexicon_url1
+  wget --no-check-certificate -P $lang_dir $lexicon_url2
+  bzcat $lang_dir/ar-ar_grapheme_lexicon_20160209.bz2 | sed '1,3d' | awk '{print $1}'  >  $lang_dir/grapheme_lexicon
+  bzcat $lang_dir/ar-ar_phoneme_lexicon_20140317.bz2 | sed '1,3d' | awk '{print $1}' >>  $lang_dir/phoneme_lexicon
+  cat download/lm/train/text | cut -d ' ' -f 2- | tr -s " " "\n" | sort -u >> $lang_dir/uniq_words
+fi
+
+
+if [ $stage -le 0 ]; then
+  echo "$0: processing lexicon text and creating lexicon... $(date)."
+  # remove vowels and  rare alef wasla
+  cat $lang_dir/uniq_words |  sed -e 's:[FNKaui\~o\`]::g' -e 's:{:}:g' | sed -r '/^\s*$/d' | sort -u > $lang_dir/grapheme_lexicon.txt
+fi
+
+echo "$0: Lexicon preparation succeeded"
diff --git a/egs/mgb2/ASR/local/prepare_lang.py b/egs/mgb2/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/mgb2/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/local/prepare_lang_bpe.py b/egs/mgb2/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/mgb2/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/local/prepare_mgb2_lexicon.py b/egs/mgb2/ASR/local/prepare_mgb2_lexicon.py
new file mode 100755
index 000000000..99e1fa34d
--- /dev/null
+++ b/egs/mgb2/ASR/local/prepare_mgb2_lexicon.py
@@ -0,0 +1,37 @@
+#!/usr/bin/env python3
+
+# Copyright      2022  Amir Hussein
+# Apache 2.0
+
+# This script prepares givel a column of words lexicon.
+
+import argparse
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        description="""Creates the list of characters and words in lexicon"""
+    )
+    parser.add_argument("input", type=str, help="""Input list of words file""")
+    parser.add_argument("output", type=str, help="""output graphemic lexicon""")
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    lex = {}
+    args = get_args()
+    with open(args.input, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            characters = list(line)
+            characters = " ".join(["V" if char == "*" else char for char in characters])
+            lex[line] = characters
+
+    with open(args.output, "w", encoding="utf-8") as fp:
+        for key in sorted(lex):
+            fp.write(key + "  " + lex[key] + "\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/local/test_prepare_lang.py b/egs/mgb2/ASR/local/test_prepare_lang.py
new file mode 120000
index 000000000..f0f864998
--- /dev/null
+++ b/egs/mgb2/ASR/local/test_prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/test_prepare_lang.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/prepare.sh b/egs/mgb2/ASR/prepare.sh
new file mode 100755
index 000000000..4ea427371
--- /dev/null
+++ b/egs/mgb2/ASR/prepare.sh
@@ -0,0 +1,234 @@
+#!/usr/bin/env bash
+# Copyright 2022 Johns Hopkins University  (Amir Hussein)
+# Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+set -eou pipefail
+nj=30
+stage=7
+stop_stage=1000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. 
+#
+#  - $dl_dir/mgb2
+#      
+#      You can download the data from 
+#
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+#
+# Note: MGB2 is not available for direct 
+# download, however you can fill out the form and  
+# download it from https://arabicspeech.org/mgb2 
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  5000
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/MGB2,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/mgb2 $dl_dir/MGB2
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare mgb2 manifest"
+  # We assume that you have downloaded the mgb2 corpus
+  # to $dl_dir/mgb2
+  mkdir -p data/manifests
+
+  lhotse prepare mgb2 $dl_dir/mgb2 data/manifests
+  
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for mgb2"
+  mkdir -p data/fbank
+  ./local/compute_fbank_mgb2.py
+   # shufling the data
+  gunzip -c data/fbank/cuts_train.jsonl.gz | shuf | gzip -c > data/fbank/cuts_train_shuf.jsonl.gz
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  ./local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  if [[ ! -e download/lm/train/text ]]; then 
+  # export train text file to build grapheme lexicon 
+  lhotse kaldi export \
+    data/manifests/mgb2_recordings_train.jsonl.gz \
+    data/manifests/mgb2_supervisions_train.jsonl.gz  \
+    download/lm/train
+  fi
+
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+  ./local/prep_mgb2_lexicon.sh 
+  python local/prepare_mgb2_lexicon.py  $dl_dir/lm/grapheme_lexicon.txt  $dl_dir/lm/lexicon.txt
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/lm/lexicon.txt |
+    sort | uniq > $lang_dir/lexicon.txt
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+fi
+
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang_phone/words.txt $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      files=$(
+        find "$dl_dir/lm/train" -name "text"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2- | sed -r '/^\s*$/d'
+      done > $lang_dir/transcript_words.txt
+    fi
+
+    ./local/train_bpe_model.py \
+      --lang-dir $lang_dir \
+      --vocab-size $vocab_size \
+      --transcript $lang_dir/transcript_words.txt
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+    fi
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare bigram P"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+      ./local/convert_transcript_words_to_tokens.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --transcript $lang_dir/transcript_words.txt \
+        --oov "<UNK>" \
+        > $lang_dir/transcript_tokens.txt
+    fi
+
+    if [ ! -f $lang_dir/P.arpa ]; then
+      ./shared/make_kn_lm.py \
+        -ngram-order 2 \
+        -text $lang_dir/transcript_tokens.txt \
+        -lm $lang_dir/P.arpa
+    fi
+
+    if [ ! -f $lang_dir/P.fst.txt ]; then
+      python3 -m kaldilm \
+        --read-symbol-table="$lang_dir/tokens.txt" \
+        --disambig-symbol='#0' \
+        --max-order=2 \
+        $lang_dir/P.arpa > $lang_dir/P.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p data/lm
+    if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+      # It is used in building HLG
+      ./shared/make_kn_lm.py \
+          -ngram-order 3 \
+          -text $lang_dir/transcript_words.txt \
+          -lm $lang_dir/G.arpa
+
+      python3 -m kaldilm \
+        --read-symbol-table="data/lang_phone/words.txt" \
+        --disambig-symbol='#0' \
+        --max-order=3 \
+        $lang_dir/G.arpa > data/lm/G_3_gram.fst.txt
+    fi
+
+    if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+      # It is used for LM rescoring
+      ./shared/make_kn_lm.py \
+          -ngram-order 4 \
+          -text $lang_dir/transcript_words.txt \
+          -lm $lang_dir/4-gram.arpa
+
+      python3 -m kaldilm \
+        --read-symbol-table="data/lang_phone/words.txt" \
+        --disambig-symbol='#0' \
+        --max-order=4 \
+        $lang_dir/4-gram.arpa > data/lm/G_4_gram.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+  done
+fi
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/__init__.py b/egs/mgb2/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/mgb2/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 120000
index 000000000..a73848de9
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1 @@
+../conformer_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/beam_search.py b/egs/mgb2/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..02d01b343
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/beam_search.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/conformer.py b/egs/mgb2/ASR/pruned_transducer_stateless5/conformer.py
new file mode 120000
index 000000000..c7c1a4b6e
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/decode.py b/egs/mgb2/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..72338bade
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,619 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins        (authors: Amir Hussein)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 18 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 200 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 18 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 200 \
+    --decoding-method beam_search \
+    --beam-size 10
+
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 18 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 10
+
+(4) fast beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 18 \
+    --avg 5 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 200 \
+    --decoding-method fast_beam_search \
+    --beam-size 10 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import MGB2AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_2000/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    MGB2AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    MGB2 = MGB2AsrDataModule(args)
+
+    test_cuts = MGB2.test_cuts()
+    dev_cuts = MGB2.dev_cuts()
+
+    test_dl = MGB2.test_dataloaders(test_cuts)
+    dev_dl = MGB2.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_all_dl = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_all_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/decoder.py b/egs/mgb2/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..6775ee67e
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/decoder.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/mgb2/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..972e44ca4
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/export.py b/egs/mgb2/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..7a5d7f680
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert args.jit is False, "Support torchscript will be added later"
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/joiner.py b/egs/mgb2/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..f5279e151
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/joiner.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/model.py b/egs/mgb2/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..7b417fd89
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/model.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/optim.py b/egs/mgb2/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..210374f22
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/optim.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/pretrained.py b/egs/mgb2/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100755
index 000000000..77ba0873b
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,344 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/scaling.py b/egs/mgb2/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..ff7bfeda9
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/scaling.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/test_model.py b/egs/mgb2/ASR/pruned_transducer_stateless5/test_model.py
new file mode 120000
index 000000000..b71d7bb81
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/test_model.py
\ No newline at end of file
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/train.py b/egs/mgb2/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..48468cfbd
--- /dev/null
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1162 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins        (authors: Amir Hussein)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 2 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --max-duration 200 \
+  --num-buckets 50
+
+# For mix precision training:
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 2 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --max-duration 200	\
+  --num-buckets 50
+
+"""
+
+# xxx
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import nvidia_smi
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import MGB2AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_2000/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need " "to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=10,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=True,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 80000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+    reduction="none",
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        inf_flag = False
+        if not torch.all(is_finite):
+            inf_flag = True
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    # info["utterances"] = feature.size(0)
+    # # averaged input duration in frames over utterances
+    # info["utt_duration"] = feature_lens.sum().item()
+    # # averaged padding proportion over utterances
+    # info["utt_pad_proportion"] = (
+    #     ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    # )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info, inf_flag
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+    with torch.no_grad():
+        for batch_idx, batch in enumerate(valid_dl):
+            loss, loss_info, inf_flag = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=False,
+            )
+            assert loss.requires_grad is False
+            tot_loss = tot_loss + loss_info
+
+        if world_size > 1:
+            tot_loss.reduce(loss.device)
+
+        loss_value = tot_loss["loss"] / tot_loss["frames"]
+        if loss_value < params.best_valid_loss:
+            params.best_valid_epoch = params.cur_epoch
+            params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch["inputs"].shape[0] == len(batch["supervisions"]["text"]):
+            params.batch_idx_train += 1
+            batch_size = len(batch["supervisions"]["text"])
+
+            try:
+                with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                    loss, loss_info, inf_flag = compute_loss(
+                        params=params,
+                        model=model,
+                        sp=sp,
+                        batch=batch,
+                        is_training=True,
+                        warmup=(params.batch_idx_train / params.model_warm_step),
+                    )
+                # summary stats
+                tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+                # NOTE: We use reduction==sum and loss is computed over utterances
+                # in the batch and there is no normalization to it so far.
+                if not inf_flag:
+                    scaler.scale(loss).backward()
+                    scheduler.step_batch(params.batch_idx_train)
+                    scaler.step(optimizer)
+                    scaler.update()
+                    optimizer.zero_grad()
+                else:
+                    continue
+            except:  # noqa
+                display_and_save_batch(batch, params=params, sp=sp)
+                raise
+
+            if params.print_diagnostics and batch_idx == 5:
+                return
+
+            if (
+                rank == 0
+                and params.batch_idx_train > 0
+                and params.batch_idx_train % params.average_period == 0
+            ):
+                update_averaged_model(
+                    params=params,
+                    model_cur=model,
+                    model_avg=model_avg,
+                )
+
+            if (
+                params.batch_idx_train > 0
+                and params.batch_idx_train % params.save_every_n == 0
+            ):
+                save_checkpoint_with_global_batch_idx(
+                    out_dir=params.exp_dir,
+                    global_batch_idx=params.batch_idx_train,
+                    model=model,
+                    model_avg=model_avg,
+                    params=params,
+                    optimizer=optimizer,
+                    scheduler=scheduler,
+                    sampler=train_dl.sampler,
+                    scaler=scaler,
+                    rank=rank,
+                )
+                remove_checkpoints(
+                    out_dir=params.exp_dir,
+                    topk=params.keep_last_k,
+                    rank=rank,
+                )
+
+            if batch_idx % params.log_interval == 0:
+                cur_lr = scheduler.get_last_lr()[0]
+                # https://silpara.medium.com/check-gpu-memory-usage-from-python-ccca503322ea
+                memory_debugging()
+                logging.info(
+                    f"Epoch {params.cur_epoch}, "
+                    f"batch {batch_idx}, loss[{loss_info}], "
+                    f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                    f"lr: {cur_lr:.2e}"
+                )
+
+                if tb_writer is not None:
+                    tb_writer.add_scalar(
+                        "train/learning_rate", cur_lr, params.batch_idx_train
+                    )
+
+                    loss_info.write_summary(
+                        tb_writer, "train/current_", params.batch_idx_train
+                    )
+                    tot_loss.write_summary(
+                        tb_writer, "train/tot_", params.batch_idx_train
+                    )
+
+            if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+                logging.info("Computing validation loss")
+                valid_info = compute_validation_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    valid_dl=valid_dl,
+                    world_size=world_size,
+                )
+                model.train()
+                logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+                if tb_writer is not None:
+                    valid_info.write_summary(
+                        tb_writer, "train/valid_", params.batch_idx_train
+                    )
+        else:
+            logging.warning(
+                f"Batch {batch_idx} mismatch in dimentions between the input and the output. Skipping ..."
+            )
+            continue
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def memory_debugging():
+    # memory nvidia debugging
+    nvidia_smi.nvmlInit()
+
+    deviceCount = nvidia_smi.nvmlDeviceGetCount()
+    for i in range(deviceCount):
+        handle = nvidia_smi.nvmlDeviceGetHandleByIndex(i)
+        info = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)
+        logging.info(
+            "Device {}: {}, Memory : ({:.2f}% free): {}(total), {} (free), {} (used)".format(
+                i,
+                nvidia_smi.nvmlDeviceGetName(handle),
+                100 * info.free / info.total,
+                info.total,
+                info.free,
+                info.used,
+            )
+        )
+
+    nvidia_smi.nvmlShutdown()
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    MGB2 = MGB2AsrDataModule(args)
+    train_cuts = MGB2.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 30 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 0.5 <= c.duration <= 30.0
+
+    def remove_short_and_long_text(c: Cut):
+        # Keep only text with charachters between 20 and 450
+
+        return 20 <= len(c.supervisions[0].text) <= 450
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.filter(remove_short_and_long_text)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = MGB2.train_dataloaders(train_cuts, sampler_state_dict=sampler_state_dict)
+
+    valid_cuts = MGB2.dev_cuts()
+    valid_dl = MGB2.test_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            # clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+
+def main():
+    parser = get_parser()
+    MGB2AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/mgb2/ASR/shared b/egs/mgb2/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/mgb2/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/README.md b/egs/multi_zh-hans/ASR/README.md
new file mode 100644
index 000000000..537816a5d
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/README.md
@@ -0,0 +1,39 @@
+
+# Introduction
+
+This recipe includes scripts for training Zipformer model using multiple Chinese datasets.
+
+# Included Training Sets
+1. THCHS-30
+2. AiShell-{1,2,4}
+3. ST-CMDS
+4. Primewords
+5. MagicData
+6. Aidatatang_200zh
+7. AliMeeting
+8. WeNetSpeech
+9. KeSpeech-ASR
+
+|Datset| Number of hours| URL|
+|---|---:|---|
+|**TOTAL**|14,106|---|
+|THCHS-30|35|https://www.openslr.org/18/|
+|AiShell-1|170|https://www.openslr.org/33/|
+|AiShell-2|1,000|http://www.aishelltech.com/aishell_2|
+|AiShell-4|120|https://www.openslr.org/111/|
+|ST-CMDS|110|https://www.openslr.org/38/|
+|Primewords|99|https://www.openslr.org/47/|
+|aidatatang_200zh|200|https://www.openslr.org/62/|
+|MagicData|755|https://www.openslr.org/68/|
+|AliMeeting|100|https://openslr.org/119/|
+|WeNetSpeech|10,000|https://github.com/wenet-e2e/WenetSpeech|
+|KeSpeech|1,542|https://github.com/KeSpeech/KeSpeech|
+
+
+# Included Test Sets
+1. Aishell-{1,2,4}
+2. Aidatatang_200zh
+3. AliMeeting
+4. MagicData
+5. KeSpeech-ASR
+6. WeNetSpeech
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/RESULTS.md b/egs/multi_zh-hans/ASR/RESULTS.md
new file mode 100644
index 000000000..15e789604
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/RESULTS.md
@@ -0,0 +1,73 @@
+## Results
+
+### Multi Chinese datasets char-based training results (Non-streaming) on zipformer model
+
+This is the [pull request #1238](https://github.com/k2-fsa/icefall/pull/1238) in icefall.
+
+#### Non-streaming (with CTC head)
+
+Best results (num of params : ~69M):
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 20 \
+  --use-fp16 1 \
+  --max-duration 600 \
+  --num-workers 8 \
+  --use-ctc 1
+```
+
+The decoding command:
+
+```
+./zipformer/decode.py \
+  --epoch 20 \
+  --avg 1 \
+  --use-ctc 1
+```
+
+Character Error Rates (CERs) listed below are produced by the checkpoint of the 20th epoch using BPE model ( # tokens is 2000, byte fallback enabled).
+
+| Datasets | aidatatang _200zh | aidatatang _200zh | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech |
+|--------------------------------|------------------------------|-------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------|
+|  Zipformer  CER   (%) | dev | test | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net |
+| CTC Decoding | 2.86 | 3.36 | 22.93 | 24.28 | 2.05 | 2.27 | 3.33 | 3.82 | 15.45 | 3.49 | 2.77 | 6.90 | 2.85 | 8.29 | 9.41 | 6.92 | 8.57 |
+| Greedy Search   | 3.36 | 3.83 | 23.90 | 25.18 | 2.77 | 3.08 | 3.70 | 4.04 | 16.13 | 3.77 | 3.15 | 6.88 | 3.14 | 8.08 | 9.04 | 7.19 | 8.17 |
+
+Pre-trained model can be found here : https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24/
+
+#### Non-streaming
+
+Best results (num of params : ~69M):
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 20 \
+  --use-fp16 1 \
+  --max-duration 600 \
+  --num-workers 8
+```
+
+The decoding command:
+
+```
+./zipformer/decode.py \
+  --epoch 20 \
+  --avg 1
+```
+
+Character Error Rates (CERs) listed below are produced by the checkpoint of the 20th epoch using greedy search and BPE model ( # tokens is 2000, byte fallback enabled).
+
+| Datasets | aidatatang _200zh | aidatatang _200zh | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech |
+|--------------------------------|------------------------------|-------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------|
+| Zipformer CER   (%) | dev | test | eval| test | dev | test | dev| test | test | dev| test | dev phase1 | dev phase2 | test | dev | test meeting | test net |
+| Greedy Search | 3.2 | 3.67 | 23.15 | 24.78 | 2.91 | 3.04 | 3.59 | 4.03 | 15.68 | 3.68 | 3.12 | 6.69 | 3.19 | 8.01 | 9.32 | 7.05 | 8.78 |
+
+
+Pre-trained model can be found here : https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/
diff --git a/egs/multi_zh-hans/ASR/local/bpe_model_to_tokens.py b/egs/multi_zh-hans/ASR/local/bpe_model_to_tokens.py
new file mode 100755
index 000000000..d078e5b98
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/bpe_model_to_tokens.py
@@ -0,0 +1,37 @@
+#!/usr/bin/env python3
+
+"""
+This script takes `bpe.model` as input and generates a file `tokens.txt`
+from it.
+
+Usage:
+./bpe_model_to_tokens.py /path/to/input/bpe.model > tokens.txt
+"""
+import argparse
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "bpe_model",
+        type=str,
+        help="Path to the input bpe.model",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    for i in range(sp.vocab_size()):
+        print(sp.id_to_piece(i), i)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/local/compile_lg.py b/egs/multi_zh-hans/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py
new file mode 100755
index 000000000..2581ee42f
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+# Copyright    2023  Xiaomi Corp.             (Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_kespeech_dev_test():
+    in_out_dir = Path("data/fbank/kespeech")
+    # number of workers in dataloader
+    num_workers = 42
+
+    # number of seconds in a batch
+    batch_duration = 600
+
+    subsets = (
+        "dev_phase1",
+        "dev_phase2",
+        "test",
+    )
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    for partition in subsets:
+        cuts_path = in_out_dir / f"kespeech-asr_cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = in_out_dir / f"kespeech-asr_cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{in_out_dir}/feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_kespeech_dev_test()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py
new file mode 100755
index 000000000..8bfbc7b50
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py
@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+# Copyright    2023  Xiaomi Corp.             (Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+    set_audio_duration_mismatch_tolerance,
+    set_caching_enabled,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--training-subset",
+        type=str,
+        default="train_phase1",
+        choices=["train_phase1", "train_phase2"],
+        help="The training subset for computing fbank feature.",
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the given subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+    return parser
+
+
+def compute_fbank_kespeech_splits(args):
+    subset = args.training_subset
+    subset = str(subset)
+    num_splits = args.num_splits
+    output_dir = f"data/fbank/kespeech/{subset}_split_{num_splits}"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    set_audio_duration_mismatch_tolerance(0.01)  # 10ms tolerance
+    set_caching_enabled(False)
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"kespeech-asr_cuts_{subset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"kespeech-asr_cuts_{subset}_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/feats_{subset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def main():
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+
+    log_filename = "log-compute_fbank_kespeech_splits"
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    log_filename = f"{log_filename}-{date_time}"
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=logging.INFO,
+        filemode="w",
+    )
+
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter(formatter))
+    logging.getLogger("").addHandler(console)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    compute_fbank_kespeech_splits(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py b/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py
new file mode 100755
index 000000000..5649d3815
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py
@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the MagicData dataset.
+It looks for manifests in the directory data/manifests/magicdata.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_magicdata(num_mel_bins: int = 80, speed_perturb: bool = False):
+    src_dir = Path("data/manifests/magicdata")
+    output_dir = Path("data/fbank")
+    num_jobs = min(30, os.cpu_count())
+
+    dataset_parts = ("train", "test", "dev")
+    prefix = "magicdata"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and speed_perturb:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--speed-perturb",
+        type=bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_magicdata(
+        num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb
+    )
diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py b/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py
new file mode 100755
index 000000000..303a16580
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py
@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the Primewords dataset.
+It looks for manifests in the directory data/manifests/primewords.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_primewords(num_mel_bins: int = 80, speed_perturb: bool = False):
+    src_dir = Path("data/manifests/primewords")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = ("train",)
+    prefix = "primewords"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and speed_perturb:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--speed-perturb",
+        type=bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_primewords(
+        num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb
+    )
diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py b/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py
new file mode 100755
index 000000000..730806954
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py
@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the ST-CMDS dataset.
+It looks for manifests in the directory data/manifests/stcmds.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_stcmds(num_mel_bins: int = 80, speed_perturb: bool = False):
+    src_dir = Path("data/manifests/stcmds")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = ("train",)
+    prefix = "stcmds"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition and speed_perturb:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--speed-perturb",
+        type=bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_stcmds(
+        num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb
+    )
diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py b/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py
new file mode 100755
index 000000000..58bb8002a
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py
@@ -0,0 +1,127 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the THCHS-30 dataset.
+It looks for manifests in the directory data/manifests/thchs30.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_thchs30(num_mel_bins: int = 80, speed_perturb: bool = False):
+    src_dir = Path("data/manifests/thchs30")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+    prefix = "thchs_30"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    (cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1))
+                    if speed_perturb
+                    else cut_set
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+    parser.add_argument(
+        "--speed-perturb",
+        type=bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_thchs30(
+        num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb
+    )
diff --git a/egs/multi_zh-hans/ASR/local/prepare_char.py b/egs/multi_zh-hans/ASR/local/prepare_char.py
new file mode 120000
index 000000000..be7da61af
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/prepare_char.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/prepare_char.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/local/prepare_for_bpe_model.py b/egs/multi_zh-hans/ASR/local/prepare_for_bpe_model.py
new file mode 100755
index 000000000..020800c15
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/prepare_for_bpe_model.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script tokenizes the training transcript by CJK characters
+# and saves the result to transcript_chars.txt, which is used
+# to train the BPE model later.
+
+import argparse
+from pathlib import Path
+
+from tqdm.auto import tqdm
+
+from icefall.utils import tokenize_by_CJK_char
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Output directory.
+        The generated transcript_chars.txt is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="WenetSpeech training transcript.",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    text = Path(args.text)
+
+    assert lang_dir.exists() and text.exists(), f"{lang_dir} or {text} does not exist!"
+
+    transcript_path = lang_dir / "transcript_chars.txt"
+
+    with open(text, "r", encoding="utf-8") as fin:
+        with open(transcript_path, "w+", encoding="utf-8") as fout:
+            for line in fin:
+                fout.write(tokenize_by_CJK_char(line) + "\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/local/prepare_lang.py b/egs/multi_zh-hans/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/local/prepare_lang_bpe.py b/egs/multi_zh-hans/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/local/preprocess_kespeech.py b/egs/multi_zh-hans/ASR/local/preprocess_kespeech.py
new file mode 100755
index 000000000..20274263f
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/preprocess_kespeech.py
@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+# Copyright    2023  Xiaomi Corp.             (Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import re
+from pathlib import Path
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall import setup_logger
+
+# Similar text filtering and normalization procedure as in:
+# https://github.com/SpeechColab/WenetSpeech/blob/main/toolkits/kaldi/wenetspeech_data_prep.sh
+
+
+def normalize_text(
+    utt: str,
+    punct_pattern=re.compile(r"<(PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    whitespace_pattern=re.compile(r"\s\s+"),
+) -> str:
+    return whitespace_pattern.sub(" ", punct_pattern.sub("", utt))
+
+
+def has_no_oov(
+    sup: SupervisionSegment,
+    oov_pattern=re.compile(r"<(SIL|MUSIC|NOISE|OTHER|SPOKEN_NOISE)>"),
+) -> bool:
+    return oov_pattern.search(sup.text) is None
+
+
+def preprocess_kespeech(speed_perturb: bool = False):
+    src_dir = Path("data/manifests/kespeech")
+    output_dir = Path("data/fbank/kespeech")
+    output_dir.mkdir(exist_ok=True)
+
+    # Note: By default, we preprocess all sub-parts.
+    # You can delete those that you don't need.
+    # For instance, if you don't want to use the test subpart, just remove
+    # the line below containing "test"
+    dataset_parts = (
+        "dev_phase1",
+        "dev_phase2",
+        "test",
+        "train_phase1",
+        "train_phase2",
+    )
+
+    logging.info("Loading manifest (may take 10 minutes)")
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix="jsonl.gz",
+        prefix="kespeech-asr",
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+    logging_threshold = 50
+    logging_count = 0
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"kespeech-asr_cuts_{partition}_raw.jsonl.gz"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        # Note this step makes the recipe different than LibriSpeech:
+        # We must filter out some utterances and remove punctuation
+        # to be consistent with Kaldi.
+        logging.info("Filtering OOV utterances from supervisions")
+        m["supervisions"] = m["supervisions"].filter(has_no_oov)
+        logging.info(f"Normalizing text in {partition}")
+        for sup in m["supervisions"]:
+            orig_text = sup.text
+            sup.text = normalize_text(sup.text)
+            if logging_count < logging_threshold and len(orig_text) != len(sup.text):
+                logging_count += 1
+                logging.info(
+                    f"\nOriginal text vs normalized text:\n{orig_text}\n{sup.text}"
+                )
+
+        # Create long-recording cut manifests.
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        )
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        if partition not in [
+            "dev_phase1",
+            "dev_phase2",
+            "test",
+        ]:
+            if speed_perturb:
+                logging.info(
+                    f"Speed perturb for {partition} with factors 0.9 and 1.1 "
+                    "(Perturbing may take 8 minutes and saving may take 20 minutes)"
+                )
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--speed-perturb",
+        type=bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+def main():
+    setup_logger(log_filename="./log-preprocess-kespeech")
+
+    args = get_args()
+    preprocess_kespeech(speed_perturb=args.speed_perturb)
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/local/text2token.py b/egs/multi_zh-hans/ASR/local/text2token.py
new file mode 120000
index 000000000..ce5cfd537
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/text2token.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/text2token.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/local/train_bpe_model.py b/egs/multi_zh-hans/ASR/local/train_bpe_model.py
new file mode 100755
index 000000000..976ea0ba8
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/train_bpe_model.py
@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# Copyright    2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    parser.add_argument(
+        "--byte-fallback",
+        type=bool,
+        default=True,
+        help="Enable byte fallback for BPE model.",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 0.98
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+            byte_fallback=args.byte_fallback,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/local/validate_bpe_lexicon.py b/egs/multi_zh-hans/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/prepare.sh b/egs/multi_zh-hans/ASR/prepare.sh
new file mode 100755
index 000000000..c09b9c1de
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/prepare.sh
@@ -0,0 +1,373 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+num_splits=100
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+vocab_sizes=(
+  2000
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+log "Dataset: musan"
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Soft link fbank of musan"
+  mkdir -p data/fbank
+  if [ -e ../../librispeech/ASR/data/fbank/.musan.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_feats) .
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_cuts.jsonl.gz) .
+    cd ../..
+  else
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 4 --stop-stage 4"
+    exit 1
+  fi
+fi
+
+log "Dataset: THCHS-30"
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare THCHS-30"
+  if [ ! -d $dl_dir/thchs30 ]; then
+    log "Downloading THCHS-30"
+    lhotse download thchs-30 $dl_dir/thchs30
+  fi
+
+  if [ ! -f data/manifests/.thchs30.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare thchs-30 $dl_dir/thchs30 data/manifests/thchs30
+    touch data/manifests/.thchs30.done
+  fi
+
+  if [ ! -f data/fbank/.thchs30.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_thchs30.py
+    touch data/fbank/.thchs30.done
+  fi
+fi
+
+log "Dataset: AISHELL-1"
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare AISHELL-1"
+  if [ -e ../../aishell/ASR/data/fbank/.aishell.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_feats_train) .
+    ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_feats_dev) .
+    ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_feats_test) .
+    ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_cuts_train.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_cuts_dev.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_cuts_test.jsonl.gz) .
+    cd ../..
+  else
+    log "Abort! Please run ../../aishell/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+fi
+
+log "Dataset: AISHELL-2"
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare AISHELL-2"
+  if [ -e ../../aishell/ASR/data/fbank/.aishell2.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_train) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_dev) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_test) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_train.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_dev.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_test.jsonl.gz) .
+    cd ../..
+  else 
+    log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi 
+fi
+
+log "Dataset: AISHELL-4"
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare AISHELL-4"
+  if [ -e ../../aishell/ASR/data/fbank/.aishell4.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_dev) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_test) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_L.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_M.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_S.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_dev.jsonl.gz) .
+    ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_test.jsonl.gz) .
+    cd ../..
+  else 
+    log "Abort! Please run ../../aishell4/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi 
+fi
+
+log "Dataset: ST-CMDS"
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare ST-CMDS"
+  if [ ! -f $dl_dir/stcmds/ST-CMDS-20170001_1-OS.tar.gz ]; then
+    log "Downloading ST-CMDS"
+    lhotse download stcmds $dl_dir/stcmds
+  fi
+
+  if [ ! -f data/manifests/.stcmds.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare stcmds $dl_dir/stcmds data/manifests/stcmds
+    touch data/manifests/.stcmds.done
+  fi
+
+  if [ ! -f data/fbank/.stcmds.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_stcmds.py
+    touch data/fbank/.stcmds.done
+  fi
+fi
+
+
+log "Dataset: Primewords"
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare Primewords"
+  if [ ! -f $dl_dir/primewords/primewords_md_2018_set1.tar.gz ]; then
+    log "Downloading Primewords"
+    lhotse download primewords $dl_dir/primewords
+  fi
+
+  if [ ! -f data/manifests/.stcmds.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare stcmds $dl_dir/primewords data/manifests/primewords
+    touch data/manifests/.primewords.done
+  fi
+
+  if [ ! -f data/fbank/.primewords.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_primewords.py
+    touch data/fbank/.primewords.done
+  fi
+fi
+
+log "Dataset: MagicData"
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare MagicData"
+  if [ ! -f $dl_dir/magicdata/train_set.tar.gz ]; then
+    log "Downloading MagicData"
+    lhotse download magicdata $dl_dir/magicdata
+  fi
+
+  if [ ! -f data/manifests/.magicdata.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare magicdata $dl_dir/magicdata data/manifests/magicdata
+    touch data/manifests/.magicdata.done
+  fi
+
+    if [ ! -f data/fbank/.magicdata.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_magicdata.py
+    touch data/fbank/.magicdata.done
+  fi
+fi
+
+log "Dataset: aidatatang_200zh"
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Prepare aidatatang_200zh"
+  if [ -e ../../aidatatang_200zh/ASR/data/fbank/.aidatatang_200zh.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_feats_train) .
+    ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_feats_dev) .
+    ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_feats_test) .
+    ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_cuts_train.jsonl.gz) .
+    ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_cuts_dev.jsonl.gz) .
+    ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_cuts_test.jsonl.gz) .
+    cd ../..
+  else
+    log "Abort! Please run ../../aidatatang_200zh/ASR/prepare.sh --stage 4 --stop-stage 4"
+    exit 1
+  fi
+fi
+
+log "Dataset: Ali-Meeting"
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Prepare Ali-Meeting"
+  if [ -e ../../alimeeting/ASR/data/fbank/.fbank.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_feats_train) .
+    ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_feats_eval) .
+    ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_feats_test) .
+    ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_cuts_train.jsonl.gz) .
+    ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_cuts_eval.jsonl.gz) .
+    ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_cuts_test.jsonl.gz) .
+    cd ../..
+  else
+    log "Abort! Please run ../../alimeeting/ASR/prepare.sh --stage 5 --stop-stage 5"
+    exit 1
+  fi
+fi
+
+log "Dataset: WenetSpeech"
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Prepare WenetSpeech"
+  if [ -e ../../wenetspeech/ASR/data/fbank/.preprocess_complete ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_DEV.jsonl.gz) .
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_L.jsonl.gz) .
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_TEST_MEETING.jsonl.gz) .
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_TEST_NET.jsonl.gz) .
+
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/L_split_1000) .
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/*.lca) .
+    ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/) ./wenetspeech
+    cd ../..
+  else
+    log "Abort! Please run ../../wenetspeech/ASR/prepare.sh"
+    exit 1
+  fi
+
+  if [ -d ../../wenetspeech/ASR/data/lang_char/ ]; then
+    cd data
+    cp -r ../../../../wenetspeech/ASR/data/lang_char .
+    cd ..
+  else
+    log "Abort! Please run ../../wenetspeech/ASR/prepare.sh"
+    exit 1
+  fi
+fi
+
+log "Dataset: KeSpeech"
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Prepare KeSpeech"
+  if [ ! -d $dl_dir/KeSpeech ]; then
+    log "Abort! Please download KeSpeech first."
+    log "KeSpeech download link: https://github.com/KeSpeech/KeSpeech"
+    exit 1
+  fi
+
+  if [ ! -f data/manifests/.kespeech.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare kespeech -j 16 $dl_dir/KeSpeech data/manifests/kespeech 
+    touch data/manifests/.kespeech.done
+  fi
+
+  if [ ! -f data/fbank/.kespeech.done ]; then
+    mkdir -p data/fbank
+
+    log "Preprocess KeSpeech manifest"
+    if [ ! -f data/fbank/.kespeech_preprocess_complete ]; then
+      python3 ./local/preprocess_kespeech.py
+      touch data/fbank/.kespeech_preprocess_complete
+    fi  
+    
+    if [ -f data/fbank/.kespeech.train_phase1.split.${num_splits}.done ]; then
+      log "Spliting KeSpeech train_phase1"
+      lhotse split ${num_splits} \
+        data/fbank/kespeech/kespeech-asr_cuts_train_phase1_raw.jsonl.gz \
+        data/fbank/kespeech/train_phase1_split_${num_splits}
+      touch data/fbank/.kespeech.train_phase1.split.${num_splits}.done
+    fi
+    
+    if [ -f data/fbank/.kespeech.train_phase2.split.${num_splits}.done ]; then
+      log "Spliting KeSpeech train_phase2"
+      lhotse split ${num_splits} \
+        data/fbank/kespeech/kespeech-asr_cuts_train_phase2_raw.jsonl.gz \
+        data/fbank/kespeech/train_phase2_split_${num_splits}
+      touch data/fbank/.kespeech.train_phase2.split.${num_splits}.done
+    fi
+    
+    log "Compute KeSpeech fbank for train_phase1"
+    ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase1
+
+    log "Compute KeSpeech fbank for train_phase2"
+    ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase2
+
+    log "Compute KeSpeech fbank for test/dev"
+    ./local/compute_fbank_kespeech_dev_test.py
+
+    touch data/fbank/.kespeech.done
+  fi
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: BPE model training (note that we use transcripts of wenetspeech only for BPE training)"
+  ./local/prepare_for_bpe_model.py --lang-dir ./data/lang_char --text ./data/lang_char/text
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    
+    mkdir -p $lang_dir
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --transcript ./data/lang_char/transcript_chars.txt \
+        --vocab-size $vocab_size
+
+      ./local/bpe_model_to_tokens.py $lang_dir/bpe.model > $lang_dir/tokens.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      cp data/lang_char/words.txt $lang_dir
+
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+    
+    if [ ! -f $lang_dir/L.fst ]; then
+      log "Converting L.pt to L.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L.pt \
+        $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+      log "Converting L_disambig.pt to L_disambig.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L_disambig.pt \
+        $lang_dir/L_disambig.fst
+    fi
+  done
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+  log "Stage 14: Prepare G (note that we use ngram lm of wenetspeech only for G preparation)"
+  
+  if [ -d ../../wenetspeech/ASR/data/lang_char/ ]; then
+    cd data
+    ln -s ../../../../wenetspeech/ASR/data/lm .
+    cd ..
+  else
+    log "Abort! Please run ../../wenetspeech/ASR/prepare.sh"
+    exit 1
+  fi
+fi
+
+if [ $stage -le 15 ] && [ $stop_stage -ge 15 ]; then
+  log "Stage 15: Compile LG"
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    python ./local/compile_lg.py --lang-dir $lang_dir
+  done
+fi
+
+
diff --git a/egs/multi_zh-hans/ASR/shared b/egs/multi_zh-hans/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py
new file mode 100644
index 000000000..02cfa1346
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1,388 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=300.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
diff --git a/egs/multi_zh-hans/ASR/zipformer/beam_search.py b/egs/multi_zh-hans/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..8e2c0a65c
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/beam_search.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/ctc_decode.py b/egs/multi_zh-hans/ASR/zipformer/ctc_decode.py
new file mode 100755
index 000000000..5143f945d
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/ctc_decode.py
@@ -0,0 +1,626 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) ctc-decoding
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from lhotse.cut import Cut
+from multi_dataset import MultiDataset
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_2000/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_2000",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    assert bpe_model is not None
+    decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = list(ref_text.replace(" ", ""))
+                hyp_words = list("".join(hyp_words))
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in ("ctc-decoding",)
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    HLG = None
+    H = k2.ctc_topo(
+        max_token=max_token_id,
+        modified=True,
+        device=device,
+    )
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+
+    G = None
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    data_module = AsrDataModule(args)
+    multi_dataset = MultiDataset(args.manifest_dir)
+
+    test_sets_cuts = multi_dataset.test_cuts()
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}"
+            )
+        return T > 0
+
+    test_sets = test_sets_cuts.keys()
+    test_dl = [
+        data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt))
+        for cuts_name in test_sets
+    ]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        logging.info(f"Start decoding test set: {test_set}")
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/zipformer/decode.py b/egs/multi_zh-hans/ASR/zipformer/decode.py
new file mode 100755
index 000000000..f501c3c30
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/decode.py
@@ -0,0 +1,828 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from multi_dataset import MultiDataset
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_2000/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_2000",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text).replace(" ", "")) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                hyp_text = "".join(hyp_words)
+                this_batch.append((cut_id, ref_text, hyp_text))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    data_module = AsrDataModule(args)
+    multi_dataset = MultiDataset(args.manifest_dir)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}"
+            )
+        return T > 0
+
+    test_sets_cuts = multi_dataset.test_cuts()
+
+    test_sets = test_sets_cuts.keys()
+    test_dl = [
+        data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt))
+        for cuts_name in test_sets
+    ]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        logging.info(f"Start decoding test set: {test_set}")
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/zipformer/decoder.py b/egs/multi_zh-hans/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/encoder_interface.py b/egs/multi_zh-hans/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..c2eaca671
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/encoder_interface.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming-ctc.py b/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming-ctc.py
new file mode 120000
index 000000000..652346001
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming-ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming-ctc.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming.py b/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming.py
new file mode 120000
index 000000000..2962eb784
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/export-onnx.py b/egs/multi_zh-hans/ASR/zipformer/export-onnx.py
new file mode 120000
index 000000000..70a15683c
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/export.py b/egs/multi_zh-hans/ASR/zipformer/export.py
new file mode 100755
index 000000000..723288191
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/export.py
@@ -0,0 +1,541 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+Note: This is a example for librispeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+(1) Export to torchscript model using torch.jit.script()
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1 \
+  --jit 1
+
+It will generate a file `jit_script.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("jit_script.pt")`.
+
+Check ./jit_pretrained.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1 \
+  --jit 1
+
+It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`.
+You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`.
+
+Check ./jit_pretrained_streaming.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+- For non-streaming model:
+
+To use the generated file with `zipformer/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_2000/bpe.model
+
+- For streaming model:
+
+To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+
+    # simulated streaming decoding
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_2000/bpe.model
+
+    # chunk-wise streaming decoding
+    ./zipformer/streaming_decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_2000/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+- non-streaming model:
+https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/
+    # You will find the pre-trained models in exp dir
+"""
+
+import argparse
+import logging
+import re
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor, nn
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, str2bool
+
+
+def num_tokens(
+    token_table: k2.SymbolTable, disambig_pattern: str = re.compile(r"^#\d+$")
+) -> int:
+    """Return the number of tokens excluding those from
+    disambiguation symbols.
+
+    Caution:
+      0 is not a token ID so it is excluded from the return value.
+    """
+    symbols = token_table.symbols
+    ans = []
+    for s in symbols:
+        if not disambig_pattern.match(s):
+            ans.append(token_table[s])
+    num_tokens = len(ans)
+    if 0 in ans:
+        num_tokens -= 1
+    return num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_2000/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class EncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+        """
+        x, x_lens = self.encoder_embed(features, feature_lengths)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+class StreamingEncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        assert len(encoder.chunk_size) == 1, encoder.chunk_size
+        assert len(encoder.left_context_frames) == 1, encoder.left_context_frames
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+
+        # The encoder_embed subsample features (T - 7) // 2
+        # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+        self.pad_length = 7 + 2 * 3
+
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor, states: List[Tensor]
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """Streaming forward for encoder_embed and encoder.
+
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+            states: a list of Tensors
+
+        Returns encoder outputs, output lengths, and updated states.
+        """
+        chunk_size = self.chunk_size
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        encoder_states = states[:-2]
+
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+        return encoder_out, encoder_out_lens, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+
+        # Wrap encoder and encoder_embed as a module
+        if params.causal:
+            model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed)
+            chunk_size = model.encoder.chunk_size
+            left_context_len = model.encoder.left_context_len
+            filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt"
+        else:
+            model.encoder = EncoderModel(model.encoder, model.encoder_embed)
+            filename = "jit_script.pt"
+
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        model.save(str(params.exp_dir / filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/multi_zh-hans/ASR/zipformer/generate_averaged_model.py b/egs/multi_zh-hans/ASR/zipformer/generate_averaged_model.py
new file mode 100755
index 000000000..68111fad7
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/generate_averaged_model.py
@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) use the checkpoint exp_dir/epoch-xxx.pt
+./zipformer/generate_averaged_model.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp
+
+It will generate a file `epoch-28-avg-15.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15.pt")`.
+
+(2) use the checkpoint exp_dir/checkpoint-iter.pt
+./zipformer/generate_averaged_model.py \
+    --iter 22000 \
+    --avg 5 \
+    --exp-dir ./zipformer/exp
+
+It will generate a file `iter-22000-avg-5.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5.pt")`.
+"""
+
+
+import argparse
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import average_checkpoints_with_averaged_model, find_checkpoints
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    print("Script started")
+
+    device = torch.device("cpu")
+    print(f"Device: {device}")
+
+    symbol_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = symbol_table["<blk>"]
+    params.unk_id = symbol_table["<unk>"]
+    params.vocab_size = len(symbol_table)
+
+    print("About to create model")
+    model = get_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg + 1
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg + 1:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        filename_start = filenames[-1]
+        filename_end = filenames[0]
+        print(
+            "Calculating the averaged model over iteration checkpoints"
+            f" from {filename_start} (excluded) to {filename_end}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+        filename = params.exp_dir / f"iter-{params.iter}-avg-{params.avg}.pt"
+        torch.save({"model": model.state_dict()}, filename)
+    else:
+        assert params.avg > 0, params.avg
+        start = params.epoch - params.avg
+        assert start >= 1, start
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        print(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+        filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+        torch.save({"model": model.state_dict()}, filename)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    print("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/zipformer/jit_pretrained.py b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_ctc.py b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_ctc.py
new file mode 120000
index 000000000..9a8da5844
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_streaming.py b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_streaming.py
new file mode 120000
index 000000000..1962351e9
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/joiner.py b/egs/multi_zh-hans/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/model.py b/egs/multi_zh-hans/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/multi_dataset.py b/egs/multi_zh-hans/ASR/zipformer/multi_dataset.py
new file mode 100644
index 000000000..b1920e62e
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/multi_dataset.py
@@ -0,0 +1,316 @@
+# Copyright      2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import glob
+import logging
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import lhotse
+from lhotse import CutSet, load_manifest_lazy
+
+
+class MultiDataset:
+    def __init__(self, fbank_dir: str):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files:
+            - aidatatang_cuts_train.jsonl.gz
+            - aishell_cuts_train.jsonl.gz
+            - aishell2_cuts_train.jsonl.gz
+            - aishell4_cuts_train_L.jsonl.gz
+            - aishell4_cuts_train_M.jsonl.gz
+            - aishell4_cuts_train_S.jsonl.gz
+            - alimeeting-far_cuts_train.jsonl.gz
+            - magicdata_cuts_train.jsonl.gz
+            - primewords_cuts_train.jsonl.gz
+            - stcmds_cuts_train.jsonl.gz
+            - thchs_30_cuts_train.jsonl.gz
+            - kespeech/kespeech-asr_cuts_train_phase1.jsonl.gz
+            - kespeech/kespeech-asr_cuts_train_phase2.jsonl.gz
+            - wenetspeech/cuts_L.jsonl.gz
+        """
+        self.fbank_dir = Path(fbank_dir)
+
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get multidataset train cuts")
+
+        # THCHS-30
+        logging.info("Loading THCHS-30 in lazy mode")
+        thchs_30_cuts = load_manifest_lazy(
+            self.fbank_dir / "thchs_30_cuts_train.jsonl.gz"
+        )
+
+        # AISHELL-1
+        logging.info("Loading Aishell-1 in lazy mode")
+        aishell_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_train.jsonl.gz"
+        )
+
+        # AISHELL-2
+        logging.info("Loading Aishell-2 in lazy mode")
+        aishell_2_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_train.jsonl.gz"
+        )
+
+        # AISHELL-4
+        logging.info("Loading Aishell-4 in lazy mode")
+        aishell_4_L_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell4_cuts_train_L.jsonl.gz"
+        )
+        aishell_4_M_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell4_cuts_train_M.jsonl.gz"
+        )
+        aishell_4_S_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell4_cuts_train_S.jsonl.gz"
+        )
+
+        # ST-CMDS
+        logging.info("Loading ST-CMDS in lazy mode")
+        stcmds_cuts = load_manifest_lazy(self.fbank_dir / "stcmds_cuts_train.jsonl.gz")
+
+        # Primewords
+        logging.info("Loading Primewords in lazy mode")
+        primewords_cuts = load_manifest_lazy(
+            self.fbank_dir / "primewords_cuts_train.jsonl.gz"
+        )
+
+        # MagicData
+        logging.info("Loading MagicData in lazy mode")
+        magicdata_cuts = load_manifest_lazy(
+            self.fbank_dir / "magicdata_cuts_train.jsonl.gz"
+        )
+
+        # Aidatatang_200zh
+        logging.info("Loading Aidatatang_200zh in lazy mode")
+        aidatatang_200zh_cuts = load_manifest_lazy(
+            self.fbank_dir / "aidatatang_cuts_train.jsonl.gz"
+        )
+
+        # Ali-Meeting
+        logging.info("Loading Ali-Meeting in lazy mode")
+        alimeeting_cuts = load_manifest_lazy(
+            self.fbank_dir / "alimeeting-far_cuts_train.jsonl.gz"
+        )
+
+        # WeNetSpeech
+        logging.info("Loading WeNetSpeech in lazy mode")
+        wenetspeech_L_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_L.jsonl.gz"
+        )
+
+        # KeSpeech
+        logging.info("Loading KeSpeech in lazy mode")
+        kespeech_1_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase1.jsonl.gz"
+        )
+        kespeech_2_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase2.jsonl.gz"
+        )
+
+        return CutSet.mux(
+            thchs_30_cuts,
+            aishell_cuts,
+            aishell_2_cuts,
+            aishell_4_L_cuts,
+            aishell_4_M_cuts,
+            aishell_4_S_cuts,
+            stcmds_cuts,
+            primewords_cuts,
+            magicdata_cuts,
+            aidatatang_200zh_cuts,
+            alimeeting_cuts,
+            wenetspeech_L_cuts,
+            kespeech_1_cuts,
+            kespeech_2_cuts,
+            weights=[
+                len(thchs_30_cuts),
+                len(aishell_cuts),
+                len(aishell_2_cuts),
+                len(aishell_4_L_cuts),
+                len(aishell_4_M_cuts),
+                len(aishell_4_S_cuts),
+                len(stcmds_cuts),
+                len(primewords_cuts),
+                len(magicdata_cuts),
+                len(aidatatang_200zh_cuts),
+                len(alimeeting_cuts),
+                len(wenetspeech_L_cuts),
+                len(kespeech_1_cuts),
+                len(kespeech_2_cuts),
+            ],
+        )
+
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get multidataset dev cuts")
+
+        # Aidatatang_200zh
+        logging.info("Loading Aidatatang_200zh DEV set in lazy mode")
+        aidatatang_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aidatatang_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL
+        logging.info("Loading Aishell DEV set in lazy mode")
+        aishell_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL-2
+        logging.info("Loading Aishell-2 DEV set in lazy mode")
+        aishell2_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_dev.jsonl.gz"
+        )
+
+        # Ali-Meeting
+        logging.info("Loading Ali-Meeting DEV set in lazy mode")
+        alimeeting_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz"
+        )
+
+        # MagicData
+        logging.info("Loading MagicData DEV set in lazy mode")
+        magicdata_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "magicdata_cuts_dev.jsonl.gz"
+        )
+
+        # KeSpeech
+        logging.info("Loading KeSpeech DEV set in lazy mode")
+        kespeech_dev_phase1_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz"
+        )
+        kespeech_dev_phase2_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz"
+        )
+
+        # WeNetSpeech
+        logging.info("Loading WeNetSpeech DEV set in lazy mode")
+        wenetspeech_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_DEV.jsonl.gz"
+        )
+
+        return wenetspeech_dev_cuts
+        # return [
+        #         aidatatang_dev_cuts,
+        #         aishell_dev_cuts,
+        #         aishell2_dev_cuts,
+        #         alimeeting_dev_cuts,
+        #         magicdata_dev_cuts,
+        #         kespeech_dev_phase1_cuts,
+        #         kespeech_dev_phase2_cuts,
+        #         wenetspeech_dev_cuts,
+        #     ]
+
+    def test_cuts(self) -> Dict[str, CutSet]:
+        logging.info("About to get multidataset test cuts")
+
+        # Aidatatang_200zh
+        logging.info("Loading Aidatatang_200zh set in lazy mode")
+        aidatatang_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aidatatang_cuts_test.jsonl.gz"
+        )
+        aidatatang_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aidatatang_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL
+        logging.info("Loading Aishell set in lazy mode")
+        aishell_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_test.jsonl.gz"
+        )
+        aishell_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL-2
+        logging.info("Loading Aishell-2 set in lazy mode")
+        aishell2_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_test.jsonl.gz"
+        )
+        aishell2_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL-4
+        logging.info("Loading Aishell-4 TEST set in lazy mode")
+        aishell4_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell4_cuts_test.jsonl.gz"
+        )
+
+        # Ali-Meeting
+        logging.info("Loading Ali-Meeting set in lazy mode")
+        alimeeting_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "alimeeting-far_cuts_test.jsonl.gz"
+        )
+        alimeeting_eval_cuts = load_manifest_lazy(
+            self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz"
+        )
+
+        # MagicData
+        logging.info("Loading MagicData set in lazy mode")
+        magicdata_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "magicdata_cuts_test.jsonl.gz"
+        )
+        magicdata_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "magicdata_cuts_dev.jsonl.gz"
+        )
+
+        # KeSpeech
+        logging.info("Loading KeSpeech set in lazy mode")
+        kespeech_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_test.jsonl.gz"
+        )
+        kespeech_dev_phase1_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz"
+        )
+        kespeech_dev_phase2_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz"
+        )
+
+        # WeNetSpeech
+        logging.info("Loading WeNetSpeech set in lazy mode")
+        wenetspeech_test_meeting_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_TEST_MEETING.jsonl.gz"
+        )
+        wenetspeech_test_net_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_TEST_NET.jsonl.gz"
+        )
+        wenetspeech_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_DEV.jsonl.gz"
+        )
+
+        return {
+            "aidatatang_test": aidatatang_test_cuts,
+            "aidatatang_dev": aidatatang_dev_cuts,
+            "alimeeting_test": alimeeting_test_cuts,
+            "alimeeting_eval": alimeeting_eval_cuts,
+            "aishell_test": aishell_test_cuts,
+            "aishell_dev": aishell_dev_cuts,
+            "aishell-2_test": aishell2_test_cuts,
+            "aishell-2_dev": aishell2_dev_cuts,
+            "aishell-4": aishell4_test_cuts,
+            "magicdata_test": magicdata_test_cuts,
+            "magicdata_dev": magicdata_dev_cuts,
+            "kespeech-asr_test": kespeech_test_cuts,
+            "kespeech-asr_dev_phase1": kespeech_dev_phase1_cuts,
+            "kespeech-asr_dev_phase2": kespeech_dev_phase2_cuts,
+            "wenetspeech-meeting_test": wenetspeech_test_meeting_cuts,
+            "wenetspeech-net_test": wenetspeech_test_net_cuts,
+            "wenetspeech_dev": wenetspeech_dev_cuts,
+        }
diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_check.py b/egs/multi_zh-hans/ASR/zipformer/onnx_check.py
new file mode 120000
index 000000000..f3dd42004
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_check.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_decode.py b/egs/multi_zh-hans/ASR/zipformer/onnx_decode.py
new file mode 120000
index 000000000..0573b88c5
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/onnx_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_decode.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming-ctc.py b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming-ctc.py
new file mode 120000
index 000000000..d623a8462
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming-ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming-ctc.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming.py b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 120000
index 000000000..cfea104c2
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained.py b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/optim.py b/egs/multi_zh-hans/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/pretrained.py b/egs/multi_zh-hans/ASR/zipformer/pretrained.py
new file mode 100755
index 000000000..69ff382da
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/pretrained.py
@@ -0,0 +1,381 @@
+#!/usr/bin/env python3
+# Copyright      2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+Note: This is a example for librispeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --epoch 23 \
+  --avg 1
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --epoch 23 \
+  --avg 1
+
+Usage of this script:
+
+- For non-streaming model:
+
+(1) greedy search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_2000/tokens.txt \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) modified beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens ./data/lang_bpe_2000/tokens.txt \
+  --method modified_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) fast beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens ./data/lang_bpe_2000/tokens.txt \
+  --method fast_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+- For streaming model:
+
+(1) greedy search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bpe_2000/tokens.txt \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) modified beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bpe_2000/tokens.txt \
+  --method modified_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) fast beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bpe_2000/tokens.txt \
+  --method fast_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+You can also use `./zipformer/exp/epoch-xx.pt`.
+
+Note: ./zipformer/exp/pretrained.pt is generated by ./zipformer/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    fast_beam_search_one_best,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from export import num_tokens
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall.utils import make_pad_mask
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+
+    params.blank_id = token_table["<blk>"]
+    params.unk_id = token_table["<unk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+
+    logging.info("Creating model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    # model forward
+    encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths)
+
+    hyps = []
+    msg = f"Using {params.method}"
+    logging.info(msg)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in hyp_tokens:
+            hyps.append(token_ids_to_words(hyp))
+    else:
+        raise ValueError(f"Unsupported method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/multi_zh-hans/ASR/zipformer/scaling.py b/egs/multi_zh-hans/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/scaling_converter.py b/egs/multi_zh-hans/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/streaming_beam_search.py b/egs/multi_zh-hans/ASR/zipformer/streaming_beam_search.py
new file mode 120000
index 000000000..b1ed54557
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/streaming_decode.py b/egs/multi_zh-hans/ASR/zipformer/streaming_decode.py
new file mode 120000
index 000000000..13fd02a78
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/streaming_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_decode.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/subsampling.py b/egs/multi_zh-hans/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/multi_zh-hans/ASR/zipformer/train.py b/egs/multi_zh-hans/ASR/zipformer/train.py
new file mode 100755
index 000000000..c1bbd2ee8
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/train.py
@@ -0,0 +1,1385 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from multi_dataset import MultiDataset
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_2000/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}"
+    )
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        loss = 0.0
+
+        if params.use_transducer:
+            s = params.simple_loss_scale
+            # take down the scale on the simple loss from 1.0 at the start
+            # to params.simple_loss scale by warm_step.
+            simple_loss_scale = (
+                s
+                if batch_idx_train >= warm_step
+                else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+            )
+            pruned_loss_scale = (
+                1.0
+                if batch_idx_train >= warm_step
+                else 0.1 + 0.9 * (batch_idx_train / warm_step)
+            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    if params.use_transducer:
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    data_module = AsrDataModule(args)
+    multi_dataset = MultiDataset(args.manifest_dir)
+
+    train_cuts = multi_dataset.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            # logging.warning(
+            #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = data_module.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = multi_dataset.dev_cuts()
+    valid_dl = data_module.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh-hans/ASR/zipformer/zipformer.py b/egs/multi_zh-hans/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/multi_zh-hans/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/README.md b/egs/multi_zh_en/ASR/README.md
new file mode 100644
index 000000000..29341571d
--- /dev/null
+++ b/egs/multi_zh_en/ASR/README.md
@@ -0,0 +1,19 @@
+# Introduction
+
+This recipe includes scripts for training Zipformer model using both English and Chinese datasets.
+
+# Included Training Sets
+
+1. LibriSpeech (English)
+2. AiShell-2 (Chinese)
+3. TAL-CSASR (Code-Switching, Chinese and English)
+   
+|Datset| Number of hours| URL|
+|---|---:|---|
+|**TOTAL**|2,547|---|
+|LibriSpeech|960|https://www.openslr.org/12/|
+|AiShell-2|1,000|http://www.aishelltech.com/aishell_2|
+|TAL-CSASR|587|https://ai.100tal.com/openData/voice|
+
+
+
diff --git a/egs/multi_zh_en/ASR/RESULTS.md b/egs/multi_zh_en/ASR/RESULTS.md
new file mode 100644
index 000000000..3562d6ac3
--- /dev/null
+++ b/egs/multi_zh_en/ASR/RESULTS.md
@@ -0,0 +1,44 @@
+## Results
+
+### Zh-En datasets bpe-based training results (Non-streaming) on Zipformer model
+
+This is the [pull request #1238](https://github.com/k2-fsa/icefall/pull/1265) in icefall.
+
+#### Non-streaming (Byte-Level BPE vocab_size=2000)
+
+Best results (num of params : ~69M):
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 35 \
+  --use-fp16 1 \
+  --max-duration 1000 \
+  --num-workers 8
+```
+
+The decoding command:
+
+```
+for method in greedy_search modified_beam_search fast_beam_search; do
+    ./zipformer/decode.py \
+    --epoch 34 \
+    --avg 19 \
+    --decoding-method $method
+done
+```
+
+Word Error Rates (WERs) listed below are produced by the checkpoint of the 20th epoch using greedy search and BPE model (# tokens is 2000).
+
+|       Datasets       | TAL-CSASR | TAL-CSASR | AiShell-2 | AiShell-2 | LibriSpeech | LibriSpeech | 
+|----------------------|-----------|-----------|-----------|-----------|-------------|-------------|
+|   Zipformer WER (%)  |    dev    |   test    |    dev    |   test    |  test-clean |  test-other | 
+|     greedy_search    |   6.65    |   6.69    |    6.57   |   7.03    |    2.43     |    5.70     |
+| modified_beam_search |   6.46    |   6.51    |    6.18   |   6.60    |    2.41     |    5.57     |
+|   fast_beam_search   |   6.57    |   6.68    |    6.40   |   6.74    |    2.40     |    5.56     |
+
+Pre-trained model can be found here : https://huggingface.co/zrjin/icefall-asr-zipformer-multi-zh-en-2023-11-22, which is trained on LibriSpeech 960-hour training set (with speed perturbation), TAL-CSASR training set (with speed perturbation) and AiShell-2 (w/o speed perturbation).
+
+
diff --git a/egs/multi_zh_en/ASR/local/compile_lg.py b/egs/multi_zh_en/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/prepare_char.py b/egs/multi_zh_en/ASR/local/prepare_char.py
new file mode 120000
index 000000000..42743b544
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/prepare_char.py
@@ -0,0 +1 @@
+../../../aishell/ASR/local/prepare_char.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/prepare_for_bpe_model.py b/egs/multi_zh_en/ASR/local/prepare_for_bpe_model.py
new file mode 100755
index 000000000..00514e6bb
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/prepare_for_bpe_model.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script tokenizes the training transcript by CJK characters
+# and saves the result to transcript_chars.txt, which is used
+# to train the BPE model later.
+
+import argparse
+from pathlib import Path
+
+from tqdm.auto import tqdm
+
+from icefall.utils import tokenize_by_CJK_char
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Output directory.
+        The generated transcript_chars.txt is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="Training transcript.",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    text = Path(args.text)
+
+    assert lang_dir.exists() and text.exists(), f"{lang_dir} or {text} does not exist!"
+
+    transcript_path = lang_dir / "transcript_chars.txt"
+
+    with open(text, "r", encoding="utf-8") as fin:
+        with open(transcript_path, "w+", encoding="utf-8") as fout:
+            for line in tqdm(fin):
+                fout.write(tokenize_by_CJK_char(line) + "\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh_en/ASR/local/prepare_lang.py b/egs/multi_zh_en/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/prepare_lang_bbpe.py b/egs/multi_zh_en/ASR/local/prepare_lang_bbpe.py
new file mode 120000
index 000000000..9a0b44642
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/prepare_lang_bbpe.py
@@ -0,0 +1 @@
+../../../aishell/ASR/local/prepare_lang_bbpe.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/prepare_lang_bpe.py b/egs/multi_zh_en/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/prepare_words.py b/egs/multi_zh_en/ASR/local/prepare_words.py
new file mode 120000
index 000000000..ef2b4eaf3
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/prepare_words.py
@@ -0,0 +1 @@
+../../../aishell2/ASR/local/prepare_words.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/text2segments.py b/egs/multi_zh_en/ASR/local/text2segments.py
new file mode 120000
index 000000000..7d68a39c3
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/text2segments.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/text2segments.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/text2token.py b/egs/multi_zh_en/ASR/local/text2token.py
new file mode 120000
index 000000000..ce5cfd537
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/text2token.py
@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/text2token.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/train_bbpe_model.py b/egs/multi_zh_en/ASR/local/train_bbpe_model.py
new file mode 120000
index 000000000..7fb4a9f9d
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/train_bbpe_model.py
@@ -0,0 +1 @@
+../../../aishell/ASR/local/train_bbpe_model.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/local/validate_bpe_lexicon.py b/egs/multi_zh_en/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/multi_zh_en/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/prepare.sh b/egs/multi_zh_en/ASR/prepare.sh
new file mode 100755
index 000000000..9f2be5a5c
--- /dev/null
+++ b/egs/multi_zh_en/ASR/prepare.sh
@@ -0,0 +1,149 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+vocab_sizes=(
+  2000
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+log "Dataset: musan"
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Soft link fbank of musan"
+  mkdir -p data/fbank
+  if [ -e ../../librispeech/ASR/data/fbank/.musan.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_feats) .
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_cuts.jsonl.gz) .
+    cd ../..
+  else
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 4 --stop-stage 4"
+    exit 1
+  fi
+fi
+
+log "Dataset: LibriSpeech"
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Soft link fbank of LibriSpeech"
+  mkdir -p data/fbank
+  if [ -e ../../librispeech/ASR/data/fbank/.librispeech.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/librispeech_cuts*) .
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/librispeech_feats*) .
+    cd ../..
+  else
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+fi
+
+log "Dataset: AiShell-2"
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Soft link fbank of AiShell-2"
+  mkdir -p data/fbank
+  if [ -e ../../aishell2/ASR/data/fbank/.aishell2.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts*) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats*) .
+    cd ../..
+  else
+    log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare Byte BPE based lang"
+  mkdir -p data/fbank
+  if [ ! -d ../../aishell2/ASR/data/lang_char ] && [ ! -d ./data/lang_char ]; then
+    log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+
+  if [ ! -d ../../librispeech/ASR/data/lang_bpe_500 ] && [ ! -d ./data/lang_bpe_500 ]; then
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 6 --stop-stage 6"
+    exit 1
+  fi
+
+  cd data/
+  if [ ! -d ./lang_char ]; then
+    ln -svf $(realpath ../../../aishell2/ASR/data/lang_char) .
+  fi
+  if [ ! -d ./lang_bpe_500 ]; then
+    ln -svf $(realpath ../../../librispeech/ASR/data/lang_bpe_500) .
+  fi
+  cd ../
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bbpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    cat data/lang_char/text data/lang_bpe_500/transcript_words.txt \
+      > $lang_dir/text
+
+    if [ ! -f $lang_dir/transcript_chars.txt ]; then
+      ./local/prepare_for_bpe_model.py \
+        --lang-dir ./$lang_dir \
+        --text $lang_dir/text
+    fi
+    
+    if [ ! -f $lang_dir/text_words_segmentation ]; then
+      python3 ./local/text2segments.py \
+        --input-file ./data/lang_char/text \
+        --output-file $lang_dir/text_words_segmentation
+      
+      cat ./data/lang_bpe_500/transcript_words.txt \
+        >> $lang_dir/text_words_segmentation
+
+      cat ./data/lang_char/text \
+        >> $lang_dir/text
+    fi
+
+    cat $lang_dir/text_words_segmentation | sed 's/ /\n/g' \
+      | sort -u | sed '/^$/d' | uniq > $lang_dir/words_no_ids.txt
+
+    if [ ! -f $lang_dir/words.txt ]; then
+      python3 ./local/prepare_words.py \
+        --input-file $lang_dir/words_no_ids.txt \
+        --output-file $lang_dir/words.txt
+    fi
+
+    if [ ! -f $lang_dir/bbpe.model ]; then
+      ./local/train_bbpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/text
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bbpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bbpe.model
+    fi
+  done 
+fi
+
diff --git a/egs/multi_zh_en/ASR/shared b/egs/multi_zh_en/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/multi_zh_en/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py b/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py
new file mode 100644
index 000000000..be6e94472
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1,385 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=300.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
diff --git a/egs/multi_zh_en/ASR/zipformer/beam_search.py b/egs/multi_zh_en/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..8e2c0a65c
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/beam_search.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/decode.py b/egs/multi_zh_en/ASR/zipformer/decode.py
new file mode 100755
index 000000000..e21e8f052
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/decode.py
@@ -0,0 +1,851 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from multi_dataset import MultiDataset
+from train import add_model_arguments, get_model, get_params
+
+from icefall import byte_encode, smart_byte_decode, tokenize_by_CJK_char
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_2000/bbpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bbpe_2000",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-tal-csasr",
+        type=str2bool,
+        default=False,
+        help="Whether to use TAL-CSASR training data.",
+    )
+
+    parser.add_argument(
+        "--use-librispeech",
+        type=str2bool,
+        default=False,
+        help="Whether to use LibriSpeech training data.",
+    )
+
+    parser.add_argument(
+        "--use-aishell2",
+        type=str2bool,
+        default=False,
+        help="Whether to use Aishell-2 training data.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(
+                byte_encode(tokenize_by_CJK_char(supervisions["text"]))
+            ),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(smart_byte_decode(sp.decode(hyp)).split())
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [tokenize_by_CJK_char(str(text)).split() for text in texts]
+        # print(texts)
+        # exit()
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    data_module = AsrDataModule(args)
+    multi_dataset = MultiDataset(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}"
+            )
+        return T > 0
+
+    test_sets_cuts = multi_dataset.test_cuts()
+
+    test_sets = test_sets_cuts.keys()
+    test_dl = [
+        data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt))
+        for cuts_name in test_sets
+    ]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        logging.info(f"Start decoding test set: {test_set}")
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh_en/ASR/zipformer/decoder.py b/egs/multi_zh_en/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/encoder_interface.py b/egs/multi_zh_en/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..c2eaca671
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/encoder_interface.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/export-onnx-streaming.py b/egs/multi_zh_en/ASR/zipformer/export-onnx-streaming.py
new file mode 120000
index 000000000..2962eb784
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/export-onnx.py b/egs/multi_zh_en/ASR/zipformer/export-onnx.py
new file mode 120000
index 000000000..70a15683c
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/export.py b/egs/multi_zh_en/ASR/zipformer/export.py
new file mode 100755
index 000000000..fbd9ce0dd
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/export.py
@@ -0,0 +1,541 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+Note: This is a example for librispeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+(1) Export to torchscript model using torch.jit.script()
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1 \
+  --jit 1
+
+It will generate a file `jit_script.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("jit_script.pt")`.
+
+Check ./jit_pretrained.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1 \
+  --jit 1
+
+It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`.
+You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`.
+
+Check ./jit_pretrained_streaming.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --epoch 20 \
+  --avg 1
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+- For non-streaming model:
+
+To use the generated file with `zipformer/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bbpe_2000/bpe.model
+
+- For streaming model:
+
+To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+
+    # simulated streaming decoding
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bbpe_2000/bpe.model
+
+    # chunk-wise streaming decoding
+    ./zipformer/streaming_decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bbpe_2000/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+- non-streaming model:
+https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/
+    # You will find the pre-trained models in exp dir
+"""
+
+import argparse
+import logging
+import re
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor, nn
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, str2bool
+
+
+def num_tokens(
+    token_table: k2.SymbolTable, disambig_pattern: str = re.compile(r"^#\d+$")
+) -> int:
+    """Return the number of tokens excluding those from
+    disambiguation symbols.
+
+    Caution:
+      0 is not a token ID so it is excluded from the return value.
+    """
+    symbols = token_table.symbols
+    ans = []
+    for s in symbols:
+        if not disambig_pattern.match(s):
+            ans.append(token_table[s])
+    num_tokens = len(ans)
+    if 0 in ans:
+        num_tokens -= 1
+    return num_tokens
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bbpe_2000/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class EncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+        """
+        x, x_lens = self.encoder_embed(features, feature_lengths)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+class StreamingEncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        assert len(encoder.chunk_size) == 1, encoder.chunk_size
+        assert len(encoder.left_context_frames) == 1, encoder.left_context_frames
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+
+        # The encoder_embed subsample features (T - 7) // 2
+        # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+        self.pad_length = 7 + 2 * 3
+
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor, states: List[Tensor]
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """Streaming forward for encoder_embed and encoder.
+
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+            states: a list of Tensors
+
+        Returns encoder outputs, output lengths, and updated states.
+        """
+        chunk_size = self.chunk_size
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        encoder_states = states[:-2]
+
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+        return encoder_out, encoder_out_lens, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+
+        # Wrap encoder and encoder_embed as a module
+        if params.causal:
+            model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed)
+            chunk_size = model.encoder.chunk_size
+            left_context_len = model.encoder.left_context_len
+            filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt"
+        else:
+            model.encoder = EncoderModel(model.encoder, model.encoder_embed)
+            filename = "jit_script.pt"
+
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        model.save(str(params.exp_dir / filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/multi_zh_en/ASR/zipformer/generate_averaged_model.py b/egs/multi_zh_en/ASR/zipformer/generate_averaged_model.py
new file mode 100755
index 000000000..68111fad7
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/generate_averaged_model.py
@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) use the checkpoint exp_dir/epoch-xxx.pt
+./zipformer/generate_averaged_model.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp
+
+It will generate a file `epoch-28-avg-15.pt` in the given `exp_dir`.
+You can later load it by `torch.load("epoch-28-avg-15.pt")`.
+
+(2) use the checkpoint exp_dir/checkpoint-iter.pt
+./zipformer/generate_averaged_model.py \
+    --iter 22000 \
+    --avg 5 \
+    --exp-dir ./zipformer/exp
+
+It will generate a file `iter-22000-avg-5.pt` in the given `exp_dir`.
+You can later load it by `torch.load("iter-22000-avg-5.pt")`.
+"""
+
+
+import argparse
+from pathlib import Path
+
+import k2
+import torch
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import average_checkpoints_with_averaged_model, find_checkpoints
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    print("Script started")
+
+    device = torch.device("cpu")
+    print(f"Device: {device}")
+
+    symbol_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = symbol_table["<blk>"]
+    params.unk_id = symbol_table["<unk>"]
+    params.vocab_size = len(symbol_table)
+
+    print("About to create model")
+    model = get_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg + 1
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg + 1:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        filename_start = filenames[-1]
+        filename_end = filenames[0]
+        print(
+            "Calculating the averaged model over iteration checkpoints"
+            f" from {filename_start} (excluded) to {filename_end}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+        filename = params.exp_dir / f"iter-{params.iter}-avg-{params.avg}.pt"
+        torch.save({"model": model.state_dict()}, filename)
+    else:
+        assert params.avg > 0, params.avg
+        start = params.epoch - params.avg
+        assert start >= 1, start
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        print(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+        filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt"
+        torch.save({"model": model.state_dict()}, filename)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    print("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh_en/ASR/zipformer/jit_pretrained.py b/egs/multi_zh_en/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/jit_pretrained_ctc.py b/egs/multi_zh_en/ASR/zipformer/jit_pretrained_ctc.py
new file mode 120000
index 000000000..9a8da5844
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/jit_pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/jit_pretrained_streaming.py b/egs/multi_zh_en/ASR/zipformer/jit_pretrained_streaming.py
new file mode 120000
index 000000000..1962351e9
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/joiner.py b/egs/multi_zh_en/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/model.py b/egs/multi_zh_en/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/multi_dataset.py b/egs/multi_zh_en/ASR/zipformer/multi_dataset.py
new file mode 100644
index 000000000..1155a3dcc
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/multi_dataset.py
@@ -0,0 +1,247 @@
+# Copyright      2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Dict
+
+from lhotse import CutSet, load_manifest_lazy
+
+
+class MultiDataset:
+    def __init__(self, args: argparse.Namespace):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files:
+            - aishell2_cuts_train.jsonl.gz
+        """
+        self.fbank_dir = Path(args.manifest_dir)
+        self.use_tal_csasr = args.use_tal_csasr
+        self.use_librispeech = args.use_librispeech
+        self.use_aishell2 = args.use_aishell2
+
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get multidataset train cuts")
+
+        # AISHELL-2
+        if self.use_aishell2:
+            logging.info("Loading Aishell-2 in lazy mode")
+            aishell_2_cuts = load_manifest_lazy(
+                self.fbank_dir / "aishell2_cuts_train.jsonl.gz"
+            )
+
+        # TAL-CSASR
+        if self.use_tal_csasr:
+            logging.info("Loading TAL-CSASR in lazy mode")
+            tal_csasr_cuts = load_manifest_lazy(
+                self.fbank_dir / "tal_csasr_cuts_train_set.jsonl.gz"
+            )
+
+        # LibriSpeech
+        if self.use_librispeech:
+            logging.info("Loading LibriSpeech in lazy mode")
+            train_clean_100_cuts = self.train_clean_100_cuts()
+            train_clean_360_cuts = self.train_clean_360_cuts()
+            train_other_500_cuts = self.train_other_500_cuts()
+
+        if self.use_tal_csasr and self.use_librispeech and self.use_aishell2:
+            return CutSet.mux(
+                aishell_2_cuts,
+                train_clean_100_cuts,
+                train_clean_360_cuts,
+                train_other_500_cuts,
+                tal_csasr_cuts,
+                weights=[
+                    len(aishell_2_cuts),
+                    len(train_clean_100_cuts),
+                    len(train_clean_360_cuts),
+                    len(train_other_500_cuts),
+                    len(tal_csasr_cuts),
+                ],
+            )
+        elif not self.use_tal_csasr and self.use_librispeech and self.use_aishell2:
+            return CutSet.mux(
+                aishell_2_cuts,
+                train_clean_100_cuts,
+                train_clean_360_cuts,
+                train_other_500_cuts,
+                weights=[
+                    len(aishell_2_cuts),
+                    len(train_clean_100_cuts),
+                    len(train_clean_360_cuts),
+                    len(train_other_500_cuts),
+                ],
+            )
+        elif self.use_tal_csasr and not self.use_librispeech and self.use_aishell2:
+            return CutSet.mux(
+                aishell_2_cuts,
+                tal_csasr_cuts,
+                weights=[
+                    len(aishell_2_cuts),
+                    len(tal_csasr_cuts),
+                ],
+            )
+        elif self.use_tal_csasr and self.use_librispeech and not self.use_aishell2:
+            return CutSet.mux(
+                train_clean_100_cuts,
+                train_clean_360_cuts,
+                train_other_500_cuts,
+                tal_csasr_cuts,
+                weights=[
+                    len(train_clean_100_cuts),
+                    len(train_clean_360_cuts),
+                    len(train_other_500_cuts),
+                    len(tal_csasr_cuts),
+                ],
+            )
+        else:
+            raise NotImplementedError(
+                f"""Not implemented for 
+                use_aishell2: {self.use_aishell2}
+                use_librispeech: {self.use_librispeech}
+                use_tal_csasr: {self.use_tal_csasr}"""
+            )
+
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get multidataset dev cuts")
+
+        # AISHELL-2
+        logging.info("Loading Aishell-2 DEV set in lazy mode")
+        aishell2_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_dev.jsonl.gz"
+        )
+
+        # LibriSpeech
+        dev_clean_cuts = self.dev_clean_cuts()
+        dev_other_cuts = self.dev_other_cuts()
+
+        logging.info("Loading TAL-CSASR set in lazy mode")
+        tal_csasr_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "tal_csasr_cuts_dev_set.jsonl.gz"
+        )
+
+        return CutSet.mux(
+            aishell2_dev_cuts,
+            dev_clean_cuts,
+            dev_other_cuts,
+            tal_csasr_dev_cuts,
+            weights=[
+                len(aishell2_dev_cuts),
+                len(dev_clean_cuts),
+                len(dev_other_cuts),
+                len(tal_csasr_dev_cuts),
+            ],
+        )
+
+    def test_cuts(self) -> Dict[str, CutSet]:
+        logging.info("About to get multidataset test cuts")
+
+        # AISHELL-2
+        if self.use_aishell2:
+            logging.info("Loading Aishell-2 set in lazy mode")
+            aishell2_test_cuts = load_manifest_lazy(
+                self.fbank_dir / "aishell2_cuts_test.jsonl.gz"
+            )
+            aishell2_dev_cuts = load_manifest_lazy(
+                self.fbank_dir / "aishell2_cuts_dev.jsonl.gz"
+            )
+
+        # LibriSpeech
+        if self.use_librispeech:
+            test_clean_cuts = self.test_clean_cuts()
+            test_other_cuts = self.test_other_cuts()
+
+        logging.info("Loading TAL-CSASR set in lazy mode")
+        tal_csasr_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "tal_csasr_cuts_test_set.jsonl.gz"
+        )
+        tal_csasr_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "tal_csasr_cuts_dev_set.jsonl.gz"
+        )
+
+        test_cuts = {
+            "tal_csasr_test": tal_csasr_test_cuts,
+            "tal_csasr_dev": tal_csasr_dev_cuts,
+        }
+
+        if self.use_aishell2:
+            test_cuts.update(
+                {
+                    "aishell-2_test": aishell2_test_cuts,
+                    "aishell-2_dev": aishell2_dev_cuts,
+                }
+            )
+        if self.use_librispeech:
+            test_cuts.update(
+                {
+                    "librispeech_test_clean": test_clean_cuts,
+                    "librispeech_test_other": test_other_cuts,
+                }
+            )
+        return test_cuts
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_train-clean-100.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_360_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-360 cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_train-clean-360.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_other_500_cuts(self) -> CutSet:
+        logging.info("About to get train-other-500 cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_train-other-500.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.fbank_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
diff --git a/egs/multi_zh_en/ASR/zipformer/onnx_check.py b/egs/multi_zh_en/ASR/zipformer/onnx_check.py
new file mode 120000
index 000000000..f3dd42004
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_check.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/onnx_decode.py b/egs/multi_zh_en/ASR/zipformer/onnx_decode.py
new file mode 120000
index 000000000..0573b88c5
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/onnx_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_decode.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/onnx_pretrained-streaming.py b/egs/multi_zh_en/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 120000
index 000000000..cfea104c2
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/onnx_pretrained.py b/egs/multi_zh_en/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/optim.py b/egs/multi_zh_en/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/pretrained.py b/egs/multi_zh_en/ASR/zipformer/pretrained.py
new file mode 100755
index 000000000..676272e1f
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/pretrained.py
@@ -0,0 +1,378 @@
+#!/usr/bin/env python3
+# Copyright      2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+Note: This is a example for librispeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --epoch 23 \
+  --avg 1
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --epoch 23 \
+  --avg 1
+
+Usage of this script:
+
+- For non-streaming model:
+
+(1) greedy search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bbpe_2000/tokens.txt \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) modified beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens ./data/lang_bbpe_2000/tokens.txt \
+  --method modified_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) fast beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens ./data/lang_bbpe_2000/tokens.txt \
+  --method fast_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+- For streaming model:
+
+(1) greedy search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bbpe_2000/tokens.txt \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) modified beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bbpe_2000/tokens.txt \
+  --method modified_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) fast beam search
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens ./data/lang_bbpe_2000/tokens.txt \
+  --method fast_beam_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+
+You can also use `./zipformer/exp/epoch-xx.pt`.
+
+Note: ./zipformer/exp/pretrained.pt is generated by ./zipformer/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    fast_beam_search_one_best,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from export import num_tokens
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall import smart_byte_decode
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to byte-level bpe model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+
+    logging.info("Creating model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    # model forward
+    encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths)
+
+    hyps = []
+    msg = f"Using {params.method}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    else:
+        raise ValueError(f"Unsupported method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/multi_zh_en/ASR/zipformer/scaling.py b/egs/multi_zh_en/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/scaling_converter.py b/egs/multi_zh_en/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/streaming_beam_search.py b/egs/multi_zh_en/ASR/zipformer/streaming_beam_search.py
new file mode 120000
index 000000000..b1ed54557
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/streaming_decode.py b/egs/multi_zh_en/ASR/zipformer/streaming_decode.py
new file mode 120000
index 000000000..13fd02a78
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/streaming_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_decode.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/subsampling.py b/egs/multi_zh_en/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/multi_zh_en/ASR/zipformer/train.py b/egs/multi_zh_en/ASR/zipformer/train.py
new file mode 100755
index 000000000..310c8fe59
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/train.py
@@ -0,0 +1,1416 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from multi_dataset import MultiDataset
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import byte_encode, diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+    tokenize_by_CJK_char,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_2000/bbpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--use-tal-csasr",
+        type=str2bool,
+        default=False,
+        help="Whether to use TAL-CSASR training data.",
+    )
+
+    parser.add_argument(
+        "--use-librispeech",
+        type=str2bool,
+        default=False,
+        help="Whether to use LibriSpeech training data.",
+    )
+
+    parser.add_argument(
+        "--use-aishell2",
+        type=str2bool,
+        default=False,
+        help="Whether to use Aishell-2 training data.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}"
+    )
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        loss = 0.0
+
+        if params.use_transducer:
+            s = params.simple_loss_scale
+            # take down the scale on the simple loss from 1.0 at the start
+            # to params.simple_loss scale by warm_step.
+            simple_loss_scale = (
+                s
+                if batch_idx_train >= warm_step
+                else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+            )
+            pruned_loss_scale = (
+                1.0
+                if batch_idx_train >= warm_step
+                else 0.1 + 0.9 * (batch_idx_train / warm_step)
+            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    if params.use_transducer:
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    data_module = AsrDataModule(args)
+    multi_dataset = MultiDataset(args)
+
+    train_cuts = multi_dataset.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    def tokenize_and_encode_text(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = byte_encode(tokenize_by_CJK_char(text))
+        c.supervisions[0].text = text
+        return c
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_cuts = train_cuts.map(tokenize_and_encode_text)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = data_module.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = multi_dataset.dev_cuts()
+    valid_dl = data_module.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/multi_zh_en/ASR/zipformer/zipformer.py b/egs/multi_zh_en/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/multi_zh_en/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/must_c/ST/local/compute_fbank_musan.py b/egs/must_c/ST/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/must_c/ST/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/must_c/ST/local/compute_fbank_must_c.py b/egs/must_c/ST/local/compute_fbank_must_c.py
new file mode 100755
index 000000000..84de099d1
--- /dev/null
+++ b/egs/must_c/ST/local/compute_fbank_must_c.py
@@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file computes fbank features of the MuST-C dataset.
+It looks for manifests in the directory "in_dir" and write
+generated features to "out_dir".
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    FeatureSet,
+    LilcomChunkyWriter,
+    load_manifest,
+)
+
+from icefall.utils import str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--in-dir",
+        type=Path,
+        required=True,
+        help="Input manifest directory",
+    )
+
+    parser.add_argument(
+        "--out-dir",
+        type=Path,
+        required=True,
+        help="Output directory where generated fbank features are saved.",
+    )
+
+    parser.add_argument(
+        "--tgt-lang",
+        type=str,
+        required=True,
+        help="Target language, e.g., zh, de, fr.",
+    )
+
+    parser.add_argument(
+        "--num-jobs",
+        type=int,
+        default=1,
+        help="Number of jobs for computing features",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="""True to enable speed perturb with factors 0.9 and 1.1 on
+        the train subset. False (by default) to disable speed perturb.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_must_c(
+    in_dir: Path,
+    out_dir: Path,
+    tgt_lang: str,
+    num_jobs: int,
+    perturb_speed: bool,
+):
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    extractor = Fbank(FbankConfig(num_mel_bins=80))
+
+    parts = ["dev", "tst-COMMON", "tst-HE", "train"]
+
+    prefix = "must_c"
+    suffix = "jsonl.gz"
+    for p in parts:
+        logging.info(f"Processing {p}")
+
+        cuts_path = f"{out_dir}/{prefix}_feats_en-{tgt_lang}_{p}"
+        if perturb_speed and p == "train":
+            cuts_path += "_sp"
+
+        cuts_path += ".jsonl.gz"
+
+        if Path(cuts_path).is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        recordings_filename = in_dir / f"{prefix}_recordings_en-{tgt_lang}_{p}.jsonl.gz"
+        supervisions_filename = (
+            in_dir / f"{prefix}_supervisions_en-{tgt_lang}_{p}_norm_rm.jsonl.gz"
+        )
+        assert recordings_filename.is_file(), recordings_filename
+        assert supervisions_filename.is_file(), supervisions_filename
+        cut_set = CutSet.from_manifests(
+            recordings=load_manifest(recordings_filename),
+            supervisions=load_manifest(supervisions_filename),
+        )
+        if perturb_speed and p == "train":
+            logging.info("Speed perturbing for the train dataset")
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+            storage_path = f"{out_dir}/{prefix}_feats_en-{tgt_lang}_{p}_sp"
+        else:
+            storage_path = f"{out_dir}/{prefix}_feats_en-{tgt_lang}_{p}"
+
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=storage_path,
+            num_jobs=num_jobs,
+            storage_type=LilcomChunkyWriter,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    args = get_args()
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    logging.info(vars(args))
+    assert args.in_dir.is_dir(), args.in_dir
+
+    compute_fbank_must_c(
+        in_dir=args.in_dir,
+        out_dir=args.out_dir,
+        tgt_lang=args.tgt_lang,
+        num_jobs=args.num_jobs,
+        perturb_speed=args.perturb_speed,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/must_c/ST/local/get_text.py b/egs/must_c/ST/local/get_text.py
new file mode 100755
index 000000000..558ab6de8
--- /dev/null
+++ b/egs/must_c/ST/local/get_text.py
@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+This file prints the text field of supervisions from cutset to the console
+"""
+
+import argparse
+
+from lhotse import load_manifest_lazy
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Input manifest",
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    assert args.manifest.is_file(), args.manifest
+
+    cutset = load_manifest_lazy(args.manifest)
+    for c in cutset:
+        for sup in c.supervisions:
+            print(sup.text)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/must_c/ST/local/get_words.py b/egs/must_c/ST/local/get_words.py
new file mode 100755
index 000000000..a61f60860
--- /dev/null
+++ b/egs/must_c/ST/local/get_words.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+This file generates words.txt from the given transcript file.
+"""
+
+import argparse
+
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "transcript",
+        type=Path,
+        help="Input transcript file",
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    assert args.transcript.is_file(), args.transcript
+
+    word_set = set()
+    with open(args.transcript) as f:
+        for line in f:
+            words = line.strip().split()
+            for w in words:
+                word_set.add(w)
+
+    # Note: reserved* should be kept in sync with ./local/prepare_lang_bpe.py
+    reserved1 = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>"]
+    reserved2 = ["#0", "<s>", "</s>"]
+
+    for w in reserved1 + reserved2:
+        assert w not in word_set, w
+
+    words = sorted(list(word_set))
+    words = reserved1 + words + reserved2
+
+    for i, w in enumerate(words):
+        print(w, i)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/must_c/ST/local/normalize_punctuation.py b/egs/must_c/ST/local/normalize_punctuation.py
new file mode 100644
index 000000000..efd47e091
--- /dev/null
+++ b/egs/must_c/ST/local/normalize_punctuation.py
@@ -0,0 +1,169 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+import re
+
+
+def normalize_punctuation(s: str, lang: str) -> str:
+    """
+    This function implements
+    https://github.com/moses-smt/mosesdecoder/blob/master/scripts/tokenizer/normalize-punctuation.perl
+
+    Args:
+      s:
+        A string to be normalized.
+      lang:
+        The language to which `s` belongs
+    Returns:
+      Return a normalized string.
+    """
+    #  s/\r//g;
+    s = re.sub("\r", "", s)
+
+    # remove extra spaces
+    # s/\(/ \(/g;
+    s = re.sub("\(", " (", s)  # add a space before (
+
+    # s/\)/\) /g; s/ +/ /g;
+    s = re.sub("\)", ") ", s)  # add a space after )
+    s = re.sub(" +", " ", s)  # convert multiple spaces to one
+
+    # s/\) ([\.\!\:\?\;\,])/\)$1/g;
+    s = re.sub("\) ([\.\!\:\?\;\,])", r")\1", s)
+
+    # s/\( /\(/g;
+    s = re.sub("\( ", "(", s)  # remove space after (
+
+    # s/ \)/\)/g;
+    s = re.sub(" \)", ")", s)  # remove space before )
+
+    # s/(\d) \%/$1\%/g;
+    s = re.sub("(\d) \%", r"\1%", s)  # remove space between a digit and %
+
+    # s/ :/:/g;
+    s = re.sub(" :", ":", s)  # remove space before :
+
+    # s/ ;/;/g;
+    s = re.sub(" ;", ";", s)  # remove space before ;
+
+    # normalize unicode punctuation
+    # s/\`/\'/g;
+    s = re.sub("`", "'", s)  # replace ` with '
+
+    # s/\'\'/ \" /g;
+    s = re.sub("''", '"', s)  #  replace '' with "
+
+    # s/„/\"/g;
+    s = re.sub("„", '"', s)  #  replace „ with "
+
+    # s/“/\"/g;
+    s = re.sub("“", '"', s)  #  replace “ with "
+
+    # s/”/\"/g;
+    s = re.sub("”", '"', s)  #  replace ” with "
+
+    # s/–/-/g;
+    s = re.sub("–", "-", s)  #  replace – with -
+
+    # s/—/ - /g; s/ +/ /g;
+    s = re.sub("—", " - ", s)
+    s = re.sub(" +", " ", s)  # convert multiple spaces to one
+
+    # s/´/\'/g;
+    s = re.sub("´", "'", s)
+
+    # s/([a-z])‘([a-z])/$1\'$2/gi;
+    s = re.sub("([a-z])‘([a-z])", r"\1'\2", s, flags=re.IGNORECASE)
+
+    # s/([a-z])’([a-z])/$1\'$2/gi;
+    s = re.sub("([a-z])’([a-z])", r"\1'\2", s, flags=re.IGNORECASE)
+
+    # s/‘/\'/g;
+    s = re.sub("‘", "'", s)
+
+    # s/‚/\'/g;
+    s = re.sub("‚", "'", s)
+
+    # s/’/\"/g;
+    s = re.sub("’", '"', s)
+
+    # s/''/\"/g;
+    s = re.sub("''", '"', s)
+
+    # s/´´/\"/g;
+    s = re.sub("´´", '"', s)
+
+    # s/…/.../g;
+    s = re.sub("…", "...", s)
+
+    # French quotes
+
+    # s/ « / \"/g;
+    s = re.sub(" « ", ' "', s)
+
+    # s/« /\"/g;
+    s = re.sub("« ", '"', s)
+
+    # s/«/\"/g;
+    s = re.sub("«", '"', s)
+
+    # s/ » /\" /g;
+    s = re.sub(" » ", '" ', s)
+
+    # s/ »/\"/g;
+    s = re.sub(" »", '"', s)
+
+    # s/»/\"/g;
+    s = re.sub("»", '"', s)
+
+    # handle pseudo-spaces
+
+    # s/ \%/\%/g;
+    s = re.sub(" %", r"%", s)
+
+    # s/nº /nº /g;
+    s = re.sub("nº ", "nº ", s)
+
+    # s/ :/:/g;
+    s = re.sub(" :", ":", s)
+
+    # s/ ºC/ ºC/g;
+    s = re.sub(" ºC", " ºC", s)
+
+    # s/ cm/ cm/g;
+    s = re.sub(" cm", " cm", s)
+
+    # s/ \?/\?/g;
+    s = re.sub(" \?", "\?", s)
+
+    # s/ \!/\!/g;
+    s = re.sub(" \!", "\!", s)
+
+    # s/ ;/;/g;
+    s = re.sub(" ;", ";", s)
+
+    # s/, /, /g; s/ +/ /g;
+    s = re.sub(", ", ", ", s)
+    s = re.sub(" +", " ", s)
+
+    if lang == "en":
+        # English "quotation," followed by comma, style
+        # s/\"([,\.]+)/$1\"/g;
+        s = re.sub('"([,\.]+)', r'\1"', s)
+    elif lang in ("cs", "cz"):
+        # Czech is confused
+        pass
+    else:
+        # German/Spanish/French "quotation", followed by comma, style
+        # s/,\"/\",/g;
+        s = re.sub(',"', '",', s)
+
+        # s/(\.+)\"(\s*[^<])/\"$1$2/g; # don't fix period at end of sentence
+        s = re.sub('(\.+)"(\s*[^<])', r'"\1\2', s)
+
+    if lang in ("de", "es", "cz", "cs", "fr"):
+        # s/(\d) (\d)/$1,$2/g;
+        s = re.sub("(\d) (\d)", r"\1,\2", s)
+    else:
+        # s/(\d) (\d)/$1.$2/g;
+        s = re.sub("(\d) (\d)", r"\1.\2", s)
+
+    return s
diff --git a/egs/must_c/ST/local/prepare_lang.py b/egs/must_c/ST/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/must_c/ST/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/must_c/ST/local/prepare_lang_bpe.py b/egs/must_c/ST/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/must_c/ST/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/must_c/ST/local/preprocess_must_c.py b/egs/must_c/ST/local/preprocess_must_c.py
new file mode 100755
index 000000000..1ba282bf4
--- /dev/null
+++ b/egs/must_c/ST/local/preprocess_must_c.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+This script normalizes transcripts from supervisions.
+
+Usage:
+  ./local/preprocess_must_c.py \
+    --manifest-dir ./data/manifests/v1.0/ \
+    --tgt-lang de
+"""
+
+import argparse
+import logging
+import re
+from functools import partial
+from pathlib import Path
+
+from lhotse.recipes.utils import read_manifests_if_cached
+from normalize_punctuation import normalize_punctuation
+from remove_non_native_characters import remove_non_native_characters
+from remove_punctuation import remove_punctuation
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--manifest-dir",
+        type=Path,
+        required=True,
+        help="Manifest directory",
+    )
+    parser.add_argument(
+        "--tgt-lang",
+        type=str,
+        required=True,
+        help="Target language, e.g., zh, de, fr.",
+    )
+    return parser.parse_args()
+
+
+def preprocess_must_c(manifest_dir: Path, tgt_lang: str):
+    normalize_punctuation_lang = partial(normalize_punctuation, lang=tgt_lang)
+    remove_non_native_characters_lang = partial(
+        remove_non_native_characters, lang=tgt_lang
+    )
+
+    prefix = "must_c"
+    suffix = "jsonl.gz"
+    parts = ["dev", "tst-COMMON", "tst-HE", "train"]
+    for p in parts:
+        logging.info(f"Processing {p}")
+        name = f"en-{tgt_lang}_{p}"
+
+        # norm: normalization
+        # rm: remove punctuation
+        dst_name = manifest_dir / f"must_c_supervisions_{name}_norm_rm.jsonl.gz"
+        if dst_name.is_file():
+            logging.info(f"{dst_name} exists - skipping")
+            continue
+
+        manifests = read_manifests_if_cached(
+            dataset_parts=name,
+            output_dir=manifest_dir,
+            prefix=prefix,
+            suffix=suffix,
+            types=("supervisions",),
+        )
+        if name not in manifests:
+            raise RuntimeError(f"Processing {p} failed.")
+
+        supervisions = manifests[name]["supervisions"]
+        supervisions = supervisions.transform_text(normalize_punctuation_lang)
+        supervisions = supervisions.transform_text(remove_punctuation)
+        supervisions = supervisions.transform_text(lambda x: x.lower())
+        supervisions = supervisions.transform_text(remove_non_native_characters_lang)
+        supervisions = supervisions.transform_text(lambda x: re.sub(" +", " ", x))
+
+        supervisions.to_file(dst_name)
+
+
+def main():
+    args = get_args()
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    logging.info(vars(args))
+    assert args.manifest_dir.is_dir(), args.manifest_dir
+
+    preprocess_must_c(
+        manifest_dir=args.manifest_dir,
+        tgt_lang=args.tgt_lang,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/must_c/ST/local/remove_non_native_characters.py b/egs/must_c/ST/local/remove_non_native_characters.py
new file mode 100755
index 000000000..f61fbd16b
--- /dev/null
+++ b/egs/must_c/ST/local/remove_non_native_characters.py
@@ -0,0 +1,21 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import re
+
+
+def remove_non_native_characters(s: str, lang: str):
+    if lang == "de":
+        # ä -> ae
+        # ö -> oe
+        # ü -> ue
+        # ß -> ss
+
+        s = re.sub("ä", "ae", s)
+        s = re.sub("ö", "oe", s)
+        s = re.sub("ü", "ue", s)
+        s = re.sub("ß", "ss", s)
+        # keep only a-z and spaces
+        # note: ' is removed
+        s = re.sub(r"[^a-z\s]", "", s)
+
+    return s
diff --git a/egs/must_c/ST/local/remove_punctuation.py b/egs/must_c/ST/local/remove_punctuation.py
new file mode 100644
index 000000000..723946ec3
--- /dev/null
+++ b/egs/must_c/ST/local/remove_punctuation.py
@@ -0,0 +1,41 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+import re
+import string
+
+
+def remove_punctuation(s: str) -> str:
+    """
+    It implements https://github.com/espnet/espnet/blob/master/utils/remove_punctuation.pl
+    """
+
+    # Remove punctuation except apostrophe
+    # s/<space>/spacemark/g;  # for scoring
+    s = re.sub("<space>", "spacemark", s)
+
+    # s/'/apostrophe/g;
+    s = re.sub("'", "apostrophe", s)
+
+    # s/[[:punct:]]//g;
+    s = s.translate(str.maketrans("", "", string.punctuation))
+    # string punctuation returns the following string
+    # !"#$%&'()*+,-./:;<=>?@[\]^_`{|}~
+    # See
+    # https://stackoverflow.com/questions/265960/best-way-to-strip-punctuation-from-a-string
+
+    # s/apostrophe/'/g;
+    s = re.sub("apostrophe", "'", s)
+
+    # s/spacemark/<space>/g;  # for scoring
+    s = re.sub("spacemark", "<space>", s)
+
+    # remove whitespace
+    # s/\s+/ /g;
+    s = re.sub("\s+", " ", s)
+
+    # s/^\s+//;
+    s = re.sub("^\s+", "", s)
+
+    # s/\s+$//;
+    s = re.sub("\s+$", "", s)
+
+    return s
diff --git a/egs/must_c/ST/local/test_normalize_punctuation.py b/egs/must_c/ST/local/test_normalize_punctuation.py
new file mode 100755
index 000000000..9079858c8
--- /dev/null
+++ b/egs/must_c/ST/local/test_normalize_punctuation.py
@@ -0,0 +1,197 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from normalize_punctuation import normalize_punctuation
+
+
+def test_normalize_punctuation():
+    #  s/\r//g;
+    s = "a\r\nb\r\n"
+    n = normalize_punctuation(s, lang="en")
+    assert "\r" not in n
+    assert len(s) - 2 == len(n), (len(s), len(n))
+
+    # s/\(/ \(/g;
+    s = "(ab (c"
+    n = normalize_punctuation(s, lang="en")
+    assert n == " (ab (c", n
+
+    # s/\)/\) /g;
+    s = "a)b c)"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a) b c) "
+
+    # s/ +/ /g;
+    s = "  a  b     c  d    "
+    n = normalize_punctuation(s, lang="en")
+    assert n == " a b c d "
+
+    # s/\) ([\.\!\:\?\;\,])/\)$1/g;
+    for i in ".!:?;,":
+        s = f"a)  {i}"
+        n = normalize_punctuation(s, lang="en")
+        assert n == f"a){i}"
+
+    # s/\( /\(/g;
+    s = "a(    b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a (b", n
+
+    # s/ \)/\)/g;
+    s = "ab    ) a"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "ab) a", n
+
+    # s/(\d) \%/$1\%/g;
+    s = "1   %a"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "1%a", n
+
+    # s/ :/:/g;
+    s = "a  :"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a:", n
+
+    # s/ ;/;/g;
+    s = "a  ;"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a;", n
+
+    # s/\`/\'/g;
+    s = "`a`"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "'a'", n
+
+    # s/\'\'/ \" /g;
+    s = "''a''"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"a"', n
+
+    # s/„/\"/g;
+    s = '„a"'
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"a"', n
+
+    # s/“/\"/g;
+    s = "“a„"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"a"', n
+
+    # s/”/\"/g;
+    s = "“a”"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"a"', n
+
+    # s/–/-/g;
+    s = "a–b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a-b", n
+
+    # s/—/ - /g; s/ +/ /g;
+    s = "a—b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a - b", n
+
+    # s/´/\'/g;
+    s = "a´b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "a'b", n
+
+    # s/([a-z])‘([a-z])/$1\'$2/gi;
+    for i in "‘’":
+        s = f"a{i}B"
+        n = normalize_punctuation(s, lang="en")
+        assert n == "a'B", n
+
+        s = f"A{i}B"
+        n = normalize_punctuation(s, lang="en")
+        assert n == "A'B", n
+
+        s = f"A{i}b"
+        n = normalize_punctuation(s, lang="en")
+        assert n == "A'b", n
+
+    # s/‘/\'/g;
+    # s/‚/\'/g;
+    for i in "‘‚":
+        s = f"a{i}b"
+        n = normalize_punctuation(s, lang="en")
+        assert n == "a'b", n
+
+    # s/’/\"/g;
+    s = "’"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"', n
+
+    # s/''/\"/g;
+    s = "''"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"', n
+
+    # s/´´/\"/g;
+    s = "´´"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"', n
+
+    # s/…/.../g;
+    s = "…"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "...", n
+
+    # s/ « / \"/g;
+    s = "a « b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == 'a "b', n
+
+    # s/« /\"/g;
+    s = "a « b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == 'a "b', n
+
+    # s/«/\"/g;
+    s = "a«b"
+    n = normalize_punctuation(s, lang="en")
+    assert n == 'a"b', n
+
+    # s/ » /\" /g;
+    s = " » "
+    n = normalize_punctuation(s, lang="en")
+    assert n == '" ', n
+
+    # s/ »/\"/g;
+    s = " »"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"', n
+
+    # s/»/\"/g;
+    s = "»"
+    n = normalize_punctuation(s, lang="en")
+    assert n == '"', n
+
+    # s/ \%/\%/g;
+    s = " %"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "%", n
+
+    # s/ :/:/g;
+    s = " :"
+    n = normalize_punctuation(s, lang="en")
+    assert n == ":", n
+
+    # s/(\d) (\d)/$1.$2/g;
+    s = "2 3"
+    n = normalize_punctuation(s, lang="en")
+    assert n == "2.3", n
+
+    # s/(\d) (\d)/$1,$2/g;
+    s = "2 3"
+    n = normalize_punctuation(s, lang="de")
+    assert n == "2,3", n
+
+
+def main():
+    test_normalize_punctuation()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/must_c/ST/local/test_remove_non_native_characters.py b/egs/must_c/ST/local/test_remove_non_native_characters.py
new file mode 100755
index 000000000..ecf8569cf
--- /dev/null
+++ b/egs/must_c/ST/local/test_remove_non_native_characters.py
@@ -0,0 +1,26 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from remove_non_native_characters import remove_non_native_characters
+
+
+def test_remove_non_native_characters():
+    s = "Ich heiße xxx好的01 fangjun".lower()
+    n = remove_non_native_characters(s, lang="de")
+    assert n == "ich heisse xxx fangjun", n
+
+    s = "äÄ".lower()
+    n = remove_non_native_characters(s, lang="de")
+    assert n == "aeae", n
+
+    s = "öÖ".lower()
+    n = remove_non_native_characters(s, lang="de")
+    assert n == "oeoe", n
+
+    s = "üÜ".lower()
+    n = remove_non_native_characters(s, lang="de")
+    assert n == "ueue", n
+
+
+if __name__ == "__main__":
+    test_remove_non_native_characters()
diff --git a/egs/must_c/ST/local/test_remove_punctuation.py b/egs/must_c/ST/local/test_remove_punctuation.py
new file mode 100755
index 000000000..a4f318550
--- /dev/null
+++ b/egs/must_c/ST/local/test_remove_punctuation.py
@@ -0,0 +1,17 @@
+#!/usr/bin/env python3
+
+from remove_punctuation import remove_punctuation
+
+
+def test_remove_punctuation():
+    s = "a,b'c!#"
+    n = remove_punctuation(s)
+    assert n == "ab'c", n
+
+    s = "  ab  "  # remove leading and trailing spaces
+    n = remove_punctuation(s)
+    assert n == "ab", n
+
+
+if __name__ == "__main__":
+    test_remove_punctuation()
diff --git a/egs/must_c/ST/local/train_bpe_model.py b/egs/must_c/ST/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/must_c/ST/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/must_c/ST/local/validate_bpe_lexicon.py b/egs/must_c/ST/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/must_c/ST/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/must_c/ST/prepare.sh b/egs/must_c/ST/prepare.sh
new file mode 100755
index 000000000..d16bb3d0b
--- /dev/null
+++ b/egs/must_c/ST/prepare.sh
@@ -0,0 +1,173 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=10
+stage=0
+stop_stage=100
+
+version=v1.0
+tgt_lang=de
+dl_dir=$PWD/download
+
+must_c_dir=$dl_dir/must-c/$version/en-$tgt_lang/data
+
+# We assume dl_dir (download dir) contains the following
+# directories and files.
+#  - $dl_dir/must-c/$version/en-$tgt_lang/data/{dev,train,tst-COMMON,tst-HE}
+#
+# Please go to https://ict.fbk.eu/must-c-releases/
+# to download and untar the dataset if you have not already done this.
+
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate
+#  data/lang_bpe_${tgt_lang}_xxx
+#  data/lang_bpe_${tgt_lang}_yyy
+# if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ ! -d $must_c_dir ]; then
+  log "$must_c_dir does not exist"
+  exit 1
+fi
+
+for d in dev train tst-COMMON tst-HE; do
+  if [ ! -d $must_c_dir/$d ]; then
+    log "$must_c_dir/$d does not exist!"
+    exit 1
+  fi
+done
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download musan"
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare must-c $version manifest for target language $tgt_lang"
+  mkdir -p data/manifests/$version
+  if [ ! -e data/manifests/$version/.${tgt_lang}.manifests.done ]; then
+    lhotse prepare must-c \
+      -j $nj \
+      --tgt-lang $tgt_lang \
+      $dl_dir/must-c/$version/ \
+      data/manifests/$version/
+
+    touch data/manifests/$version/.${tgt_lang}.manifests.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Text normalization for $version with target language $tgt_lang"
+  if [ ! -f ./data/manifests/$version/.$tgt_lang.norm.done ]; then
+    ./local/preprocess_must_c.py \
+      --manifest-dir ./data/manifests/$version/ \
+      --tgt-lang $tgt_lang
+    touch ./data/manifests/$version/.$tgt_lang.norm.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for $version with target language $tgt_lang"
+  mkdir -p data/fbank/$version/
+  if [ ! -e data/fbank/$version/.$tgt_lang.done ]; then
+    ./local/compute_fbank_must_c.py \
+      --in-dir ./data/manifests/$version/ \
+      --out-dir ./data/fbank/$version/ \
+      --tgt-lang $tgt_lang \
+      --num-jobs $nj
+
+    ./local/compute_fbank_must_c.py \
+      --in-dir ./data/manifests/$version/ \
+      --out-dir ./data/fbank/$version/ \
+      --tgt-lang $tgt_lang \
+      --num-jobs $nj \
+      --perturb-speed 1
+
+    touch data/fbank/$version/.$tgt_lang.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare BPE based lang for $version with target language $tgt_lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}/$version/$tgt_lang/
+    mkdir -p $lang_dir
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      ./local/get_text.py ./data/fbank/$version/must_c_feats_en-${tgt_lang}_train.jsonl.gz > $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/words.txt ]; then
+      ./local/get_words.py $lang_dir/transcript_words.txt > $lang_dir/words.txt
+    fi
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+  done
+fi
diff --git a/egs/must_c/ST/shared b/egs/must_c/ST/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/must_c/ST/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/peoples_speech/ASR/local/compute_fbank_musan.py b/egs/peoples_speech/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/peoples_speech/ASR/local/compute_fbank_peoples_speech_splits.py b/egs/peoples_speech/ASR/local/compute_fbank_peoples_speech_splits.py
new file mode 100755
index 000000000..c2ab3d07d
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/compute_fbank_peoples_speech_splits.py
@@ -0,0 +1,154 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.             (Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+    set_audio_duration_mismatch_tolerance,
+    set_caching_enabled,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the train subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_peoples_speech_splits(args):
+    subsets = ("dirty", "dirty_sa", "clean", "clean_sa")
+    num_splits = args.num_splits
+    output_dir = f"data/fbank/peoples_speech_train_split"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = 8
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    set_audio_duration_mismatch_tolerance(0.01)  # 10ms tolerance
+    set_caching_enabled(False)
+
+    for partition in subsets:
+        for i in range(start, stop):
+            idx = f"{i + 1}".zfill(num_digits)
+            logging.info(f"Processing {partition}: {idx}")
+
+            cuts_path = output_dir / f"peoples_speech_cuts_{partition}.{idx}.jsonl.gz"
+            if cuts_path.is_file():
+                logging.info(f"{cuts_path} exists - skipping")
+                continue
+
+            raw_cuts_path = (
+                output_dir / f"peoples_speech_cuts_{partition}_raw.{idx}.jsonl.gz"
+            )
+
+            logging.info(f"Loading {raw_cuts_path}")
+            cut_set = CutSet.from_file(raw_cuts_path)
+
+            logging.info("Splitting cuts into smaller chunks.")
+            cut_set = cut_set.trim_to_supervisions(
+                keep_overlapping=False, min_duration=None
+            )
+
+            logging.info("Computing features")
+            cut_set = cut_set.compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=f"{output_dir}/peoples_speech_feats_{partition}_{idx}",
+                num_workers=args.num_workers,
+                batch_duration=args.batch_duration,
+                storage_type=LilcomChunkyWriter,
+                overwrite=True,
+            )
+
+            logging.info(f"Saving to {cuts_path}")
+            cut_set.to_file(cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_peoples_speech_splits(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/peoples_speech/ASR/local/compute_fbank_peoples_speech_valid_test.py b/egs/peoples_speech/ASR/local/compute_fbank_peoples_speech_valid_test.py
new file mode 100755
index 000000000..89f43a674
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/compute_fbank_peoples_speech_valid_test.py
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the People's Speech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_peoples_speech_valid_test():
+    src_dir = Path(f"data/manifests")
+    output_dir = Path(f"data/fbank")
+    num_workers = 42
+    batch_duration = 600
+
+    subsets = ("validation", "test")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    for partition in subsets:
+        cuts_path = output_dir / f"peoples_speech_cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping.")
+            continue
+
+        raw_cuts_path = output_dir / f"peoples_speech_cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/peoples_speech_feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_peoples_speech_valid_test()
diff --git a/egs/peoples_speech/ASR/local/filter_cuts.py b/egs/peoples_speech/ASR/local/filter_cuts.py
new file mode 120000
index 000000000..27aca1729
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/filter_cuts.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/filter_cuts.py
\ No newline at end of file
diff --git a/egs/peoples_speech/ASR/local/prepare_lang_bpe.py b/egs/peoples_speech/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/peoples_speech/ASR/local/preprocess_peoples_speech.py b/egs/peoples_speech/ASR/local/preprocess_peoples_speech.py
new file mode 100755
index 000000000..c5417049f
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/preprocess_peoples_speech.py
@@ -0,0 +1,123 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import re
+from pathlib import Path
+from typing import Optional
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    return parser.parse_args()
+
+
+def normalize_text(utt: str) -> str:
+    utt = re.sub(r"[{0}]+".format("-"), " ", utt)
+    return re.sub(r"[^a-zA-Z\s]", "", utt).upper()
+
+
+def preprocess_peoples_speech(dataset: Optional[str] = None):
+    src_dir = Path(f"data/manifests")
+    output_dir = Path(f"data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    if dataset is None:
+        dataset_parts = (
+            "validation",
+            "test",
+            "dirty",
+            "dirty_sa",
+            "clean",
+            "clean_sa",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    logging.info("Loading manifest, it may takes 8 minutes")
+    prefix = f"peoples_speech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix=suffix,
+        prefix=prefix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"{prefix}_cuts_{partition}_raw.{suffix}"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        logging.info(f"Normalizing text in {partition}")
+        i = 0
+        for sup in m["supervisions"]:
+            text = str(sup.text)
+            orig_text = text
+            sup.text = normalize_text(sup.text)
+            text = str(sup.text)
+            if i < 10 and len(orig_text) != len(text):
+                logging.info(
+                    f"\nOriginal text vs normalized text:\n{orig_text}\n{text}"
+                )
+                i += 1
+
+        # Create long-recording cut manifests.
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        ).resample(16000)
+
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    preprocess_peoples_speech(dataset=args.dataset)
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/peoples_speech/ASR/local/train_bpe_model.py b/egs/peoples_speech/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/peoples_speech/ASR/local/validate_bpe_lexicon.py b/egs/peoples_speech/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/peoples_speech/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/peoples_speech/ASR/prepare.sh b/egs/peoples_speech/ASR/prepare.sh
new file mode 100755
index 000000000..3787858d9
--- /dev/null
+++ b/egs/peoples_speech/ASR/prepare.sh
@@ -0,0 +1,247 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=32
+stage=-1
+stop_stage=100
+
+# Split data/set to a number of pieces
+# This is to avoid OOM during feature extraction.
+num_per_split=4000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/peoples_speech
+#      This directory contains the following files downloaded from
+#       https://huggingface.co/datasets/MLCommons/peoples_speech
+# 
+#     - test
+#     - train
+#     - validation
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/peoples_speech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/peoples_speech $dl_dir/peoples_speech
+  #
+  if [ ! -d $dl_dir/peoples_speech/train ]; then
+    git lfs install
+    git clone https://huggingface.co/datasets/MLCommons/peoples_speech
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare People's Speech manifest"
+  # We assume that you have downloaded the People's Speech corpus
+  # to $dl_dir/peoples_speech
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.peoples_speech.done ]; then
+    lhotse prepare peoples-speech -j $nj $dl_dir/peoples_speech data/manifests
+    touch data/manifests/.peoples_speech.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Preprocess People's Speech manifest"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.preprocess_complete ]; then
+    ./local/preprocess_peoples_speech.py
+    touch data/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for valid and test subsets of People's Speech"
+  if [ ! -e data/fbank/.peoples_speech_valid_test.done ]; then
+    ./local/compute_fbank_peoples_speech_valid_test.py
+    touch data/fbank/.peoples_speech_valid_test.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Split train subset into pieces"
+  split_dir=data/fbank/peoples_speech_train_split
+  if [ ! -e $split_dir/.peoples_speech_dirty_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_dirty_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_dirty_split.done
+  fi
+
+  if [ ! -e $split_dir/.peoples_speech_dirty_sa_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_dirty_sa_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_dirty_sa_split.done
+  fi
+
+  if [ ! -e $split_dir/.peoples_speech_clean_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_clean_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_clean_split.done
+  fi
+
+  if [ ! -e $split_dir/.peoples_speech_clean_sa_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_clean_sa_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_clean_sa_split.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compute features for train subset of People's Speech"
+  if [ ! -e data/fbank/.peoples_speech_train.done ]; then
+    ./local/compute_fbank_peoples_speech_splits.py \
+      --num-workers $nj \
+      --batch-duration 600 \
+      --start 0 \
+      --num-splits 2000
+    touch data/fbank/.peoples_speech_train.done
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      file=$(
+        find "data/fbank/peoples_speech_cuts_dirty_raw.jsonl.gz"
+	find "data/fbank/peoples_speech_cuts_dirty_sa_raw.jsonl.gz"
+	find "data/fbank/peoples_speech_cuts_clean_raw.jsonl.gz"
+	find "data/fbank/peoples_speech_cuts_clean_sa_raw.jsonl.gz"
+      )
+      gunzip -c ${file} | awk -F '"' '{print $30}' > $lang_dir/transcript_words.txt
+
+      # Ensure space only appears once
+      sed -i 's/\t/ /g' $lang_dir/transcript_words.txt
+      sed -i 's/ +/ /g' $lang_dir/transcript_words.txt
+    fi
+ 
+    if [ ! -f $lang_dir/words.txt ]; then
+      cat $lang_dir/transcript_words.txt | sed 's/ /\n/g' \
+        | sort -u | sed '/^$/d' > $lang_dir/words.txt
+      (echo '!SIL'; echo '<SPOKEN_NOISE>'; echo '<UNK>'; ) |
+        cat - $lang_dir/words.txt | sort | uniq | awk '
+        BEGIN {
+          print "<eps> 0";
+        }
+        {
+          if ($1 == "<s>") {
+            print "<s> is in the vocabulary!" | "cat 1>&2"
+            exit 1;
+          }
+          if ($1 == "</s>") {
+            print "</s> is in the vocabulary!" | "cat 1>&2"
+            exit 1;
+          }
+          printf("%s %d\n", $1, NR);
+        }
+        END {
+          printf("#0 %d\n", NR+1);
+          printf("<s> %d\n", NR+2);
+          printf("</s> %d\n", NR+3);
+        }' > $lang_dir/words || exit 1;
+      mv $lang_dir/words $lang_dir/words.txt
+    fi
+ 
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+  
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+      log "Converting L.pt to L.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L.pt \
+        $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+      log "Converting L_disambig.pt to L_disambig.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L_disambig.pt \
+        $lang_dir/L_disambig.fst
+    fi
+  done
+fi
diff --git a/egs/peoples_speech/ASR/shared b/egs/peoples_speech/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/peoples_speech/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/ptb/LM/README.md b/egs/ptb/LM/README.md
new file mode 100644
index 000000000..7629a950d
--- /dev/null
+++ b/egs/ptb/LM/README.md
@@ -0,0 +1,18 @@
+## Description
+
+(Note: the experiments here are only about language modeling)
+
+ptb is short for Penn Treebank.
+
+
+About the Penn Treebank corpus:
+  - This corpus is free for research purposes
+  - ptb.train.txt: train set
+  - ptb.valid.txt: development set (should be used just for tuning hyper-parameters, but not for training)
+  - ptb.test.txt: test set for reporting perplexity
+
+You can download the dataset from one of the following URLs:
+
+- https://github.com/townie/PTB-dataset-from-Tomas-Mikolov-s-webpage
+- http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
+- https://deepai.org/dataset/penn-treebank
diff --git a/egs/ptb/LM/local/prepare_lm_training_data.py b/egs/ptb/LM/local/prepare_lm_training_data.py
new file mode 120000
index 000000000..abc00d421
--- /dev/null
+++ b/egs/ptb/LM/local/prepare_lm_training_data.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
diff --git a/egs/ptb/LM/local/sort_lm_training_data.py b/egs/ptb/LM/local/sort_lm_training_data.py
new file mode 100755
index 000000000..bed3856e4
--- /dev/null
+++ b/egs/ptb/LM/local/sort_lm_training_data.py
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input the filename of LM training data
+generated by ./local/prepare_lm_training_data.py and sorts
+it by sentence length.
+
+Sentence length equals to the number of BPE tokens in a sentence.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import numpy as np
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--in-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/sorted_lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-statistics",
+        type=str,
+        help="Statistics about LM training data., data/bpe_500/statistics.txt",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    in_lm_data = Path(args.in_lm_data)
+    out_lm_data = Path(args.out_lm_data)
+    assert in_lm_data.is_file(), f"{in_lm_data}"
+    if out_lm_data.is_file():
+        logging.warning(f"{out_lm_data} exists - skipping")
+        return
+    data = torch.load(in_lm_data)
+    words2bpe = data["words"]
+    sentences = data["sentences"]
+    sentence_lengths = data["sentence_lengths"]
+
+    num_sentences = sentences.dim0
+    assert num_sentences == sentence_lengths.numel(), (
+        num_sentences,
+        sentence_lengths.numel(),
+    )
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+
+    sorted_sentences = sentences[indices.to(torch.int32)]
+    sorted_sentence_lengths = sentence_lengths[indices]
+
+    # Check that sentences are ordered by length
+    assert num_sentences == sorted_sentences.dim0, (
+        num_sentences,
+        sorted_sentences.dim0,
+    )
+
+    cur = None
+    for i in range(num_sentences):
+        word_ids = sorted_sentences[i]
+        token_ids = words2bpe[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+        if cur is not None:
+            assert cur >= token_ids.numel(), (cur, token_ids.numel())
+
+        cur = token_ids.numel()
+        assert cur == sorted_sentence_lengths[i]
+
+    data["sentences"] = sorted_sentences
+    data["sentence_lengths"] = sorted_sentence_lengths
+    torch.save(data, args.out_lm_data)
+    logging.info(f"Saved to {args.out_lm_data}")
+
+    statistics = Path(args.out_statistics)
+
+    # Write statistics
+    num_words = sorted_sentences.numel()
+    num_tokens = sentence_lengths.sum().item()
+    max_sentence_length = sentence_lengths[indices[0]]
+    min_sentence_length = sentence_lengths[indices[-1]]
+
+    step = 10
+    hist, bins = np.histogram(
+        sentence_lengths.numpy(),
+        bins=np.arange(1, max_sentence_length + step, step),
+    )
+
+    histogram = np.stack((bins[:-1], hist)).transpose()
+
+    with open(statistics, "w") as f:
+        f.write(f"num_sentences: {num_sentences}\n")
+        f.write(f"num_words: {num_words}\n")
+        f.write(f"num_tokens: {num_tokens}\n")
+        f.write(f"max_sentence_length: {max_sentence_length}\n")
+        f.write(f"min_sentence_length: {min_sentence_length}\n")
+        f.write("histogram:\n")
+        f.write("  bin  count  percent\n")
+        for row in histogram:
+            f.write(
+                f"{int(row[0]):>5} {int(row[1]):>5}   "
+                f"{100.*row[1]/num_sentences:.3f}%\n"
+            )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/ptb/LM/local/test_prepare_lm_training_data.py b/egs/ptb/LM/local/test_prepare_lm_training_data.py
new file mode 100755
index 000000000..3790045fa
--- /dev/null
+++ b/egs/ptb/LM/local/test_prepare_lm_training_data.py
@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+
+
+def main():
+    lm_training_data = Path("./data/bpe_500/lm_data.pt")
+    bpe_model = Path("./data/bpe_500/bpe.model")
+    if not lm_training_data.exists():
+        logging.warning(f"{lm_training_data} does not exist - skipping")
+        return
+
+    if not bpe_model.exists():
+        logging.warning(f"{bpe_model} does not exist - skipping")
+        return
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(bpe_model))
+
+    data = torch.load(lm_training_data)
+    words2bpe = data["words"]
+    sentences = data["sentences"]
+
+    ss = []
+    unk = sp.decode(sp.unk_id()).strip()
+    for i in range(10):
+        s = sp.decode(words2bpe[sentences[i]].values.tolist())
+        s = s.replace(unk, "<unk>")
+        ss.append(s)
+
+    for s in ss:
+        print(s)
+    # You can compare the output with the first 10 lines of ptb.train.txt
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/ptb/LM/local/train_bpe_model.py b/egs/ptb/LM/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/ptb/LM/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/ptb/LM/prepare.sh b/egs/ptb/LM/prepare.sh
new file mode 100755
index 000000000..69fab999a
--- /dev/null
+++ b/egs/ptb/LM/prepare.sh
@@ -0,0 +1,126 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+dl_dir=$PWD/download
+# The following files will be downloaded to $dl_dir
+#  - ptb.train.txt
+#  - ptb.valid.txt
+#  - ptb.test.txt
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/bpe_xxx, data/bpe_yyy
+# if the array contains xxx, yyy
+vocab_sizes=(
+  500
+  # 1000
+  # 2000
+  # 5000
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+mkdir -p $dl_dir
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download data"
+
+  # Caution: The downloaded data has already been normalized for LM training.
+
+  if [ ! -f $dl_dir/.complete ]; then
+    url=http://raw.githubusercontent.com/townie/PTB-dataset-from-Tomas-Mikolov-s-webpage/master/data
+    wget --directory-prefix $dl_dir $url/ptb.train.txt
+    wget --directory-prefix $dl_dir $url/ptb.valid.txt
+    wget --directory-prefix $dl_dir $url/ptb.test.txt
+    touch $dl_dir/.complete
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Train BPE model"
+
+  # Caution: You have to use the same bpe model for training your acoustic model
+  # Caution: You have to use the same bpe model for training your acoustic model
+  # Caution: You have to use the same bpe model for training your acoustic model
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    ./local/train_bpe_model.py \
+      --lang-dir $lang_dir \
+      --vocab-size $vocab_size \
+      --transcript $dl_dir/ptb.train.txt
+  done
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Generate LM training data"
+  # Note: ptb.train.txt has already been normalized
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/ptb.train.txt \
+      --lm-archive $out_dir/lm_data.pt
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/ptb.valid.txt \
+      --lm-archive $out_dir/lm_data-valid.pt
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/ptb.test.txt \
+      --lm-archive $out_dir/lm_data-test.pt
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Sort LM training data"
+  # Sort LM training data generated in stage 1
+  # by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of BPE tokens
+  # in a sentence.
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data.pt \
+      --out-lm-data $out_dir/sorted_lm_data.pt \
+      --out-statistics $out_dir/statistics.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-valid.pt \
+      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+      --out-statistics $out_dir/statistics-valid.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-test.pt \
+      --out-lm-data $out_dir/sorted_lm_data-test.pt \
+      --out-statistics $out_dir/statistics-test.txt
+  done
+fi
diff --git a/egs/ptb/LM/rnn_lm b/egs/ptb/LM/rnn_lm
new file mode 120000
index 000000000..87f29771e
--- /dev/null
+++ b/egs/ptb/LM/rnn_lm
@@ -0,0 +1 @@
+../../../icefall/rnn_lm
\ No newline at end of file
diff --git a/egs/ptb/LM/shared b/egs/ptb/LM/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/ptb/LM/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/ptb/LM/train-rnn-lm.sh b/egs/ptb/LM/train-rnn-lm.sh
new file mode 100755
index 000000000..cb70b7856
--- /dev/null
+++ b/egs/ptb/LM/train-rnn-lm.sh
@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+
+# Please run ./prepare.sh first
+
+stage=-1
+stop_stage=100
+
+# Number of GPUs to use for training
+world_size=1
+
+# Number of epochs to train
+num_epochs=20
+
+# Use this epoch for computing ppl
+use_epoch=19
+
+# number of models to average for computing ppl
+use_avg=2
+
+exp_dir=./my-rnnlm-exp
+
+. shared/parse_options.sh || exit 1
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Training RNN LM"
+
+  ./rnn_lm/train.py \
+    --exp-dir $exp_dir \
+    --start-epoch 0 \
+    --num-epochs $num_epochs \
+    --world-size $world_size \
+    --use-fp16 0 \
+    --vocab-size 500 \
+    --lm-data ./data/lm_training_bpe_500/sorted_lm_data.pt \
+    --lm-data-valid ./data/lm_training_bpe_500/sorted_lm_data-valid.pt \
+    --embedding-dim 800 \
+    --hidden-dim 200 \
+    --num-layers 2 \
+    --tie-weights false \
+    --batch-size 50
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Computing perplexity"
+
+  ./rnn_lm/compute_perplexity.py \
+    --exp-dir $exp_dir \
+    --epoch $use_epoch \
+    --avg $use_avg \
+    --vocab-size 500 \
+    --lm-data ./data/lm_training_bpe_500/sorted_lm_data-test.pt \
+    --embedding-dim 800 \
+    --hidden-dim 200 \
+    --num-layers 2 \
+    --tie-weights false \
+    --batch-size 50
+fi
diff --git a/egs/reazonspeech/ASR/corpus b/egs/reazonspeech/ASR/corpus
new file mode 120000
index 000000000..d002acbb2
--- /dev/null
+++ b/egs/reazonspeech/ASR/corpus
@@ -0,0 +1 @@
+/mnt/syno128/volume1/fujimotos/workdir/20231211-k2/icefall/egs/reazonspeech/ASR/corpus/
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/decode_greedy.sh b/egs/reazonspeech/ASR/decode_greedy.sh
new file mode 100755
index 000000000..fe86db1f5
--- /dev/null
+++ b/egs/reazonspeech/ASR/decode_greedy.sh
@@ -0,0 +1,14 @@
+num_epochs=30
+for ((i=$num_epochs; i>=1; i--));
+do
+    for ((j=1; j<=$i; j++));
+    do
+        python3 ./pruned_transducer_stateless7_streaming/decode.py \
+            --exp-dir exp \
+            --lang data/lang_char \
+            --epoch $i \
+            --avg $j \
+            --max-duration 180 \
+            --decoding-method greedy_search
+    done
+done
diff --git a/egs/reazonspeech/ASR/decode_modified_beam.sh b/egs/reazonspeech/ASR/decode_modified_beam.sh
new file mode 100755
index 000000000..2011c74c2
--- /dev/null
+++ b/egs/reazonspeech/ASR/decode_modified_beam.sh
@@ -0,0 +1,14 @@
+num_epochs=30
+for ((i=$num_epochs; i>=1; i--));
+do
+    for ((j=1; j<=$i; j++));
+    do
+        python3 ./pruned_transducer_stateless7_streaming/decode.py \
+            --exp-dir exp \
+            --lang data/lang_char \
+            --epoch $i \
+            --avg $j \
+            --max-duration 180 \
+            --decoding-method modified_beam_search
+    done
+done
diff --git a/egs/reazonspeech/ASR/local/compute_fbank_reazonspeech.py b/egs/reazonspeech/ASR/local/compute_fbank_reazonspeech.py
new file mode 100644
index 000000000..00b18a13b
--- /dev/null
+++ b/egs/reazonspeech/ASR/local/compute_fbank_reazonspeech.py
@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+# Copyright    2023  The University of Electro-Communications  (Author: Teo Wen Shen)  # noqa
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import List, Tuple
+
+import torch
+
+# fmt: off
+from lhotse import (  # See the following for why LilcomChunkyWriter is preferred; https://github.com/k2-fsa/icefall/pull/404; https://github.com/lhotse-speech/lhotse/pull/527
+    CutSet,
+    Fbank,
+    FbankConfig,
+    LilcomChunkyWriter,
+    RecordingSet,
+    SupervisionSet,
+)
+
+# fmt: on
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+RNG_SEED = 42
+concat_params = {"gap": 1.0, "maxlen": 10.0}
+
+
+def make_cutset_blueprints(
+    manifest_dir: Path,
+) -> List[Tuple[str, CutSet]]:
+    cut_sets = []
+
+    # Create test dataset
+    logging.info("Creating test cuts.")
+    cut_sets.append(("test", CutSet.from_manifests(
+        recordings=RecordingSet.from_file(
+            manifest_dir / "reazonspeech_recordings_test.jsonl.gz"
+        ),
+        supervisions=SupervisionSet.from_file(
+            manifest_dir / "reazonspeech_supervisions_test.jsonl.gz"
+        ),
+    )))
+
+    # Create valid dataset
+    logging.info("Creating valid cuts.")
+    cut_sets.append(("valid", CutSet.from_manifests(
+        recordings=RecordingSet.from_file(
+            manifest_dir / "reazonspeech_recordings_valid.jsonl.gz"
+        ),
+        supervisions=SupervisionSet.from_file(
+            manifest_dir / "reazonspeech_supervisions_valid.jsonl.gz"
+        ),
+    )))
+
+    # Create train dataset
+    logging.info("Creating train cuts.")
+    cut_sets.append(("train", CutSet.from_manifests(
+        recordings=RecordingSet.from_file(
+            manifest_dir / "reazonspeech_recordings_train.jsonl.gz"
+        ),
+        supervisions=SupervisionSet.from_file(
+            manifest_dir / "reazonspeech_supervisions_train.jsonl.gz"
+        ),
+    )))
+    return cut_sets
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument("-m", "--manifest-dir", type=Path)
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    extractor = Fbank(FbankConfig(num_mel_bins=80))
+    num_jobs = min(16, os.cpu_count())
+
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    if (args.manifest_dir / ".reazonspeech-fbank.done").exists():
+        logging.info(
+            "Previous fbank computed for ReazonSpeech found. "
+            f"Delete {args.manifest_dir / '.reazonspeech-fbank.done'} to allow recomputing fbank."
+        )
+        return
+    else:
+        cut_sets = make_cutset_blueprints(args.manifest_dir)
+        for part, cut_set in cut_sets:
+            logging.info(f"Processing {part}")
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                num_jobs=num_jobs,
+                storage_path=(args.manifest_dir / f"feats_{part}").as_posix(),
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(args.manifest_dir / f"reazonspeech_cuts_{part}.jsonl.gz")
+
+        logging.info("All fbank computed for ReazonSpeech.")
+        (args.manifest_dir / ".reazonspeech-fbank.done").touch()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/local/display_manifest_statistics.py b/egs/reazonspeech/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..48e9dee8d
--- /dev/null
+++ b/egs/reazonspeech/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2022  The University of Electro-Communications (author: Teo Wen Shen)  # noqa
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest
+
+ARGPARSE_DESCRIPTION = """
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in
+pruned_transducer_stateless5/train.py for usage.
+"""
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description=ARGPARSE_DESCRIPTION,
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+
+    parser.add_argument("--manifest-dir", type=Path, help="Path to cutset manifests")
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_parser()
+
+    for part in ["train", "valid"]:
+        path = args.manifest_dir / f"reazonspeech_cuts_{part}.jsonl.gz"
+        cuts: CutSet = load_manifest(path)
+
+        print("\n---------------------------------\n")
+        print(path.name + ":")
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/local/prepare_lang_char.py b/egs/reazonspeech/ASR/local/prepare_lang_char.py
new file mode 100644
index 000000000..44ec0ea71
--- /dev/null
+++ b/egs/reazonspeech/ASR/local/prepare_lang_char.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+# Copyright    2022  The University of Electro-Communications  (Author: Teo Wen Shen)  # noqa
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+
+    parser.add_argument(
+        "train_cut", metavar="train-cut", type=Path, help="Path to the train cut"
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help=(
+            "Name of lang dir. "
+            "If not set, this will default to lang_char_{trans-mode}"
+        ),
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    logging.basicConfig(
+        format=("%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"),
+        level=logging.INFO,
+    )
+
+    sysdef_string = set(["<blk>", "<unk>", "<sos/eos>", " "])
+
+    token_set = set()
+    logging.info(f"Creating vocabulary from {args.train_cut}.")
+    train_cut: CutSet = CutSet.from_file(args.train_cut)
+    for cut in train_cut:
+        for sup in cut.supervisions:
+            token_set.update(sup.text)
+
+    token_set = ["<blk>"] + sorted(token_set - sysdef_string) + ["<unk>", "<sos/eos>"]
+    args.lang_dir.mkdir(parents=True, exist_ok=True)
+    (args.lang_dir / "tokens.txt").write_text(
+        "\n".join(f"{t}\t{i}" for i, t in enumerate(token_set))
+    )
+
+    (args.lang_dir / "lang_type").write_text("char")
+    logging.info("Done.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/local/utils/asr_datamodule.py b/egs/reazonspeech/ASR/local/utils/asr_datamodule.py
new file mode 100644
index 000000000..84ed9647b
--- /dev/null
+++ b/egs/reazonspeech/ASR/local/utils/asr_datamodule.py
@@ -0,0 +1,349 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class ReazonSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=False,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        transforms = []
+        input_transforms = []
+
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.info("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "reazonspeech_cuts_train.jsonl.gz")
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get valid cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "reazonspeech_cuts_valid.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "reazonspeech_cuts_test.jsonl.gz")
diff --git a/egs/reazonspeech/ASR/local/utils/tokenizer.py b/egs/reazonspeech/ASR/local/utils/tokenizer.py
new file mode 100644
index 000000000..c9be72be1
--- /dev/null
+++ b/egs/reazonspeech/ASR/local/utils/tokenizer.py
@@ -0,0 +1,253 @@
+import argparse
+from pathlib import Path
+from typing import Callable, List, Union
+
+import sentencepiece as spm
+from k2 import SymbolTable
+
+
+class Tokenizer:
+    text2word: Callable[[str], List[str]]
+
+    @staticmethod
+    def add_arguments(parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(title="Lang related options")
+
+        group.add_argument("--lang", type=Path, help="Path to lang directory.")
+
+        group.add_argument(
+            "--lang-type",
+            type=str,
+            default=None,
+            help=(
+                "Either 'bpe' or 'char'. If not provided, it expects lang_dir/lang_type to exists. "
+                "Note: 'bpe' directly loads sentencepiece.SentencePieceProcessor"
+            ),
+        )
+
+    @staticmethod
+    def Load(lang_dir: Path, lang_type="", oov="<unk>"):
+
+        if not lang_type:
+            assert (lang_dir / "lang_type").exists(), "lang_type not specified."
+            lang_type = (lang_dir / "lang_type").read_text().strip()
+
+        tokenizer = None
+
+        if lang_type == "bpe":
+            assert (
+                lang_dir / "bpe.model"
+            ).exists(), f"No BPE .model could be found in {lang_dir}."
+            tokenizer = spm.SentencePieceProcessor()
+            tokenizer.Load(str(lang_dir / "bpe.model"))
+        elif lang_type == "char":
+            tokenizer = CharTokenizer(lang_dir, oov=oov)
+        else:
+            raise NotImplementedError(f"{lang_type} not supported at the moment.")
+
+        return tokenizer
+
+    load = Load
+
+    def PieceToId(self, piece: str) -> int:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    piece_to_id = PieceToId
+
+    def IdToPiece(self, id: int) -> str:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    id_to_piece = IdToPiece
+
+    def GetPieceSize(self) -> int:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    get_piece_size = GetPieceSize
+
+    def __len__(self) -> int:
+        return self.get_piece_size()
+
+    def EncodeAsIdsBatch(self, input: List[str]) -> List[List[int]]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def EncodeAsPiecesBatch(self, input: List[str]) -> List[List[str]]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def EncodeAsIds(self, input: str) -> List[int]:
+        return self.EncodeAsIdsBatch([input])[0]
+
+    def EncodeAsPieces(self, input: str) -> List[str]:
+        return self.EncodeAsPiecesBatch([input])[0]
+
+    def Encode(
+        self, input: Union[str, List[str]], out_type=int
+    ) -> Union[List, List[List]]:
+        if not input:
+            return []
+
+        if isinstance(input, list):
+            if out_type is int:
+                return self.EncodeAsIdsBatch(input)
+            if out_type is str:
+                return self.EncodeAsPiecesBatch(input)
+
+        if out_type is int:
+            return self.EncodeAsIds(input)
+        if out_type is str:
+            return self.EncodeAsPieces(input)
+
+    encode = Encode
+
+    def DecodeIdsBatch(self, input: List[List[int]]) -> List[str]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def DecodePiecesBatch(self, input: List[List[str]]) -> List[str]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def DecodeIds(self, input: List[int]) -> str:
+        return self.DecodeIdsBatch([input])[0]
+
+    def DecodePieces(self, input: List[str]) -> str:
+        return self.DecodePiecesBatch([input])[0]
+
+    def Decode(
+        self,
+        input: Union[int, List[int], List[str], List[List[int]], List[List[str]]],
+    ) -> Union[List[str], str]:
+
+        if not input:
+            return ""
+
+        if isinstance(input, int):
+            return self.id_to_piece(input)
+        elif isinstance(input, str):
+            raise TypeError(
+                "Unlike spm.SentencePieceProcessor, cannot decode from type str."
+            )
+
+        if isinstance(input[0], list):
+            if not input[0] or isinstance(input[0][0], int):
+                return self.DecodeIdsBatch(input)
+
+            if isinstance(input[0][0], str):
+                return self.DecodePiecesBatch(input)
+
+        if isinstance(input[0], int):
+            return self.DecodeIds(input)
+        if isinstance(input[0], str):
+            return self.DecodePieces(input)
+
+        raise RuntimeError("Unknown input type")
+
+    decode = Decode
+
+    def SplitBatch(self, input: List[str]) -> List[List[str]]:
+        raise NotImplementedError(
+            "You need to implement this function in the child class."
+        )
+
+    def Split(self, input: Union[List[str], str]) -> Union[List[List[str]], List[str]]:
+        if isinstance(input, list):
+            return self.SplitBatch(input)
+        elif isinstance(input, str):
+            return self.SplitBatch([input])[0]
+        raise RuntimeError("Unknown input type")
+
+    split = Split
+
+
+class CharTokenizer(Tokenizer):
+    def __init__(self, lang_dir: Path, oov="<unk>", sep=""):
+        assert (
+            lang_dir / "tokens.txt"
+        ).exists(), f"tokens.txt could not be found in {lang_dir}."
+        token_table = SymbolTable.from_file(lang_dir / "tokens.txt")
+        assert (
+            "#0" not in token_table
+        ), "This tokenizer does not support disambig symbols."
+        self._id2sym = token_table._id2sym
+        self._sym2id = token_table._sym2id
+        self.oov = oov
+        self.oov_id = self._sym2id[oov]
+        self.sep = sep
+        if self.sep:
+            self.text2word = lambda x: x.split(self.sep)
+        else:
+            self.text2word = lambda x: list(x.replace(" ", ""))
+
+    def piece_to_id(self, piece: str) -> int:
+        try:
+            return self._sym2id[piece]
+        except KeyError:
+            return self.oov_id
+
+    def id_to_piece(self, id: int) -> str:
+        return self._id2sym[id]
+
+    def get_piece_size(self) -> int:
+        return len(self._sym2id)
+
+    def EncodeAsIdsBatch(self, input: List[str]) -> List[List[int]]:
+        return [[self.piece_to_id(i) for i in self.text2word(text)] for text in input]
+
+    def EncodeAsPiecesBatch(self, input: List[str]) -> List[List[str]]:
+        return [
+            [i if i in self._sym2id else self.oov for i in self.text2word(text)]
+            for text in input
+        ]
+
+    def DecodeIdsBatch(self, input: List[List[int]]) -> List[str]:
+        return [self.sep.join(self.id_to_piece(i) for i in text) for text in input]
+
+    def DecodePiecesBatch(self, input: List[List[str]]) -> List[str]:
+        return [self.sep.join(text) for text in input]
+
+    def SplitBatch(self, input: List[str]) -> List[List[str]]:
+        return [self.text2word(text) for text in input]
+
+
+def test_CharTokenizer():
+    test_single_string = "こんにちは"
+    test_multiple_string = [
+        "今日はいい天気ですよね",
+        "諏訪湖は綺麗でしょう",
+        "这在词表外",
+        "分かち 書き に し た 文章 です",
+        "",
+    ]
+    test_empty_string = ""
+    sp = Tokenizer.load(Path("lang_char"), "char", oov="<unk>")
+    splitter = sp.split
+    print(sp.encode(test_single_string, out_type=str))
+    print(sp.encode(test_single_string, out_type=int))
+    print(sp.encode(test_multiple_string, out_type=str))
+    print(sp.encode(test_multiple_string, out_type=int))
+    print(sp.encode(test_empty_string, out_type=str))
+    print(sp.encode(test_empty_string, out_type=int))
+    print(sp.decode(sp.encode(test_single_string, out_type=str)))
+    print(sp.decode(sp.encode(test_single_string, out_type=int)))
+    print(sp.decode(sp.encode(test_multiple_string, out_type=str)))
+    print(sp.decode(sp.encode(test_multiple_string, out_type=int)))
+    print(sp.decode(sp.encode(test_empty_string, out_type=str)))
+    print(sp.decode(sp.encode(test_empty_string, out_type=int)))
+    print(splitter(test_single_string))
+    print(splitter(test_multiple_string))
+    print(splitter(test_empty_string))
+
+
+if __name__ == "__main__":
+    test_CharTokenizer()
diff --git a/egs/reazonspeech/ASR/local/validate_manifest.py b/egs/reazonspeech/ASR/local/validate_manifest.py
new file mode 100644
index 000000000..7f67c64b6
--- /dev/null
+++ b/egs/reazonspeech/ASR/local/validate_manifest.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+- Supervision time bounds are within cut time bounds
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def validate_one_supervision_per_cut(c: Cut):
+    if len(c.supervisions) != 1:
+        raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
+
+
+def validate_supervision_and_cut_time_bounds(c: Cut):
+    s = c.supervisions[0]
+
+    # Removed because when the cuts were trimmed from supervisions,
+    # the start time of the supervision can be lesser than cut start time.
+    # https://github.com/lhotse-speech/lhotse/issues/813
+    # if s.start < c.start:
+    #     raise ValueError(
+    #         f"{c.id}: Supervision start time {s.start} is less "
+    #         f"than cut start time {c.start}"
+    #     )
+
+    if s.end > c.end:
+        raise ValueError(
+            f"{c.id}: Supervision end time {s.end} is larger "
+            f"than cut end time {c.end}"
+        )
+
+
+def main():
+    args = get_args()
+
+    manifest = Path(args.manifest)
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    for c in cut_set:
+        validate_one_supervision_per_cut(c)
+        validate_supervision_and_cut_time_bounds(c)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/reazonspeech/ASR/prepare.sh b/egs/reazonspeech/ASR/prepare.sh
new file mode 100755
index 000000000..f8e54f58c
--- /dev/null
+++ b/egs/reazonspeech/ASR/prepare.sh
@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=8
+stage=-1
+stop_stage=100
+
+reazonspeech_dir=corpus
+reazonspeech_manifest_dir=data
+
+. shared/parse_options.sh || exit 1
+
+mkdir -p data
+
+log() {
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+    log "Stage 1: Prepare ReazonSpeech manifest"
+    if [ ! -e $reazonspeech_manifest_dir/.reazonspeech.done ]; then
+        lhotse prepare reazonspeech $reazonspeech_dir $reazonspeech_manifest_dir
+        touch $reazonspeech_manifest_dir/.reazonspeech.done
+    fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+    log "Stage 2: Compute ReazonSpeech fbank"
+    if [ ! -e $reazonspeech_manifest_dir/.reazonspeech-validated.done ]; then
+        python local/compute_fbank_reazonspeech.py --manifest-dir $reazonspeech_manifest_dir
+        python local/validate_manifest.py --manifest $reazonspeech_manifest_dir/reazonspeech_cuts_train.jsonl.gz
+        python local/validate_manifest.py --manifest $reazonspeech_manifest_dir/reazonspeech_cuts_valid.jsonl.gz
+        python local/validate_manifest.py --manifest $reazonspeech_manifest_dir/reazonspeech_cuts_test.jsonl.gz
+        touch $reazonspeech_manifest_dir/.reazonspeech-validated.done
+    fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3: Prepare ReazonSpeech lang_char"
+    python local/prepare_lang_char.py $reazonspeech_manifest_dir/reazonspeech_cuts_train.jsonl.gz
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+    log "Stage 4: Show manifest statistics"
+    python local/display_manifest_statistics.py --manifest-dir $reazonspeech_manifest_dir > $reazonspeech_manifest_dir/manifest_statistics.txt
+    cat $reazonspeech_manifest_dir/manifest_statistics.txt
+fi
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
new file mode 120000
index 000000000..a48591198
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
@@ -0,0 +1 @@
+../local/utils/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
new file mode 120000
index 000000000..d7349b0a3
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/beam_search.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode.py
new file mode 100755
index 000000000..446890daa
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode.py
@@ -0,0 +1,846 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --lang data/lang_char \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --lang data/lang_char \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --lang data/lang_char \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --lang data/lang_char \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --lang data/lang_char \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --lang data/lang_char \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --lang data/lang_char \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import ReazonSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from tokenizer import Tokenizer
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--gpu",
+        type=int,
+        default=0,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--res-dir",
+        type=Path,
+        default=None,
+        help="The path to save results.",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        help="The lang dir. It should contain at least a word table.",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-graph",
+        type=str,
+        default="",
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--pad-feature",
+        type=int,
+        default=30,
+        help="""
+        Number of frames to pad at the end.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: Tokenizer,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.pad_feature:
+        feature_lens += params.pad_feature
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.pad_feature),
+            value=LOG_EPS,
+        )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(sp.text2word(hyp))
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.text2word(sp.decode(hyp)))
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: Tokenizer,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = sp.text2word(ref_text)
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    return test_set_wers
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    ReazonSpeechAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    if not params.res_dir:
+        params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", params.gpu)
+
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> and <unk> are defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    decoding_graph = None
+    word_table = None
+
+    if params.decoding_graph:
+        decoding_graph = k2.Fsa.from_dict(
+            torch.load(params.decoding_graph, map_location=device)
+        )
+    elif "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    reazonspeech_corpus = ReazonSpeechAsrDataModule(args)
+
+    for subdir in ["valid"]:
+        results_dict = decode_dataset(
+            dl=reazonspeech_corpus.test_dataloaders(getattr(reazonspeech_corpus, f"{subdir}_cuts")()),
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+        tot_err = save_results(
+            params=params,
+            test_set_name=subdir,
+            results_dict=results_dict,
+        )
+        with (
+            params.res_dir
+            / (
+                f"{subdir}-{params.decode_chunk_len}_{params.beam_size}"
+                f"_{params.avg}_{params.epoch}.cer"
+            )
+        ).open("w") as fout:
+            if len(tot_err) == 1:
+                fout.write(f"{tot_err[0][1]}")
+            else:
+                fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
new file mode 120000
index 000000000..ca8fed319
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decoder.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
new file mode 120000
index 000000000..1ce277aa6
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decoder.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
new file mode 100755
index 000000000..072679cfc
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py
@@ -0,0 +1,1261 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import math
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import ReazonSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from tokenizer import Tokenizer
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer_for_ncnn_export_only import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+LOG_EPS = math.log(1e-10)
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--pad-feature",
+        type=int,
+        default=0,
+        help="""
+        Number of frames to pad at the end.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+        is_pnnx=True,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.pad_feature:
+        feature_lens += params.pad_feature
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.pad_feature),
+            value=LOG_EPS,
+        )
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: Tokenizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except Exception as e:  # noqa
+            logging.error(e, exc_info=True)
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise e
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            log_mode = logging.info
+            log_mode(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            log_mode(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, master_port=params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(args.lang, args.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 0.3 or c.duration > 30.0:
+            logging.debug(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.info(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    reazonspeech_corpus = ReazonSpeechAsrDataModule(args)
+    train_cuts = reazonspeech_corpus.train_cuts()
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = reazonspeech_corpus.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = reazonspeech_corpus.valid_cuts()
+    valid_dl = reazonspeech_corpus.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: Tokenizer,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: Tokenizer,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    raise RuntimeError("Please don't use this file directly!")
+    parser = get_parser()
+    ReazonSpeechAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
new file mode 120000
index 000000000..cb673b3eb
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/export.py
new file mode 100644
index 000000000..666bcc831
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/export.py
@@ -0,0 +1,250 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from tokenizer import Tokenizer
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/joiner.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
new file mode 120000
index 000000000..482ebcfef
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/joiner.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/model.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/model.py
new file mode 120000
index 000000000..16c2bf28d
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/model.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/optim.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/optim.py
new file mode 120000
index 000000000..522bbaff9
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/optim.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
new file mode 100644
index 000000000..932026868
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --lang data/lang_char \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7_streaming/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7_streaming/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_streaming/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from tokenizer import Tokenizer
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
new file mode 120000
index 000000000..a7ef73bcb
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/scaling.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
new file mode 120000
index 000000000..566c317ff
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
new file mode 120000
index 000000000..92c3904af
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
new file mode 120000
index 000000000..2adf271c1
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
new file mode 100755
index 000000000..4c18c7563
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
@@ -0,0 +1,597 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding_method greedy_search \
+  --lang data/lang_char \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+import torch.nn as nn
+from asr_datamodule import ReazonSpeechAsrDataModule
+from decode import save_results
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from tokenizer import Tokenizer
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--gpu",
+        type=int,
+        default=0,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-graph",
+        type=str,
+        default="",
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    parser.add_argument(
+        "--res-dir",
+        type=Path,
+        default=None,
+        help="The path to save results.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: Tokenizer,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].custom[params.transcript_mode]
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        sp.text2word(decode_streams[i].ground_truth),
+                        sp.text2word(sp.decode(decode_streams[i].decoding_result())),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    sp.text2word(decode_streams[i].ground_truth),
+                    sp.text2word(sp.decode(decode_streams[i].decoding_result())),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    ReazonSpeechAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if not params.res_dir:
+        params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", params.gpu)
+
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(params.lang, params.lang_type)
+
+    # <blk> and <unk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_graph:
+        decoding_graph = k2.Fsa.from_dict(
+            torch.load(params.decoding_graph, map_location=device)
+        )
+    elif params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    args.return_cuts = True
+    reazonspeech_corpus = ReazonSpeechAsrDataModule(args)
+
+    for subdir in ["valid"]:
+        results_dict = decode_dataset(
+            cuts=getattr(reazonspeech_corpus, f"{subdir}_cuts")(),
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+        tot_err = save_results(
+            params=params, test_set_name=subdir, results_dict=results_dict
+        )
+
+        with (
+            params.res_dir
+            / (
+                f"{subdir}-{params.decode_chunk_len}"
+                f"_{params.avg}_{params.epoch}.cer"
+            )
+        ).open("w") as fout:
+            if len(tot_err) == 1:
+                fout.write(f"{tot_err[0][1]}")
+            else:
+                fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/tokenizer.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/tokenizer.py
new file mode 120000
index 000000000..958c99e85
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/tokenizer.py
@@ -0,0 +1 @@
+../local/utils/tokenizer.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/train.py
new file mode 100755
index 000000000..32cd4c576
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -0,0 +1,1259 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --lang data/lang_char \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import math
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import ReazonSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from tokenizer import Tokenizer
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+LOG_EPS = math.log(1e-10)
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--pad-feature",
+        type=int,
+        default=0,
+        help="""
+        Number of frames to pad at the end.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.pad_feature:
+        feature_lens += params.pad_feature
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.pad_feature),
+            value=LOG_EPS,
+        )
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: Tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: Tokenizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except Exception as e:  # noqa
+            logging.error(e, exc_info=True)
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise e
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            log_mode = logging.info
+            log_mode(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            log_mode(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, master_port=params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = Tokenizer.load(args.lang, args.lang_type)
+
+    # <blk> is defined in local/prepare_lang_char.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 0.3 or c.duration > 30.0:
+            logging.debug(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.info(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    reazonspeech_corpus = ReazonSpeechAsrDataModule(args)
+    train_cuts = reazonspeech_corpus.train_cuts()
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = reazonspeech_corpus.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = reazonspeech_corpus.valid_cuts()
+    valid_dl = reazonspeech_corpus.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: Tokenizer,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: Tokenizer,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    ReazonSpeechAsrDataModule.add_arguments(parser)
+    Tokenizer.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
new file mode 120000
index 000000000..ec183baa7
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
new file mode 120000
index 000000000..d301e1f9b
--- /dev/null
+++ b/egs/reazonspeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer_for_ncnn_export_only.py
\ No newline at end of file
diff --git a/egs/reazonspeech/ASR/shared b/egs/reazonspeech/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/reazonspeech/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/README.md b/egs/spgispeech/ASR/README.md
new file mode 100644
index 000000000..f60408cc1
--- /dev/null
+++ b/egs/spgispeech/ASR/README.md
@@ -0,0 +1,32 @@
+# SPGISpeech
+
+SPGISpeech consists of 5,000 hours of recorded company earnings calls and their respective
+transcriptions. The original calls were split into slices ranging from 5 to 15 seconds in
+length to allow easy training for speech recognition systems. Calls represent a broad
+cross-section of international business English; SPGISpeech contains approximately 50,000
+speakers, one of the largest numbers of any speech corpus, and offers a variety of L1 and
+L2 English accents. The format of each WAV file is single channel, 16kHz, 16 bit audio.
+
+Transcription text represents the output of several stages of manual post-processing.
+As such, the text contains polished English orthography following a detailed style guide,
+including proper casing, punctuation, and denormalized non-standard words such as numbers
+and acronyms, making SPGISpeech suited for training fully formatted end-to-end models.
+
+Official reference:
+
+O’Neill, P.K., Lavrukhin, V., Majumdar, S., Noroozi, V., Zhang, Y., Kuchaiev, O., Balam,
+J., Dovzhenko, Y., Freyberg, K., Shulman, M.D., Ginsburg, B., Watanabe, S., & Kucsko, G.
+(2021). SPGISpeech: 5, 000 hours of transcribed financial audio for fully formatted
+end-to-end speech recognition. ArXiv, abs/2104.02014.
+
+ArXiv link: https://arxiv.org/abs/2104.02014
+
+## Performance Record
+
+| Decoding method           | val WER    | val CER |
+|---------------------------|------------|---------|
+| greedy search             | 2.40       |  0.99   |
+| modified beam search      | 2.24       |  0.91   |
+| fast beam search          | 2.35       |  0.97   |
+
+See [RESULTS](/egs/spgispeech/ASR/RESULTS.md) for details.
diff --git a/egs/spgispeech/ASR/RESULTS.md b/egs/spgispeech/ASR/RESULTS.md
new file mode 100644
index 000000000..de9e35c5a
--- /dev/null
+++ b/egs/spgispeech/ASR/RESULTS.md
@@ -0,0 +1,73 @@
+## Results
+
+### SPGISpeech BPE training results (Pruned Transducer)
+
+#### 2022-05-11
+
+#### Conformer encoder + embedding decoder
+
+Conformer encoder + non-current decoder. The decoder
+contains only an embedding layer, a Conv1d (with kernel size 2) and a linear
+layer (to transform tensor dim).
+
+The WERs are
+
+|                           | dev | val | comment                                  |
+|---------------------------|------------|------------|------------------------------------------|
+| greedy search             | 2.46       | 2.40       | --avg-last-n 10 --max-duration 500 |
+| modified beam search      | 2.28       | 2.24       | --avg-last-n 10 --max-duration 500 --beam-size 4 |
+| fast beam search          | 2.38       | 2.35       | --avg-last-n 10 --max-duration 500 --beam-size 4 --max-contexts 4 --max-states 8 |
+
+**NOTE:** SPGISpeech transcripts can be prepared in `ortho` or `norm` ways, which refer to whether the
+transcripts are orthographic or normalized. These WERs correspond to the normalized transcription
+scenario.
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 8 \
+  --num-epochs 20 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --max-duration 200 \
+  --prune-range 5 \
+  --lr-factor 5 \
+  --lm-scale 0.25 \
+  --use-fp16 True
+```
+
+The decoding command is:
+```
+# greedy search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+# modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+# fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+Pretrained model is available at <https://huggingface.co/desh2608/icefall-asr-spgispeech-pruned-transducer-stateless2>
+
+The tensorboard training log can be found at
+<https://tensorboard.dev/experiment/ExSoBmrPRx6liMTGLu0Tgw/#scalars>
diff --git a/egs/spgispeech/ASR/local/__init__.py b/egs/spgispeech/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/spgispeech/ASR/local/compile_hlg.py b/egs/spgispeech/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/spgispeech/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/local/compute_fbank_musan.py b/egs/spgispeech/ASR/local/compute_fbank_musan.py
new file mode 100755
index 000000000..9bea28a41
--- /dev/null
+++ b/egs/spgispeech/ASR/local/compute_fbank_musan.py
@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, LilcomChunkyWriter, combine
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    manifests = read_manifests_if_cached(
+        prefix="musan", dataset_parts=dataset_parts, output_dir=src_dir
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    musan_cuts_path = src_dir / "cuts_musan.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    # create chunks of Musan with duration 5 - 10 seconds
+    musan_cuts = (
+        CutSet.from_manifests(
+            recordings=combine(part["recordings"] for part in manifests.values())
+        )
+        .cut_into_windows(10.0)
+        .filter(lambda c: c.duration > 5)
+        .compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=output_dir / "feats_musan",
+            batch_duration=500,
+            num_workers=4,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+    )
+
+    logging.info(f"Saving to {musan_cuts_path}")
+    musan_cuts.to_file(musan_cuts_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()
diff --git a/egs/spgispeech/ASR/local/compute_fbank_spgispeech.py b/egs/spgispeech/ASR/local/compute_fbank_spgispeech.py
new file mode 100755
index 000000000..20ff6d7ab
--- /dev/null
+++ b/egs/spgispeech/ASR/local/compute_fbank_spgispeech.py
@@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the SPGISpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import LilcomChunkyWriter, load_manifest_lazy
+from lhotse.features.kaldifeat import (
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    KaldifeatFrameOptions,
+    KaldifeatMelOptions,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        default=20,
+        help="Number of splits for the train set.",
+    )
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Start index of the train set split.",
+    )
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop index of the train set split.",
+    )
+    parser.add_argument(
+        "--test",
+        action="store_true",
+        help="If set, only compute features for the dev and val set.",
+    )
+    parser.add_argument(
+        "--train",
+        action="store_true",
+        help="If set, only compute features for the train set.",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_spgispeech(args):
+    assert args.train or args.test, "Either train or test must be set."
+
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    sampling_rate = 16000
+    num_mel_bins = 80
+
+    extractor = KaldifeatFbank(
+        KaldifeatFbankConfig(
+            frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate),
+            mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins),
+            device="cuda",
+        )
+    )
+
+    if args.train:
+        logging.info("Processing train")
+        cut_set = load_manifest_lazy(src_dir / "cuts_train_raw.jsonl.gz")
+        chunk_size = len(cut_set) // args.num_splits
+        cut_sets = cut_set.split_lazy(
+            output_dir=src_dir / f"cuts_train_raw_split{args.num_splits}",
+            chunk_size=chunk_size,
+        )
+        start = args.start
+        stop = min(args.stop, args.num_splits) if args.stop > 0 else args.num_splits
+        num_digits = len(str(args.num_splits))
+        for i in range(start, stop):
+            idx = f"{i + 1}".zfill(num_digits)
+            cuts_train_idx_path = src_dir / f"cuts_train_{idx}.jsonl.gz"
+            logging.info(f"Processing train split {i}")
+            cs = cut_sets[i].compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_train_{idx}",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+                overwrite=True,
+            )
+            cs.to_file(cuts_train_idx_path)
+
+    if args.test:
+        for partition in ["dev", "val"]:
+            if (output_dir / f"cuts_{partition}.jsonl.gz").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = load_manifest_lazy(src_dir / f"cuts_{partition}_raw.jsonl.gz")
+            cut_set = cut_set.compute_and_store_features_batch(
+                extractor=extractor,
+                storage_path=output_dir / f"feats_{partition}",
+                manifest_path=src_dir / f"cuts_{partition}.jsonl.gz",
+                batch_duration=500,
+                num_workers=4,
+                storage_type=LilcomChunkyWriter,
+                overwrite=True,
+            )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_spgispeech(args)
diff --git a/egs/spgispeech/ASR/local/prepare_lang.py b/egs/spgispeech/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/spgispeech/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/local/prepare_lang_bpe.py b/egs/spgispeech/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/spgispeech/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/local/prepare_splits.py b/egs/spgispeech/ASR/local/prepare_splits.py
new file mode 100755
index 000000000..508d4acd8
--- /dev/null
+++ b/egs/spgispeech/ASR/local/prepare_splits.py
@@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+# Copyright    2022  Johns Hopkins University        (authors: Desh Raj)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file splits the training set into train and dev sets.
+"""
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def split_spgispeech_train():
+    src_dir = Path("data/manifests")
+
+    manifests = read_manifests_if_cached(
+        dataset_parts=["train", "val"],
+        output_dir=src_dir,
+        prefix="spgispeech",
+        suffix="jsonl.gz",
+        lazy=True,
+    )
+    assert manifests is not None
+
+    train_dev_cuts = CutSet.from_manifests(
+        recordings=manifests["train"]["recordings"],
+        supervisions=manifests["train"]["supervisions"],
+    )
+    dev_cuts = train_dev_cuts.subset(first=4000)
+    train_cuts = train_dev_cuts.filter(lambda c: c not in dev_cuts)
+
+    # Add speed perturbation
+    train_cuts = (
+        train_cuts + train_cuts.perturb_speed(0.9) + train_cuts.perturb_speed(1.1)
+    )
+
+    # Write the manifests to disk.
+    train_cuts.to_file(src_dir / "cuts_train_raw.jsonl.gz")
+    dev_cuts.to_file(src_dir / "cuts_dev_raw.jsonl.gz")
+
+    # Also write the val set to disk.
+    val_cuts = CutSet.from_manifests(
+        recordings=manifests["val"]["recordings"],
+        supervisions=manifests["val"]["supervisions"],
+    )
+    val_cuts.to_file(src_dir / "cuts_val_raw.jsonl.gz")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    split_spgispeech_train()
diff --git a/egs/spgispeech/ASR/local/train_bpe_model.py b/egs/spgispeech/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/spgispeech/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/prepare.sh b/egs/spgispeech/ASR/prepare.sh
new file mode 100755
index 000000000..8331f94d5
--- /dev/null
+++ b/egs/spgispeech/ASR/prepare.sh
@@ -0,0 +1,199 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=20
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/spgispeech
+#      You can find train.csv, val.csv, train, and val in this directory, which belong
+#      to the SPGISpeech dataset.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/spgispeech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/spgispeech $dl_dir/spgispeech
+  #
+  if [ ! -d $dl_dir/spgispeech/train.csv ]; then
+    lhotse download spgispeech $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare SPGISpeech manifest (may take ~1h)"
+  # We assume that you have downloaded the SPGISpeech corpus
+  # to $dl_dir/spgispeech. We perform text normalization for the transcripts.
+  mkdir -p data/manifests
+  lhotse prepare spgispeech -j $nj --normalize-text $dl_dir/spgispeech data/manifests
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+  lhotse combine data/manifests/recordings_{music,speech,noise}.json data/manifests/recordings_musan.jsonl.gz
+  lhotse cut simple -r data/manifests/recordings_musan.jsonl.gz data/manifests/cuts_musan_raw.jsonl.gz
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Split train into train and dev and create cut sets."
+  python local/prepare_splits.py
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank features for spgispeech dev and val"
+  mkdir -p data/fbank
+  python local/compute_fbank_spgispeech.py --test
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank features for train"
+  mkdir -p data/fbank
+  python local/compute_fbank_spgispeech.py --train --num-splits 20
+
+  log "Combine features from train splits (may take ~1h)"
+  if [ ! -f data/manifests/cuts_train.jsonl.gz ]; then
+    pieces=$(find data/manifests -name "cuts_train_[0-9]*.jsonl.gz")
+    lhotse combine $pieces data/manifests/cuts_train.jsonl.gz
+  fi
+  gunzip -c data/manifests/cuts_train.jsonl.gz | shuf | gzip -c > data/manifests/cuts_train_shuf.jsonl.gz
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compute fbank features for musan"
+  mkdir -p data/fbank
+  python local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Dump transcripts for LM training"
+  mkdir -p data/lm
+  gunzip -c data/manifests/cuts_train_raw.jsonl.gz \
+    | jq '.supervisions[0].text' \
+    | sed 's:"::g' \
+    > data/lm/transcript_words.txt
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    # Add special words to words.txt
+    echo "<eps> 0" > $lang_dir/words.txt
+    echo "!SIL 1" >> $lang_dir/words.txt
+    echo "<UNK> 2" >> $lang_dir/words.txt
+
+    # Add regular words to words.txt
+    gunzip -c data/manifests/cuts_train_raw.jsonl.gz \
+      | jq '.supervisions[0].text' \
+      | sed 's:"::g' \
+      | sed 's: :\n:g' \
+      | sort \
+      | uniq \
+      | sed '/^$/d' \
+      | awk '{print $0,NR+2}' \
+      >> $lang_dir/words.txt
+
+    # Add remaining special word symbols expected by LM scripts.
+    num_words=$(cat $lang_dir/words.txt | wc -l)
+    echo "<s> ${num_words}" >> $lang_dir/words.txt
+    num_words=$(cat $lang_dir/words.txt | wc -l)
+    echo "</s> ${num_words}" >> $lang_dir/words.txt
+    num_words=$(cat $lang_dir/words.txt | wc -l)
+    echo "#0 ${num_words}" >> $lang_dir/words.txt
+
+    ./local/train_bpe_model.py \
+      --lang-dir $lang_dir \
+      --vocab-size $vocab_size \
+      --transcript data/lm/transcript_words.txt
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+    fi
+  done
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Train LM"
+  lm_dir=data/lm
+
+  if [ ! -f $lm_dir/G.arpa ]; then
+    ./shared/make_kn_lm.py \
+      -ngram-order 3 \
+      -text $lm_dir/transcript_words.txt \
+      -lm $lm_dir/G.arpa
+  fi
+
+  if [ ! -f $lm_dir/G_3_gram.fst.txt ]; then
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $lm_dir/G.arpa > $lm_dir/G_3_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+  done
+fi
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/__init__.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 100644
index 000000000..cf70fc0f8
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1,353 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class SPGISpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/manifests"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=100.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the BucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=8,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=2,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+            )
+
+        logging.info("Using DynamicBucketingSampler.")
+        train_sampler = DynamicBucketingSampler(
+            cuts_train,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
+        )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+        )
+        sampler = DynamicBucketingSampler(
+            cuts, max_duration=self.args.max_duration, shuffle=False
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get SPGISpeech train cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_train_shuf.jsonl.gz")
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get SPGISpeech dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def val_cuts(self) -> CutSet:
+        logging.info("About to get SPGISpeech val cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_val.jsonl.gz")
+
+
+def test():
+    parser = argparse.ArgumentParser()
+    SPGISpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    adm = SPGISpeechAsrDataModule(args)
+
+    cuts = adm.train_cuts()
+    dl = adm.train_dataloaders(cuts)
+    for i, batch in tqdm(enumerate(dl)):
+        if i == 100:
+            break
+
+    cuts = adm.dev_cuts()
+    dl = adm.valid_dataloaders(cuts)
+    for i, batch in tqdm(enumerate(dl)):
+        if i == 100:
+            break
+
+
+if __name__ == "__main__":
+    test()
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/conformer.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/conformer.py
new file mode 120000
index 000000000..a65957180
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/decode.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..4434aae62
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,566 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --iter 696000 \
+        --avg 10 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import SPGISpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        wers_filename = params.res_dir / f"wers-{test_set_name}-{params.suffix}.txt"
+        with open(wers_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        # we also compute CER for spgispeech dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        cers_filename = params.res_dir / f"cers-{test_set_name}-{params.suffix}.txt"
+        with open(cers_filename, "w") as f:
+            cer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_cers[key] = cer
+
+        logging.info("Wrote detailed error stats to {}".format(wers_filename))
+
+    test_set_wers = {k: v for k, v in sorted(test_set_wers.items(), key=lambda x: x[1])}
+    test_set_cers = {k: v for k, v in sorted(test_set_cers.items(), key=lambda x: x[1])}
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER\tCER", file=f)
+        for key in test_set_wers:
+            print(
+                "{}\t{}\t{}".format(key, test_set_wers[key], test_set_cers[key]),
+                file=f,
+            )
+
+    s = "\nFor {}, WER/CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key in test_set_wers:
+        s += "{}\t{}\t{}{}\n".format(key, test_set_wers[key], test_set_cers[key], note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    SPGISpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    spgispeech = SPGISpeechAsrDataModule(args)
+
+    dev_cuts = spgispeech.dev_cuts()
+    val_cuts = spgispeech.val_cuts()
+
+    dev_dl = spgispeech.test_dataloaders(dev_cuts)
+    val_dl = spgispeech.test_dataloaders(val_cuts)
+
+    test_sets = ["dev", "val"]
+    test_dl = [dev_dl, val_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/decoder.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py
new file mode 100755
index 000000000..68763808a
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,197 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --avg-last-n 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/spgispeech/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/joiner.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/model.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/optim.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/scaling.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/train.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/train.py
new file mode 100755
index 000000000..a9146a0fe
--- /dev/null
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,1003 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --use_fp16 1 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import SPGISpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=4,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=10,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    spgispeech = SPGISpeechAsrDataModule(args)
+
+    train_cuts = spgispeech.train_cuts()
+
+    # Ideally we should filter utterances that are too long or too short,
+    # but SPGISpeech contains regular length utterances so we don't need to
+    # do that. Here are the statistics of the training data (obtained by
+    # `train_cuts.describe()`):
+
+    # Cuts count: 5886320
+    # Total duration (hours): 15070.1
+    # Speech duration (hours): 15070.1 (100.0%)
+    # ***
+    # Duration statistics (seconds):
+    # mean    9.2
+    # std     2.8
+    # min     4.6
+    # 25%     6.9
+    # 50%     8.9
+    # 75%     11.2
+    # 99%     16.0
+    # 99.5%   16.3
+    # 99.9%   16.6
+    # max     16.7
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = spgispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = spgispeech.dev_cuts()
+    valid_dl = spgispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+                display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+
+def main():
+    parser = get_parser()
+    SPGISpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/spgispeech/ASR/shared b/egs/spgispeech/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/spgispeech/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/swbd/ASR/.gitignore b/egs/swbd/ASR/.gitignore
new file mode 100644
index 000000000..11d674922
--- /dev/null
+++ b/egs/swbd/ASR/.gitignore
@@ -0,0 +1,2 @@
+switchboard_word_alignments.tar.gz
+./swb_ms98_transcriptions/
diff --git a/egs/swbd/ASR/README.md b/egs/swbd/ASR/README.md
new file mode 100644
index 000000000..13b27815a
--- /dev/null
+++ b/egs/swbd/ASR/README.md
@@ -0,0 +1,25 @@
+# Switchboard
+
+The Switchboard-1 Telephone Speech Corpus (LDC97S62) consists of approximately 260 hours of speech and was originally collected by Texas Instruments in 1990-1, under DARPA sponsorship. The first release of the corpus was published by NIST and distributed by the LDC in 1992-3. Since that release, a number of corrections have been made to the data files as presented on the original CD-ROM set and all copies of the first pressing have been distributed.
+
+Switchboard is a collection of about 2,400 two-sided telephone conversations among 543 speakers (302 male, 241 female) from all areas of the United States. A computer-driven robot operator system handled the calls, giving the caller appropriate recorded prompts, selecting and dialing another person (the callee) to take part in a conversation, introducing a topic for discussion and recording the speech from the two subjects into separate channels until the conversation was finished. About 70 topics were provided, of which about 50 were used frequently. Selection of topics and callees was constrained so that: (1) no two speakers would converse together more than once and (2) no one spoke more than once on a given topic.
+
+(The above introduction is from the [LDC Switchboard-1 Release 2 webpage](https://catalog.ldc.upenn.edu/LDC97S62).)
+
+
+## Performance Record
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    33.37   |  35.06 |
+
+See [RESULTS](/egs/swbd/ASR/RESULTS.md) for details.
+
+## Credit
+
+The training script for `conformer_ctc` comes from the LibriSpeech `conformer_ctc` recipe in icefall.
+
+A lot of the scripts for data processing are from the first-gen Kaldi and the ESPNet project, tailored by myself to incorporate with Lhotse and Icefall.
+
+Some of the scripts for text normalization are from stale pull requests of [Piotr Żelasko](https://github.com/pzelasko) and [Nagendra Goel](https://github.com/ngoel17).
+
+The `sclite_scoring.py` is from the GigaSpeech recipe for post processing and glm-like scoring, which is definitely not an elegant stuff to do.
diff --git a/egs/swbd/ASR/RESULTS.md b/egs/swbd/ASR/RESULTS.md
new file mode 100644
index 000000000..f3a22c444
--- /dev/null
+++ b/egs/swbd/ASR/RESULTS.md
@@ -0,0 +1,113 @@
+## Results
+### Switchboard BPE training results (Conformer-CTC)
+
+#### 2023-09-04
+
+The best WER, as of 2023-09-04, for the Switchboard is below
+
+Results using attention decoder are given as:
+
+|                                |  eval2000-swbd  |  eval2000-callhome  | eval2000-avg |
+|--------------------------------|-----------------|---------------------|--------------|
+|         `conformer_ctc`        |      9.48       |         17.73       |    13.67     | 
+
+Decoding results and models can be found here:
+https://huggingface.co/zrjin/icefall-asr-swbd-conformer-ctc-2023-8-26
+#### 2023-06-27
+
+The best WER, as of 2023-06-27, for the Switchboard is below
+
+Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    30.80   |  32.29 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+
+##### eval2000
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.9       |       1.1       |
+
+##### rt03
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.9       |       1.9       |
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+cd egs/swbd/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1"
+./conformer_ctc/train.py \
+  --max-duration 120 \
+  --num-workers 8 \
+  --enable-musan False \
+  --world-size 2 \
+  --num-epochs 100
+```
+
+and the following command for decoding:
+
+```bash
+./conformer_ctc/decode.py \
+  --epoch 99 \
+  --avg 10 \
+  --max-duration 50
+```
+
+#### 2023-06-26
+
+The best WER, as of 2023-06-26, for the Switchboard is below
+
+Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    33.37   |  35.06 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+
+##### eval2000
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.3       |       2.5       |
+
+##### rt03
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.7       |       1.3       |
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+cd egs/swbd/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1"
+./conformer_ctc/train.py \
+  --max-duration 120 \
+  --num-workers 8 \
+  --enable-musan False \
+  --world-size 2 \
+```
+
+and the following command for decoding:
+
+```bash
+./conformer_ctc/decode.py \
+  --epoch 55 \
+  --avg 1 \
+  --max-duration 50
+```
+
+For your reference, the nbest oracle WERs are:
+
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    25.64   |  26.84 |
diff --git a/egs/swbd/ASR/conformer_ctc/__init__.py b/egs/swbd/ASR/conformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/swbd/ASR/conformer_ctc/asr_datamodule.py b/egs/swbd/ASR/conformer_ctc/asr_datamodule.py
new file mode 100644
index 000000000..ce8634a1d
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/asr_datamodule.py
@@ -0,0 +1,416 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+# Modified by Zengrui Jin for the SwitchBoard corpus
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class SwitchBoardAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train dataloader,
+    but there can be multiple test dataloaders (e.g. SwitchBoard rt03
+    and eval2000).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=50,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+                buffer_size=50000,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_all_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "swbd_cuts_all.jsonl.gz"
+        ).subset(last=166844)
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "swbd_cuts_all.jsonl.gz"
+        ).subset(first=300)
+
+    @lru_cache()
+    def test_eval2000_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get eval2000 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "eval2000" / "eval2000_cuts_all.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_rt03_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get rt03 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "swbd_cuts_rt03.jsonl.gz")
diff --git a/egs/swbd/ASR/conformer_ctc/conformer.py b/egs/swbd/ASR/conformer_ctc/conformer.py
new file mode 120000
index 000000000..d1f4209d7
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/conformer.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/decode.py b/egs/swbd/ASR/conformer_ctc/decode.py
new file mode 100755
index 000000000..2bbade374
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/decode.py
@@ -0,0 +1,853 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+# Modified by Zengrui Jin for the SwitchBoard corpus
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import SwitchBoardAsrDataModule
+from conformer import Conformer
+
+from sclite_scoring import asr_text_post_processing
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=98,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<unk>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[int], List[int]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    if test_set_name == "test-eval2000":
+        subsets = {"callhome": "en_", "swbd": "sw_", "avg": "*"}
+    elif test_set_name == "test-rt03":
+        subsets = {"fisher": "fsh_", "swbd": "sw_", "avg": "*"}
+    else:
+        raise NotImplementedError(f"No implementation for testset {test_set_name}")
+    for subset, prefix in subsets.items():
+        test_set_wers = dict()
+        for key, results in results_dict.items():
+            recog_path = params.exp_dir / f"recogs-{test_set_name}-{subset}-{key}.txt"
+            results = post_processing(results)
+            results = (
+                sorted(list(filter(lambda x: x[0].startswith(prefix), results)))
+                if subset != "avg"
+                else sorted(results)
+            )
+            store_transcripts(filename=recog_path, texts=results)
+            if enable_log:
+                logging.info(f"The transcripts are stored in {recog_path}")
+
+            # The following prints out WERs, per-word error statistics and aligned
+            # ref/hyp pairs.
+            errs_filename = params.exp_dir / f"errs-{test_set_name}-{subset}-{key}.txt"
+            with open(errs_filename, "w") as f:
+                wer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{subset}-{key}",
+                    results,
+                    enable_log=enable_log,
+                    sclite_mode=True,
+                )
+                test_set_wers[key] = wer
+
+            if enable_log:
+                logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+        errs_info = params.exp_dir / f"wer-summary-{test_set_name}-{subset}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tWER", file=f)
+            for key, val in test_set_wers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}-{}, WER of different settings are:\n".format(
+            test_set_name, subset
+        )
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_wers:
+            s += "{}\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    SwitchBoardAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        model = load_averaged_model(
+            params.exp_dir, model, params.epoch, params.avg, device
+        )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        if params.rnn_lm_avg == 1:
+            load_checkpoint(
+                f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                rnn_lm_model,
+            )
+            rnn_lm_model.to(device)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    switchboard = SwitchBoardAsrDataModule(args)
+
+    test_eval2000_cuts = switchboard.test_eval2000_cuts().trim_to_supervisions(
+        keep_all_channels=True
+    )
+    # test_rt03_cuts = switchboard.test_rt03_cuts().trim_to_supervisions(
+    #     keep_all_channels=True
+    # )
+
+    test_eval2000_dl = switchboard.test_dataloaders(test_eval2000_cuts)
+    # test_rt03_dl = switchboard.test_dataloaders(test_rt03_cuts)
+
+    # test_sets = ["test-eval2000", "test-rt03"]
+    # test_dl = [test_eval2000_dl, test_rt03_dl]
+    test_sets = ["test-eval2000"]
+    test_dl = [test_eval2000_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/export.py b/egs/swbd/ASR/conformer_ctc/export.py
new file mode 100755
index 000000000..1bb6277ad
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/export.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from conformer import Conformer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=98,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""It contains language related input files such as "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            "num_decoder_layers": 6,
+        }
+    )
+    return params
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/label_smoothing.py b/egs/swbd/ASR/conformer_ctc/label_smoothing.py
new file mode 120000
index 000000000..e9d239fff
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/label_smoothing.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/pretrained.py b/egs/swbd/ASR/conformer_ctc/pretrained.py
new file mode 120000
index 000000000..526bc9678
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/pretrained.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/sclite_scoring.py b/egs/swbd/ASR/conformer_ctc/sclite_scoring.py
new file mode 100755
index 000000000..0383c4d71
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/sclite_scoring.py
@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+# Copyright 2021 Jiayu Du
+# Copyright 2022 Johns Hopkins University (Author: Guanbo Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import os
+
+conversational_filler = [
+    "UH",
+    "UHH",
+    "UM",
+    "EH",
+    "MM",
+    "HM",
+    "AH",
+    "HUH",
+    "HA",
+    "ER",
+    "OOF",
+    "HEE",
+    "ACH",
+    "EEE",
+    "EW",
+    "MHM",
+    "HUM",
+    "AW",
+    "OH",
+    "HMM",
+    "UMM",
+]
+unk_tags = ["<UNK>", "<unk>"]
+switchboard_garbage_utterance_tags = [
+    "[LAUGHTER]",
+    "[NOISE]",
+    "[VOCALIZED-NOISE]",
+    "[SILENCE]",
+]
+non_scoring_words = (
+    conversational_filler + unk_tags + switchboard_garbage_utterance_tags
+)
+
+
+def asr_text_post_processing(text: str) -> str:
+    # 1. convert to uppercase
+    text = text.upper()
+
+    # 2. remove non-scoring words from evaluation
+    remaining_words = []
+    text_split = text.split()
+    word_to_skip = 0
+    for idx, word in enumerate(text_split):
+        if word_to_skip > 0:
+            word_to_skip -= 1
+            continue
+        if word in non_scoring_words:
+            continue
+        elif word == "CANCELLED":
+            remaining_words.append("CANCELED")
+            continue
+        elif word == "AIRFLOW":
+            remaining_words.append("AIR")
+            remaining_words.append("FLOW")
+            continue
+        elif word == "PHD":
+            remaining_words.append("P")
+            remaining_words.append("H")
+            remaining_words.append("D")
+            continue
+        elif word == "UCLA":
+            remaining_words.append("U")
+            remaining_words.append("C")
+            remaining_words.append("L")
+            remaining_words.append("A")
+            continue
+        elif word == "ONTO":
+            remaining_words.append("ON")
+            remaining_words.append("TO")
+            continue
+        elif word == "DAY":
+            try:
+                if text_split[idx + 1] == "CARE":
+                    remaining_words.append("DAYCARE")
+                word_to_skip = 1
+            except:
+                remaining_words.append(word)
+            continue
+        remaining_words.append(word)
+
+    return " ".join(remaining_words)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="This script evaluates GigaSpeech ASR result via"
+        "SCTK's tool sclite"
+    )
+    parser.add_argument(
+        "ref",
+        type=str,
+        help="sclite's standard transcription(trn) reference file",
+    )
+    parser.add_argument(
+        "hyp",
+        type=str,
+        help="sclite's standard transcription(trn) hypothesis file",
+    )
+    parser.add_argument(
+        "work_dir",
+        type=str,
+        help="working dir",
+    )
+    args = parser.parse_args()
+
+    if not os.path.isdir(args.work_dir):
+        os.mkdir(args.work_dir)
+
+    REF = os.path.join(args.work_dir, "REF")
+    HYP = os.path.join(args.work_dir, "HYP")
+    RESULT = os.path.join(args.work_dir, "RESULT")
+
+    for io in [(args.ref, REF), (args.hyp, HYP)]:
+        with open(io[0], "r", encoding="utf8") as fi:
+            with open(io[1], "w+", encoding="utf8") as fo:
+                for line in fi:
+                    line = line.strip()
+                    if line:
+                        cols = line.split()
+                        text = asr_text_post_processing(" ".join(cols[0:-1]))
+                        uttid_field = cols[-1]
+                        print(f"{text} {uttid_field}", file=fo)
+
+    # GigaSpeech's uttid comforms to swb
+    os.system(f"sclite -r {REF} trn -h {HYP} trn -i swb | tee {RESULT}")
diff --git a/egs/swbd/ASR/conformer_ctc/subsampling.py b/egs/swbd/ASR/conformer_ctc/subsampling.py
new file mode 120000
index 000000000..16354dc73
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/test_label_smoothing.py b/egs/swbd/ASR/conformer_ctc/test_label_smoothing.py
new file mode 100755
index 000000000..5d4438fd1
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/test_label_smoothing.py
@@ -0,0 +1,52 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from distutils.version import LooseVersion
+
+import torch
+from label_smoothing import LabelSmoothingLoss
+
+torch_ver = LooseVersion(torch.__version__)
+
+
+def test_with_torch_label_smoothing_loss():
+    if torch_ver < LooseVersion("1.10.0"):
+        print(f"Current torch version: {torch_ver}")
+        print("Please use torch >= 1.10 to run this test - skipping")
+        return
+    torch.manual_seed(20211105)
+    x = torch.rand(20, 30, 5000)
+    tgt = torch.randint(low=-1, high=x.size(-1), size=x.shape[:2])
+    for reduction in ["none", "sum", "mean"]:
+        custom_loss_func = LabelSmoothingLoss(
+            ignore_index=-1, label_smoothing=0.1, reduction=reduction
+        )
+        custom_loss = custom_loss_func(x, tgt)
+
+        torch_loss_func = torch.nn.CrossEntropyLoss(
+            ignore_index=-1, reduction=reduction, label_smoothing=0.1
+        )
+        torch_loss = torch_loss_func(x.reshape(-1, x.size(-1)), tgt.reshape(-1))
+        assert torch.allclose(custom_loss, torch_loss)
+
+
+def main():
+    test_with_torch_label_smoothing_loss()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/test_subsampling.py b/egs/swbd/ASR/conformer_ctc/test_subsampling.py
new file mode 100755
index 000000000..81fa234dd
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/test_subsampling.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+
+def test_conv2d_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = Conv2dSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
+
+
+def test_vgg_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = VggSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
diff --git a/egs/swbd/ASR/conformer_ctc/test_transformer.py b/egs/swbd/ASR/conformer_ctc/test_transformer.py
new file mode 120000
index 000000000..8b0990ec6
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/test_transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/test_transformer.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/train.py b/egs/swbd/ASR/conformer_ctc/train.py
new file mode 100755
index 000000000..7f1eebbcf
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/train.py
@@ -0,0 +1,814 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+# Modified by Zengrui Jin for the SwitchBoard corpus
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./conformer_ctc/train.py \
+     --exp-dir ./conformer_ctc/exp \
+     --world-size 4 \
+     --max-duration 200 \
+     --num-epochs 20
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import SwitchBoardAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=98,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Normalization for the input features, can be a
+                              boolean indicating whether to do batch
+                              normalization, or a float which means just scaling
+                              the input features with this float value.
+                              If given a float value, we will remove batchnorm
+                              layer in `ConvolutionModule` as well.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    if isinstance(graph_compiler, BpeCtcTrainingGraphCompiler):
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+    elif isinstance(graph_compiler, CtcTrainingGraphCompiler):
+        # Works with a phone lexicon
+        decoding_graph = graph_compiler.compile(texts)
+    else:
+        raise ValueError(f"Unsupported type of graph compiler: {type(graph_compiler)}")
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    if "lang_bpe" in str(params.lang_dir):
+        graph_compiler = BpeCtcTrainingGraphCompiler(
+            params.lang_dir,
+            device=device,
+            sos_token="<sos/eos>",
+            eos_token="<sos/eos>",
+        )
+    elif "lang_phone" in str(params.lang_dir):
+        assert params.att_rate == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. Set --att-rate=0 "
+            "for pure CTC training when using a phone-based lang dir."
+        )
+        assert params.num_decoder_layers == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. "
+            "Set --num-decoder-layers=0 for pure CTC training when using "
+            "a phone-based lang dir."
+        )
+        graph_compiler = CtcTrainingGraphCompiler(
+            lexicon,
+            device=device,
+        )
+        # Manually add the sos/eos ID with their default values
+        # from the BPE recipe which we're adapting here.
+        graph_compiler.sos_id = 1
+        graph_compiler.eos_id = 1
+    else:
+        raise ValueError(
+            f"Unsupported type of lang dir (we expected it to have "
+            f"'lang_bpe' or 'lang_phone' in its name): {params.lang_dir}"
+        )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    switchboard = SwitchBoardAsrDataModule(args)
+
+    train_cuts = switchboard.train_all_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = switchboard.train_dataloaders(train_cuts)
+
+    valid_cuts = switchboard.dev_cuts()
+    valid_dl = switchboard.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    SwitchBoardAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/transformer.py b/egs/swbd/ASR/conformer_ctc/transformer.py
new file mode 120000
index 000000000..1c3f43fcf
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/transformer.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/compile_hlg.py b/egs/swbd/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/swbd/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/compile_lg.py b/egs/swbd/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/swbd/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/compute_fbank_eval2000.py b/egs/swbd/ASR/local/compute_fbank_eval2000.py
new file mode 100755
index 000000000..d446e8ff3
--- /dev/null
+++ b/egs/swbd/ASR/local/compute_fbank_eval2000.py
@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Modified     2023  The Chinese University of Hong Kong (author: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the SwitchBoard dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_switchboard(
+    dir_name: str,
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path(f"data/manifests/{dir_name}")
+    output_dir = Path(f"data/fbank/{dir_name}")
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("all",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = dir_name
+    suffix = "jsonl.gz"
+    manifests = {
+        "eval2000": "data/manifests/eval2000/eval2000_cuts_all_trimmed.jsonl.gz",
+    }
+    assert manifests is not None
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=16000))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        partition = "all"
+        cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+        print(cuts_filename)
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{prefix} already exists - skipping.")
+            return
+        logging.info(f"Processing {prefix}")
+        cut_set = CutSet.from_file(manifests[prefix]).resample(16000)
+
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+        cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+        cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_switchboard(
+        dir_name="eval2000",
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/swbd/ASR/local/compute_fbank_swbd.py b/egs/swbd/ASR/local/compute_fbank_swbd.py
new file mode 100755
index 000000000..dd82220c0
--- /dev/null
+++ b/egs/swbd/ASR/local/compute_fbank_swbd.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Modified     2023  The Chinese University of Hong Kong (author: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the SwitchBoard dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    parser.add_argument(
+        "--split-index",
+        type=int,
+        required=True,
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_switchboard(
+    dir_name: str,
+    split_index: int,
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path(f"data/manifests/{dir_name}")
+    output_dir = Path(f"data/fbank/{dir_name}_split16")
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("all",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = dir_name
+    suffix = "jsonl.gz"
+    split_dir = Path("data/manifests/swbd_split16/")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=16000))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        partition = "all"
+        cuts_filename = (
+            f"{prefix}_cuts_{partition}.{str(split_index).zfill(2)}.{suffix}"
+        )
+        print(cuts_filename)
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{prefix} already exists - skipping.")
+            return
+        logging.info(f"Processing {prefix}")
+        cut_set = (
+            CutSet.from_file(
+                split_dir
+                / f"swbd_train_all_trimmed.{str(split_index).zfill(2)}.jsonl.gz"
+            )
+            .resample(16000)
+            .to_eager()
+            .filter(lambda c: c.duration > 2.0)
+        )
+
+        if bpe_model:
+            cut_set = filter_cuts(cut_set, sp)
+        if perturb_speed:
+            logging.info(f"Doing speed perturb")
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}_{str(split_index).zfill(2)}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False,
+            min_duration=None,
+        )
+        cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+
+    compute_fbank_switchboard(
+        dir_name="swbd",
+        split_index=args.split_index,
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/swbd/ASR/local/convert_transcript_words_to_tokens.py b/egs/swbd/ASR/local/convert_transcript_words_to_tokens.py
new file mode 100755
index 000000000..a8d5117c9
--- /dev/null
+++ b/egs/swbd/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+"""
+Convert a transcript file containing words to a corpus file containing tokens
+for LM training with the help of a lexicon.
+
+If the lexicon contains phones, the resulting LM will be a phone LM; If the
+lexicon contains word pieces, the resulting LM will be a word piece LM.
+
+If a word has multiple pronunciations, the one that appears first in the lexicon
+is kept; others are removed.
+
+If the input transcript is:
+
+    hello zoo world hello
+    world zoo
+    foo zoo world hellO
+
+and if the lexicon is
+
+    <UNK> SPN
+    hello h e l l o 2
+    hello h e l l o
+    world w o r l d
+    zoo z o o
+
+Then the output is
+
+    h e l l o 2 z o o w o r l d h e l l o 2
+    w o r l d z o o
+    SPN z o o w o r l d SPN
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+from generate_unique_lexicon import filter_multiple_pronunications
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="The input transcript file."
+        "We assume that the transcript file consists of "
+        "lines. Each line consists of space separated words.",
+    )
+    parser.add_argument("--lexicon", type=str, help="The input lexicon file.")
+    parser.add_argument("--oov", type=str, default="<UNK>", help="The OOV word.")
+
+    return parser.parse_args()
+
+
+def process_line(lexicon: Dict[str, List[str]], line: str, oov_token: str) -> None:
+    """
+    Args:
+      lexicon:
+        A dict containing pronunciations. Its keys are words and values
+        are pronunciations (i.e., tokens).
+      line:
+        A line of transcript consisting of space(s) separated words.
+      oov_token:
+        The pronunciation of the oov word if a word in `line` is not present
+        in the lexicon.
+    Returns:
+      Return None.
+    """
+    s = ""
+    words = line.strip().split()
+    for i, w in enumerate(words):
+        tokens = lexicon.get(w, oov_token)
+        s += " ".join(tokens)
+        s += " "
+    print(s.strip())
+
+
+def main():
+    args = get_args()
+    assert Path(args.lexicon).is_file()
+    assert Path(args.transcript).is_file()
+    assert len(args.oov) > 0
+
+    # Only the first pronunciation of a word is kept
+    lexicon = filter_multiple_pronunications(read_lexicon(args.lexicon))
+
+    lexicon = dict(lexicon)
+
+    assert args.oov in lexicon
+
+    oov_token = lexicon[args.oov]
+
+    with open(args.transcript) as f:
+        for line in f:
+            process_line(lexicon=lexicon, line=line, oov_token=oov_token)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/dict.patch b/egs/swbd/ASR/local/dict.patch
new file mode 100644
index 000000000..12c63d612
--- /dev/null
+++ b/egs/swbd/ASR/local/dict.patch
@@ -0,0 +1,380 @@
+1d0
+<  file: $SWB/data/dictionary/sw-ms98-dict.text
+8645a8646
+> uh-hum ah m hh ah m
+9006c9007
+< April ey p r ih l
+---
+> April ey p r ax l
+9144d9144
+< B ay zh aa n iy z
+9261c9261
+< Battle b ae t el
+---
+> Battle b ae t ax l
+10014a10015
+> Chevy sh eh v iy
+10211a10213
+> Colorado k ao l ax r aa d ow
+10212a10215
+> Colorado' k ao l ax r aa d ow z
+10370c10373
+< Creek k r ih k
+---
+> Creek k r iy k
+10889a10893
+> Eleven ax l eh v ih n
+10951c10955
+< Erie ih r iy
+---
+> Erie iy r iy
+11183c11187
+< Forever f ax r eh v er
+---
+> Forever f er eh v er
+11231a11236
+> Friday f r ay d iy
+11744a11750
+> History hh ih s t r iy
+12004a12011,12012
+> Israel ih z r ih l
+> Israel's ih z r ih l z
+12573a12582
+> Lincoln l ih ng k ih n
+12574a12584
+> Lincolns l ih ng k ih n z
+13268c13278
+< NAACP eh ey ey s iy p iy
+---
+> NAACP eh n ey ey s iy p iy
+13286c13296
+< NIT eh ay t iy
+---
+> NIT eh n ay t iy
+13292c13302
+< NTSC eh t iy eh s s iy
+---
+> NTSC eh n t iy eh s s iy
+14058a14069
+> Quarter k ow r t er
+14059a14071
+> Quarterback k ow r t er b ae k
+14060a14073
+> Quarters k ow r t er z
+14569a14583
+> Science s ay n s
+15087a15102
+> Sunday s ah n d iy
+15088a15104
+> Sunday's s ah n d iy z
+15089a15106
+> Sundays s ah n d iy z
+15290,15291c15307,15308
+< Texan t eh k sh ih n
+< Texan's t eh k sh ih n s
+---
+> Texan t eh k s ih n
+> Texan's t eh k s ih n s
+15335a15353
+> Thousands th aw z ih n z
+15739c15757
+< Waco w ae k ow
+---
+> Waco w ey k ow
+15841a15860
+> Weekends w iy k eh n z
+16782a16802
+> acceptable eh k s eh p ax b ax l
+16833a16854
+> accounting ax k aw n ih ng
+16948a16970
+> address ax d r eh s
+17281a17304
+> already aa r d iy
+17315a17339
+> am m
+17709a17734
+> asked ae s t
+17847a17873
+> attorney ih t er n iy
+17919a17946
+> autopilot ao t ow p ay l ih t
+17960a17988
+> awfully ao f l iy
+18221a18250
+> basketball b ae s k ax b ao l
+18222a18252
+> basketball's b ae s k ax b ao l z
+18302a18333
+> become b ah k ah m
+18303a18335
+> becomes b iy k ah m z
+18344a18377
+> began b ax g en n
+18817c18850
+< bottle b aa t el
+---
+> bottle b aa t ax l
+19332,19333c19365,19367
+< camera's k ae m ax r ax z
+< cameras k ae m ax r ax z
+---
+> camera k ae m r ax
+> camera's k ae m r ax z
+> cameras k ae m r ax z
+19411a19446
+> capital k ae p ax l
+19505a19541
+> carrying k ae r ih ng
+20316a20353,20354
+> combination k aa m ih n ey sh ih n
+> combinations k aa m ih n ey sh ih n z
+20831a20870
+> contracts k aa n t r ae k s
+21010a21050
+> costs k ao s
+21062a21103
+> county k aw n iy
+21371a21413
+> cultural k ao l ch ax r ax l
+21372a21415
+> culturally k ao l ch ax r ax l iy
+21373a21417
+> culture k ao l ch er
+21375a21420
+> cultures k ao l ch er z
+21543a21589
+> data d ey t ax
+22097a22144
+> differently d ih f ax r ih n t l iy
+22972a23020
+> effects ax f eh k t s
+23016a23065
+> election ax l eh k sh ih n
+23018a23068
+> elections ax l eh k sh ih n z
+23052a23103
+> eleven ax l eh v ih n
+23242a23294
+> enjoyable ae n jh oy ax b ax l
+23248a23301
+> enjoys ae n jh oy z
+23293a23347
+> entire ih n t ay r
+23295a23350,23351
+> entirely ih n t ay r l iy
+> entirety ih n t ay r t iy
+23745a23802
+> extra eh k s t er
+23818a23876
+> facts f ae k s
+24508c24566
+< forever f ax r eh v er
+---
+> forever f er eh v er
+24514c24572
+< forget f ow r g eh t
+---
+> forget f er r g eh t
+24521a24580
+> forgot f er r g aa t
+24522a24582
+> forgotten f er r g aa t ax n
+24563a24624
+> forward f ow er d
+24680a24742
+> frightening f r ay t n ih ng
+24742a24805
+> full-time f ax l t ay m
+24862a24926
+> garage g r aa jh
+25218a25283
+> grandmother g r ae m ah dh er
+25790a25856
+> heavily hh eh v ax l iy
+25949a26016
+> history hh ih s t r iy
+26038a26106
+> honestly aa n ax s t l iy
+26039a26108
+> honesty aa n ax s t iy
+26099a26169
+> horror hh ow r
+26155a26226
+> houses hh aw z ih z
+26184c26255
+< huh-uh hh ah hh ah
+---
+> huh-uh ah hh ah
+26189c26260
+< hum-um hh m hh m
+---
+> hum-um ah m hh ah m
+26236a26308
+> hunting hh ah n ih ng
+26307a26380,26381
+> ideal ay d iy l
+> idealist ay d iy l ih s t
+26369a26444
+> imagine m ae jh ih n
+26628a26704
+> individuals ih n d ih v ih jh ax l z
+26968a27045
+> interest ih n t r ih s t
+27184a27262
+> it'd ih d
+27702a27781
+> lead l iy d
+28378a28458
+> mandatory m ae n d ih t ow r iy
+28885a28966
+> minute m ih n ih t
+29167a29249
+> mountains m aw t n z
+29317a29400
+> mysteries m ih s t r iy z
+29318a29402
+> mystery m ih s t r iy
+29470a29555
+> nervous n er v ih s
+29578,29580c29663,29665
+< nobody n ow b aa d iy
+< nobody'll n ow b aa d iy l
+< nobody's n ow b aa d iy z
+---
+> nobody n ow b ah d iy
+> nobody'll n ow b ah d iy l
+> nobody's n ow b ah d iy z
+29712a29798
+> nuclear n uw k l iy r
+29938a30025
+> onto aa n t ax
+30051a30139
+> originally ax r ih jh ax l iy
+30507a30596
+> particularly p er t ih k y ax l iy
+30755a30845
+> perfectly p er f ih k l iy
+30820a30911
+> personally p er s n ax l iy
+30915a31007
+> physically f ih z ih k l iy
+30986a31079
+> pilot p ay l ih t
+30987a31081
+> pilot's p ay l ih t s
+31227a31322
+> police p l iy s
+31513a31609
+> prefer p er f er
+31553a31650
+> prepare p r ax p ey r
+31578a31676
+> prescription p er s k r ih p sh ih n
+31579a31678
+> prescriptions p er s k r ih p sh ih n z
+31770a31870
+> products p r aa d ax k s
+31821a31922
+> projects p r aa jh eh k s
+31908a32010
+> protect p er t eh k t
+31909a32012
+> protected p er t eh k t ih d
+31911a32015
+> protection p er t eh k sh ih n
+31914a32019
+> protection p er t eh k t ih v
+32149a32255
+> quarter k ow r t er
+32414a32521
+> read r iy d
+32785a32893
+> rehabilitation r iy ax b ih l ih t ey sh ih n
+33150a33259
+> resource r ih s ow r s
+33151a33261
+> resources r iy s ow r s ih z
+33539c33649
+< roots r uh t s
+---
+> roots r uw t s
+33929a34040
+> science s ay n s
+34315a34427
+> seventy s eh v ih n iy
+34319,34320c34431,34432
+< severe s ax v iy r
+< severely s ax v iy r l iy
+---
+> severe s ih v iy r
+> severely s ih v iy r l iy
+35060a35173
+> software s ao f w ey r
+35083a35197
+> solid s ao l ih d
+35084a35199
+> solidly s ao l ih d l iy
+35750a35866
+> stood s t ih d
+35854a35971
+> strictly s t r ih k l iy
+35889c36006
+< stronger s t r ao ng er
+---
+> stronger s t r ao ng g er
+36192a36310,36311
+> supposed s p ow z
+> supposed s p ow s
+36510a36630
+> tastes t ey s
+36856a36977
+> thoroughly th er r l iy
+36866a36988
+> thousands th aw z ih n z
+37081c37203
+< toots t uh t s
+---
+> toots t uw t s
+37157a37280
+> toward t w ow r d
+37158a37282
+> towards t w ow r d z
+37564a37689
+> twenties t w eh n iy z
+37565a37691
+> twentieth t w eh n iy ih th
+37637a37764
+> unacceptable ah n ae k s eh p ax b ax l
+37728a37856
+> understand ah n d er s t ae n
+37860a37989
+> unless ih n l eh s
+38040a38170
+> use y uw z
+38049a38180
+> uses y uw z ih z
+38125a38257
+> various v ah r iy ih s
+38202a38335
+> versus v er s ih z
+38381c38514
+< wacko w ae k ow
+---
+> wacko w ey k ow
+38455c38588
+< wanna w aa n ax
+---
+> wanna w ah n ax
+38675c38808
+< whatnot w ah t n aa t
+---
+> whatnot w aa t n aa t
+38676a38810
+> whatsoever w aa t s ow eh v er
+38890c39024
+< wok w aa k
+---
+> wok w ao k
+38910a39045
+> wondering w ah n d r ih ng
diff --git a/egs/swbd/ASR/local/display_manifest_statistics.py b/egs/swbd/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..9aa204863
--- /dev/null
+++ b/egs/swbd/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    #  path = "./data/fbank/swbd_cuts_rt03.jsonl.gz"
+    path = "./data/fbank/eval2000/eval2000_cuts_all.jsonl.gz"
+    # path = "./data/fbank/swbd_cuts_all.jsonl.gz"
+
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Training Cut statistics:
+╒═══════════════════════════╤═══════════╕
+│ Cuts count:               │ 167244    │
+├───────────────────────────┼───────────┤
+│ Total duration (hh:mm:ss) │ 281:01:26 │
+├───────────────────────────┼───────────┤
+│ mean                      │ 6.0       │
+├───────────────────────────┼───────────┤
+│ std                       │ 3.3       │
+├───────────────────────────┼───────────┤
+│ min                       │ 2.0       │
+├───────────────────────────┼───────────┤
+│ 25%                       │ 3.2       │
+├───────────────────────────┼───────────┤
+│ 50%                       │ 5.2       │
+├───────────────────────────┼───────────┤
+│ 75%                       │ 8.3       │
+├───────────────────────────┼───────────┤
+│ 99%                       │ 14.4      │
+├───────────────────────────┼───────────┤
+│ 99.5%                     │ 14.7      │
+├───────────────────────────┼───────────┤
+│ 99.9%                     │ 15.0      │
+├───────────────────────────┼───────────┤
+│ max                       │ 57.5      │
+├───────────────────────────┼───────────┤
+│ Recordings available:     │ 167244    │
+├───────────────────────────┼───────────┤
+│ Features available:       │ 167244    │
+├───────────────────────────┼───────────┤
+│ Supervisions available:   │ 167244    │
+╘═══════════════════════════╧═══════════╛
+Speech duration statistics:
+╒══════════════════════════════╤═══════════╤══════════════════════╕
+│ Total speech duration        │ 281:01:26 │ 100.00% of recording │
+├──────────────────────────────┼───────────┼──────────────────────┤
+│ Total speaking time duration │ 281:01:26 │ 100.00% of recording │
+├──────────────────────────────┼───────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00  │ 0.00% of recording   │
+╘══════════════════════════════╧═══════════╧══════════════════════╛
+
+Eval2000 Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 4473     │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 03:37:13 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 2.9      │
+├───────────────────────────┼──────────┤
+│ std                       │ 2.6      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.1      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 1.2      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 2.1      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 4.0      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 12.6     │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 13.7     │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 14.7     │
+├───────────────────────────┼──────────┤
+│ max                       │ 15.5     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 4473     │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 4473     │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 4473     │
+╘═══════════════════════════╧══════════╛
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 03:37:13 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 03:37:13 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+"""
diff --git a/egs/swbd/ASR/local/extend_segments.pl b/egs/swbd/ASR/local/extend_segments.pl
new file mode 100755
index 000000000..e8b4894d5
--- /dev/null
+++ b/egs/swbd/ASR/local/extend_segments.pl
@@ -0,0 +1,99 @@
+#!/usr/bin/env perl
+use warnings; #sed replacement for -w perl parameter
+
+if (@ARGV != 1 || !($ARGV[0] =~ m/^-?\d+\.?\d*$/ &&  $ARGV[0] >= 0)) {
+  print STDERR "Usage: extend_segments.pl time-in-seconds <segments >segments.extended \n" .
+       "e.g. extend_segments.pl 0.25 <segments.1 >segments.2\n" .
+       "This command modifies a segments file, with lines like\n" .
+       " <utterance-id> <recording-id> <start-time> <end-time>\n" .
+       "by extending the beginning and end of each segment by a certain\n" .
+       "length of time.  This script makes sure the output segments do not\n" .
+       "overlap as a result of this time-extension, and that there are no\n" .
+       "negative times in the output.\n";
+  exit 1;
+}
+
+$extend = $ARGV[0];
+
+@all_lines = ();
+
+while (<STDIN>) {
+  chop;
+  @A = split(" ", $_);
+  if (@A != 4) {
+    die "invalid line in segments file: $_";
+  }
+  $line = @all_lines;  # current number of lines.
+  ($utt_id, $reco_id, $start_time, $end_time) = @A;
+
+  push @all_lines, [ $utt_id, $reco_id, $start_time, $end_time ]; # anonymous array.
+  if (! defined $lines_for_reco{$reco_id}) {
+    $lines_for_reco{$reco_id} = [ ];  # push new anonymous array.
+  }
+  push @{$lines_for_reco{$reco_id}}, $line;
+}
+
+foreach $reco_id (keys %lines_for_reco) {
+  $ref = $lines_for_reco{$reco_id};
+  @line_numbers = sort { ${$all_lines[$a]}[2] <=> ${$all_lines[$b]}[2] } @$ref;
+
+
+  {
+    # handle start of earliest segment as a special case.
+    $l0 = $line_numbers[0];
+    $tstart = ${$all_lines[$l0]}[2] - $extend;
+    if ($tstart < 0.0) { $tstart = 0.0; }
+    ${$all_lines[$l0]}[2] = $tstart;
+  }
+  {
+    # handle end of latest segment as a special case.
+    $lN = $line_numbers[$#line_numbers];
+    $tend = ${$all_lines[$lN]}[3] + $extend;
+    ${$all_lines[$lN]}[3] = $tend;
+  }
+  for ($i = 0; $i < $#line_numbers; $i++) {
+    $ln = $line_numbers[$i];
+    $ln1 = $line_numbers[$i+1];
+    $tend = ${$all_lines[$ln]}[3]; # end of earlier segment.
+    $tstart = ${$all_lines[$ln1]}[2]; # start of later segment.
+    if ($tend > $tstart) {
+      $utt1 = ${$all_lines[$ln]}[0];
+      $utt2 = ${$all_lines[$ln1]}[0];
+      print STDERR "Warning: for utterances $utt1 and $utt2, segments " .
+        "already overlap; leaving these times unchanged.\n";
+    } else {
+      $my_extend = $extend;
+      $max_extend =  0.5 * ($tstart - $tend);
+      if ($my_extend > $max_extend) { $my_extend = $max_extend; }
+      $tend += $my_extend;
+      $tstart -= $my_extend;
+      ${$all_lines[$ln]}[3] = $tend;
+      ${$all_lines[$ln1]}[2] = $tstart;
+    }
+  }
+}
+
+# leave the numbering of the lines unchanged.
+for ($l = 0; $l < @all_lines; $l++) {
+  $ref = $all_lines[$l];
+  ($utt_id, $reco_id, $start_time, $end_time) = @$ref;
+  printf("%s %s %.2f %.2f\n", $utt_id, $reco_id, $start_time, $end_time);
+}
+
+__END__
+
+# testing below.
+
+# ( echo a1 A 0 1; echo a2 A 3 4; echo b1 B 0 1; echo b2 B 2 3 ) | local/extend_segments.pl 1.0
+a1 A 0.00 2.00
+a2 A 2.00 5.00
+b1 B 0.00 1.50
+b2 B 1.50 4.00
+# ( echo a1 A 0 2; echo a2 A 1 3 ) | local/extend_segments.pl 1.0
+Warning: for utterances a1 and a2, segments already overlap; leaving these times unchanged.
+a1 A 0.00 2.00
+a2 A 1.00 4.00
+# ( echo a1 A 0 2; echo a2 A 5 6; echo a3 A 3 4 ) | local/extend_segments.pl 1.0
+a1 A 0.00 2.50
+a2 A 4.50 7.00
+a3 A 2.50 4.50
diff --git a/egs/swbd/ASR/local/filter_cuts.py b/egs/swbd/ASR/local/filter_cuts.py
new file mode 100755
index 000000000..fbcc9e24a
--- /dev/null
+++ b/egs/swbd/ASR/local/filter_cuts.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script removes short and long utterances from a cutset.
+
+Caution:
+  You may need to tune the thresholds for your own dataset.
+
+Usage example:
+
+  python3 ./local/filter_cuts.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --in-cuts data/fbank/librispeech_cuts_test-clean.jsonl.gz \
+    --out-cuts data/fbank-filtered/librispeech_cuts_test-clean.jsonl.gz
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=Path,
+        help="Path to the bpe.model",
+    )
+
+    parser.add_argument(
+        "--in-cuts",
+        type=Path,
+        help="Path to the input cutset",
+    )
+
+    parser.add_argument(
+        "--out-cuts",
+        type=Path,
+        help="Path to the output cutset",
+    )
+
+    return parser.parse_args()
+
+
+def filter_cuts(cut_set: CutSet, sp: spm.SentencePieceProcessor):
+    total = 0  # number of total utterances before removal
+    removed = 0  # number of removed utterances
+
+    def remove_short_and_long_utterances(c: Cut):
+        """Return False to exclude the input cut"""
+        nonlocal removed, total
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ./display_manifest_statistics.py
+        #
+        # You should use ./display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        total += 1
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            removed += 1
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./pruned_transducer_stateless2/conformer.py, the
+        # conv module uses the following expression
+        # for subsampling
+        if c.num_frames is None:
+            num_frames = c.duration * 100  # approximate
+        else:
+            num_frames = c.num_frames
+
+        T = ((num_frames - 1) // 2 - 1) // 2
+        # Note: for ./lstm_transducer_stateless/lstm.py, the formula is
+        #  T = ((num_frames - 3) // 2 - 1) // 2
+
+        # Note: for ./pruned_transducer_stateless7/zipformer.py, the formula is
+        # T = ((num_frames - 7) // 2 + 1) // 2
+
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            removed += 1
+            return False
+
+        return True
+
+    # We use to_eager() here so that we can print out the value of total
+    # and removed below.
+    ans = cut_set.filter(remove_short_and_long_utterances).to_eager()
+    ratio = removed / total * 100
+    logging.info(
+        f"Removed {removed} cuts from {total} cuts. {ratio:.3f}% data is removed."
+    )
+    return ans
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    if args.out_cuts.is_file():
+        logging.info(f"{args.out_cuts} already exists - skipping")
+        return
+
+    assert args.in_cuts.is_file(), f"{args.in_cuts} does not exist"
+    assert args.bpe_model.is_file(), f"{args.bpe_model} does not exist"
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    cut_set = load_manifest_lazy(args.in_cuts)
+    assert isinstance(cut_set, CutSet)
+
+    cut_set = filter_cuts(cut_set, sp)
+    logging.info(f"Saving to {args.out_cuts}")
+    args.out_cuts.parent.mkdir(parents=True, exist_ok=True)
+    cut_set.to_file(args.out_cuts)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/swbd/ASR/local/filter_empty_text.py b/egs/swbd/ASR/local/filter_empty_text.py
new file mode 100755
index 000000000..6b3316800
--- /dev/null
+++ b/egs/swbd/ASR/local/filter_empty_text.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+# Copyright    2023  The Chinese University of Hong Kong (author: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from pathlib import Path
+import logging
+from typing import List
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--kaldi-data-dir",
+        type=Path,
+        required=True,
+        help="Path to the kaldi data dir",
+    )
+
+    return parser.parse_args()
+
+
+def load_segments(path: Path):
+    segments = {}
+    with open(path, "r") as f:
+        lines = f.readlines()
+        for line in lines:
+            line = line.strip()
+            utt_id, rec_id, start, end = line.split()
+            segments[utt_id] = line
+    return segments
+
+
+def filter_text(path: Path):
+    with open(path, "r") as f:
+        lines = f.readlines()
+        return list(filter(lambda x: len(x.strip().split()) > 1, lines))
+
+
+def write_segments(path: Path, texts: List[str]):
+    with open(path, "w") as f:
+        f.writelines(texts)
+
+
+def main():
+    args = get_args()
+    orig_text_dict = filter_text(args.kaldi_data_dir / "text")
+    write_segments(args.kaldi_data_dir / "text", orig_text_dict)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
+
+    logging.info("Empty lines filtered")
diff --git a/egs/swbd/ASR/local/format_acronyms_dict.py b/egs/swbd/ASR/local/format_acronyms_dict.py
new file mode 100755
index 000000000..fa598dd03
--- /dev/null
+++ b/egs/swbd/ASR/local/format_acronyms_dict.py
@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+
+# Copyright 2015  Minhua Wu
+# Apache 2.0
+
+# convert acronyms in swbd dict to fisher convention
+# IBM to i._b._m.
+# BBC to b._b._c.
+# BBCs to b._b._c.s
+# BBC's to b._b._c.'s
+
+import argparse
+import re
+
+__author__ = "Minhua Wu"
+
+parser = argparse.ArgumentParser(description="format acronyms to a._b._c.")
+parser.add_argument("-i", "--input", help="Input lexicon", required=True)
+parser.add_argument("-o", "--output", help="Output lexicon", required=True)
+parser.add_argument(
+    "-L", "--Letter", help="Input single letter pronunciation", required=True
+)
+parser.add_argument("-M", "--Map", help="Output acronyms mapping", required=True)
+args = parser.parse_args()
+
+
+fin_lex = open(args.input, "r")
+fin_Letter = open(args.Letter, "r")
+fout_lex = open(args.output, "w")
+fout_map = open(args.Map, "w")
+
+# Initialise single letter dictionary
+dict_letter = {}
+for single_letter_lex in fin_Letter:
+    items = single_letter_lex.split()
+    dict_letter[items[0]] = single_letter_lex[len(items[0]) + 1 :].strip()
+fin_Letter.close()
+# print dict_letter
+
+for lex in fin_lex:
+    items = lex.split()
+    word = items[0]
+    lexicon = lex[len(items[0]) + 1 :].strip()
+    # find acronyms from words with only letters and '
+    pre_match = re.match(r"^[A-Za-z]+$|^[A-Za-z]+\'s$|^[A-Za-z]+s$", word)
+    if pre_match:
+        # find if words in the form of xxx's is acronym
+        if word[-2:] == "'s" and (lexicon[-1] == "s" or lexicon[-1] == "z"):
+            actual_word = word[:-2]
+            actual_lexicon = lexicon[:-2]
+            acronym_lexicon = ""
+            for w in actual_word:
+                acronym_lexicon = acronym_lexicon + dict_letter[w.upper()] + " "
+            if acronym_lexicon.strip() == actual_lexicon:
+                acronym_mapped = ""
+                acronym_mapped_back = ""
+                for w in actual_word[:-1]:
+                    acronym_mapped = acronym_mapped + w.lower() + "._"
+                    acronym_mapped_back = acronym_mapped_back + w.lower() + " "
+                acronym_mapped = acronym_mapped + actual_word[-1].lower() + ".'s"
+                acronym_mapped_back = (
+                    acronym_mapped_back + actual_word[-1].lower() + "'s"
+                )
+                fout_map.write(
+                    word + "\t" + acronym_mapped + "\t" + acronym_mapped_back + "\n"
+                )
+                fout_lex.write(acronym_mapped + " " + lexicon + "\n")
+            else:
+                fout_lex.write(lex)
+
+        # find if words in the form of xxxs is acronym
+        elif word[-1] == "s" and (lexicon[-1] == "s" or lexicon[-1] == "z"):
+            actual_word = word[:-1]
+            actual_lexicon = lexicon[:-2]
+            acronym_lexicon = ""
+            for w in actual_word:
+                acronym_lexicon = acronym_lexicon + dict_letter[w.upper()] + " "
+            if acronym_lexicon.strip() == actual_lexicon:
+                acronym_mapped = ""
+                acronym_mapped_back = ""
+                for w in actual_word[:-1]:
+                    acronym_mapped = acronym_mapped + w.lower() + "._"
+                    acronym_mapped_back = acronym_mapped_back + w.lower() + " "
+                acronym_mapped = acronym_mapped + actual_word[-1].lower() + ".s"
+                acronym_mapped_back = (
+                    acronym_mapped_back + actual_word[-1].lower() + "'s"
+                )
+                fout_map.write(
+                    word + "\t" + acronym_mapped + "\t" + acronym_mapped_back + "\n"
+                )
+                fout_lex.write(acronym_mapped + " " + lexicon + "\n")
+            else:
+                fout_lex.write(lex)
+
+        # find if words in the form of xxx (not ended with 's or s) is acronym
+        elif word.find("'") == -1 and word[-1] != "s":
+            acronym_lexicon = ""
+            for w in word:
+                acronym_lexicon = acronym_lexicon + dict_letter[w.upper()] + " "
+            if acronym_lexicon.strip() == lexicon:
+                acronym_mapped = ""
+                acronym_mapped_back = ""
+                for w in word[:-1]:
+                    acronym_mapped = acronym_mapped + w.lower() + "._"
+                    acronym_mapped_back = acronym_mapped_back + w.lower() + " "
+                acronym_mapped = acronym_mapped + word[-1].lower() + "."
+                acronym_mapped_back = acronym_mapped_back + word[-1].lower()
+                fout_map.write(
+                    word + "\t" + acronym_mapped + "\t" + acronym_mapped_back + "\n"
+                )
+                fout_lex.write(acronym_mapped + " " + lexicon + "\n")
+            else:
+                fout_lex.write(lex)
+        else:
+            fout_lex.write(lex)
+
+    else:
+        fout_lex.write(lex)
diff --git a/egs/swbd/ASR/local/generate_unique_lexicon.py b/egs/swbd/ASR/local/generate_unique_lexicon.py
new file mode 100755
index 000000000..3459c2f5a
--- /dev/null
+++ b/egs/swbd/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input a lexicon.txt and output a new lexicon,
+in which each word has a unique pronunciation.
+
+The way to do this is to keep only the first pronunciation of a word
+in lexicon.txt.
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        This file will generate a new file uniq_lexicon.txt
+        in it.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def filter_multiple_pronunications(
+    lexicon: List[Tuple[str, List[str]]]
+) -> List[Tuple[str, List[str]]]:
+    """Remove multiple pronunciations of words from a lexicon.
+
+    If a word has more than one pronunciation in the lexicon, only
+    the first one is kept, while other pronunciations are removed
+    from the lexicon.
+
+    Args:
+      lexicon:
+        The input lexicon, containing a list of (word, [p1, p2, ..., pn]),
+        where "p1, p2, ..., pn" are the pronunciations of the "word".
+    Returns:
+      Return a new lexicon where each word has a unique pronunciation.
+    """
+    seen = set()
+    ans = []
+
+    for word, tokens in lexicon:
+        if word in seen:
+            continue
+        seen.add(word)
+        ans.append((word, tokens))
+    return ans
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    lexicon_filename = lang_dir / "lexicon.txt"
+
+    in_lexicon = read_lexicon(lexicon_filename)
+
+    out_lexicon = filter_multiple_pronunications(in_lexicon)
+
+    write_lexicon(lang_dir / "uniq_lexicon.txt", out_lexicon)
+
+    logging.info(f"Number of entries in lexicon.txt: {len(in_lexicon)}")
+    logging.info(f"Number of entries in uniq_lexicon.txt: {len(out_lexicon)}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/swbd/ASR/local/map_acronyms_transcripts.py b/egs/swbd/ASR/local/map_acronyms_transcripts.py
new file mode 100755
index 000000000..ba02aaec3
--- /dev/null
+++ b/egs/swbd/ASR/local/map_acronyms_transcripts.py
@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+
+# Copyright 2015  Minhua Wu
+# Apache 2.0
+
+# convert acronyms in swbd transcript to fisher convention
+# according to first two columns in the input acronyms mapping
+
+import argparse
+import re
+
+__author__ = "Minhua Wu"
+
+parser = argparse.ArgumentParser(description="format acronyms to a._b._c.")
+parser.add_argument("-i", "--input", help="Input transcripts", required=True)
+parser.add_argument("-o", "--output", help="Output transcripts", required=True)
+parser.add_argument("-M", "--Map", help="Input acronyms mapping", required=True)
+args = parser.parse_args()
+
+fin_map = open(args.Map, "r")
+dict_acronym = {}
+dict_acronym_noi = {}  # Mapping of acronyms without I, i
+for pair in fin_map:
+    items = pair.split("\t")
+    dict_acronym[items[0]] = items[1]
+    dict_acronym_noi[items[0]] = items[1]
+fin_map.close()
+del dict_acronym_noi["I"]
+del dict_acronym_noi["i"]
+
+
+fin_trans = open(args.input, "r")
+fout_trans = open(args.output, "w")
+for line in fin_trans:
+    items = line.split()
+    L = len(items)
+    # First pass mapping to map I as part of acronym
+    for i in range(L):
+        if items[i] == "I":
+            x = 0
+            while i - 1 - x >= 0 and re.match(r"^[A-Z]$", items[i - 1 - x]):
+                x += 1
+
+            y = 0
+            while i + 1 + y < L and re.match(r"^[A-Z]$", items[i + 1 + y]):
+                y += 1
+
+            if x + y > 0:
+                for bias in range(-x, y + 1):
+                    items[i + bias] = dict_acronym[items[i + bias]]
+
+    # Second pass mapping (not mapping 'i' and 'I')
+    for i in range(len(items)):
+        if items[i] in dict_acronym_noi.keys():
+            items[i] = dict_acronym_noi[items[i]]
+    sentence = " ".join(items[1:])
+    fout_trans.write(items[0] + " " + sentence.lower() + "\n")
+
+fin_trans.close()
+fout_trans.close()
diff --git a/egs/swbd/ASR/local/normalize_and_filter_supervisions.py b/egs/swbd/ASR/local/normalize_and_filter_supervisions.py
new file mode 100755
index 000000000..20ab90caf
--- /dev/null
+++ b/egs/swbd/ASR/local/normalize_and_filter_supervisions.py
@@ -0,0 +1,283 @@
+#!/usr/bin/env python3
+# Copyright    2023      (authors: Nagendra Goel https://github.com/ngoel17)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import re
+from typing import Tuple
+
+from lhotse import SupervisionSegment, SupervisionSet
+from lhotse.serialization import load_manifest_lazy_or_eager
+from tqdm import tqdm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("input_sups")
+    parser.add_argument("output_sups")
+    return parser.parse_args()
+
+
+# replacement function to convert lowercase letter to uppercase
+def to_upper(match_obj):
+    if match_obj.group() is not None:
+        return match_obj.group().upper()
+
+
+def insert_groups_and_capitalize_3(match):
+    return f"{match.group(1)} {match.group(2)} {match.group(3)}".upper()
+
+
+def insert_groups_and_capitalize_2(match):
+    return f"{match.group(1)} {match.group(2)}".upper()
+
+
+def insert_groups_and_capitalize_1(match):
+    return f"{match.group(1)}".upper()
+
+
+def insert_groups_and_capitalize_1s(match):
+    return f"{match.group(1)}".upper() + "'s"
+
+
+class FisherSwbdNormalizer:
+    """Note: the functions "normalize" and "keep" implement the logic
+    similar to Kaldi's data prep scripts for Fisher and SWBD: One
+    notable difference is that we don't change [cough], [lipsmack],
+    etc. to [noise].  We also don't implement all the edge cases of
+    normalization from Kaldi (hopefully won't make too much
+    difference).
+    """
+
+    def __init__(self) -> None:
+        self.remove_regexp_before = re.compile(
+            r"|".join(
+                [
+                    # special symbols
+                    r"\[\[skip.*\]\]",
+                    r"\[skip.*\]",
+                    r"\[pause.*\]",
+                    r"\[silence\]",
+                    r"<b_aside>",
+                    r"<e_aside>",
+                    r"_1",
+                ]
+            )
+        )
+
+        # tuples of (pattern, replacement)
+        # note: Kaldi replaces sighs, coughs, etc with [noise].
+        #       We don't do that here.
+        #       We also lowercase the text as the first operation.
+        self.replace_regexps: Tuple[re.Pattern, str] = [
+            # SWBD:
+            # [LAUGHTER-STORY] -> STORY
+            (re.compile(r"\[laughter-(.*?)\]"), r"\1"),
+            # [WEA[SONABLE]-/REASONABLE]
+            (re.compile(r"\[\S+/(\S+)\]"), r"\1"),
+            # -[ADV]AN[TAGE]- -> AN
+            (re.compile(r"-?\[.*?\](\w+)\[.*?\]-?"), r"\1-"),
+            # ABSOLUTE[LY]- -> ABSOLUTE-
+            (re.compile(r"(\w+)\[.*?\]-?"), r"\1-"),
+            # [AN]Y- -> Y-
+            # -[AN]Y- -> Y-
+            (re.compile(r"-?\[.*?\](\w+)-?"), r"\1-"),
+            # special tokens
+            (re.compile(r"\[laugh.*?\]"), r"[laughter]"),
+            (re.compile(r"\[sigh.*?\]"), r"[sigh]"),
+            (re.compile(r"\[cough.*?\]"), r"[cough]"),
+            (re.compile(r"\[mn.*?\]"), r"[vocalized-noise]"),
+            (re.compile(r"\[breath.*?\]"), r"[breath]"),
+            (re.compile(r"\[lipsmack.*?\]"), r"[lipsmack]"),
+            (re.compile(r"\[sneeze.*?\]"), r"[sneeze]"),
+            # abbreviations
+            (
+                re.compile(
+                    r"(\w)\.(\w)\.(\w)",
+                ),
+                insert_groups_and_capitalize_3,
+            ),
+            (
+                re.compile(
+                    r"(\w)\.(\w)",
+                ),
+                insert_groups_and_capitalize_2,
+            ),
+            (
+                re.compile(
+                    r"([a-h,j-z])\.",
+                ),
+                insert_groups_and_capitalize_1,
+            ),
+            (
+                re.compile(
+                    r"\._",
+                ),
+                r" ",
+            ),
+            (
+                re.compile(
+                    r"_(\w)",
+                ),
+                insert_groups_and_capitalize_1,
+            ),
+            (
+                re.compile(
+                    r"(\w)\.s",
+                ),
+                insert_groups_and_capitalize_1s,
+            ),
+            (
+                re.compile(
+                    r"([A-Z])\'s",
+                ),
+                insert_groups_and_capitalize_1s,
+            ),
+            (
+                re.compile(
+                    r"(\s\w\b|^\w\b)",
+                ),
+                insert_groups_and_capitalize_1,
+            ),
+            # words between apostrophes
+            (re.compile(r"'(\S*?)'"), r"\1"),
+            # dangling dashes (2 passes)
+            (re.compile(r"\s-\s"), r" "),
+            (re.compile(r"\s-\s"), r" "),
+            # special symbol with trailing dash
+            (re.compile(r"(\[.*?\])-"), r"\1"),
+            # Just remove all dashes
+            (re.compile(r"-"), r" "),
+        ]
+
+        # unwanted symbols in the transcripts
+        self.remove_regexp_after = re.compile(
+            r"|".join(
+                [
+                    # remaining punctuation
+                    r"\.",
+                    r",",
+                    r"\?",
+                    r"{",
+                    r"}",
+                    r"~",
+                    r"_\d",
+                ]
+            )
+        )
+
+        self.post_fixes = [
+            # Fix an issue related to [VOCALIZED NOISE] after dash removal
+            (re.compile(r"\[vocalized noise\]"), "[vocalized-noise]"),
+        ]
+
+        self.whitespace_regexp = re.compile(r"\s+")
+
+    def normalize(self, text: str) -> str:
+        text = text.lower()
+
+        # first remove
+        text = self.remove_regexp_before.sub("", text)
+
+        # then replace
+        for pattern, sub in self.replace_regexps:
+            text = pattern.sub(sub, text)
+
+        # then remove
+        text = self.remove_regexp_after.sub("", text)
+
+        # post fixes
+        for pattern, sub in self.post_fixes:
+            text = pattern.sub(sub, text)
+
+        # then clean up whitespace
+        text = self.whitespace_regexp.sub(" ", text).strip()
+
+        return text.upper()
+
+
+def keep(sup: SupervisionSegment) -> bool:
+    if "((" in sup.text:
+        return False
+
+    if "<german" in sup.text:
+        return False
+
+    return True
+
+
+def main():
+    args = get_args()
+    sups = load_manifest_lazy_or_eager(args.input_sups)
+    assert isinstance(sups, SupervisionSet)
+
+    normalizer = FisherSwbdNormalizer()
+
+    tot, skip = 0, 0
+    with SupervisionSet.open_writer(args.output_sups) as writer:
+        for sup in tqdm(sups, desc="Normalizing supervisions"):
+            tot += 1
+
+            if not keep(sup):
+                skip += 1
+                continue
+
+            sup.text = normalizer.normalize(sup.text).upper()
+            if not sup.text:
+                skip += 1
+                continue
+
+            writer.write(sup)
+    print(f"tot: {tot}, skip: {skip}")
+
+
+def test():
+    normalizer = FisherSwbdNormalizer()
+    for text in [
+        "[laughterr] [SILENCE]",
+        "[laugh] oh this is great [silence] <B_ASIDE> yes",
+        "[laugh] oh this is [laught] this is great [silence] <B_ASIDE> yes",
+        "i don't kn- - know A.B.C's",
+        "so x. corp is good?",
+        "'absolutely yes",
+        "absolutely' yes",
+        "'absolutely' yes",
+        "'absolutely' yes 'aight",
+        "ABSOLUTE[LY]",
+        "ABSOLUTE[LY]-",
+        "[AN]Y",
+        "[AN]Y-",
+        "[ADV]AN[TAGE]",
+        "[ADV]AN[TAGE]-",
+        "-[ADV]AN[TAGE]",
+        "-[ADV]AN[TAGE]-",
+        "[WEA[SONABLE]-/REASONABLE]",
+        "[VOCALIZED-NOISE]-",
+        "~BULL",
+        "Frank E Peretti P E R E T T I",
+        "yeah yeah like Double O Seven he's supposed to do it",
+        "P A P E R paper",
+        "[noise] okay_1 um let me see [laughter] i've been sitting here awhile",
+    ]:
+        print(text)
+        print(normalizer.normalize(text))
+        print()
+
+
+if __name__ == "__main__":
+    test()
+    # exit()
+    main()
diff --git a/egs/swbd/ASR/local/normalize_eval2000.py b/egs/swbd/ASR/local/normalize_eval2000.py
new file mode 100755
index 000000000..7316193d0
--- /dev/null
+++ b/egs/swbd/ASR/local/normalize_eval2000.py
@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+# Copyright    2023      (authors: Nagendra Goel https://github.com/ngoel17)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import re
+from typing import Tuple
+
+from lhotse import SupervisionSegment, SupervisionSet
+from lhotse.serialization import load_manifest_lazy_or_eager
+from tqdm import tqdm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("input_sups")
+    parser.add_argument("output_sups")
+    return parser.parse_args()
+
+
+def remove_punctutation_and_other_symbol(text: str) -> str:
+    text = text.replace("--", " ")
+    text = text.replace("//", " ")
+    text = text.replace(".", " ")
+    text = text.replace("?", " ")
+    text = text.replace("~", " ")
+    text = text.replace(",", " ")
+    text = text.replace(";", " ")
+    text = text.replace("(", " ")
+    text = text.replace(")", " ")
+    text = text.replace("&", " ")
+    text = text.replace("%", " ")
+    text = text.replace("*", " ")
+    text = text.replace("{", " ")
+    text = text.replace("}", " ")
+    return text
+
+
+def eval2000_clean_eform(text: str, eform_count) -> str:
+    string_to_remove = []
+    piece = text.split('">')
+    for i in range(0, len(piece)):
+        s = piece[i] + '">'
+        res = re.search(r"<contraction e_form(.*?)\">", s)
+        if res is not None:
+            res_rm = res.group(1)
+            string_to_remove.append(res_rm)
+    for p in string_to_remove:
+        eform_string = p
+        text = text.replace(eform_string, " ")
+    eform_1 = "<contraction e_form"
+    text = text.replace(eform_1, " ")
+    eform_2 = '">'
+    text = text.replace(eform_2, " ")
+    # print("TEXT final: ", text)
+    return text
+
+
+def replace_silphone(text: str) -> str:
+    text = text.replace("[/BABY CRYING]", " ")
+    text = text.replace("[/CHILD]", " ")
+    text = text.replace("[[DISTORTED]]", " ")
+    text = text.replace("[/DISTORTION]", " ")
+    text = text.replace("[[DRAWN OUT]]", " ")
+    text = text.replace("[[DRAWN-OUT]]", " ")
+    text = text.replace("[[FAINT]]", " ")
+    text = text.replace("[SMACK]", " ")
+    text = text.replace("[[MUMBLES]]", " ")
+    text = text.replace("[[HIGH PITCHED SQUEAKY VOICE]]", " ")
+    text = text.replace("[[IN THE LAUGH]]", "[LAUGHTER]")
+    text = text.replace("[[LAST WORD SPOKEN WITH A LAUGH]]", "[LAUGHTER]")
+    text = text.replace("[[PART OF FIRST SYLLABLE OF PREVIOUS WORD CUT OFF]]", " ")
+    text = text.replace("[[PREVIOUS WORD SPOKEN WITH A LAUGH]]", " ")
+    text = text.replace("[[PREVIOUS TWO WORDS SPOKEN WHILE LAUGHING]]", " ")
+    text = text.replace("[[PROLONGED]]", " ")
+    text = text.replace("[/RUNNING WATER]", " ")
+    text = text.replace("[[SAYS LAUGHING]]", "[LAUGHTER]")
+    text = text.replace("[[SINGING]]", " ")
+    text = text.replace("[[SPOKEN WHILE LAUGHING]]", "[LAUGHTER]")
+    text = text.replace("[/STATIC]", " ")
+    text = text.replace("['THIRTIETH' DRAWN OUT]", " ")
+    text = text.replace("[/VOICES]", " ")
+    text = text.replace("[[WHISPERED]]", " ")
+    text = text.replace("[DISTORTION]", " ")
+    text = text.replace("[DISTORTION, HIGH VOLUME ON WAVES]", " ")
+    text = text.replace("[BACKGROUND LAUGHTER]", "[LAUGHTER]")
+    text = text.replace("[CHILD'S VOICE]", " ")
+    text = text.replace("[CHILD SCREAMS]", " ")
+    text = text.replace("[CHILD VOICE]", " ")
+    text = text.replace("[CHILD YELLING]", " ")
+    text = text.replace("[CHILD SCREAMING]", " ")
+    text = text.replace("[CHILD'S VOICE IN BACKGROUND]", " ")
+    text = text.replace("[CHANNEL NOISE]", " ")
+    text = text.replace("[CHANNEL ECHO]", " ")
+    text = text.replace("[ECHO FROM OTHER CHANNEL]", " ")
+    text = text.replace("[ECHO OF OTHER CHANNEL]", " ")
+    text = text.replace("[CLICK]", " ")
+    text = text.replace("[DISTORTED]", " ")
+    text = text.replace("[BABY CRYING]", " ")
+    text = text.replace("[METALLIC KNOCKING SOUND]", " ")
+    text = text.replace("[METALLIC SOUND]", " ")
+
+    text = text.replace("[PHONE JIGGLING]", " ")
+    text = text.replace("[BACKGROUND SOUND]", " ")
+    text = text.replace("[BACKGROUND VOICE]", " ")
+    text = text.replace("[BACKGROUND VOICES]", " ")
+    text = text.replace("[BACKGROUND NOISE]", " ")
+    text = text.replace("[CAR HORNS IN BACKGROUND]", " ")
+    text = text.replace("[CAR HORNS]", " ")
+    text = text.replace("[CARNATING]", " ")
+    text = text.replace("[CRYING CHILD]", " ")
+    text = text.replace("[CHOPPING SOUND]", " ")
+    text = text.replace("[BANGING]", " ")
+    text = text.replace("[CLICKING NOISE]", " ")
+    text = text.replace("[CLATTERING]", " ")
+    text = text.replace("[ECHO]", " ")
+    text = text.replace("[KNOCK]", " ")
+    text = text.replace("[NOISE-GOOD]", "[NOISE]")
+    text = text.replace("[RIGHT]", " ")
+    text = text.replace("[SOUND]", " ")
+    text = text.replace("[SQUEAK]", " ")
+    text = text.replace("[STATIC]", " ")
+    text = text.replace("[[SAYS WITH HIGH-PITCHED SCREAMING LAUGHTER]]", " ")
+    text = text.replace("[UH]", "UH")
+    text = text.replace("[MN]", "[VOCALIZED-NOISE]")
+    text = text.replace("[VOICES]", " ")
+    text = text.replace("[WATER RUNNING]", " ")
+    text = text.replace("[SOUND OF TWISTING PHONE CORD]", " ")
+    text = text.replace("[SOUND OF SOMETHING FALLING]", " ")
+    text = text.replace("[SOUND]", " ")
+    text = text.replace("[NOISE OF MOVING PHONE]", " ")
+    text = text.replace("[SOUND OF RUNNING WATER]", " ")
+    text = text.replace("[CHANNEL]", " ")
+    text = text.replace("[SILENCE]", " ")
+    text = text.replace("-[W]HERE", "WHERE")
+    text = text.replace("Y[OU]I-", "YOU I")
+    text = text.replace("-[A]ND", "AND")
+    text = text.replace("JU[ST]", "JUST")
+    text = text.replace("{BREATH}", " ")
+    text = text.replace("{BREATHY}", " ")
+    text = text.replace("{CHANNEL NOISE}", " ")
+    text = text.replace("{CLEAR THROAT}", " ")
+
+    text = text.replace("{CLEARING THROAT}", " ")
+    text = text.replace("{CLEARS THROAT}", " ")
+    text = text.replace("{COUGH}", " ")
+    text = text.replace("{DRAWN OUT}", " ")
+    text = text.replace("{EXHALATION}", " ")
+    text = text.replace("{EXHALE}", " ")
+    text = text.replace("{GASP}", " ")
+    text = text.replace("{HIGH SQUEAL}", " ")
+    text = text.replace("{INHALE}", " ")
+    text = text.replace("{LAUGH}", "[LAUGHTER]")
+    text = text.replace("{LAUGH}", "[LAUGHTER]")
+    text = text.replace("{LAUGH}", "[LAUGHTER]")
+    text = text.replace("{LIPSMACK}", " ")
+    text = text.replace("{LIPSMACK}", " ")
+
+    text = text.replace("{NOISE OF DISGUST}", " ")
+    text = text.replace("{SIGH}", " ")
+    text = text.replace("{SNIFF}", " ")
+    text = text.replace("{SNORT}", " ")
+    text = text.replace("{SHARP EXHALATION}", " ")
+    text = text.replace("{BREATH LAUGH}", " ")
+
+    text = text.replace("[LAUGHTER]", " ")
+    text = text.replace("[NOISE]", " ")
+    text = text.replace("[VOCALIZED-NOISE]", " ")
+    text = text.replace("-", " ")
+    return text
+
+
+def remove_languagetag(text: str) -> str:
+    langtag = re.findall(r"<(.*?)>", text)
+    for t in langtag:
+        text = text.replace(t, " ")
+    text = text.replace("<", " ")
+    text = text.replace(">", " ")
+    return text
+
+
+def eval2000_normalizer(text: str) -> str:
+    # print("TEXT original: ",text)
+    eform_count = text.count("contraction e_form")
+    # print("eform corunt:", eform_count)
+    if eform_count > 0:
+        text = eval2000_clean_eform(text, eform_count)
+    text = text.upper()
+    text = remove_languagetag(text)
+    text = replace_silphone(text)
+    text = remove_punctutation_and_other_symbol(text)
+    text = text.replace("IGNORE_TIME_SEGMENT_IN_SCORING", " ")
+    text = text.replace("IGNORE_TIME_SEGMENT_SCORING", " ")
+    spaces = re.findall(r"\s+", text)
+    for sp in spaces:
+        text = text.replace(sp, " ")
+    text = text.strip()
+    # text = self.whitespace_regexp.sub(" ", text).strip()
+    # print(text)
+    return text
+
+
+def main():
+    args = get_args()
+    sups = load_manifest_lazy_or_eager(args.input_sups)
+    assert isinstance(sups, SupervisionSet)
+
+    tot, skip = 0, 0
+    with SupervisionSet.open_writer(args.output_sups) as writer:
+        for sup in tqdm(sups, desc="Normalizing supervisions"):
+            tot += 1
+            sup.text = eval2000_normalizer(sup.text)
+            if not sup.text:
+                skip += 1
+                continue
+            writer.write(sup)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/prepare_lang.py b/egs/swbd/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/swbd/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/prepare_lang_bpe.py b/egs/swbd/ASR/local/prepare_lang_bpe.py
new file mode 100755
index 000000000..d82a085ec
--- /dev/null
+++ b/egs/swbd/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1,274 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.utils import str2bool
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_lexicon(
+    model_file: str, words: List[str], oov: str
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    words_pieces_ids: List[List[int]] = sp.encode(words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
+
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bpe.model"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = [
+        "<eps>",
+        "!SIL",
+        "<SPOKEN_NOISE>",
+        args.oov,
+        "#0",
+        "<s>",
+        "</s>",
+    ]
+
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_lexicon(model_file, words, args.oov)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/prepare_lm_training_data.py b/egs/swbd/ASR/local/prepare_lm_training_data.py
new file mode 120000
index 000000000..abc00d421
--- /dev/null
+++ b/egs/swbd/ASR/local/prepare_lm_training_data.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/rt03_data_prep.sh b/egs/swbd/ASR/local/rt03_data_prep.sh
new file mode 100755
index 000000000..8a5f64324
--- /dev/null
+++ b/egs/swbd/ASR/local/rt03_data_prep.sh
@@ -0,0 +1,107 @@
+#!/usr/bin/env bash
+
+# RT-03 data preparation (conversational telephone speech part only)
+# Adapted from Arnab Ghoshal's script for Hub-5 Eval 2000 by Peng Qi
+
+# To be run from one directory above this script.
+
+# Expects the standard directory layout for RT-03
+
+if [ $# -ne 1 ]; then
+    echo "Usage: $0 <rt03-dir>"
+    echo "e.g.: $0 /export/corpora/LDC/LDC2007S10"
+    echo "See comments in the script for more details"
+    exit 1
+fi
+
+sdir=$1
+[ ! -d $sdir/data/audio/eval03/english/cts ] &&
+    echo Expecting directory $sdir/data/audio/eval03/english/cts to be present && exit 1
+[ ! -d $sdir/data/references/eval03/english/cts ] &&
+    echo Expecting directory $tdir/data/references/eval03/english/cts to be present && exit 1
+
+dir=data/local/rt03
+mkdir -p $dir
+
+rtroot=$sdir
+tdir=$sdir/data/references/eval03/english/cts
+sdir=$sdir/data/audio/eval03/english/cts
+
+find -L $sdir -iname '*.sph' | sort >$dir/sph.flist
+sed -e 's?.*/??' -e 's?.sph??' $dir/sph.flist | paste - $dir/sph.flist \
+    >$dir/sph.scp
+
+sph2pipe=sph2pipe
+! command -v "${sph2pipe}" &>/dev/null &&
+    echo "Could not execute the sph2pipe program at $sph2pipe" && exit 1
+
+awk -v sph2pipe=$sph2pipe '{
+    printf("%s-A %s -f wav -p -c 1 %s |\n", $1, sph2pipe, $2);
+    printf("%s-B %s -f wav -p -c 2 %s |\n", $1, sph2pipe, $2);
+}' <$dir/sph.scp | sort >$dir/wav.scp || exit 1
+#side A - channel 1, side B - channel 2
+
+# Get segments file...
+# segments file format is: utt-id side-id start-time end-time, e.g.:
+# sw02001-A_000098-001156 sw02001-A 0.98 11.56
+#pem=$sdir/english/hub5e_00.pem
+#[ ! -f $pem ] && echo "No such file $pem" && exit 1;
+# pem file has lines like:
+# en_4156 A unknown_speaker 301.85 302.48
+
+#grep -v ';;' $pem \
+cat $tdir/*.stm | grep -v ';;' | grep -v inter_segment_gap |
+    awk '{
+    spk=$1"-"(($2==1)?"A":"B");
+    utt=sprintf("%s_%06d-%06d",spk,$4*100,$5*100);
+    print utt,spk,$4,$5;}' |
+    sort -u >$dir/segments
+
+# stm file has lines like:
+# en_4156 A en_4156_A 357.64 359.64 <O,en,F,en-F>  HE IS A POLICE OFFICER
+# TODO(arnab): We should really be lowercasing this since the Edinburgh
+# recipe uses lowercase. This is not used in the actual scoring.
+#grep -v ';;' $tdir/reference/hub5e00.english.000405.stm \
+cat $tdir/*.stm | grep -v ';;' | grep -v inter_segment_gap |
+    awk '{
+    spk=$1"-"(($2==1)?"A":"B");
+    utt=sprintf("%s_%06d-%06d",spk,$4*100,$5*100);
+    printf utt; for(n=7;n<=NF;n++) printf(" %s", $n); print ""; }' |
+    sort >$dir/text.all
+
+# We'll use the stm file for sclite scoring.  There seem to be various errors
+# in the stm file that upset hubscr.pl, and we fix them here.
+cat $tdir/*.stm |
+    sed -e 's:((:(:' -e 's:<B_ASIDE>::g' -e 's:<E_ASIDE>::g' |
+    grep -v inter_segment_gap |
+    awk '{
+    printf $1; if ($1==";;") printf(" %s",$2); else printf(($2==1)?" A":" B"); for(n=3;n<=NF;n++) printf(" %s", $n); print ""; }' \
+        >$dir/stm
+#$tdir/reference/hub5e00.english.000405.stm >  $dir/stm
+cp $rtroot/data/trans_rules/en20030506.glm $dir/glm
+
+# next line uses command substitution
+# Just checking that the segments are the same in pem vs. stm.
+! cmp <(awk '{print $1}' $dir/text.all) <(awk '{print $1}' $dir/segments) &&
+    echo "Segments from pem file and stm file do not match." && exit 1
+
+grep -v IGNORE_TIME_SEGMENT_ $dir/text.all >$dir/text
+
+# create an utt2spk file that assumes each conversation side is
+# a separate speaker.
+awk '{print $1,$2;}' $dir/segments >$dir/utt2spk
+utils/utt2spk_to_spk2utt.pl $dir/utt2spk >$dir/spk2utt
+
+# cp $dir/segments $dir/segments.tmp
+# awk '{x=$3-0.05; if (x<0.0) x=0.0; y=$4+0.05; print $1, $2, x, y; }' \
+#   $dir/segments.tmp > $dir/segments
+
+awk '{print $1}' $dir/wav.scp |
+    perl -ane '$_ =~ m:^(\S+)-([AB])$: || die "bad label $_";
+    print "$1-$2 $1 $2\n"; ' \
+        >$dir/reco2file_and_channel || exit 1
+
+./utils/fix_data_dir.sh $dir
+
+echo Data preparation and formatting completed for RT-03
+echo "(but not MFCC extraction)"
diff --git a/egs/swbd/ASR/local/sort_lm_training_data.py b/egs/swbd/ASR/local/sort_lm_training_data.py
new file mode 100755
index 000000000..bed3856e4
--- /dev/null
+++ b/egs/swbd/ASR/local/sort_lm_training_data.py
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input the filename of LM training data
+generated by ./local/prepare_lm_training_data.py and sorts
+it by sentence length.
+
+Sentence length equals to the number of BPE tokens in a sentence.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import numpy as np
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--in-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/sorted_lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-statistics",
+        type=str,
+        help="Statistics about LM training data., data/bpe_500/statistics.txt",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    in_lm_data = Path(args.in_lm_data)
+    out_lm_data = Path(args.out_lm_data)
+    assert in_lm_data.is_file(), f"{in_lm_data}"
+    if out_lm_data.is_file():
+        logging.warning(f"{out_lm_data} exists - skipping")
+        return
+    data = torch.load(in_lm_data)
+    words2bpe = data["words"]
+    sentences = data["sentences"]
+    sentence_lengths = data["sentence_lengths"]
+
+    num_sentences = sentences.dim0
+    assert num_sentences == sentence_lengths.numel(), (
+        num_sentences,
+        sentence_lengths.numel(),
+    )
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+
+    sorted_sentences = sentences[indices.to(torch.int32)]
+    sorted_sentence_lengths = sentence_lengths[indices]
+
+    # Check that sentences are ordered by length
+    assert num_sentences == sorted_sentences.dim0, (
+        num_sentences,
+        sorted_sentences.dim0,
+    )
+
+    cur = None
+    for i in range(num_sentences):
+        word_ids = sorted_sentences[i]
+        token_ids = words2bpe[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+        if cur is not None:
+            assert cur >= token_ids.numel(), (cur, token_ids.numel())
+
+        cur = token_ids.numel()
+        assert cur == sorted_sentence_lengths[i]
+
+    data["sentences"] = sorted_sentences
+    data["sentence_lengths"] = sorted_sentence_lengths
+    torch.save(data, args.out_lm_data)
+    logging.info(f"Saved to {args.out_lm_data}")
+
+    statistics = Path(args.out_statistics)
+
+    # Write statistics
+    num_words = sorted_sentences.numel()
+    num_tokens = sentence_lengths.sum().item()
+    max_sentence_length = sentence_lengths[indices[0]]
+    min_sentence_length = sentence_lengths[indices[-1]]
+
+    step = 10
+    hist, bins = np.histogram(
+        sentence_lengths.numpy(),
+        bins=np.arange(1, max_sentence_length + step, step),
+    )
+
+    histogram = np.stack((bins[:-1], hist)).transpose()
+
+    with open(statistics, "w") as f:
+        f.write(f"num_sentences: {num_sentences}\n")
+        f.write(f"num_words: {num_words}\n")
+        f.write(f"num_tokens: {num_tokens}\n")
+        f.write(f"max_sentence_length: {max_sentence_length}\n")
+        f.write(f"min_sentence_length: {min_sentence_length}\n")
+        f.write("histogram:\n")
+        f.write("  bin  count  percent\n")
+        for row in histogram:
+            f.write(
+                f"{int(row[0]):>5} {int(row[1]):>5}   "
+                f"{100.*row[1]/num_sentences:.3f}%\n"
+            )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/swbd/ASR/local/swbd1_data_prep.sh b/egs/swbd/ASR/local/swbd1_data_prep.sh
new file mode 100755
index 000000000..159359491
--- /dev/null
+++ b/egs/swbd/ASR/local/swbd1_data_prep.sh
@@ -0,0 +1,128 @@
+#!/usr/bin/env bash
+
+# Switchboard-1 training data preparation customized for Edinburgh
+# Author:  Arnab Ghoshal (Jan 2013)
+
+# To be run from one directory above this script.
+
+## The input is some directory containing the switchboard-1 release 2
+## corpus (LDC97S62).  Note: we don't make many assumptions about how
+## you unpacked this.  We are just doing a "find" command to locate
+## the .sph files.
+
+## The second input is optional, which should point to a directory containing
+## Switchboard transcriptions/documentations (specifically, the conv.tab file).
+## If specified, the script will try to use the actual speaker PINs provided
+## with the corpus instead of the conversation side ID (Kaldi default). We
+## will be using "find" to locate this file so we don't make any assumptions
+## on the directory structure. (Peng Qi, Aug 2014)
+
+#check existing directories
+if [ $# != 1 -a $# != 2 ]; then
+    echo "Usage: swbd1_data_prep.sh /path/to/SWBD [/path/to/SWBD_DOC]"
+    exit 1
+fi
+
+SWBD_DIR=$1
+
+dir=data/local/train
+mkdir -p $dir
+
+# Audio data directory check
+if [ ! -d $SWBD_DIR ]; then
+    echo "Error: run.sh requires a directory argument"
+    exit 1
+fi
+
+sph2pipe=sph2pipe
+! command -v "${sph2pipe}" &>/dev/null &&
+    echo "Could not execute the sph2pipe program at $sph2pipe" && exit 1
+
+# Option A: SWBD dictionary file check
+[ ! -f ./swb_ms98_transcriptions/sw-ms98-dict.text ] &&
+    echo "SWBD dictionary file does not exist" && exit 1
+
+# find sph audio files
+find -L $SWBD_DIR -iname '*.sph' | sort >$dir/sph.flist
+
+n=$(cat $dir/sph.flist | wc -l)
+[ $n -ne 2435 ] && [ $n -ne 2438 ] &&
+    echo Warning: expected 2435 or 2438 data data files, found $n
+
+# (1a) Transcriptions preparation
+# make basic transcription file (add segments info)
+# **NOTE: In the default Kaldi recipe, everything is made uppercase, while we
+# make everything lowercase here. This is because we will be using SRILM which
+# can optionally make everything lowercase (but not uppercase) when mapping
+# LM vocabs.
+awk '{
+name=substr($1,1,6); gsub("^sw","sw0",name); side=substr($1,7,1);
+stime=$2; etime=$3;
+printf("%s-%s_%06.0f-%06.0f",
+name, side, int(100*stime+0.5), int(100*etime+0.5));
+for(i=4;i<=NF;i++) printf(" %s", $i); printf "\n"
+}' ./swb_ms98_transcriptions/*/*/*-trans.text >$dir/transcripts1.txt
+
+# test if trans. file is sorted
+export LC_ALL=C
+sort -c $dir/transcripts1.txt || exit 1 # check it's sorted.
+
+# Remove SILENCE, <B_ASIDE> and <E_ASIDE>.
+
+# Note: we have [NOISE], [VOCALIZED-NOISE], [LAUGHTER], [SILENCE].
+# removing [SILENCE], and the <B_ASIDE> and <E_ASIDE> markers that mark
+# speech to somone; we will give phones to the other three (NSN, SPN, LAU).
+# There will also be a silence phone, SIL.
+# **NOTE: modified the pattern matches to make them case insensitive
+cat $dir/transcripts1.txt |
+    perl -ane 's:\s\[SILENCE\](\s|$):$1:gi;
+    s/<B_ASIDE>//gi;
+    s/<E_ASIDE>//gi;
+    print;' |
+    awk '{if(NF > 1) { print; } } ' >$dir/transcripts2.txt
+
+# **NOTE: swbd1_map_words.pl has been modified to make the pattern matches
+# case insensitive
+local/swbd1_map_words.pl -f 2- $dir/transcripts2.txt >$dir/text
+
+# format acronyms in text
+python3 local/map_acronyms_transcripts.py -i $dir/text -o $dir/text_map \
+    -M data/local/dict_nosp/acronyms.map
+mv $dir/text_map $dir/text
+
+# (1c) Make segment files from transcript
+#segments file format is: utt-id side-id start-time end-time, e.g.:
+#sw02001-A_000098-001156 sw02001-A 0.98 11.56
+awk '{
+segment=$1;
+split(segment,S,"[_-]");
+side=S[2]; audioname=S[1]; startf=S[3]; endf=S[4];
+print segment " " audioname "-" side " " startf/100 " " endf/100
+}' <$dir/text >$dir/segments
+
+sed -e 's?.*/??' -e 's?.sph??' $dir/sph.flist | paste - $dir/sph.flist \
+    >$dir/sph.scp
+
+awk -v sph2pipe=$sph2pipe '{
+printf("%s-A %s -f wav -p -c 1 %s |\n", $1, sph2pipe, $2);
+printf("%s-B %s -f wav -p -c 2 %s |\n", $1, sph2pipe, $2);
+}' <$dir/sph.scp | sort >$dir/wav.scp || exit 1
+#side A - channel 1, side B - channel 2
+
+# this file reco2file_and_channel maps recording-id (e.g. sw02001-A)
+# to the file name sw02001 and the A, e.g.
+# sw02001-A  sw02001 A
+# In this case it's trivial, but in other corpora the information might
+# be less obvious.  Later it will be needed for ctm scoring.
+awk '{print $1}' $dir/wav.scp |
+    perl -ane '$_ =~ m:^(\S+)-([AB])$: || die "bad label $_";
+    print "$1-$2 $1 $2\n"; ' \
+        >$dir/reco2file_and_channel || exit 1
+
+awk '{spk=substr($1,1,9); print $1 " " spk}' $dir/segments >$dir/utt2spk ||
+    exit 1
+sort -k 2 $dir/utt2spk | utils/utt2spk_to_spk2utt.pl >$dir/spk2utt || exit 1
+
+echo Switchboard-1 data preparation succeeded.
+
+utils/fix_data_dir.sh data/local/train
diff --git a/egs/swbd/ASR/local/swbd1_map_words.pl b/egs/swbd/ASR/local/swbd1_map_words.pl
new file mode 100755
index 000000000..4fb8d4ffe
--- /dev/null
+++ b/egs/swbd/ASR/local/swbd1_map_words.pl
@@ -0,0 +1,52 @@
+#!/usr/bin/env perl
+
+# Modified from swbd_map_words.pl in Kaldi s5 recipe to make pattern
+# matches case-insensitive --Arnab (Jan 2013)
+
+if ($ARGV[0] eq "-f") {
+  shift @ARGV;
+  $field_spec = shift @ARGV;
+  if ($field_spec =~ m/^\d+$/) {
+    $field_begin = $field_spec - 1; $field_end = $field_spec - 1;
+  }
+  if ($field_spec =~ m/^(\d*)[-:](\d*)/) { # accept e.g. 1:10 as a courtesy (properly, 1-10)
+    if ($1 ne "") {
+      $field_begin = $1 - 1;    # Change to zero-based indexing.
+    }
+    if ($2 ne "") {
+      $field_end = $2 - 1;      # Change to zero-based indexing.
+    }
+  }
+  if (!defined $field_begin && !defined $field_end) {
+    die "Bad argument to -f option: $field_spec";
+  }
+}
+
+
+while (<>) {
+  @A = split(" ", $_);
+  for ($n = 0; $n < @A; $n++) {
+    $a = $A[$n];
+    if ( (!defined $field_begin || $n >= $field_begin)
+         && (!defined $field_end || $n <= $field_end)) {
+      # e.g. [LAUGHTER-STORY] -> STORY;
+      $a =~ s:(|\-)^\[LAUGHTER-(.+)\](|\-)$:$1$2$3:i;
+      # $1 and $3 relate to preserving trailing "-"
+      $a =~ s:^\[(.+)/.+\](|\-)$:$1$2:; # e.g. [IT'N/ISN'T] -> IT'N ... note,
+      # 1st part may include partial-word stuff, which we process further below,
+      # e.g. [LEM[GUINI]-/LINGUINI]
+      # the (|\_) at the end is to accept and preserve trailing -'s.
+      $a =~ s:^(|\-)\[[^][]+\](.+)$:-$2:; # e.g. -[AN]Y , note \047 is quote;
+      # let the leading - be optional on input, as sometimes omitted.
+      $a =~ s:^(.+)\[[^][]+\](|\-)$:$1-:; # e.g. AB[SOLUTE]- -> AB-;
+      # let the trailing - be optional on input, as sometimes omitted.
+      $a =~ s:([^][]+)\[.+\]$:$1:; # e.g. EX[SPECIALLY]-/ESPECIALLY] -> EX-
+      # which is a  mistake in the input.
+      $a =~ s:^\{(.+)\}$:$1:;                 # e.g. {YUPPIEDOM} -> YUPPIEDOM
+      $a =~ s:[A-Z]\[([^][])+\][A-Z]:$1-$3:i; # e.g. AMMU[N]IT- -> AMMU-IT-
+      $a =~ s:_\d$::;                         # e.g. THEM_1 -> THEM
+    }
+    $A[$n] = $a;
+  }
+  print join(" ", @A) . "\n";
+}
diff --git a/egs/swbd/ASR/local/swbd1_prepare_dict.sh b/egs/swbd/ASR/local/swbd1_prepare_dict.sh
new file mode 100755
index 000000000..eff5fb5f1
--- /dev/null
+++ b/egs/swbd/ASR/local/swbd1_prepare_dict.sh
@@ -0,0 +1,101 @@
+#!/usr/bin/env bash
+
+# Formatting the Mississippi State dictionary for use in Edinburgh. Differs
+# from the one in Kaldi s5 recipe in that it uses lower-case --Arnab (Jan 2013)
+
+# To be run from one directory above this script.
+
+#check existing directories
+[ $# != 0 ] && echo "Usage: local/swbd1_data_prep.sh" && exit 1
+
+srcdir=. # This is where we downloaded some stuff..
+dir=./data/local/dict_nosp
+mkdir -p $dir
+srcdict=$srcdir/swb_ms98_transcriptions/sw-ms98-dict.text
+
+# assume swbd_p1_data_prep.sh was done already.
+[ ! -f "$srcdict" ] && echo "$0: No such file $srcdict" && exit 1
+
+cp $srcdict $dir/lexicon0.txt || exit 1
+chmod a+w $dir/lexicon0.txt
+patch <local/dict.patch $dir/lexicon0.txt || exit 1
+
+#(2a) Dictionary preparation:
+# Pre-processing (remove comments)
+grep -v '^#' $dir/lexicon0.txt | awk 'NF>0' | sort >$dir/lexicon1.txt || exit 1
+
+cat $dir/lexicon1.txt | awk '{ for(n=2;n<=NF;n++){ phones[$n] = 1; }} END{for (p in phones) print p;}' |
+    grep -v sil >$dir/nonsilence_phones.txt || exit 1
+
+(
+    echo sil
+    echo spn
+    echo nsn
+    echo lau
+) >$dir/silence_phones.txt
+
+echo sil >$dir/optional_silence.txt
+
+# No "extra questions" in the input to this setup, as we don't
+# have stress or tone.
+echo -n >$dir/extra_questions.txt
+
+cp local/MSU_single_letter.txt $dir/
+# Add to the lexicon the silences, noises etc.
+# Add single letter lexicon
+# The original swbd lexicon does not have precise single letter lexicion
+# e.g. it does not have entry of W
+(
+    echo '!SIL SIL'
+    echo '[VOCALIZED-NOISE] spn'
+    echo '[NOISE] nsn'
+    echo '[LAUGHTER] lau'
+    echo '<UNK> spn'
+) |
+    cat - $dir/lexicon1.txt $dir/MSU_single_letter.txt >$dir/lexicon2.txt || exit 1
+
+# Map the words in the lexicon.  That is-- for each word in the lexicon, we map it
+# to a new written form.  The transformations we do are:
+# remove laughter markings, e.g.
+# [LAUGHTER-STORY] -> STORY
+# Remove partial-words, e.g.
+# -[40]1K W AH N K EY
+# becomes -1K
+# and
+# -[AN]Y IY
+# becomes
+# -Y
+# -[A]B[OUT]- B
+# becomes
+# -B-
+# Also, curly braces, which appear to be used for "nonstandard"
+# words or non-words, are removed, e.g.
+# {WOLMANIZED} W OW L M AX N AY Z D
+# -> WOLMANIZED
+# Also, mispronounced words, e.g.
+#  [YEAM/YEAH] Y AE M
+# are changed to just e.g. YEAM, i.e. the orthography
+# of the mispronounced version.
+# Note-- this is only really to be used in training.  The main practical
+# reason is to avoid having tons of disambiguation symbols, which
+# we otherwise would get because there are many partial words with
+# the same phone sequences (most problematic: S).
+# Also, map
+# THEM_1 EH M -> THEM
+# so that multiple pronunciations just have alternate entries
+# in the lexicon.
+
+local/swbd1_map_words.pl -f 1 $dir/lexicon2.txt | sort -u \
+    >$dir/lexicon3.txt || exit 1
+
+python3 local/format_acronyms_dict.py -i $dir/lexicon3.txt -o $dir/lexicon4.txt \
+    -L $dir/MSU_single_letter.txt -M $dir/acronyms_raw.map
+cat $dir/acronyms_raw.map | sort -u >$dir/acronyms.map
+
+(echo 'i ay') | cat - $dir/lexicon4.txt | tr '[A-Z]' '[a-z]' | sort -u >$dir/lexicon5.txt
+
+pushd $dir >&/dev/null
+ln -sf lexicon5.txt lexicon.txt # This is the final lexicon.
+popd >&/dev/null
+rm $dir/lexiconp.txt 2>/dev/null
+echo Prepared input dictionary and phone-sets for Switchboard phase 1.
diff --git a/egs/swbd/ASR/local/train_bpe_model.py b/egs/swbd/ASR/local/train_bpe_model.py
new file mode 100755
index 000000000..9b4e28635
--- /dev/null
+++ b/egs/swbd/ASR/local/train_bpe_model.py
@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    user_defined_symbols += ["[LAUGHTER]", "[NOISE]", "[VOCALIZED-NOISE]"]
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/validate_bpe_lexicon.py b/egs/swbd/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/swbd/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/prepare.sh b/egs/swbd/ASR/prepare.sh
new file mode 100755
index 000000000..6b6f4ff86
--- /dev/null
+++ b/egs/swbd/ASR/prepare.sh
@@ -0,0 +1,463 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. Most of them can't be downloaded automatically
+# as they are not publically available and require a license purchased
+# from the LDC.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=./download
+# swbd1_dir="/export/corpora3/LDC/LDC97S62"
+swbd1_dir=./download/LDC97S62/
+
+# eval2000_dir contains the following files and directories
+# downloaded from LDC website:
+#  - LDC2002S09
+#       - hub5e_00
+#  - LDC2002T43
+#       - reference
+eval2000_dir="/export/corpora2/LDC/eval2000"
+
+rt03_dir="/export/corpora/LDC/LDC2007S10"
+fisher_dir="/export/corpora3/LDC/LDC2004T19"
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+    # 5000
+    # 2000
+    1000
+    500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+    # This function is from espnet
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "swbd1_dir: $swbd1_dir"
+log "eval2000_dir: $eval2000_dir"
+log "rt03_dir: $rt03_dir"
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+    log "Stage 1: Prepare SwitchBoard manifest"
+    # We assume that you have downloaded the SwitchBoard corpus
+    # to respective dirs
+    mkdir -p data/manifests
+    if [ ! -e data/manifests/.swbd.done ]; then
+        lhotse prepare switchboard --absolute-paths 1 --omit-silence $swbd1_dir data/manifests/swbd
+        ./local/normalize_and_filter_supervisions.py \
+            data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+            data/manifests/swbd/swbd_supervisions_all_norm.jsonl.gz
+        mv data/manifests/swbd/swbd_supervisions_all.jsonl.gz data/manifests/swbd/swbd_supervisions_orig.jsonl.gz
+        mv data/manifests/swbd/swbd_supervisions_all_norm.jsonl.gz data/manifests/swbd/swbd_supervisions_all.jsonl.gz
+
+        lhotse cut simple \
+            -r data/manifests/swbd/swbd_recordings_all.jsonl.gz \
+            -s data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+            data/manifests/swbd/swbd_train_all.jsonl.gz
+        lhotse cut trim-to-supervisions \
+            --discard-overlapping \
+            --discard-extra-channels \
+            data/manifests/swbd/swbd_train_all.jsonl.gz \
+            data/manifests/swbd/swbd_train_all_trimmed.jsonl.gz
+
+        num_splits=16
+        mkdir -p data/manifests/swbd_split${num_splits}
+        lhotse split ${num_splits} \
+            data/manifests/swbd/swbd_train_all_trimmed.jsonl.gz \
+            data/manifests/swbd_split${num_splits}
+
+        lhotse prepare eval2000 --absolute-paths 1 $eval2000_dir data/manifests/eval2000
+        ./local/normalize_eval2000.py \
+            data/manifests/eval2000/eval2000_supervisions_unnorm.jsonl.gz \
+            data/manifests/eval2000/eval2000_supervisions_all.jsonl.gz
+        
+        lhotse cut simple \
+            -r data/manifests/eval2000/eval2000_recordings_all.jsonl.gz \
+            -s data/manifests/eval2000/eval2000_supervisions_all.jsonl.gz \
+            data/manifests/eval2000/eval2000_cuts_all.jsonl.gz
+
+        lhotse cut trim-to-supervisions \
+            --discard-overlapping \
+            --discard-extra-channels \
+            data/manifests/eval2000/eval2000_cuts_all.jsonl.gz \
+            data/manifests/eval2000/eval2000_cuts_all_trimmed.jsonl.gz
+
+        sed -e 's:((:(:' -e 's:<B_ASIDE>::g' -e 's:<E_ASIDE>::g' \
+            $eval2000_dir/LDC2002T43/reference/hub5e00.english.000405.stm >  data/manifests/eval2000/stm
+        cp $eval2000_dir/LDC2002T43/reference/en20000405_hub5.glm  $dir/glm
+
+        # ./local/rt03_data_prep.sh $rt03_dir
+
+        # normalize eval2000 and rt03 texts by
+        # 1) convert upper to lower
+        # 2) remove tags (%AH) (%HESITATION) (%UH)
+        # 3) remove <B_ASIDE> <E_ASIDE>
+        # 4) remove "(" or ")"
+        # for x in  rt03; do
+        #     cp data/local/${x}/text data/local/${x}/text.org
+        #     paste -d "" \
+        #         <(cut -f 1 -d" " data/local/${x}/text.org) \
+        #         <(awk '{$1=""; print tolower($0)}' data/local/${x}/text.org | perl -pe 's| \(\%.*\)||g' | perl -pe 's| \<.*\>||g' | sed -e "s/(//g" -e "s/)//g") |
+        #         sed -e 's/\s\+/ /g' >data/local/${x}/text
+        #     rm data/local/${x}/text.org
+        # done
+
+        # lhotse fix data/manifests_rt03/swbd_recordings_rt03.jsonl.gz data/manifests_rt03/swbd_supervisions_rt03.jsonl.gz data/manifests
+
+        touch data/manifests/.swbd.done
+    fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+    log "Stage 2: Prepare musan manifest"
+    # We assume that you have downloaded the musan corpus
+    # to $dl_dir/musan
+    mkdir -p data/manifests
+    if [ ! -e data/manifests/.musan.done ]; then
+        lhotse prepare musan $dl_dir/musan data/manifests
+        touch data/manifests/.musan.done
+    fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3 I: Compute fbank for SwitchBoard"
+    if [ ! -e data/fbank/.swbd.done ]; then
+        mkdir -p data/fbank/swbd_split${num_splits}/
+        for index in $(seq 1 16); do
+            ./local/compute_fbank_swbd.py --split-index ${index} &
+        done
+        wait
+        pieces=$(find data/fbank/swbd_split${num_splits} -name "swbd_cuts_all.*.jsonl.gz")
+        lhotse combine $pieces data/fbank/swbd_cuts_all.jsonl.gz
+        touch data/fbank/.swbd.done
+    fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3 II: Compute fbank for eval2000"
+    if [ ! -e data/fbank/.eval2000.done ]; then
+        mkdir -p data/fbank/eval2000/
+        ./local/compute_fbank_eval2000.py 
+        touch data/fbank/.eval2000.done
+    fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+    log "Stage 4: Compute fbank for musan"
+    mkdir -p data/fbank
+    if [ ! -e data/fbank/.musan.done ]; then
+        ./local/compute_fbank_musan.py
+        touch data/fbank/.musan.done
+    fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+    log "Stage 5: Prepare phone based lang"
+    lang_dir=data/lang_phone
+    mkdir -p $lang_dir
+
+    if ! which jq; then
+      echo "This script is intended to be used with jq but you have not installed jq
+      Note: in Linux, you can install jq with the following command:
+      1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+      2. chmod +x ./jq
+      3. cp jq /usr/bin" && exit 1
+    fi
+    if [ ! -f $lang_dir/text ] || [ ! -s $lang_dir/text ]; then
+        log "Prepare text."
+        gunzip -c data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+        | jq '.text' | sed 's/"//g'  > $lang_dir/text
+    fi
+
+    log "Prepare dict"
+    ./local/swbd1_prepare_dict.sh 
+    cut -f 2- -d" " $lang_dir/text >${lang_dir}/input.txt
+    # [noise] nsn
+    # !sil sil
+    # <unk> spn
+    cat data/local/dict_nosp/lexicon.txt | sed 's/-//g' | sed 's/\[vocalizednoise\]/\[vocalized-noise\]/g' |
+        sort | uniq >$lang_dir/lexicon_lower.txt
+
+    cat $lang_dir/lexicon_lower.txt | tr a-z A-Z > $lang_dir/lexicon.txt
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+        ./local/prepare_lang.py --lang-dir $lang_dir
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+        log "Converting L.pt to L.fst"
+        ./shared/convert-k2-to-openfst.py \
+            --olabels aux_labels \
+            $lang_dir/L.pt \
+            $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+        log "Converting L_disambig.pt to L_disambig.fst"
+        ./shared/convert-k2-to-openfst.py \
+            --olabels aux_labels \
+            $lang_dir/L_disambig.pt \
+            $lang_dir/L_disambig.fst
+    fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+    log "Stage 6: Prepare BPE based lang"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+        mkdir -p $lang_dir
+        # We reuse words.txt from phone based lexicon
+        # so that the two can share G.pt later.
+        cp data/lang_phone/words.txt $lang_dir
+
+        if [ ! -f $lang_dir/transcript_words.txt ]; then
+            log "Generate data for BPE training"
+
+            cat data/lang_phone/text | cut -d " " -f 2- >$lang_dir/transcript_words.txt
+        fi
+
+        if [ ! -f $lang_dir/bpe.model ]; then
+            ./local/train_bpe_model.py \
+                --lang-dir $lang_dir \
+                --vocab-size $vocab_size \
+                --transcript $lang_dir/transcript_words.txt
+        fi
+
+        if [ ! -f $lang_dir/L_disambig.pt ]; then
+            ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+            log "Validating $lang_dir/lexicon.txt"
+            ./local/validate_bpe_lexicon.py \
+                --lexicon $lang_dir/lexicon.txt \
+                --bpe-model $lang_dir/bpe.model
+        fi
+
+        if [ ! -f $lang_dir/L.fst ]; then
+            log "Converting L.pt to L.fst"
+            ./shared/convert-k2-to-openfst.py \
+                --olabels aux_labels \
+                $lang_dir/L.pt \
+                $lang_dir/L.fst
+        fi
+
+        if [ ! -f $lang_dir/L_disambig.fst ]; then
+            log "Converting L_disambig.pt to L_disambig.fst"
+            ./shared/convert-k2-to-openfst.py \
+                --olabels aux_labels \
+                $lang_dir/L_disambig.pt \
+                $lang_dir/L_disambig.fst
+        fi
+    done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+    log "Stage 7: Prepare bigram token-level P for MMI training"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+
+        if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+            ./local/convert_transcript_words_to_tokens.py \
+                --lexicon $lang_dir/lexicon.txt \
+                --transcript $lang_dir/transcript_words.txt \
+                --oov "<UNK>" \
+                >$lang_dir/transcript_tokens.txt
+        fi
+
+        if [ ! -f $lang_dir/P.arpa ]; then
+            ./shared/make_kn_lm.py \
+                -ngram-order 2 \
+                -text $lang_dir/transcript_tokens.txt \
+                -lm $lang_dir/P.arpa
+        fi
+
+        if [ ! -f $lang_dir/P.fst.txt ]; then
+            python3 -m kaldilm \
+                --read-symbol-table="$lang_dir/tokens.txt" \
+                --disambig-symbol='#0' \
+                --max-order=2 \
+                $lang_dir/P.arpa >$lang_dir/P.fst.txt
+        fi
+    done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+    log "Stage 8: Prepare G"
+    lang_dir=data/lang_phone
+    # We assume you have installed kaldilm, if not, please install
+    # it using: pip install kaldilm
+
+    mkdir -p data/lm
+    if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+        # It is used in building HLG
+        ./shared/make_kn_lm.py \
+            -ngram-order 3 \
+            -text ${lang_dir}/input.txt \
+            -lm data/lm/3-gram.arpa
+        python3 -m kaldilm \
+            --read-symbol-table="data/lang_phone/words.txt" \
+            --disambig-symbol='#0' \
+            --max-order=3 \
+            data/lm/3-gram.arpa >data/lm/G_3_gram.fst.txt
+    fi
+
+    if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+        # It is used for LM rescoring
+        ./shared/make_kn_lm.py \
+            -ngram-order 4 \
+            -text ${lang_dir}/input.txt \
+            -lm data/lm/4-gram.arpa
+        python3 -m kaldilm \
+            --read-symbol-table="data/lang_phone/words.txt" \
+            --disambig-symbol='#0' \
+            --max-order=4 \
+            data/lm/4-gram.arpa >data/lm/G_4_gram.fst.txt
+    fi
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+    log "Stage 9: Compile HLG"
+    ./local/compile_hlg.py --lang-dir data/lang_phone
+
+    # Note If ./local/compile_hlg.py throws OOM,
+    # please switch to the following command
+    #
+    # ./local/compile_hlg_using_openfst.py --lang-dir data/lang_phone
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+        ./local/compile_hlg.py --lang-dir $lang_dir
+
+        # Note If ./local/compile_hlg.py throws OOM,
+        # please switch to the following command
+        #
+        # ./local/compile_hlg_using_openfst.py --lang-dir $lang_dir
+    done
+fi
+
+# Compile LG for RNN-T fast_beam_search decoding
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+    log "Stage 10: Compile LG"
+    ./local/compile_lg.py --lang-dir data/lang_phone
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+        ./local/compile_lg.py --lang-dir $lang_dir
+    done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+    log "Stage 11: Generate LM training data"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        log "Processing vocab_size == ${vocab_size}"
+        lang_dir=data/lang_bpe_${vocab_size}
+        out_dir=data/lm_training_bpe_${vocab_size}
+        mkdir -p $out_dir
+
+        if [ ! -f $out_dir/train.txt ]; then
+              tail -n 250000 data/lang_phone/input.txt > $out_dir/train.txt
+        fi
+
+        ./local/prepare_lm_training_data.py \
+            --bpe-model $lang_dir/bpe.model \
+            --lm-data data/lang_phone/input.txt \
+            --lm-archive $out_dir/lm_data.pt
+    done
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+    log "Stage 12: Generate LM validation data"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        log "Processing vocab_size == ${vocab_size}"
+        out_dir=data/lm_training_bpe_${vocab_size}
+        mkdir -p $out_dir
+
+        if [ ! -f $out_dir/valid.txt ]; then
+            head -n 14332 data/lang_phone/input.txt > $out_dir/valid.txt
+        fi
+
+        lang_dir=data/lang_bpe_${vocab_size}
+        ./local/prepare_lm_training_data.py \
+            --bpe-model $lang_dir/bpe.model \
+            --lm-data $out_dir/valid.txt \
+            --lm-archive $out_dir/lm_data-valid.pt
+    done
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+    log "Stage 13: Generate LM test data"
+    testsets=(eval2000)
+
+    for testset in ${testsets[@]}; do
+        for vocab_size in ${vocab_sizes[@]}; do
+            log "Processing vocab_size == ${vocab_size}"
+            out_dir=data/lm_training_bpe_${vocab_size}
+            mkdir -p $out_dir
+
+            if [ ! -f $out_dir/${testset}.txt ]; then
+                gunzip -c data/manifests/${testset}/eval2000_supervisions_all.jsonl.gz \
+                    | jq '.text' | sed 's/"//g' > $out_dir/${testset}.txt
+            fi
+
+            lang_dir=data/lang_bpe_${vocab_size}
+            ./local/prepare_lm_training_data.py \
+                --bpe-model $lang_dir/bpe.model \
+                --lm-data $out_dir/${testset}.txt \
+                --lm-archive $out_dir/lm_data-${testset}.pt
+        done
+    done
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+    log "Stage 14: Sort LM training data"
+    testsets=(eval2000)
+    # Sort LM training data by sentence length in descending order
+    # for ease of training.
+    #
+    # Sentence length equals to the number of BPE tokens
+    # in a sentence.
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        out_dir=data/lm_training_bpe_${vocab_size}
+        mkdir -p $out_dir
+        ./local/sort_lm_training_data.py \
+            --in-lm-data $out_dir/lm_data.pt \
+            --out-lm-data $out_dir/sorted_lm_data.pt \
+            --out-statistics $out_dir/statistics.txt
+        for testset in ${testsets[@]}; do
+            ./local/sort_lm_training_data.py \
+                --in-lm-data $out_dir/lm_data-${testset}.pt \
+                --out-lm-data $out_dir/sorted_lm_data-${testset}.pt \
+                --out-statistics $out_dir/statistics-test-${testset}.txt
+        done
+    done
+fi
diff --git a/egs/swbd/ASR/shared b/egs/swbd/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/swbd/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/swbd/ASR/utils/filter_scp.pl b/egs/swbd/ASR/utils/filter_scp.pl
new file mode 100755
index 000000000..b76d37f41
--- /dev/null
+++ b/egs/swbd/ASR/utils/filter_scp.pl
@@ -0,0 +1,87 @@
+#!/usr/bin/env perl
+# Copyright 2010-2012 Microsoft Corporation
+#                     Johns Hopkins University (author: Daniel Povey)
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# This script takes a list of utterance-ids or any file whose first field
+# of each line is an utterance-id, and filters an scp
+# file (or any file whose "n-th" field is an utterance id), printing
+# out only those lines whose "n-th" field is in id_list. The index of
+# the "n-th" field is 1, by default, but can be changed by using
+# the -f <n> switch
+
+$exclude = 0;
+$field = 1;
+$shifted = 0;
+
+do {
+  $shifted=0;
+  if ($ARGV[0] eq "--exclude") {
+    $exclude = 1;
+    shift @ARGV;
+    $shifted=1;
+  }
+  if ($ARGV[0] eq "-f") {
+    $field = $ARGV[1];
+    shift @ARGV; shift @ARGV;
+    $shifted=1
+  }
+} while ($shifted);
+
+if(@ARGV < 1 || @ARGV > 2) {
+  die "Usage: filter_scp.pl [--exclude] [-f <field-to-filter-on>] id_list [in.scp] > out.scp \n" .
+      "Prints only the input lines whose f'th field (default: first) is in 'id_list'.\n" .
+      "Note: only the first field of each line in id_list matters.  With --exclude, prints\n" .
+      "only the lines that were *not* in id_list.\n" .
+      "Caution: previously, the -f option was interpreted as a zero-based field index.\n" .
+      "If your older scripts (written before Oct 2014) stopped working and you used the\n" .
+      "-f option, add 1 to the argument.\n" .
+      "See also: utils/filter_scp.pl .\n";
+}
+
+
+$idlist = shift @ARGV;
+open(F, "<$idlist") || die "Could not open id-list file $idlist";
+while(<F>) {
+  @A = split;
+  @A>=1 || die "Invalid id-list file line $_";
+  $seen{$A[0]} = 1;
+}
+
+if ($field == 1) { # Treat this as special case, since it is common.
+  while(<>) {
+    $_ =~ m/\s*(\S+)\s*/ || die "Bad line $_, could not get first field.";
+    # $1 is what we filter on.
+    if ((!$exclude && $seen{$1}) || ($exclude && !defined $seen{$1})) {
+      print $_;
+    }
+  }
+} else {
+  while(<>) {
+    @A = split;
+    @A > 0 || die "Invalid scp file line $_";
+    @A >= $field || die "Invalid scp file line $_";
+    if ((!$exclude && $seen{$A[$field-1]}) || ($exclude && !defined $seen{$A[$field-1]})) {
+      print $_;
+    }
+  }
+}
+
+# tests:
+# the following should print "foo 1"
+# ( echo foo 1; echo bar 2 ) | utils/filter_scp.pl <(echo foo)
+# the following should print "bar 2".
+# ( echo foo 1; echo bar 2 ) | utils/filter_scp.pl -f 2 <(echo 2)
diff --git a/egs/swbd/ASR/utils/fix_data_dir.sh b/egs/swbd/ASR/utils/fix_data_dir.sh
new file mode 100755
index 000000000..ca0972ca8
--- /dev/null
+++ b/egs/swbd/ASR/utils/fix_data_dir.sh
@@ -0,0 +1,197 @@
+#!/bin/bash
+
+# This script makes sure that only the segments present in
+# all of "feats.scp", "wav.scp" [if present], segments [if present]
+# text, and utt2spk are present in any of them.
+# It puts the original contents of data-dir into
+# data-dir/.backup
+
+cmd="$@"
+
+utt_extra_files=
+spk_extra_files=
+
+. utils/parse_options.sh
+
+if [ $# != 1 ]; then
+  echo "Usage: utils/data/fix_data_dir.sh <data-dir>"
+  echo "e.g.: utils/data/fix_data_dir.sh data/train"
+  echo "This script helps ensure that the various files in a data directory"
+  echo "are correctly sorted and filtered, for example removing utterances"
+  echo "that have no features (if feats.scp is present)"
+  exit 1
+fi
+
+data=$1
+
+if [ -f $data/images.scp ]; then
+  image/fix_data_dir.sh $cmd
+  exit $?
+fi
+
+mkdir -p $data/.backup
+
+[ ! -d $data ] && echo "$0: no such directory $data" && exit 1;
+
+[ ! -f $data/utt2spk ] && echo "$0: no such file $data/utt2spk" && exit 1;
+
+set -e -o pipefail -u
+
+tmpdir=$(mktemp -d /tmp/kaldi.XXXX);
+trap 'rm -rf "$tmpdir"' EXIT HUP INT PIPE TERM
+
+export LC_ALL=C
+
+function check_sorted {
+  file=$1
+  sort -k1,1 -u <$file >$file.tmp
+  if ! cmp -s $file $file.tmp; then
+    echo "$0: file $1 is not in sorted order or not unique, sorting it"
+    mv $file.tmp $file
+  else
+    rm $file.tmp
+  fi
+}
+
+for x in utt2spk spk2utt feats.scp text segments wav.scp cmvn.scp vad.scp \
+    reco2file_and_channel spk2gender utt2lang utt2uniq utt2dur reco2dur utt2num_frames; do
+  if [ -f $data/$x ]; then
+    cp $data/$x $data/.backup/$x
+    check_sorted $data/$x
+  fi
+done
+
+
+function filter_file {
+  filter=$1
+  file_to_filter=$2
+  cp $file_to_filter ${file_to_filter}.tmp
+  utils/filter_scp.pl $filter ${file_to_filter}.tmp > $file_to_filter
+  if ! cmp ${file_to_filter}.tmp  $file_to_filter >&/dev/null; then
+    length1=$(cat ${file_to_filter}.tmp | wc -l)
+    length2=$(cat ${file_to_filter} | wc -l)
+    if [ $length1 -ne $length2 ]; then
+      echo "$0: filtered $file_to_filter from $length1 to $length2 lines based on filter $filter."
+    fi
+  fi
+  rm $file_to_filter.tmp
+}
+
+function filter_recordings {
+  # We call this once before the stage when we filter on utterance-id, and once
+  # after.
+
+  if [ -f $data/segments ]; then
+  # We have a segments file -> we need to filter this and the file wav.scp, and
+  # reco2file_and_utt, if it exists, to make sure they have the same list of
+  # recording-ids.
+
+    if [ ! -f $data/wav.scp ]; then
+      echo "$0: $data/segments exists but not $data/wav.scp"
+      exit 1;
+    fi
+    awk '{print $2}' < $data/segments | sort | uniq > $tmpdir/recordings
+    n1=$(cat $tmpdir/recordings | wc -l)
+    [ ! -s $tmpdir/recordings ] && \
+      echo "Empty list of recordings (bad file $data/segments)?" && exit 1;
+    utils/filter_scp.pl $data/wav.scp $tmpdir/recordings > $tmpdir/recordings.tmp
+    mv $tmpdir/recordings.tmp $tmpdir/recordings
+
+
+    cp $data/segments{,.tmp}; awk '{print $2, $1, $3, $4}' <$data/segments.tmp >$data/segments
+    filter_file $tmpdir/recordings $data/segments
+    cp $data/segments{,.tmp}; awk '{print $2, $1, $3, $4}' <$data/segments.tmp >$data/segments
+    rm $data/segments.tmp
+
+    filter_file $tmpdir/recordings $data/wav.scp
+    [ -f $data/reco2file_and_channel ] && filter_file $tmpdir/recordings $data/reco2file_and_channel
+    [ -f $data/reco2dur ] && filter_file $tmpdir/recordings $data/reco2dur
+    true
+  fi
+}
+
+function filter_speakers {
+  # throughout this program, we regard utt2spk as primary and spk2utt as derived, so...
+  utils/utt2spk_to_spk2utt.pl $data/utt2spk > $data/spk2utt
+
+  cat $data/spk2utt | awk '{print $1}' > $tmpdir/speakers
+  for s in cmvn.scp spk2gender; do
+    f=$data/$s
+    if [ -f $f ]; then
+      filter_file $f $tmpdir/speakers
+    fi
+  done
+
+  filter_file $tmpdir/speakers $data/spk2utt
+  utils/spk2utt_to_utt2spk.pl $data/spk2utt > $data/utt2spk
+
+  for s in cmvn.scp spk2gender $spk_extra_files; do
+    f=$data/$s
+    if [ -f $f ]; then
+      filter_file $tmpdir/speakers $f
+    fi
+  done
+}
+
+function filter_utts {
+  cat $data/utt2spk | awk '{print $1}' > $tmpdir/utts
+
+  ! cat $data/utt2spk | sort | cmp - $data/utt2spk && \
+    echo "utt2spk is not in sorted order (fix this yourself)" && exit 1;
+
+  ! cat $data/utt2spk | sort -k2 | cmp - $data/utt2spk && \
+    echo "utt2spk is not in sorted order when sorted first on speaker-id " && \
+    echo "(fix this by making speaker-ids prefixes of utt-ids)" && exit 1;
+
+  ! cat $data/spk2utt | sort | cmp - $data/spk2utt && \
+    echo "spk2utt is not in sorted order (fix this yourself)" && exit 1;
+
+  if [ -f $data/utt2uniq ]; then
+    ! cat $data/utt2uniq | sort | cmp - $data/utt2uniq && \
+      echo "utt2uniq is not in sorted order (fix this yourself)" && exit 1;
+  fi
+
+  maybe_wav=
+  maybe_reco2dur=
+  [ ! -f $data/segments ] && maybe_wav=wav.scp # wav indexed by utts only if segments does not exist.
+  [ -s $data/reco2dur ] && [ ! -f $data/segments ] && maybe_reco2dur=reco2dur # reco2dur indexed by utts
+  for x in feats.scp text segments utt2lang $maybe_wav; do
+    if [ -f $data/$x ]; then
+      utils/filter_scp.pl $data/$x $tmpdir/utts > $tmpdir/utts.tmp
+      mv $tmpdir/utts.tmp $tmpdir/utts
+    fi
+  done
+  [ ! -s $tmpdir/utts ] && echo "fix_data_dir.sh: no utterances remained: not proceeding further." && \
+    rm $tmpdir/utts && exit 1;
+
+
+  if [ -f $data/utt2spk ]; then
+    new_nutts=$(cat $tmpdir/utts | wc -l)
+    old_nutts=$(cat $data/utt2spk | wc -l)
+    if [ $new_nutts -ne $old_nutts ]; then
+      echo "fix_data_dir.sh: kept $new_nutts utterances out of $old_nutts"
+    else
+      echo "fix_data_dir.sh: kept all $old_nutts utterances."
+    fi
+  fi
+
+  for x in utt2spk utt2uniq feats.scp vad.scp text segments utt2lang utt2dur utt2num_frames $maybe_wav $maybe_reco2dur $utt_extra_files; do
+    if [ -f $data/$x ]; then
+      cp $data/$x $data/.backup/$x
+      if ! cmp -s $data/$x <( utils/filter_scp.pl $tmpdir/utts $data/$x ) ; then
+        utils/filter_scp.pl $tmpdir/utts $data/.backup/$x > $data/$x
+      fi
+    fi
+  done
+
+}
+
+filter_recordings
+filter_speakers
+filter_utts
+filter_speakers
+filter_recordings
+
+utils/utt2spk_to_spk2utt.pl $data/utt2spk > $data/spk2utt
+
+echo "fix_data_dir.sh: old files are kept in $data/.backup"
diff --git a/egs/swbd/ASR/utils/parse_options.sh b/egs/swbd/ASR/utils/parse_options.sh
new file mode 100755
index 000000000..34476fdb3
--- /dev/null
+++ b/egs/swbd/ASR/utils/parse_options.sh
@@ -0,0 +1,97 @@
+#!/bin/bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### No we process the command line options
+###
+while true; do
+  [ -z "${1:-}" ] && break;  # break if there are no arguments
+  case "$1" in
+    # If the enclosing script is called with --help option, print the help
+    # message and exit.  Scripts should put help messages in $help_message
+    --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
+      else printf "$help_message\n" 1>&2 ; fi;
+      exit 0 ;;
+    --*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
+      exit 1 ;;
+    # If the first command-line argument begins with "--" (e.g. --foo-bar),
+    # then work out the variable name as $name, which will equal "foo_bar".
+    --*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
+      # Next we test whether the variable in question is undefned-- if so it's
+      # an invalid option and we die.  Note: $0 evaluates to the name of the
+      # enclosing script.
+      # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
+      # is undefined.  We then have to wrap this test inside "eval" because
+      # foo_bar is itself inside a variable ($name).
+      eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+
+      oldval="`eval echo \\$$name`";
+      # Work out whether we seem to be expecting a Boolean argument.
+      if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      eval $name=\"$2\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
+        exit 1;
+      fi
+      shift 2;
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.
diff --git a/egs/swbd/ASR/utils/spk2utt_to_utt2spk.pl b/egs/swbd/ASR/utils/spk2utt_to_utt2spk.pl
new file mode 100755
index 000000000..23992f25d
--- /dev/null
+++ b/egs/swbd/ASR/utils/spk2utt_to_utt2spk.pl
@@ -0,0 +1,27 @@
+#!/usr/bin/env perl
+# Copyright 2010-2011 Microsoft Corporation
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+while(<>){ 
+    @A = split(" ", $_);
+    @A > 1 || die "Invalid line in spk2utt file: $_";
+    $s = shift @A;
+    foreach $u ( @A ) {
+        print "$u $s\n";
+    }
+}
+
+
diff --git a/egs/swbd/ASR/utils/utt2spk_to_spk2utt.pl b/egs/swbd/ASR/utils/utt2spk_to_spk2utt.pl
new file mode 100755
index 000000000..6e0e438ca
--- /dev/null
+++ b/egs/swbd/ASR/utils/utt2spk_to_spk2utt.pl
@@ -0,0 +1,38 @@
+#!/usr/bin/env perl
+# Copyright 2010-2011 Microsoft Corporation
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+# converts an utt2spk file to a spk2utt file.
+# Takes input from the stdin or from a file argument;
+# output goes to the standard out.
+
+if ( @ARGV > 1 ) {
+    die "Usage: utt2spk_to_spk2utt.pl [ utt2spk ] > spk2utt";
+}
+
+while(<>){ 
+    @A = split(" ", $_);
+    @A == 2 || die "Invalid line in utt2spk file: $_";
+    ($u,$s) = @A;
+    if(!$seen_spk{$s}) {
+        $seen_spk{$s} = 1;
+        push @spklist, $s;
+    }
+    push (@{$spk_hash{$s}}, "$u");
+}
+foreach $s (@spklist) {
+    $l = join(' ',@{$spk_hash{$s}});
+    print "$s $l\n";
+}
diff --git a/egs/tal_csasr/ASR/README.md b/egs/tal_csasr/ASR/README.md
new file mode 100644
index 000000000..a705a2f44
--- /dev/null
+++ b/egs/tal_csasr/ASR/README.md
@@ -0,0 +1,19 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with TAL_CSASR.
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/tal_csasr/ASR/RESULTS.md b/egs/tal_csasr/ASR/RESULTS.md
new file mode 100644
index 000000000..e696279bd
--- /dev/null
+++ b/egs/tal_csasr/ASR/RESULTS.md
@@ -0,0 +1,133 @@
+## Results
+
+#### Pruned transducer stateless 7 (zipformer)
+
+See <https://github.com/k2-fsa/icefall/pull/1033>
+
+[./pruned_transducer_stateless7_bbpe](./pruned_transducer_stateless7_bbpe)
+
+**Note**: The modeling units are byte level BPEs
+
+The best results I have gotten are:
+
+Vocab size | greedy (dev & test) | modified beam search (dev & test) |  |
+-- | -- | -- | --
+500  | 6.88 & 6.98 | 6.87 & 6.94 | --epoch 35 --avg 26
+
+The training command:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_bbpe/train.py \
+  --world-size 4 \
+  --start-epoch 1 \
+  --num-epochs 35 \
+  --use-fp16 1 \
+  --max-duration 800 \
+  --bbpe-model data/lang_bbpe_500/bbpe.model \
+  --exp-dir pruned_transducer_stateless7_bbpe/exp \
+  --master-port 12535
+```
+
+The decoding command:
+
+```
+ ./pruned_transducer_stateless7_bbpe/decode.py \
+   --epoch 35 \
+   --avg 26 \
+   --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+   --max-sym-per-frame 1 \
+   --bpe-model data/lang_bbpe_500/bbpe.model \
+   --max-duration 2000 \
+   --decoding-method greedy_search  # modified_beam_search
+```
+
+The pretrained model is available at:  https://huggingface.co/pkufool/icefall_asr_tal_csasr_pruned_transducer_stateless7_bbpe
+
+
+### TAL_CSASR Mix Chars and BPEs training results (Pruned Transducer Stateless5)
+
+#### 2022-06-22
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/428.
+
+The WERs are
+
+|decoding-method | epoch(iter) | avg | dev | test |
+|--|--|--|--|--|
+|greedy_search | 30 | 24 | 7.49 | 7.58|
+|modified_beam_search | 30 | 24 | 7.33 | 7.38|
+|fast_beam_search | 30 | 24 | 7.32 | 7.42|
+|greedy_search(use-averaged-model=True) | 30 | 24 | 7.30 | 7.39|
+|modified_beam_search(use-averaged-model=True) | 30 | 24 | 7.15 | 7.22|
+|fast_beam_search(use-averaged-model=True) | 30 | 24 | 7.18 | 7.27|
+|greedy_search | 348000 | 30 | 7.46 | 7.54|
+|modified_beam_search | 348000 | 30 | 7.24 | 7.36|
+|fast_beam_search | 348000 | 30 | 7.25 | 7.39 |
+
+The results (CER(%) and WER(%)) for Chinese CER and English WER respectivly (zh: Chinese, en: English):
+|decoding-method | epoch(iter) | avg | dev | dev_zh | dev_en | test | test_zh | test_en |
+|--|--|--|--|--|--|--|--|--|
+|greedy_search(use-averaged-model=True) | 30 | 24 | 7.30 | 6.48 | 19.19 |7.39| 6.66 | 19.13|
+|modified_beam_search(use-averaged-model=True) | 30 | 24 | 7.15 | 6.35 | 18.95 | 7.22| 6.50 | 18.70 |
+|fast_beam_search(use-averaged-model=True) | 30 | 24 | 7.18 | 6.39| 18.90 |  7.27| 6.55 | 18.77|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 6 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir data/lang_char \
+  --max-duration 90
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/KaACzXOVR0OM6cy0qbN5hw/#scalars
+
+The decoding command is:
+```
+epoch=30
+avg=24
+use_average_model=True
+
+## greedy search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 800 \
+  --use-averaged-model $use_average_model
+
+## modified beam search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 800 \
+  --decoding-method modified_beam_search \
+  --beam-size 4 \
+  --use-averaged-model $use_average_model
+
+## fast beam search
+./pruned_transducer_stateless5/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 1500 \
+  --decoding-method fast_beam_search \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8 \
+  --use-averaged-model $use_average_model
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_tal-csasr_pruned_transducer_stateless5>
diff --git a/egs/tal_csasr/ASR/local/__init__.py b/egs/tal_csasr/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tal_csasr/ASR/local/compute_fbank_musan.py b/egs/tal_csasr/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/local/compute_fbank_tal_csasr.py b/egs/tal_csasr/ASR/local/compute_fbank_tal_csasr.py
new file mode 100755
index 000000000..602e50d29
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/compute_fbank_tal_csasr.py
@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the tal_csasr dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_tal_csasr(num_mel_bins: int = 80):
+    src_dir = Path("data/manifests/tal_csasr")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train_set",
+        "dev_set",
+        "test_set",
+    )
+    prefix = "tal_csasr"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_tal_csasr(num_mel_bins=args.num_mel_bins)
diff --git a/egs/tal_csasr/ASR/local/display_manifest_statistics.py b/egs/tal_csasr/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..7521bb55b
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,96 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest
+
+
+def main():
+    paths = [
+        "./data/fbank/tal_csasr_cuts_train_set.jsonl.gz",
+        "./data/fbank/tal_csasr_cuts_dev_set.jsonl.gz",
+        "./data/fbank/tal_csasr_cuts_test_set.jsonl.gz",
+    ]
+
+    for path in paths:
+        print(f"Displaying the statistics for {path}")
+        cuts = load_manifest(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Displaying the statistics for ./data/fbank/tal_csasr_cuts_train_set.jsonl.gz
+Cuts count: 1050000
+Total duration (hours): 1679.0
+Speech duration (hours): 1679.0 (100.0%)
+***
+Duration statistics (seconds):
+mean    5.8
+std     4.1
+min     0.3
+25%     2.8
+50%     4.4
+75%     7.3
+99%     18.0
+99.5%   18.8
+99.9%   20.8
+max     36.5
+Displaying the statistics for ./data/fbank/tal_csasr_cuts_dev_set.jsonl.gz
+Cuts count: 5000
+Total duration (hours): 8.0
+Speech duration (hours): 8.0 (100.0%)
+***
+Duration statistics (seconds):
+mean    5.8
+std     4.0
+min     0.5
+25%     2.8
+50%     4.5
+75%     7.4
+99%     17.0
+99.5%   17.7
+99.9%   19.5
+max     21.5
+Displaying the statistics for ./data/fbank/tal_csasr_cuts_test_set.jsonl.gz
+Cuts count: 15000
+Total duration (hours): 23.6
+Speech duration (hours): 23.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    5.7
+std     4.0
+min     0.5
+25%     2.8
+50%     4.4
+75%     7.2
+99%     17.2
+99.5%   17.9
+99.9%   19.6
+max     32.3
+"""
diff --git a/egs/tal_csasr/ASR/local/prepare_char.py b/egs/tal_csasr/ASR/local/prepare_char.py
new file mode 100755
index 000000000..499937462
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/prepare_char.py
@@ -0,0 +1,264 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text_with_bpe,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(
+    token_sym_table: Dict[str, int],
+    words: List[str],
+    bpe_model=None,
+) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    sp = ""
+    if bpe_model is not None:
+        sp = spm.SentencePieceProcessor()
+        sp.load(str(bpe_model))
+
+    lexicon = []
+    zhPattern = re.compile(r"([\u4e00-\u9fa5])")
+    for word in words:
+        match = zhPattern.search(word)
+        tokens = []
+        if match:
+            tokens = list(word.strip(" \t"))
+        else:
+            tokens = sp.encode_as_pieces(word.strip(" \t"))
+
+        if contain_oov(token_sym_table, tokens):
+            continue
+        lexicon.append((word, tokens))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([\t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = line.split(" ")
+            for char in chars:
+                if char not in tokens:
+                    tokens[char] = len(tokens)
+
+    return tokens
+
+
+def main():
+    lang_dir = Path("data/lang_char")
+    text_file = lang_dir / "text_with_bpe"
+    bpe_model = lang_dir / "bpe.model"
+    words_file = lang_dir / "words.txt"
+
+    word_sym_table = k2.SymbolTable.from_file(words_file)
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words, bpe_model=bpe_model)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/prepare_lang.py b/egs/tal_csasr/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..c8cf9b881
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/prepare_lang.py
@@ -0,0 +1,388 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone")
+    return parser.parse_args()
+
+
+def main():
+    out_dir = Path(get_args().lang_dir)
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/prepare_words.py b/egs/tal_csasr/ASR/local/prepare_words.py
new file mode 100755
index 000000000..41ab3b2cb
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/prepare_words.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input words.txt without ids:
+    - words_no_ids.txt
+and generates the new words.txt with related ids.
+    - words.txt
+"""
+
+
+import argparse
+import logging
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare words.txt",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input",
+        default="data/lang_char/words_no_ids.txt",
+        type=str,
+        help="the words file without ids for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output",
+        default="data/lang_char/words.txt",
+        type=str,
+        help="the words file with ids for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input
+    output_file = args.output
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    add_words = ["<eps> 0", "!SIL 1", "<SPOKEN_NOISE> 2", "<UNK> 3"]
+    new_lines.extend(add_words)
+
+    logging.info("Starting reading the input file")
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        idx = 4 + i
+        new_line = str(x.strip("\n")) + " " + str(idx)
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the words.txt")
+    f_out = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/test_prepare_lang.py b/egs/tal_csasr/ASR/local/test_prepare_lang.py
new file mode 100755
index 000000000..74e025ad7
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/test_prepare_lang.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/text2segments.py b/egs/tal_csasr/ASR/local/text2segments.py
new file mode 100644
index 000000000..3df727c67
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/text2segments.py
@@ -0,0 +1,83 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input "text", which refers to the transcript file for
+WenetSpeech:
+    - text
+and generates the output file text_word_segmentation which is implemented
+with word segmenting:
+    - text_words_segmentation
+"""
+
+
+import argparse
+
+import jieba
+from tqdm import tqdm
+
+jieba.enable_paddle()
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Chinese Word Segmentation for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/text",
+        type=str,
+        help="the input text file for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/text_words_segmentation",
+        type=str,
+        help="the text implemented with words segmenting for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        x = lines[i].rstrip()
+        seg_list = jieba.cut(x, use_paddle=True)
+        new_line = " ".join(seg_list)
+        new_lines.append(new_line)
+
+    f_new = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_new.write(line)
+        f_new.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/text2token.py b/egs/tal_csasr/ASR/local/text2token.py
new file mode 100755
index 000000000..85047c367
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/text2token.py
@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument("--space", default="<space>", type=str, help="space symbol")
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument("text", type=str, default=False, nargs="?", help="input text")
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer)
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = "".join(a_flat)
+        print(a_chars)
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/text_normalize.py b/egs/tal_csasr/ASR/local/text_normalize.py
new file mode 100755
index 000000000..e97b3a5a3
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/text_normalize.py
@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# Copyright    2022  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input "text_full", which includes all transcript files
+for a dataset:
+    - text_full
+and generates the output file text_normalize which is implemented
+to normalize text:
+    - text
+"""
+
+
+import argparse
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Normalizing for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input",
+        default="data/lang_char/text_full",
+        type=str,
+        help="the input text file",
+    )
+    parser.add_argument(
+        "--output",
+        default="data/lang_char/text",
+        type=str,
+        help="the text implemented with normalizer",
+    )
+
+    return parser
+
+
+def text_normalize(str_line: str):
+    line = str_line.strip().rstrip("\n")
+    line = line.replace("<sil>", "")
+    line = line.replace("<%>", "")
+    line = line.replace("<->", "")
+    line = line.replace("<$>", "")
+    line = line.replace("<#>", "")
+    line = line.replace("<_>", "")
+    line = line.replace("<space>", "")
+    line = line.replace("`", "")
+    line = line.replace("'", "")
+    line = line.replace("&", "")
+    line = line.replace(",", "")
+    line = line.replace("Ａ", "A")
+    line = line.replace("Ｃ", "C")
+    line = line.replace("Ｄ", "D")
+    line = line.replace("Ｅ", "E")
+    line = line.replace("Ｇ", "G")
+    line = line.replace("Ｈ", "H")
+    line = line.replace("Ｉ", "I")
+    line = line.replace("Ｎ", "N")
+    line = line.replace("Ｕ", "U")
+    line = line.replace("Ｗ", "W")
+    line = line.replace("Ｙ", "Y")
+    line = line.replace("ａ", "A")
+    line = line.replace("ｂ", "B")
+    line = line.replace("ｃ", "C")
+    line = line.replace("ｋ", "K")
+    line = line.replace("ｔ", "T")
+    line = line.replace("，", "")
+    line = line.replace("丶", "")
+    line = line.replace("。", "")
+    line = line.replace("、", "")
+    line = line.replace("？", "")
+    line = line.replace("·", "")
+    line = line.replace("*", "")
+    line = line.replace("！", "")
+    line = line.replace("$", "")
+    line = line.replace("+", "")
+    line = line.replace("-", "")
+    line = line.replace("\\", "")
+    line = line.replace("?", "")
+    line = line.replace("￥", "")
+    line = line.replace("%", "")
+    line = line.replace(".", "")
+    line = line.replace("<", "")
+    line = line.replace("＆", "")
+    line = line.replace("~", "")
+    line = line.replace("=", "")
+    line = line.replace(":", "")
+    line = line.replace("!", "")
+    line = line.replace("/", "")
+    line = line.replace("‘", "")
+    line = line.replace("’", "")
+    line = line.replace("“", "")
+    line = line.replace("”", "")
+    line = line.replace("[", "")
+    line = line.replace("]", "")
+    line = line.replace("@", "")
+    line = line.replace("#", "")
+    line = line.replace("：", "")
+    line = line.replace("；", "")
+    line = line.replace("…", "")
+    line = line.replace("《", "")
+    line = line.replace("》", "")
+    line = line.upper()
+
+    return line
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input
+    output_file = args.output
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        new_line = text_normalize(lines[i])
+        new_lines.append(new_line)
+
+    f_new = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_new.write(line)
+        f_new.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/tokenize_with_bpe_model.py b/egs/tal_csasr/ASR/local/tokenize_with_bpe_model.py
new file mode 100644
index 000000000..d7fd838f2
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/tokenize_with_bpe_model.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+#
+# Copyright    2021 Mobvoi Inc.          (authors: Binbin Zhang)
+# Copyright    2022  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script takes as input text (it includes Chinese and English):
+    - text
+and generates the text_with_bpe.
+    - text_with_bpe
+"""
+
+
+import argparse
+import logging
+
+import sentencepiece as spm
+from tqdm import tqdm
+
+from icefall.utils import tokenize_by_bpe_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare text_with_bpe",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input",
+        default="data/lang_char/text",
+        type=str,
+        help="the text includes Chinese and English words",
+    )
+    parser.add_argument(
+        "--output",
+        default="data/lang_char/text_with_bpe",
+        type=str,
+        help="the text_with_bpe tokenized by bpe model",
+    )
+    parser.add_argument(
+        "--bpe-model",
+        default="data/lang_char/bpe.model",
+        type=str,
+        help="the bpe model for processing the English parts",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input
+    output_file = args.output
+    bpe_model = args.bpe_model
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+
+    logging.info("Starting reading the text")
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        txt_tokens = tokenize_by_bpe_model(sp, x)
+        new_line = txt_tokens.replace("/", " ")
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the text_with_bpe")
+    f_out = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/local/train_bbpe_model.py b/egs/tal_csasr/ASR/local/train_bbpe_model.py
new file mode 120000
index 000000000..7fb4a9f9d
--- /dev/null
+++ b/egs/tal_csasr/ASR/local/train_bbpe_model.py
@@ -0,0 +1 @@
+../../../aishell/ASR/local/train_bbpe_model.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/prepare.sh b/egs/tal_csasr/ASR/prepare.sh
new file mode 100755
index 000000000..2de4ac8f5
--- /dev/null
+++ b/egs/tal_csasr/ASR/prepare.sh
@@ -0,0 +1,198 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/TALCS_corpus
+#      You can find three directories:train_set, dev_set, and test_set.
+#      You can get it from https://ai.100tal.com/dataset
+#     - dev_set
+#     - test_set
+#     - train_set
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bbpe_xxx,
+# data/lang_bbpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 2000
+  1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+  # Before you run this script, you must get the TAL_CSASR dataset
+  # from https://ai.100tal.com/dataset
+  if [ ! -d $dl_dir/tal_csasr/TALCS_corpus ]; then
+    mv $dl_dir/TALCS_corpus $dl_dir/tal_csasr
+  fi
+
+  # If you have pre-downloaded it to /path/to/TALCS_corpus,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/TALCS_corpus $dl_dir/tal_csasr
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/musan
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare tal_csasr manifest"
+  # We assume that you have downloaded the TALCS_corpus
+  # to $dl_dir/tal_csasr
+  if [ ! -f data/manifests/tal_csasr/.manifests.done ]; then
+    mkdir -p data/manifests/tal_csasr
+    lhotse prepare tal-csasr $dl_dir/tal_csasr data/manifests/tal_csasr
+    touch data/manifests/tal_csasr/.manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for tal_csasr"
+  if [ ! -f data/fbank/.tal_csasr.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_tal_csasr.py
+    touch data/fbank/.tal_csasr.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+
+  # Download BPE models trained with LibriSpeech
+  # Here we use the BPE model with 5000 units trained with Librispeech.
+  # You can also use other BPE models if available.
+  if [ ! -f $lang_char_dir/bpe.model ]; then
+    wget -O $lang_char_dir/bpe.model \
+      https://huggingface.co/luomingshuang/bpe_models_trained_with_Librispeech/resolve/main/lang_bpe_500/bpe.model
+  fi
+
+  # we extract text from manifests rather than the label.txt in corpus, because
+  # the texts in manifests have been normalized in lhotse.
+  if [ ! -f $lang_char_dir/text ]; then
+    gunzip -c data/manifests/tal_csasr/tal_csasr_supervisions_train_set.jsonl.gz \
+      | grep -o 'text":\s[^,]*' | sed 's/text": "//g;s/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text_train
+
+    gunzip -c data/manifests/tal_csasr/tal_csasr_supervisions_dev_set.jsonl.gz \
+      | grep -o 'text":\s[^,]*' | sed 's/text": "//g;s/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text_dev
+
+    gunzip -c data/manifests/tal_csasr/tal_csasr_supervisions_test_set.jsonl.gz \
+      | grep -o 'text":\s[^,]*' | sed 's/text": "//g;s/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text_test
+
+    for r in text_train text_dev text_test ; do
+      cat $lang_char_dir/$r >> $lang_char_dir/text
+    done
+  fi
+
+  # Prepare words.txt
+  # We assume you have installed jieba, if not, please install
+  # it using: pip install jieba
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    python -m jieba $lang_char_dir/text | sed 's/\///g;s/\s\+/ /g' > $lang_char_dir/text.seg
+
+   (echo '<eps> 0'; echo '!SIL 1'; echo '<SPOKEN_NOISE> 2'; echo '<UNK> 3';) \
+     > $lang_char_dir/words.txt
+
+   cat $lang_char_dir/text.seg | sed 's/ /\n/g' | sort -u | sed '/^$/d' \
+      | awk '{print $1" "NR+3}' >> $lang_char_dir/words.txt
+
+   num_lines=$(< $lang_char_dir/words.txt wc -l)
+    (echo "#0 $num_lines"; echo "<s> $(($num_lines + 1))"; echo "</s> $(($num_lines + 2))";) \
+      >> $lang_char_dir/words.txt
+  fi
+
+  # Tokenize text with BPE model
+  python ./local/tokenize_with_bpe_model.py \
+    --input $lang_char_dir/text \
+    --output $lang_char_dir/text_with_bpe \
+    --bpe-model $lang_char_dir/bpe.model
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    python local/prepare_char.py
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 7: Prepare Byte BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bbpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp $lang_char_dir/words.txt $lang_dir
+    cp $lang_char_dir/text $lang_dir
+
+    if [ ! -f $lang_dir/bbpe.model ]; then
+      ./local/train_bbpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/text
+    fi
+  done
+fi
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/__init__.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 100644
index 000000000..5269a1778
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1,424 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 for AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class TAL_CSASRAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=300,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                num_cuts_for_bins_estimate=20000,
+                buffer_size=60000,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            rank=0,
+            world_size=1,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            rank=0,
+            world_size=1,
+            shuffle=False,
+        )
+        logging.info("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tal_csasr_cuts_train_set.jsonl.gz"
+        )
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tal_csasr_cuts_dev_set.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tal_csasr_cuts_test_set.jsonl.gz"
+        )
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/beam_search.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..ed78bd4bb
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/conformer.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/conformer.py
new file mode 120000
index 000000000..b2af4e1df
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..3485d4005
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,740 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+import re
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TAL_CSASRAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from local.text_normalize import text_normalize
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    sp: spm.SentencePieceProcessor = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    zh_hyps = []
+    en_hyps = []
+    pattern = re.compile(r"([\u4e00-\u9fff])")
+    en_letter = "[\u0041-\u005a|\u0061-\u007a]+"  # English letters
+    zh_char = "[\u4e00-\u9fa5]+"  # Chinese chars
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            hyps.append(chars_new)
+            zh_hyps.append(zh_text)
+            en_hyps.append(en_text)
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            hyps.append(chars_new)
+            zh_hyps.append(zh_text)
+            en_hyps.append(en_text)
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            hyps.append(chars_new)
+            zh_hyps.append(zh_text)
+            en_hyps.append(en_text)
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            for i in range(encoder_out.size(0)):
+                hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+                chars = pattern.split(hyp.upper())
+                chars_new = []
+                zh_text = []
+                en_text = []
+                for char in chars:
+                    if char != "":
+                        tokens = char.strip().split(" ")
+                        chars_new.extend(tokens)
+                        for token in tokens:
+                            zh_text.extend(re.findall(zh_char, token))
+                            en_text.extend(re.findall(en_letter, token))
+                hyps.append(chars_new)
+                zh_hyps.append(zh_text)
+                en_hyps.append(en_text)
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": (hyps, zh_hyps, en_hyps)}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): (hyps, zh_hyps, en_hyps)
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": (hyps, zh_hyps, en_hyps)}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+    sp: spm.SentencePieceProcessor = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    zh_results = defaultdict(list)
+    en_results = defaultdict(list)
+    pattern = re.compile(r"([\u4e00-\u9fff])")
+    en_letter = "[\u0041-\u005a|\u0061-\u007a]+"  # English letters
+    zh_char = "[\u4e00-\u9fa5]+"  # Chinese chars
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        zh_texts = []
+        en_texts = []
+        for i in range(len(texts)):
+            text = texts[i]
+            chars = pattern.split(text.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            zh_texts.append(zh_text)
+            en_texts.append(en_text)
+            texts[i] = chars_new
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+            sp=sp,
+        )
+
+        for name, hyps_texts in hyps_dict.items():
+            this_batch = []
+            this_batch_zh = []
+            this_batch_en = []
+            # print(hyps_texts)
+            hyps, zh_hyps, en_hyps = hyps_texts
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            for cut_id, hyp_words, ref_text in zip(cut_ids, zh_hyps, zh_texts):
+                this_batch_zh.append((cut_id, ref_text, hyp_words))
+
+            for cut_id, hyp_words, ref_text in zip(cut_ids, en_hyps, en_texts):
+                this_batch_en.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+            zh_results[name + "_zh"].extend(this_batch_zh)
+            en_results[name + "_en"].extend(this_batch_en)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results, zh_results, en_results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TAL_CSASRAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    bpe_model = params.lang_dir + "/bpe.model"
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    def text_normalize_for_cut(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = text.strip("\n").strip("\t")
+        c.supervisions[0].text = text_normalize(text)
+        return c
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tal_csasr = TAL_CSASRAsrDataModule(args)
+
+    dev_cuts = tal_csasr.valid_cuts()
+    dev_cuts = dev_cuts.map(text_normalize_for_cut)
+    dev_dl = tal_csasr.valid_dataloaders(dev_cuts)
+
+    test_cuts = tal_csasr.test_cuts()
+    test_cuts = test_cuts.map(text_normalize_for_cut)
+    test_dl = tal_csasr.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict, zh_results_dict, en_results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            sp=sp,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set + "-zh",
+            results_dict=zh_results_dict,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set + "-en",
+            results_dict=en_results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/decoder.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..8a5e07bd5
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..2fc10439b
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/export.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..bc33dd160
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,282 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#           2022 Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --lang-dir ./data/lang_char \
+  --epoch 30 \
+  --avg 24 \
+  --use-averaged-model True
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tal_csasr/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 30 \
+        --avg 24 \
+        --max-duration 800 \
+        --decoding-method greedy_search \
+        --lang-dir ./data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    bpe_model = params.lang_dir + "/bpe.model"
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/joiner.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..f31b5fd9b
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/model.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..be059ba7c
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/optim.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..661206562
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/pretrained.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100755
index 000000000..3305f5bd3
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,367 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#                2022  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir ./data/lang_char \
+    --decoding-method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir ./data/lang_char \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --lang-dir ./data/lang_char \
+    --decoding-method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+import re
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    bpe_model = params.lang_dir + "/bpe.model"
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_di = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    pattern = re.compile(r"([\u4e00-\u9fff])")
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            for char in chars:
+                if char != "":
+                    chars_new.extend(char.strip().split(" "))
+            hyps.append(chars_new)
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            for char in chars:
+                if char != "":
+                    chars_new.extend(char.strip().split(" "))
+            hyps.append(chars_new)
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            for char in chars:
+                if char != "":
+                    chars_new.extend(char.strip().split(" "))
+            hyps.append(chars_new)
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            for char in chars:
+                if char != "":
+                    chars_new.extend(char.strip().split(" "))
+            hyps.append(chars_new)
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/scaling.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..be7b111c6
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../../../../egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/test_model.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/test_model.py
new file mode 100755
index 000000000..9aad32014
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/test_model.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless4/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 18
+    params.dim_feedforward = 1024
+    params.encoder_dim = 256
+    params.nhead = 4
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..d03970265
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1086 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TAL_CSASRAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from local.text_normalize import text_normalize
+from local.tokenize_with_bpe_model import tokenize_by_bpe_model
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 100,
+            "valid_interval": 2000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 1000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids_with_bpe(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+        # print(batch["supervisions"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                batch=batch,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        scaler.scale(loss).backward()
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    bpe_model = params.lang_dir + "/bpe.model"
+    import sentencepiece as spm
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    tal_csasr = TAL_CSASRAsrDataModule(args)
+    train_cuts = tal_csasr.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    def text_normalize_for_cut(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = text.strip("\n").strip("\t")
+        text = text_normalize(text)
+        text = tokenize_by_bpe_model(sp, text)
+        c.supervisions[0].text = text
+        return c
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.map(text_normalize_for_cut)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = tal_csasr.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = tal_csasr.valid_cuts()
+    valid_cuts = valid_cuts.map(text_normalize_for_cut)
+    valid_dl = tal_csasr.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    return
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    TAL_CSASRAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/__init__.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/asr_datamodule.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/asr_datamodule.py
new file mode 120000
index 000000000..c473a600a
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless5/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/beam_search.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/beam_search.py
new file mode 120000
index 000000000..4eef3d295
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless5/beam_search.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/decode.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/decode.py
new file mode 100755
index 000000000..885778965
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/decode.py
@@ -0,0 +1,815 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_bbpe/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TAL_CSASRAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import (
+    LmScorer,
+    NgramLm,
+    byte_encode,
+    smart_byte_decode,
+    tokenize_by_CJK_char,
+)
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_500/bbpe.model",
+        help="Path to the byte BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bbpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+        If you use fast_beam_search_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.25,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--ilme-scale",
+        type=float,
+        default=0.2,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for the internal language model estimation.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "fast_beam_search_LG":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            subtract_ilme=True,
+            ilme_scale=params.ilme_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        ref_texts = []
+        for tx in supervisions["text"]:
+            ref_texts.append(byte_encode(tokenize_by_CJK_char(tx)))
+
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(ref_texts),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(smart_byte_decode(sp.decode(hyp)).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+            key += f"_ilme_scale_{params.ilme_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = tokenize_by_CJK_char(ref_text).split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TAL_CSASRAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+            params.suffix += f"-ilme-scale-{params.ilme_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bbpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tal_csasr = TAL_CSASRAsrDataModule(args)
+
+    test_cuts = tal_csasr.test_cuts()
+    dev_cuts = tal_csasr.valid_cuts()
+
+    test_dl = tal_csasr.test_dataloaders(test_cuts)
+    dev_dl = tal_csasr.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/decoder.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/encoder_interface.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/encoder_interface.py
new file mode 120000
index 000000000..083f693ef
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless5/encoder_interface.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/export.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/export.py
new file mode 100755
index 000000000..862509d3f
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/export.py
@@ -0,0 +1,320 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `pruned_transducer_stateless7_bbpe/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless7_bbpe/decode.py \
+        --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bbpe_500/bbpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/pkufool/icefall_asr_tal_csasr_pruned_transducer_stateless7_bbpe
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/pkufool/icefall_asr_tal_csasr_pruned_transducer_stateless7_bbpe
+    # You will find the pre-trained model in icefall_asr_tal_csasr_pruned_transducer_stateless7_bbpe/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_bbpe/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_500/bbpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py
new file mode 100755
index 000000000..a23e2a04f
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py
@@ -0,0 +1,273 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless7_bbpe/jit_pretrained.py \
+  --nn-model-filename ./pruned_transducer_stateless7_bbpe/exp/cpu_jit.pt \
+  --bpe-model ./data/lang_bbpe_500/bbpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+from icefall import smart_byte_decode
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.decoder.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(args.nn_model_filename)
+
+    model.eval()
+
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = smart_byte_decode(sp.decode(hyp))
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/joiner.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/model.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/optim.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/pretrained.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/pretrained.py
new file mode 100755
index 000000000..f365986f6
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/pretrained.py
@@ -0,0 +1,355 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7_bbpe/export.py \
+  --exp-dir ./pruned_transducer_stateless7_bbpe/exp \
+  --bpe-model data/lang_bbpe_500/bbpe.model \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7_bbpe/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt \
+    --bpe-model ./data/lang_bbpe_500/bbpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7_bbpe/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7_bbpe/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7_bbpe/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import smart_byte_decode
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(smart_byte_decode(hyp).split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(smart_byte_decode(sp.decode(hyp)).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/scaling.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/scaling_converter.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/test_model.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/test_model.py
new file mode 120000
index 000000000..7ceac5d10
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/test_model.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py
new file mode 100755
index 000000000..aee3972cd
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py
@@ -0,0 +1,1249 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_bbpe/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_bbpe/exp \
+  --max-duration 400
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_bbpe/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_bbpe/exp \
+  --max-duration 800
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TAL_CSASRAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut, CutSet
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import byte_encode, diagnostics, tokenize_by_CJK_char
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_bbpe/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bbpe_500/bbpe.model",
+        help="Path to the Byte BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10.0,
+            "allowed_excess_duration_ratio": 0.1,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bbpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    tal_csasr = TAL_CSASRAsrDataModule(args)
+    train_cuts = tal_csasr.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    def tokenize_text_in_cut(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = byte_encode(tokenize_by_CJK_char(text))
+        c.supervisions[0].text = text
+        return c
+
+    logging.info(f"Filtering short and long utterances.")
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    logging.info(f"Tokenizing and encoding texts in train cuts.")
+    train_cuts = train_cuts.map(tokenize_text_in_cut)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = tal_csasr.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = tal_csasr.valid_cuts()
+
+    logging.info(f"Tokenizing and encoding texts in valid cuts.")
+    valid_cuts = valid_cuts.map(tokenize_text_in_cut)
+
+    valid_dl = tal_csasr.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    TAL_CSASRAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/zipformer.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/shared b/egs/tal_csasr/ASR/shared
new file mode 120000
index 000000000..e9461a6d7
--- /dev/null
+++ b/egs/tal_csasr/ASR/shared
@@ -0,0 +1 @@
+../../librispeech/ASR/shared
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/README.md b/egs/tedlium3/ASR/README.md
new file mode 100644
index 000000000..0740258a7
--- /dev/null
+++ b/egs/tedlium3/ASR/README.md
@@ -0,0 +1,18 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with TedLium3.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                  | Encoder   | Decoder            | Comment                     |
+|----------------------------------|-----------|--------------------|-----------------------------|
+| `transducer_stateless`           | Conformer | Embedding + Conv1d |                             |
+| `pruned_transducer_stateless`    | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/tedlium3/ASR/RESULTS.md b/egs/tedlium3/ASR/RESULTS.md
new file mode 100644
index 000000000..bd8a5b43f
--- /dev/null
+++ b/egs/tedlium3/ASR/RESULTS.md
@@ -0,0 +1,341 @@
+## Results
+
+### TedLium3 BPE training results (Zipformer)
+
+#### 2023-06-15 (Regular transducer)
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/1125.
+
+Number of model parameters: 65549011, i.e., 65.5 M
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 6.74       | 6.16       | --epoch 50, --avg 22, --max-duration 500 |
+|      beam search (beam size 4)     | 6.56       | 5.95       | --epoch 50, --avg 22, --max-duration 500 |
+| modified beam search (beam size 4) | 6.54       | 6.00       | --epoch 50, --avg 22, --max-duration 500 |
+| fast beam search (set as default)  | 6.91       | 6.28       | --epoch 50, --avg 22, --max-duration 500 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer/train.py \
+  --use-fp16 true \
+  --world-size 4 \
+  --num-epochs 50 \
+  --start-epoch 0 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/AKXbJha0S9aXyfmuvG4h5A/#scalars
+
+The decoding command is:
+```
+epoch=50
+avg=22
+
+## greedy search
+./zipformer/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir zipformer/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 500
+
+## beam search
+./zipformer/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir zipformer/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 500 \
+  --decoding-method beam_search \
+  --beam-size 4
+
+## modified beam search
+./zipformer/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir zipformer/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 500 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./zipformer/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir ./zipformer/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 1500 \
+  --decoding-method fast_beam_search \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/desh2608/icefall-asr-tedlium3-zipformer>
+
+#### 2023-06-26 (Modified transducer)
+
+```
+./zipformer/train.py \
+  --use-fp16 true \
+  --world-size 4 \
+  --num-epochs 50 \
+  --start-epoch 0 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000 \
+  --rnnt-type modified
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/3d4bYmbJTGiWQQaW88CVEQ/#scalars
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 6.32       | 5.83       | --epoch 50, --avg 22, --max-duration 500 |
+| modified beam search (beam size 4) | 6.16       | 5.79       | --epoch 50, --avg 22, --max-duration 500 |
+| fast beam search (set as default)  | 6.30       | 5.89       | --epoch 50, --avg 22, --max-duration 500 |
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/desh2608/icefall-asr-tedlium3-zipformer>.
+
+### TedLium3 BPE training results (Conformer-CTC 2)
+
+#### [conformer_ctc2](./conformer_ctc2)
+
+See <https://github.com/k2-fsa/icefall/pull/696> for more details.
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/5NQQiqOqSqazfn4w2yeWEQ/>
+
+You can find a pretrained model and decoding results at:
+<https://huggingface.co/videodanchik/icefall-asr-tedlium3-conformer-ctc2>
+
+Number of model parameters: 101141699, i.e., 101.14 M
+
+The WERs are
+
+|                          | dev        | test        | comment             |
+|--------------------------|------------|-------------|---------------------|
+| ctc decoding             | 6.45       | 5.96        | --epoch 38 --avg 26 |
+| 1best                    | 5.92       | 5.51        | --epoch 38 --avg 26 |
+| whole lattice rescoring  | 5.96       | 5.47        | --epoch 38 --avg 26 |
+| attention decoder        | 5.60       | 5.33        | --epoch 38 --avg 26 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc2/train.py \
+    --world-size 4 \
+    --num-epochs 40 \
+    --exp-dir conformer_ctc2/exp \
+    --max-duration 350 \
+    --use-fp16 true
+```
+
+The decoding command is:
+```
+epoch=38
+avg=26
+
+## ctc decoding
+./conformer_ctc2/decode.py \
+  --method ctc-decoding \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+
+## 1best
+./conformer_ctc2/decode.py \
+  --method 1best \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+
+## whole lattice rescoring
+./conformer_ctc2/decode.py \
+  --method whole-lattice-rescoring \
+  --exp-dir conformer_ctc2/exp \
+  --lm-path data/lm/G_4_gram_big.pt \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+
+## attention decoder
+./conformer_ctc2/decode.py \
+  --method attention-decoder \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_500 \
+  --result-dir conformer_ctc2/exp \
+  --max-duration 500 \
+  --epoch $epoch \
+  --avg $avg
+```
+
+### TedLium3 BPE training results (Pruned Transducer)
+
+#### 2022-03-21
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/261.
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 7.27       | 6.69       | --epoch 29, --avg 13, --max-duration 100 |
+|      beam search (beam size 4)     | 6.70       | 6.04       | --epoch 29, --avg 13, --max-duration 100 |
+| modified beam search (beam size 4) | 6.77       | 6.14       | --epoch 29, --avg 13, --max-duration 100 |
+| fast beam search (set as default)  | 7.14       | 6.50       | --epoch 29, --avg 13, --max-duration 1500|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --max-duration 300
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/VpA8b7SZQ7CEjZs9WZ5HNA/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=13
+
+## greedy search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100
+
+## beam search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method beam_search \
+  --beam-size 4
+
+## modified beam search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_tedlium3_pruned_transducer_stateless>
+
+### TedLium3 BPE training results (Transducer)
+
+#### Conformer encoder + embedding decoder
+
+##### 2022-03-21
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/233
+And the SpecAugment codes from this PR https://github.com/lhotse-speech/lhotse/pull/604
+
+Conformer encoder + non-current decoder. The decoder
+contains only an embedding layer and a Conv1d (with kernel size 2).
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 7.19       | 6.70       | --epoch 29, --avg 11, --max-duration 100 |
+|      beam search (beam size 4)     | 7.02       | 6.36       | --epoch 29, --avg 11, --max-duration 100 |
+| modified beam search (beam size 4) | 6.91       | 6.33       | --epoch 29, --avg 11, --max-duration 100 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 300
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/4ks15jYHR4uMyvpW7Nz76Q/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=11
+
+## greedy search
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100
+
+## beam search
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method beam_search \
+  --beam-size 4
+
+## modified beam search
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_tedlium3_transducer_stateless>
diff --git a/egs/tedlium3/ASR/conformer_ctc2/__init__.py b/egs/tedlium3/ASR/conformer_ctc2/__init__.py
new file mode 100755
index 000000000..e69de29bb
diff --git a/egs/tedlium3/ASR/conformer_ctc2/asr_datamodule.py b/egs/tedlium3/ASR/conformer_ctc2/asr_datamodule.py
new file mode 120000
index 000000000..49b2ee483
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/attention.py b/egs/tedlium3/ASR/conformer_ctc2/attention.py
new file mode 100644
index 000000000..178cd7e62
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/attention.py
@@ -0,0 +1,201 @@
+# Copyright    2022  Behavox LLC.        (author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple, Union
+
+import torch
+from scaling import ScaledLinear
+
+
+class MultiheadAttention(torch.nn.Module):
+    """Allows the model to jointly attend to information
+    from different representation subspaces. This is a modified
+    version of the original version of multihead attention
+    (see Attention Is All You Need <https://arxiv.org/abs/1706.03762>)
+    with replacement of input / output projection layers
+    with newly introduced ScaleLinear layer
+    (see https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py).
+
+    Args:
+        embed_dim:
+          total dimension of the model.
+        num_heads:
+          number of parallel attention heads. Note that embed_dim will be split
+          across num_heads, i.e. each head will have dimension (embed_dim // num_heads).
+        dropout:
+          dropout probability on attn_output_weights. (default=0.0).
+        bias:
+          if specified, adds bias to input / output projection layers (default=True).
+        add_bias_kv:
+          if specified, adds bias to the key and value sequences at dim=0 (default=False).
+        add_zero_attn:
+          if specified, adds a new batch of zeros to the key and value sequences
+          at dim=1 (default=False).
+        batch_first:
+          if True, then the input and output tensors are provided as
+          (batch, seq, feature), otherwise (seq, batch, feature) (default=False).
+
+    Examples::
+        >>> multihead_attn = MultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        bias: bool = True,
+        add_bias_kv: bool = False,
+        add_zero_attn: bool = False,
+        batch_first: bool = False,
+        device: Union[torch.device, str, None] = None,
+        dtype: Union[torch.dtype, str, None] = None,
+    ) -> None:
+
+        super().__init__()
+
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.batch_first = batch_first
+
+        if embed_dim % num_heads != 0:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads. "
+                "Got embedding dim vs number 0f heads: "
+                f"{embed_dim} vs {num_heads}"
+            )
+
+        self.head_dim = embed_dim // num_heads
+
+        self.in_proj = ScaledLinear(
+            embed_dim,
+            3 * embed_dim,
+            bias=bias,
+            device=device,
+            dtype=dtype,
+        )
+        self.out_proj = ScaledLinear(
+            embed_dim,
+            embed_dim,
+            bias=bias,
+            initial_scale=0.25,
+            device=device,
+            dtype=dtype,
+        )
+
+        if add_bias_kv:
+            self.bias_k = torch.nn.Parameter(
+                torch.empty((1, 1, embed_dim), device=device, dtype=dtype)
+            )
+            self.bias_v = torch.nn.Parameter(
+                torch.empty((1, 1, embed_dim), device=device, dtype=dtype)
+            )
+        else:
+            self.register_parameter("bias_k", None)
+            self.register_parameter("bias_v", None)
+
+        self.add_zero_attn = add_zero_attn
+
+        self._reset_parameters()
+
+    def _reset_parameters(self) -> None:
+        if self.bias_k is not None:
+            torch.nn.init.xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            torch.nn.init.xavier_normal_(self.bias_v)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_padding_mask: Optional[torch.Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+            query:
+              Query embeddings of shape (L, N, E_q) when batch_first=False or (N, L, E_q)
+              when batch_first=True, where L is the target sequence length, N is the batch size,
+              and E_q is the query embedding dimension embed_dim. Queries are compared against
+              key-value pairs to produce the output. See "Attention Is All You Need" for more details.
+            key:
+              Key embeddings of shape (S, N, E_k) when batch_first=False or (N, S, E_k) when
+              batch_first=True, where S is the source sequence length, N is the batch size, and
+              E_k is the key embedding dimension kdim. See "Attention Is All You Need" for more details.
+            value:
+              Value embeddings of shape (S, N, E_v) when batch_first=False or (N, S, E_v) when
+              batch_first=True, where S is the source sequence length, N is the batch size, and
+              E_v is the value embedding dimension vdim. See "Attention Is All You Need" for more details.
+            key_padding_mask:
+              If specified, a mask of shape (N, S) indicating which elements within key
+              to ignore for the purpose of attention (i.e. treat as "padding").
+              Binary and byte masks are supported. For a binary mask, a True value indicates
+              that the corresponding key value will be ignored for the purpose of attention.
+              For a byte mask, a non-zero value indicates that the corresponding key value will be ignored.
+            need_weights:
+              If specifid, returns attn_output_weights in addition to attn_outputs (default=True).
+            attn_mask:
+              If specified, a 2D or 3D mask preventing attention to certain positions. Must be of shape
+              (L, S) or (N * num_heads, L, S), where N is the batch size, L is the target sequence length,
+              and S is the source sequence length. A 2D mask will be broadcasted across the batch while
+              a 3D mask allows for a different mask for each entry in the batch.
+              Binary, byte, and float masks are supported. For a binary mask, a True value indicates
+              that the corresponding position is not allowed to attend. For a byte mask, a non-zero
+              value indicates that the corresponding position is not allowed to attend. For a float mask,
+              the mask values will be added to the attention weight.
+
+        Returns:
+            attn_output:
+              Attention outputs of shape (L, N, E) when batch_first=False or (N, L, E) when batch_first=True,
+              where L is the target sequence length, N is the batch size, and E is the embedding dimension
+              embed_dim.
+            attn_output_weights:
+              Attention output weights of shape (N, L, S), where N is the batch size, L is the target sequence
+              length, and S is the source sequence length. Only returned when need_weights=True.
+        """
+        if self.batch_first:
+            query, key, value = [x.transpose(1, 0) for x in (query, key, value)]
+
+        (
+            attn_output,
+            attn_output_weights,
+        ) = torch.nn.functional.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            self.embed_dim,
+            self.num_heads,
+            in_proj_weight=self.in_proj.get_weight(),
+            in_proj_bias=self.in_proj.get_bias(),
+            bias_k=self.bias_k,
+            bias_v=self.bias_v,
+            add_zero_attn=self.add_zero_attn,
+            dropout_p=self.dropout,
+            out_proj_weight=self.out_proj.get_weight(),
+            out_proj_bias=self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+        if self.batch_first:
+            return attn_output.transpose(1, 0), attn_output_weights
+        return attn_output, attn_output_weights
diff --git a/egs/tedlium3/ASR/conformer_ctc2/combiner.py b/egs/tedlium3/ASR/conformer_ctc2/combiner.py
new file mode 100644
index 000000000..ff526029d
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/combiner.py
@@ -0,0 +1,244 @@
+# Copyright    2022  Behavox LLC.        (author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List
+
+import torch
+
+
+class RandomCombine(torch.nn.Module):
+    """
+    This module combines a list of Tensors, all with the same shape, to
+    produce a single output of that same shape which, in training time,
+    is a random combination of all the inputs; but which in test time
+    will be just the last input.
+    The idea is that the list of Tensors will be a list of outputs of multiple
+    conformer layers.  This has a similar effect as iterated loss. (See:
+    DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER
+    NETWORKS).
+    """
+
+    def __init__(
+        self,
+        num_inputs: int,
+        final_weight: float = 0.5,
+        pure_prob: float = 0.5,
+        stddev: float = 2.0,
+    ) -> None:
+        """
+        Args:
+          num_inputs:
+            The number of tensor inputs, which equals the number of layers'
+            outputs that are fed into this module.  E.g. in an 18-layer neural
+            net if we output layers 16, 12, 18, num_inputs would be 3.
+          final_weight:
+            The amount of weight or probability we assign to the
+            final layer when randomly choosing layers or when choosing
+            continuous layer weights.
+          pure_prob:
+            The probability, on each frame, with which we choose
+            only a single layer to output (rather than an interpolation)
+          stddev:
+            A standard deviation that we add to log-probs for computing
+            randomized weights.
+        The method of choosing which layers, or combinations of layers, to use,
+        is conceptually as follows::
+            With probability `pure_prob`::
+               With probability `final_weight`: choose final layer,
+               Else: choose random non-final layer.
+            Else::
+               Choose initial log-weights that correspond to assigning
+               weight `final_weight` to the final layer and equal
+               weights to other layers; then add Gaussian noise
+               with variance `stddev` to these log-weights, and normalize
+               to weights (note: the average weight assigned to the
+               final layer here will not be `final_weight` if stddev>0).
+        """
+        super().__init__()
+        assert 0 <= pure_prob <= 1, pure_prob
+        assert 0 < final_weight < 1, final_weight
+        assert num_inputs >= 1, num_inputs
+
+        self.num_inputs = num_inputs
+        self.final_weight = final_weight
+        self.pure_prob = pure_prob
+        self.stddev = stddev
+
+        self.final_log_weight = (
+            torch.tensor((final_weight / (1 - final_weight)) * (self.num_inputs - 1))
+            .log()
+            .item()
+        )
+
+    def forward(self, inputs: List[torch.Tensor]) -> torch.Tensor:
+        """Forward function.
+        Args:
+          inputs:
+            A list of Tensor, e.g. from various layers of a transformer.
+            All must be the same shape, of (*, num_channels)
+        Returns:
+          A Tensor of shape (*, num_channels). In test mode
+          this is just the final input.
+        """
+        num_inputs = self.num_inputs
+        assert len(inputs) == num_inputs, f"{len(inputs)}, {num_inputs}"
+        if not self.training or torch.jit.is_scripting() or len(inputs) == 1:
+            return inputs[-1]
+
+        # Shape of weights: (*, num_inputs)
+        num_channels = inputs[0].shape[-1]
+        num_frames = inputs[0].numel() // num_channels
+
+        ndim = inputs[0].ndim
+        # stacked_inputs: (num_frames, num_channels, num_inputs)
+        stacked_inputs = torch.stack(inputs, dim=ndim).reshape(
+            (num_frames, num_channels, num_inputs)
+        )
+
+        # weights: (num_frames, num_inputs)
+        weights = self._get_random_weights(
+            inputs[0].dtype, inputs[0].device, num_frames
+        )
+
+        weights = weights.reshape(num_frames, num_inputs, 1)
+        # ans: (num_frames, num_channels, 1)
+        ans = torch.matmul(stacked_inputs, weights)
+        # ans: (*, num_channels)
+
+        ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,))
+
+        return ans
+
+    def _get_random_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), such that
+          `ans.sum(dim=1)` is all ones.
+        """
+        pure_prob = self.pure_prob
+        if pure_prob == 0.0:
+            return self._get_random_mixed_weights(dtype, device, num_frames)
+        elif pure_prob == 1.0:
+            return self._get_random_pure_weights(dtype, device, num_frames)
+        else:
+            p = self._get_random_pure_weights(dtype, device, num_frames)
+            m = self._get_random_mixed_weights(dtype, device, num_frames)
+            return torch.where(
+                torch.rand(num_frames, 1, device=device) < self.pure_prob, p, m
+            )
+
+    def _get_random_pure_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A one-hot tensor of shape `(num_frames, self.num_inputs)`, with
+          exactly one weight equal to 1.0 on each frame.
+        """
+        final_prob = self.final_weight
+
+        # final contains self.num_inputs - 1 in all elements
+        final = torch.full((num_frames,), self.num_inputs - 1, device=device)
+        # nonfinal contains random integers in [0..num_inputs - 2], these are for non-final weights.
+        nonfinal = torch.randint(self.num_inputs - 1, (num_frames,), device=device)
+
+        indexes = torch.where(
+            torch.rand(num_frames, device=device) < final_prob, final, nonfinal
+        )
+        ans = torch.nn.functional.one_hot(indexes, num_classes=self.num_inputs).to(
+            dtype=dtype
+        )
+        return ans
+
+    def _get_random_mixed_weights(
+        self, dtype: torch.dtype, device: torch.device, num_frames: int
+    ) -> torch.Tensor:
+        """Return a tensor of random one-hot weights, of shape
+        `(num_frames, self.num_inputs)`,
+        Args:
+          dtype:
+            The data-type desired for the answer, e.g. float, double.
+          device:
+            The device needed for the answer.
+          num_frames:
+            The number of sets of weights desired.
+        Returns:
+          A tensor of shape (num_frames, self.num_inputs), which elements
+          in [0..1] that sum to one over the second axis, i.e.
+          `ans.sum(dim=1)` is all ones.
+        """
+        logprobs = (
+            torch.randn(num_frames, self.num_inputs, dtype=dtype, device=device)
+            * self.stddev
+        )
+        logprobs[:, -1] += self.final_log_weight
+        return logprobs.softmax(dim=1)
+
+
+def _test_random_combine(
+    final_weight: float,
+    pure_prob: float,
+    stddev: float,
+) -> None:
+    print(
+        f"_test_random_combine: final_weight={final_weight}, "
+        f"pure_prob={pure_prob}, stddev={stddev}"
+    )
+    num_inputs = 3
+    num_channels = 50
+    m = RandomCombine(
+        num_inputs=num_inputs,
+        final_weight=final_weight,
+        pure_prob=pure_prob,
+        stddev=stddev,
+    )
+
+    x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)]
+
+    y = m(x)
+    assert y.shape == x[0].shape
+    assert torch.allclose(y, x[0])  # .. since actually all ones.
+
+
+def _test_random_combine_main() -> None:
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.0)
+    _test_random_combine(0.999, 0, 0.0)
+    _test_random_combine(0.5, 0, 0.3)
+    _test_random_combine(0.5, 1, 0.3)
+    _test_random_combine(0.5, 0.5, 0.3)
+
+
+if __name__ == "__main__":
+    _test_random_combine_main()
diff --git a/egs/tedlium3/ASR/conformer_ctc2/conformer.py b/egs/tedlium3/ASR/conformer_ctc2/conformer.py
new file mode 100644
index 000000000..fad2f371f
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/conformer.py
@@ -0,0 +1,1033 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corp.                              (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from combiner import RandomCombine
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        aux_layer_period: int = 3,
+    ) -> None:
+        """
+        Args:
+          num_features (int):
+            number of input features.
+          num_classes (int):
+            number of output classes.
+          subsampling_factor (int):
+            subsampling factor of encoder;
+            currently, subsampling_factor MUST be 4.
+          d_model (int):
+            attention dimension, also the output dimension.
+          nhead (int):
+            number of heads in multi-head attention;
+            must satisfy d_model // nhead == 0.
+          dim_feedforward (int):
+            feedforward dimention.
+          num_encoder_layers (int):
+            number of encoder layers.
+          num_decoder_layers (int):
+            number of decoder layers.
+          dropout (float):
+            dropout rate.
+          layer_dropout (float):
+            layer-dropout rate.
+          cnn_module_kernel (int):
+            kernel size of convolution module.
+          aux_layer_period (int):
+            determines the auxiliary encoder layers.
+        """
+
+        super().__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+            cnn_module_kernel=cnn_module_kernel,
+        )
+
+        # aux_layers from 1/3
+        self.encoder = ConformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            aux_layers=list(
+                range(
+                    num_encoder_layers // 3,
+                    num_encoder_layers - 1,
+                    aux_layer_period,
+                )
+            ),
+        )
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            the input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+          warmup:
+            a floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          torch.Tensor: Predictor tensor of dimension (S, N, C).
+          torch.Tensor: Mask tensor of dimension (N, S)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, S, C) -> (S, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+
+        x = self.encoder(
+            x, pos_emb, src_key_padding_mask=mask, warmup=warmup
+        )  # (S, N, C)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Examples:
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        bypass_scale: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        """
+        Args:
+          d_model:
+            the number of expected features in the input (required).
+          nhead:
+            the number of heads in the multiheadattention models (required).
+          dim_feedforward:
+            the dimension of the feedforward network model (default=2048).
+          dropout:
+            the dropout value (default=0.1).
+          bypass_scale:
+            a scale on the layer's output, used in bypass (resnet-type) skip-connection;
+            when the layer is bypassed the final output will be a
+            weighted sum of the layer's input and layer's output with weights
+            (1.0-bypass_scale) and bypass_scale correspondingly (default=0.1).
+          layer_dropout:
+            the probability to bypass the layer (default=0.075).
+          cnn_module_kernel (int):
+            kernel size of convolution module (default=31).
+        """
+        super().__init__()
+
+        if bypass_scale < 0.0 or bypass_scale > 1.0:
+            raise ValueError("bypass_scale should be between 0.0 and 1.0")
+
+        if layer_dropout < 0.0 or layer_dropout > 1.0:
+            raise ValueError("layer_dropout should be between 0.0 and 1.0")
+
+        self.bypass_scale = bypass_scale
+        self.layer_dropout = layer_dropout
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        pos_emb: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+          src:
+            the sequence to the encoder layer of shape (S, N, C) (required).
+          pos_emb:
+            positional embedding tensor of shape (N, 2*S-1, C) (required).
+          src_mask:
+            the mask for the src sequence of shape (S, S) (optional).
+          src_key_padding_mask:
+            the mask for the src keys per batch of shape (N, S) (optional).
+          warmup:
+            controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+
+        Returns:
+            Output tensor of the shape (S, N, C), where
+            S is the source sequence length,
+            N is the batch size,
+            C is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(self.bypass_scale + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else self.bypass_scale
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src = src + self.dropout(self.conv_module(src))
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class ConformerEncoder(nn.Module):
+    """
+    ConformerEncoder is a stack of N encoder layers
+
+    Examples:
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        aux_layers: List[int],
+    ) -> None:
+
+        """
+        Args:
+          encoder_layer:
+            an instance of the ConformerEncoderLayer() class (required).
+          num_layers:
+            the number of sub-encoder-layers in the encoder (required).
+          aux_layers:
+            list of indexes of sub-encoder-layers outputs to be combined (required).
+        """
+
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert len(set(aux_layers)) == len(aux_layers)
+
+        assert num_layers - 1 not in aux_layers
+        self.aux_layers = aux_layers + [num_layers - 1]
+
+        self.combiner = RandomCombine(
+            num_inputs=len(self.aux_layers),
+            final_weight=0.5,
+            pure_prob=0.333,
+            stddev=2.0,
+        )
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layers in turn.
+
+        Args:
+          src:
+            the sequence to the encoder of shape (S, N, C) (required).
+          pos_emb:
+            positional embedding tensor of shape (N, 2*S-1, C) (required).
+          mask:
+            the mask for the src sequence of shape (S, S) (optional).
+          src_key_padding_mask:
+            the mask for the src keys per batch of shape (N, S) (optional).
+          warmup:
+            controls selective bypass of layer; if < 1.0, we will
+            bypass the layer more frequently (default=1.0).
+
+        Returns:
+          Output tensor of the shape (S, N, C), where
+          S is the source sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        output = src
+
+        outputs = []
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+            if i in self.aux_layers:
+                outputs.append(output)
+
+        output = self.combiner(outputs)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """
+    Relative positional encoding module.
+
+    See: Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """
+        Construct an PositionalEncoding object.
+
+        Args:
+          d_model: Embedding dimension.
+          dropout_rate: Dropout rate.
+          max_len: Maximum input length.
+
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """
+        Reset the positional encodings.
+
+        Args:
+          x:
+            input tensor (N, T, C), where
+            T is the source sequence length,
+            N is the batch size.
+            C is the feature number.
+
+        """
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            input tensor (N, T, C).
+
+        Returns:
+          torch.Tensor: Encoded tensor (N, T, C).
+          torch.Tensor: Encoded tensor (N, 2*T-1, C), where
+          T is the source sequence length,
+          N is the batch size.
+          C is the feature number.
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    """
+    Multi-Head Attention layer with relative position encoding
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context".
+
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        """
+        Args:
+          embed_dim:
+            total dimension of the model.
+          num_heads:
+            parallel attention heads.
+          dropout:
+            a Dropout layer on attn_output_weights. Default: 0.0.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        pos_emb: torch.Tensor,
+        key_padding_mask: Optional[torch.Tensor] = None,
+        need_weights: bool = False,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          query, key, value: map a query and a set of key-value pairs to an output.
+          pos_emb: Positional embedding tensor
+          key_padding_mask: if provided, specified padding elements in the key will
+                            be ignored by the attention. When given a binary mask
+                            and a value is True, the corresponding value on the attention
+                            layer will be ignored. When given a byte mask and a value is
+                            non-zero, the corresponding value on the attention layer will be ignored.
+          need_weights: output attn_output_weights.
+          attn_mask: 2D or 3D mask that prevents attention to certain positions.
+                     A 2D mask will be broadcasted for all the batches while a 3D
+                     mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+          - Inputs:
+          - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+          - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+          - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+          - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+          - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+          - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+          - Outputs:
+          - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+          - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Compute relative positional encoding.
+
+        Args:
+          x:
+            input tensor (batch, head, time1, 2*time1-1).
+            time1 means the length of query vector.
+
+        Returns:
+          torch.Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        pos_emb: torch.Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: torch.Tensor,
+        in_proj_bias: torch.Tensor,
+        dropout_p: float,
+        out_proj_weight: torch.Tensor,
+        out_proj_bias: torch.Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[torch.Tensor] = None,
+        need_weights: bool = False,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          query, key, value: map a query and a set of key-value pairs to an output.
+          pos_emb: Positional embedding tensor
+          embed_dim_to_check: total dimension of the model.
+          num_heads: parallel attention heads.
+          in_proj_weight, in_proj_bias: input projection weight and bias.
+          dropout_p: probability of an element to be zeroed.
+          out_proj_weight, out_proj_bias: the output projection weight and bias.
+          training: apply dropout if is ``True``.
+          key_padding_mask: if provided, specified padding elements in the key will
+                            be ignored by the attention. This is an binary mask.
+                            When the value is True, the corresponding value on the
+                            attention layer will be filled with -inf.
+          need_weights: output attn_output_weights.
+          attn_mask: 2D or 3D mask that prevents attention to certain positions.
+                     A 2D mask will be broadcasted for all the batches while a 3D
+                     mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+          Inputs:
+          - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+          - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+          - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+          - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+          - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+          - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+          Outputs:
+          - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+          - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    f"attn_mask's dimension {attn_mask.dim()} is not supported"
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [bsz * num_heads, tgt_len, src_len]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """
+        ConvolutionModule in Conformer model.
+        Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+        Construct a ConvolutionModule object.
+
+        Args:
+          channels (int):
+            the number of channels of conv layers.
+          kernel_size (int):
+            kernerl size of conv layers.
+          bias (bool):
+            whether to use bias in conv layers (default=True).
+        """
+        super().__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Compute convolution module.
+
+        Args:
+          x:
+            input tensor of shape (T, N, C).
+
+        Returns:
+          torch.Tensor: Output tensor (T, N, C), where
+          T is the source sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
diff --git a/egs/tedlium3/ASR/conformer_ctc2/decode.py b/egs/tedlium3/ASR/conformer_ctc2/decode.py
new file mode 100755
index 000000000..28d39de70
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/decode.py
@@ -0,0 +1,896 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from train import add_model_arguments
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) ctc-greedy-search. It only use CTC output and a sentence piece
+              model for decoding. It produces the same results with ctc-decoding.
+            - (2) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (3) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (6) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-path",
+        type=str,
+        default="data/lm/G_4_gram.pt",
+        help="""The n-gram LM dir for rescoring.
+        It should contain either lm_fname.pt or lm_fname.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--result-dir",
+        type=str,
+        default="conformer_ctc2/exp/results",
+        help="Directory to store results.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+    """
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            # parameters for decoding
+            "search_beam": 15,
+            "output_beam": 8,
+            "min_active_states": 10,
+            "max_active_states": 7000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def ctc_greedy_search(
+    ctc_probs: torch.Tensor,
+    mask: torch.Tensor,
+) -> List[List[int]]:
+    """Apply CTC greedy search
+    Args:
+      ctc_probs (torch.Tensor): (batch, max_len, num_bpe)
+      mask (torch.Tensor): (batch, max_len)
+    Returns:
+      best path result
+    """
+
+    _, max_index = ctc_probs.max(2)  # (B, maxlen)
+    max_index = max_index.masked_fill_(mask, 0)  # (B, maxlen)
+
+    ret_hyps = []
+    for hyp in max_index:
+        hyp = torch.unique_consecutive(hyp)
+        hyp = hyp[hyp > 0].tolist()
+        ret_hyps.append(hyp)
+    return ret_hyps
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        unk = bpe_model.decode(bpe_model.unk_id()).strip()
+        hyps = [[w for w in s.split() if w != unk] for s in hyps]
+        key = "ctc-decoding"
+
+        return {key: hyps}
+
+    if params.method == "ctc-greedy-search":
+        hyps = ctc_greedy_search(nnet_output, memory_key_padding_mask)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(hyps)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        unk = bpe_model.decode(bpe_model.unk_id()).strip()
+        hyps = [[w for w in s.split() if w != unk] for s in hyps]
+        key = "ctc-greedy-search"
+
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<unk>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [
+            [word_table[i] for i in ids if word_table[i] != "<unk>"] for ids in hyps
+        ]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method == "nbest":
+        best_path = nbest_decoding(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            use_double_scores=params.use_double_scores,
+            nbest_scale=params.nbest_scale,
+        )
+        key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [
+            [word_table[i] for i in ids if word_table[i] != "<unk>"] for ids in hyps
+        ]
+        return {key: hyps}
+
+    assert params.method in [
+        "1best",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "1best":
+        best_path_dict = one_best_decoding(
+            lattice=lattice,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.method}")
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [
+                [word_table[i] for i in ids if word_table[i] != "<unk>"] for ids in hyps
+            ]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+) -> None:
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.result_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.result_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.result_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main() -> None:
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_path = Path(args.lm_path)
+    args.result_dir = Path(args.result_dir)
+
+    args.result_dir.mkdir(exist_ok=True)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    if params.method in ("ctc-decoding", "ctc-greedy-search"):
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in ("nbest-rescoring", "whole-lattice-rescoring"):
+        assert params.lm_path.suffix in (".pt", ".txt")
+
+        if params.lm_path.is_file() and params.lm_path.suffix == ".pt":
+            logging.info(f"Loading pre-compiled {params.lm_path.name}")
+            d = torch.load(params.lm_path, map_location=device)
+            G = k2.Fsa.from_dict(d)
+        elif not params.lm_path.is_file() and params.lm_path.suffix == ".txt":
+            raise FileNotFoundError(f"No such language model file: '{params.lm_path}'")
+        else:
+            # here we pass only if LM filename ends with '.pt' and doesn't exist
+            # or if LM filename ends '.txt' and exists.
+            if (
+                not params.lm_path.is_file()
+                and params.lm_path.suffix == ".pt"
+                and not (
+                    params.lm_path.parent / f"{params.lm_path.stem}.fst.txt"
+                ).is_file()
+            ):
+                raise FileNotFoundError(
+                    f"No such language model file: '{params.lm_path}'\n"
+                    "'.fst.txt' representation of the language model was "
+                    "not found either."
+                )
+            else:
+                # whatever params.lm_path.name we got lm_name.pt or lm_name.fst.txt
+                # we are going to load lm_name.fst.txt here
+                params.lm_path = params.lm_path.parent / params.lm_path.name.replace(
+                    ".pt", ".fst.txt"
+                )
+                logging.info(f"Loading {params.lm_path.name}")
+                logging.warning("It may take 8 minutes.")
+                with open(params.lm_path) as f:
+                    first_word_disambig_id = lexicon.word_table["#0"]
+
+                    G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                    # G.aux_labels is not needed in later computations, so
+                    # remove it here.
+                    del G.aux_labels
+                    # CAUTION: The following line is crucial.
+                    # Arcs entering the back-off state have label equal to #0.
+                    # We have to change it to 0 here.
+                    G.labels[G.labels >= first_word_disambig_id] = 0
+                    # See https://github.com/k2-fsa/k2/issues/874
+                    # for why we need to set G.properties to None
+                    G.__dict__["_properties"] = None
+                    G = k2.Fsa.from_fsas([G]).to(device)
+                    G = k2.arc_sort(G)
+                    # Save a dummy value so that it can be loaded in C++.
+                    # See https://github.com/pytorch/pytorch/issues/67902
+                    # for why we need to do this.
+                    G.dummy = 1
+
+                    torch.save(
+                        G.as_dict(),
+                        params.lm_path.parent
+                        / params.lm_path.name.replace(".fst.txt", ".pt"),
+                    )
+
+        if params.method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.dim_model,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+
+    valid_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [valid_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+# when we import add_model_arguments from train.py
+# we enforce torch.set_num_interop_threads(1) in it,
+# so we ended up with setting num_interop_threads to one
+# two times: in train.py and decode.py which cause an error,
+# that is why added an additional if statement.
+if torch.get_num_interop_threads() != 1:
+    torch.set_num_interop_threads(1)
+
+# The flag below controls whether to allow TF32 on matmul. This flag defaults to False
+# in PyTorch 1.12 and later.
+torch.backends.cuda.matmul.allow_tf32 = True
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/conformer_ctc2/export.py b/egs/tedlium3/ASR/conformer_ctc2/export.py
new file mode 100755
index 000000000..009bea230
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/export.py
@@ -0,0 +1,294 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Behavox LLC (Author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./conformer_ctc2/export.py \
+  --exp-dir ./conformer_ctc2/exp \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `conformer_ctc2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tedlium3/ASR
+    ./conformer_ctc2/decode.py \
+        --exp-dir ./conformer_ctc2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from conformer import Conformer
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help=(
+            "Number of checkpoints to average. Automatically select "
+            "consecutive checkpoints before the checkpoint specified by "
+            "'--epoch' and '--iter'"
+        ),
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help=(
+            "Whether to load averaged model. Currently it only supports "
+            "using --epoch. If True, it would decode with the averaged model "
+            "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+            "Actually only the models with epoch number of `epoch-avg` and "
+            "`epoch` are loaded for averaging. "
+        ),
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+    """
+    # parameters for conformer
+    params = AttributeDict({"subsampling_factor": 4, "feature_dim": 80})
+    return params
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info(params)
+
+    logging.info("About to create model")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.dim_model,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                "Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tedlium3/ASR/conformer_ctc2/label_smoothing.py b/egs/tedlium3/ASR/conformer_ctc2/label_smoothing.py
new file mode 120000
index 000000000..e9d239fff
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/label_smoothing.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/local b/egs/tedlium3/ASR/conformer_ctc2/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/lstmp.py b/egs/tedlium3/ASR/conformer_ctc2/lstmp.py
new file mode 120000
index 000000000..b82e115fc
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/lstmp.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/optim.py b/egs/tedlium3/ASR/conformer_ctc2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/scaling.py b/egs/tedlium3/ASR/conformer_ctc2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/scaling_converter.py b/egs/tedlium3/ASR/conformer_ctc2/scaling_converter.py
new file mode 120000
index 000000000..db93d155b
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/subsampling.py b/egs/tedlium3/ASR/conformer_ctc2/subsampling.py
new file mode 120000
index 000000000..8c91f2336
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc2/subsampling.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/train.py b/egs/tedlium3/ASR/conformer_ctc2/train.py
new file mode 100755
index 000000000..fc3e3b2d9
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/train.py
@@ -0,0 +1,1061 @@
+#!/usr/bin/env python3
+# Copyright    2022  Behavox LLC.        (authors: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir conformer_ctc/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./conformer_ctc/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir conformer_ctc/exp \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser) -> None:
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward module dimension of the conformer model.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer multiheadattention modules.",
+    )
+
+    parser.add_argument(
+        "--dim-model",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer model.",
+    )
+
+
+def get_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt" and "bpe.model"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="Number of epochs that affects how rapidly the learning rate decreases.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for ctc loss
+            "beam_size": 10,
+            "reduction": "none",
+            "use_double_scores": True,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: torch.nn.Module,
+    model_avg: torch.nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that is used for training.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[torch.nn.Module, DDP],
+    model_avg: Optional[torch.nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used for training.
+      scheduler:
+        The learning rate scheduler used for training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[torch.nn.Module, DDP],
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(
+            feature, supervisions, warmup=warmup
+        )
+
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=params.subsampling_factor
+        )
+
+        token_ids = convert_texts_into_ids(texts, graph_compiler.sp)
+        decoding_graph = graph_compiler.compile(token_ids)
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            nnet_output,
+            supervision_segments,
+            allow_truncate=params.subsampling_factor - 1,
+        )
+
+        ctc_loss = k2.ctc_loss(
+            decoding_graph=decoding_graph,
+            dense_fsa_vec=dense_fsa_vec,
+            output_beam=params.beam_size,
+            reduction=params.reduction,
+            use_double_scores=params.use_double_scores,
+        )
+
+        if params.att_rate > 0.0:
+            with torch.set_grad_enabled(is_training):
+                mmodel = model.module if hasattr(model, "module") else model
+                # Note: We need to generate an unsorted version of token_ids
+                # `encode_supervisions()` called above sorts text, but
+                # encoder_memory and memory_mask are not sorted, so we
+                # use an unsorted version `supervisions["text"]` to regenerate
+                # the token_ids
+                #
+                # See https://github.com/k2-fsa/icefall/issues/97
+                # for more details
+                unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+                att_loss = mmodel.decoder_forward(
+                    encoder_memory,
+                    memory_mask,
+                    token_ids=unsorted_token_ids,
+                    sos_id=graph_compiler.sos_id,
+                    eos_id=graph_compiler.eos_id,
+                    warmup=warmup,
+                )
+        else:
+            att_loss = torch.tensor([0])
+
+        ctc_loss_is_finite = torch.isfinite(ctc_loss)
+        att_loss_is_finite = torch.isfinite(att_loss)
+        if torch.any(~ctc_loss_is_finite) or torch.any(~att_loss_is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"ctc_loss: {ctc_loss}\n"
+                f"att_loss: {att_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=graph_compiler.sp)
+            ctc_loss = ctc_loss[ctc_loss_is_finite]
+            att_loss = att_loss[att_loss_is_finite]
+
+            # If the batch contains more than 10 utterances AND
+            # if either all ctc_loss or att_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~ctc_loss_is_finite) or torch.all(~att_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        ctc_loss = ctc_loss.sum()
+        att_loss = att_loss.sum()
+
+        if params.att_rate > 0.0:
+            loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+        else:
+            loss = ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    # info["frames"] is an approximate number for two reasons:
+    # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+    # (2) If some utterances in the batch lead to inf/nan loss, they
+    #     are filtered out.
+    info["frames"] = (
+        torch.div(feature_lens, params.subsampling_factor, rounding_mode="floor")
+        .sum()
+        .item()
+    )
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate > 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[torch.nn.Module, DDP],
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch in valid_dl:
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[torch.nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[torch.nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=graph_compiler.sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    if "lang_bpe" not in str(params.lang_dir):
+        raise ValueError(
+            f"Unsupported type of lang dir (we expected it to have "
+            f"'lang_bpe' in its name): {params.lang_dir}"
+        )
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.dim_model,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[torch.nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.Eve(model.parameters(), lr=params.initial_lr)
+    scheduler = optim.Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and checkpoints.get("optimizer") is not None:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if checkpoints and checkpoints.get("scheduler") is not None:
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    tedlium = TedLiumAsrDataModule(args)
+
+    train_cuts = tedlium.train_cuts()
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = tedlium.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = tedlium.dev_cuts()
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+
+    if (
+        params.start_epoch <= 1
+        and params.start_batch <= 0
+        and not params.print_diagnostics
+    ):
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+        train_dl.dataset.epoch = epoch - 1
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[torch.nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=graph_compiler.sp)
+            raise
+
+
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# The flag below controls whether to allow TF32 on matmul. This flag defaults to False
+# in PyTorch 1.12 and later.
+torch.backends.cuda.matmul.allow_tf32 = True
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/conformer_ctc2/transformer.py b/egs/tedlium3/ASR/conformer_ctc2/transformer.py
new file mode 100644
index 000000000..9dbf32e48
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/transformer.py
@@ -0,0 +1,1093 @@
+# Copyright    2021  University of Chinese Academy of Sciences (author: Han Zhu)
+# Copyright    2022  Xiaomi Corp.                              (author: Quandong Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from attention import MultiheadAttention
+from combiner import RandomCombine
+from label_smoothing import LabelSmoothingLoss
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledEmbedding,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        aux_layer_period: int = 3,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            the input dimension of the model.
+          num_classes:
+            the output dimension of the model.
+          subsampling_factor:
+            number of output frames is num_in_frames // subsampling_factor;
+            currently, subsampling_factor MUST be 4.
+          d_model:
+            attention dimension.
+          nhead:
+            number of heads in multi-head attention;
+            must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            the output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            number of encoder layers.
+          num_decoder_layers:
+            number of decoder layers.
+          dropout:
+            dropout in encoder/decoder.
+          layer_dropout:
+            layer-dropout rate.
+          aux_layer_period:
+            determines the auxiliary encoder layers.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+        # aux_layers from 1/3
+        self.encoder = TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            aux_layers=list(
+                range(
+                    num_encoder_layers // 3,
+                    num_encoder_layers - 1,
+                    aux_layer_period,
+                )
+            ),
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), ScaledLinear(d_model, num_classes, bias=True)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = ScaledEmbedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+            )
+
+            self.decoder = TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                aux_layers=[],
+            )
+
+            self.decoder_output_layer = ScaledLinear(
+                d_model, self.decoder_num_class, bias=True
+            )
+
+            self.decoder_criterion = LabelSmoothingLoss(reduction="none")
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        supervision: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, S, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+          warmup:
+            a floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up". It is used
+            to turn modules on sequentially.
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, S, C)
+            - Encoder output with shape (S, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, S).
+              It is None if `supervision` is None.
+        """
+
+        encoder_memory, memory_key_padding_mask = self.run_encoder(
+            x, supervision, warmup
+        )
+
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, S, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+          warmup:
+            a floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up". It is used
+            to turn modules on sequentially.
+
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (S, N, C)
+            - encoder padding mask, with shape (N, S).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, S, C) -> (S, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask, warmup=warmup)  # (S, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            the output tensor from the transformer encoder;
+            its shape is (S, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, S, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (S, N, C) -> (N, S, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, S, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder of shape (S, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder of shape (N, S).
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+          warmup:
+            a floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up". It is used
+            to turn modules on sequentially.
+
+        Returns:
+          A scalar, the **sum** of label smoothing loss over utterances
+          in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+            warmup=warmup,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder of shape (S, N, C).
+          memory_key_padding_mask:
+            The padding mask from the encoder of shape (N, S).
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+          warmup:
+            a floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up". It is used
+            to turn modules on sequentially.
+
+        Returns:
+          A 2-D tensor of shape (len(token_ids), max_token_length)
+          representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, С) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+            warmup=warmup,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+
+    Example:
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        bypass_scale: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        """
+        Args:
+          d_model:
+            the number of expected features in the input (required).
+          nhead:
+            the number of heads in the multiheadattention models (required).
+          dim_feedforward:
+            the dimension of the feedforward network model (default=2048).
+          dropout:
+            the dropout value (default=0.1).
+          bypass_scale:
+            a scale on the layer's output, used in bypass (resnet-type) skip-connection;
+            when the layer is bypassed the final output will be a
+            weighted sum of the layer's input and layer's output with weights
+            (1.0-bypass_scale) and bypass_scale correspondingly (default=0.1).
+          layer_dropout:
+            the probability to bypass the layer (default=0.075).
+        """
+
+        super().__init__()
+
+        if bypass_scale < 0.0 or bypass_scale > 1.0:
+            raise ValueError("bypass_scale should be between 0.0 and 1.0")
+
+        if layer_dropout < 0.0 or layer_dropout > 1.0:
+            raise ValueError("layer_dropout should be between 0.0 and 1.0")
+
+        self.bypass_scale = bypass_scale
+        self.layer_dropout = layer_dropout
+
+        self.self_attn = MultiheadAttention(d_model, nhead)
+        # Implementation of Feedforward model
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+          src:
+            the sequence to the encoder layer of shape (S, N, C) (required).
+          src_mask:
+            the mask for the src sequence of shape (S, S) (optional).
+          src_key_padding_mask:
+            the mask for the src keys per batch of shape (N, S) (optional)
+          warmup:
+            controls selective bypass of layers; if < 1.0, we will
+            bypass the layer more frequently (default=1.0).
+
+        Returns:
+          Output tensor of the shape (S, N, C), where
+          S is the source sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        src_orig = src
+
+        warmup_scale = min(self.bypass_scale + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else self.bypass_scale
+            )
+        else:
+            alpha = 1.0
+
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1.0 - alpha) * src_orig
+
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """Modified from torch.nn.TransformerDecoderLayer.
+
+    Example:
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        bypass_scale: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+
+        """
+        Args:
+          d_model:
+            the number of expected features in the input (required).
+          nhead:
+            the number of heads in the multiheadattention models (required).
+          dim_feedforward:
+            the dimension of the feedforward network model (default=2048).
+          dropout:
+            the dropout value (default=0.1).
+          bypass_scale:
+            a scale on the layer's output, used in bypass (resnet-type) skip-connection;
+            when the layer is bypassed, the final output will be a
+            weighted sum of the layer's input and layer's output with weights
+            (1.0-bypass_scale) and bypass_scale correspondingly (default=0.1).
+          layer_dropout:
+            the probability to bypass the layer (default=0.075).
+        """
+
+        super().__init__()
+
+        if bypass_scale < 0.0 or bypass_scale > 1.0:
+            raise ValueError("bypass_scale should be between 0.0 and 1.0")
+
+        if layer_dropout < 0.0 or layer_dropout > 1.0:
+            raise ValueError("layer_dropout should be between 0.0 and 1.0")
+
+        self.bypass_scale = bypass_scale
+        self.layer_dropout = layer_dropout
+
+        self.self_attn = MultiheadAttention(d_model, nhead)
+        self.src_attn = MultiheadAttention(d_model, nhead)
+
+        # Implementation of Feedforward model
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer of shape (T, N, C) (required).
+          memory:
+            the sequence from the last layer of the encoder of shape (S, N, C) (required).
+          tgt_mask:
+            the mask for the tgt sequence of shape (T, T) (optional).
+          memory_mask:
+            the mask for the memory sequence of shape (T, S) (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch of shape (N, T) (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch of shape (N, S) (optional).
+          warmup: controls selective bypass of layers; if < 1.0, we will
+            bypass the layer more frequently (default=1.0).
+
+        Returns:
+          Output tensor of the shape (T, N, C), where
+          S is the source sequence length,
+          T is the target sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        tgt_orig = tgt
+
+        warmup_scale = min(self.bypass_scale + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else self.bypass_scale
+            )
+        else:
+            alpha = 1.0
+
+        tgt_att = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(tgt_att)
+
+        src_att = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(src_att)
+
+        tgt = tgt + self.dropout(self.feed_forward(tgt))
+
+        tgt = self.norm_final(self.balancer(tgt))
+
+        if alpha != 1.0:
+            tgt = alpha * tgt + (1.0 - alpha) * tgt_orig
+
+        return tgt
+
+
+class TransformerEncoder(nn.Module):
+    """TransformerEncoder is a stack of N encoder layers
+
+    Examples:
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        aux_layers: List[int],
+    ) -> None:
+        """
+        Args:
+          encoder_layer:
+            an instance of the TransformerEncoderLayer() class (required).
+          num_layers:
+            the number of sub-encoder-layers in the encoder (required).
+          aux_layers:
+            list of indexes of sub-encoder-layers outputs to be combined (required).
+        """
+
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert len(set(aux_layers)) == len(aux_layers)
+
+        assert num_layers - 1 not in aux_layers
+        self.aux_layers = aux_layers + [num_layers - 1]
+
+        self.combiner = RandomCombine(
+            num_inputs=len(self.aux_layers),
+            final_weight=0.5,
+            pure_prob=0.333,
+            stddev=2.0,
+        )
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """Pass the input through the encoder layers in turn.
+
+        Args:
+          src:
+            the input to the encoder of shape (S, N, C) (required).
+          mask:
+            the mask for the src sequence of shape (S, S) (optional).
+          src_key_padding_mask:
+            the mask for the src keys per batch of shape (N, S) (optional).
+          warmup:
+            controls selective bypass of layer; if < 1.0, we will
+            bypass the layer more frequently (default=1.0).
+
+        Returns:
+          Output tensor of the shape (S, N, C), where
+          S is the source sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        output = src
+
+        outputs = []
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+            if i in self.aux_layers:
+                outputs.append(output)
+
+        output = self.combiner(outputs)
+
+        return output
+
+
+class TransformerDecoder(nn.Module):
+    """TransformerDecoder is a stack of N decoder layers
+
+    Examples:
+        >>> decoder_layer = TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> transformer_decoder = TransformerDecoder(decoder_layer, num_layers=6)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = transformer_decoder(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        decoder_layer: nn.Module,
+        num_layers: int,
+        aux_layers: List[int],
+    ) -> None:
+        """
+        Args:
+          decoder_layer:
+            an instance of the TransformerDecoderLayer() class (required).
+          num_layers:
+            the number of decoder layers in the decoder (required).
+          aux_layers:
+            list of indexes of decoder layer outputs to be combined (required).
+        """
+
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(decoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert len(set(aux_layers)) == len(aux_layers)
+
+        assert num_layers - 1 not in aux_layers
+        self.aux_layers = aux_layers + [num_layers - 1]
+
+        self.combiner = RandomCombine(
+            num_inputs=len(self.aux_layers),
+            final_weight=0.5,
+            pure_prob=0.333,
+            stddev=2.0,
+        )
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """Pass the input (and mask) through the decoder layers in turn.
+
+        Args:
+          tgt:
+            the sequence to the decoder of shape (T, N, C) (required).
+          memory:
+            the sequence from the last layer of the encoder of shape (S, N, C) (required).
+          tgt_mask:
+            the mask for the tgt sequence of shape (T, T) (optional).
+          memory_mask:
+            the mask for the memory sequence of shape (T, S) (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch of shape (N, T)  (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch of shape (N, S) (optional).
+          warmup:
+            controls selective bypass of layer; if < 1.0, we will
+            bypass the layer more frequently (default=1.0).
+
+        Returns:
+          Output tensor of the shape (T, N, C), where
+          S is the source sequence length,
+          T is the target sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        output = tgt
+
+        outputs = []
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                memory,
+                tgt_mask=tgt_mask,
+                memory_mask=memory_mask,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                memory_key_padding_mask=memory_key_padding_mask,
+                warmup=warmup,
+            )
+
+            if i in self.aux_layers:
+                outputs.append(output)
+
+        output = self.combiner(outputs)
+
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note:
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model: Embedding dimension.
+          dropout: Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+            T is the target sequence length,
+            N is the batch size,
+            C is the feature number.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x: Input of shape is (N, T, C)
+
+        Returns:
+          A tensor of the same shape (N, T, C),
+          T is the target sequence length,
+          N is the batch size,
+          C is the feature number.
+
+        """
+        self.extend_pe(x)
+        x = x + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO:
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+      Mask tensor of dimension (batch_size, input_length),
+      True denotes the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+        A bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask tensor of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-lists of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-lists, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/tedlium3/ASR/local/__init__.py b/egs/tedlium3/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tedlium3/ASR/local/compile_hlg.py b/egs/tedlium3/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/tedlium3/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/local/compute_fbank_musan.py b/egs/tedlium3/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/tedlium3/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/local/compute_fbank_tedlium.py b/egs/tedlium3/ASR/local/compute_fbank_tedlium.py
new file mode 100755
index 000000000..733ebf235
--- /dev/null
+++ b/egs/tedlium3/ASR/local/compute_fbank_tedlium.py
@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 	       2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This file computes fbank features of the TedLium3 dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_tedlium():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+
+    prefix = "tedlium"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cur_num_jobs = num_jobs if ex is None else 80
+            cur_num_jobs = min(cur_num_jobs, len(cut_set))
+
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=cur_num_jobs,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            # Split long cuts into many short and un-overlapping cuts
+            cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_tedlium()
diff --git a/egs/tedlium3/ASR/local/convert_transcript_words_to_bpe_ids.py b/egs/tedlium3/ASR/local/convert_transcript_words_to_bpe_ids.py
new file mode 100644
index 000000000..19ba8d24b
--- /dev/null
+++ b/egs/tedlium3/ASR/local/convert_transcript_words_to_bpe_ids.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+
+# Copyright    2022 Xiaomi Corporation  (Author: Mingshuang Luo)
+"""
+Convert a transcript based on words to a list of BPE ids.
+
+For example, if we use 2 as the encoding id of <unk>
+Note: it, inserts a space token before each <unk>
+
+texts = ['this is a <unk> day']
+spm_ids = [[38, 33, 6, 15, 2, 316]]
+
+texts = ['<unk> this is a sunny day']
+spm_ids = [[15, 2, 38, 33, 6, 118, 11, 11, 21, 316]]
+
+texts = ['<unk>']
+spm_ids = [[15, 2]]
+
+"""
+
+import argparse
+import logging
+from typing import List
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--texts", type=List[str], help="The input transcripts list.")
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    return parser.parse_args()
+
+
+def convert_texts_into_ids(
+    texts: List[str],
+    sp: spm.SentencePieceProcessor,
+) -> List[List[int]]:
+    """
+    Args:
+      texts:
+        A string list of transcripts, such as ['Today is Monday', 'It's sunny'].
+      sp:
+        A sentencepiece BPE model.
+    Returns:
+      Return an integer list of bpe ids.
+    """
+    y = []
+    for text in texts:
+        if "<unk>" in text:
+            id_segments = sp.encode(text.split("<unk>"), out_type=int)
+
+            y_ids = []
+            for i in range(len(id_segments)):
+                y_ids += id_segments[i]
+                if i < len(id_segments) - 1:
+                    y_ids += [sp.piece_to_id("▁"), sp.unk_id()]
+        else:
+            y_ids = sp.encode(text, out_type=int)
+        y.append(y_ids)
+
+    return y
+
+
+def main():
+    args = get_args()
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    y = convert_texts_into_ids(texts=args.texts, sp=sp)
+
+    logging.info(f"The input texts: {args.texts}")
+    logging.info(f"The encoding ids: {y}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/local/display_manifest_statistics.py b/egs/tedlium3/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..52e152389
--- /dev/null
+++ b/egs/tedlium3/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						   Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    path = "./data/fbank/tedlium_cuts_train.jsonl.gz"
+    path = "./data/fbank/tedlium_cuts_dev.jsonl.gz"
+    path = "./data/fbank/tedlium_cuts_test.jsonl.gz"
+
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+## train
+Cuts count: 804789
+Total duration (hours): 1370.6
+Speech duration (hours): 1370.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.1
+std     3.1
+min     0.5
+25%     3.7
+50%     6.0
+75%     8.3
+99.5%   14.9
+99.9%   16.6
+max     33.3
+
+## dev
+Cuts count: 507
+Total duration (hours): 1.6
+Speech duration (hours): 1.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    11.3
+std     5.7
+min     0.5
+25%     7.5
+50%     10.6
+75%     14.4
+99.5%   29.8
+99.9%   37.7
+max     39.9
+
+## test
+Cuts count: 1155
+Total duration (hours): 2.6
+Speech duration (hours): 2.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    8.2
+std     4.3
+min     0.3
+25%     4.6
+50%     8.2
+75%     10.9
+99.5%   22.1
+99.9%   26.7
+max     32.5
+"""
diff --git a/egs/tedlium3/ASR/local/prepare_lang_bpe.py b/egs/tedlium3/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/tedlium3/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/local/prepare_transcripts.py b/egs/tedlium3/ASR/local/prepare_transcripts.py
new file mode 100755
index 000000000..d4ccdd1e3
--- /dev/null
+++ b/egs/tedlium3/ASR/local/prepare_transcripts.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (author: Mingshuang Luo)
+# Copyright    2022  Behavox LLC.        (author: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes input text file and removes all words
+that iclude any character out of English alphabet.
+
+"""
+import argparse
+import logging
+import re
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input-text-path",
+        type=str,
+        help="Input text file path.",
+    )
+    parser.add_argument(
+        "--output-text-path",
+        type=str,
+        help="Output text file path.",
+    )
+
+    return parser.parse_args()
+
+
+def prepare_transcripts(input_text_path: Path, output_text_path: Path) -> None:
+    """
+    Args:
+      input_text_path:
+        The input data text file path, e.g., data/lang/train_orig.txt.
+      output_text_path:
+        The output data text file path, e.g., data/lang/train.txt.
+
+    Return:
+      Saved text file in output_text_path.
+    """
+
+    foreign_chr_check = re.compile(r"[^a-z']")
+
+    logging.info(f"Loading {input_text_path.name}")
+    with open(input_text_path, "r", encoding="utf8") as f:
+        texts = {t.rstrip("\n") for t in f}
+
+    texts = {
+        " ".join([w for w in t.split() if foreign_chr_check.search(w) is None])
+        for t in texts
+    }
+
+    with open(output_text_path, "w+", encoding="utf8") as f:
+        for t in texts:
+            f.write(f"{t}\n")
+
+
+def main() -> None:
+    args = get_args()
+    input_text_path = Path(args.input_text_path)
+    output_text_path = Path(args.output_text_path)
+
+    logging.info(f"Generating {output_text_path.name}")
+    prepare_transcripts(input_text_path, output_text_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/tedlium3/ASR/local/prepare_words.py b/egs/tedlium3/ASR/local/prepare_words.py
new file mode 100755
index 000000000..a37d0f08f
--- /dev/null
+++ b/egs/tedlium3/ASR/local/prepare_words.py
@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+# Copyright    2022  Behavox LLC.        (authors: Daniil Kulko)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input supervisions json dir "data/manifests"
+consisting of tedlium_supervisions_train.json and does the following:
+
+1. Generate words.txt.
+
+"""
+import argparse
+import logging
+import re
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="Output directory.",
+    )
+
+    return parser.parse_args()
+
+
+def prepare_words(lang_dir: str) -> None:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang.
+
+    Return:
+      The words.txt file.
+    """
+
+    words_orig_path = Path(lang_dir) / "words_orig.txt"
+    words_path = Path(lang_dir) / "words.txt"
+
+    foreign_chr_check = re.compile(r"[^a-z']")
+
+    logging.info(f"Loading {words_orig_path.name}")
+    with open(words_orig_path, "r", encoding="utf8") as f:
+        words = {w for w_compl in f for w in w_compl.strip("-\n").split("_")}
+    words = {w for w in words if foreign_chr_check.search(w) is None and w != ""}
+    words.add("<unk>")
+    words = ["<eps>", "!SIL"] + sorted(words) + ["#0", "<s>", "</s>"]
+
+    with open(words_path, "w+", encoding="utf8") as f:
+        for idx, word in enumerate(words):
+            f.write(f"{word} {idx}\n")
+
+
+def main() -> None:
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    logging.info("Generating words.txt")
+    prepare_words(lang_dir)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/tedlium3/ASR/local/train_bpe_model.py b/egs/tedlium3/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/tedlium3/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/prepare.sh b/egs/tedlium3/ASR/prepare.sh
new file mode 100755
index 000000000..2f58ca0ee
--- /dev/null
+++ b/egs/tedlium3/ASR/prepare.sh
@@ -0,0 +1,212 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=0
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/tedlium3
+#      You can find data, doc, legacy, LM, etc, inside it.
+#      You can download them from https://www.openslr.org/51
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  5000
+  2000
+  1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/tedlium3,
+  # you can create a symlink
+  #
+  # ln -sfv /path/to/tedlium3 $dl_dir/tedlium3
+  #
+  if [ ! -d $dl_dir/tedlium3 ]; then
+    lhotse download tedlium $dl_dir
+    mv $dl_dir/TEDLIUM_release-3 $dl_dir/tedlium3
+  fi
+
+  # Download big and small 4 gram lanuage models
+  if [ ! -d $dl_dir/lm ]; then
+    wget --continue http://kaldi-asr.org/models/5/4gram_small.arpa.gz -P $dl_dir/lm
+    wget --continue http://kaldi-asr.org/models/5/4gram_big.arpa.gz -P $dl_dir/lm
+    gzip -d $dl_dir/lm/4gram_small.arpa.gz $dl_dir/lm/4gram_big.arpa.gz
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #ln -sfv /path/to/musan $dl_dir/musan
+
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare tedlium3 manifests"
+  if [ ! -f data/manifests/.tedlium3.done ]; then
+    # We assume that you have downloaded the tedlium3 corpus
+    # to $dl_dir/tedlium3
+    mkdir -p data/manifests
+    lhotse prepare tedlium $dl_dir/tedlium3 data/manifests
+    touch data/manifests/.tedlium3.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifests"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -e data/manifests/.musan.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for tedlium3"
+
+  if [ ! -e data/fbank/.tedlium3.done ]; then
+    mkdir -p data/fbank
+
+    python3 ./local/compute_fbank_tedlium.py
+
+    gunzip -c data/fbank/tedlium_cuts_train.jsonl.gz | shuf | \
+    gzip -c > data/fbank/tedlium_cuts_train-shuf.jsonl.gz
+    mv data/fbank/tedlium_cuts_train-shuf.jsonl.gz \
+       data/fbank/tedlium_cuts_train.jsonl.gz
+
+    touch data/fbank/.tedlium3.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -e data/fbank/.musan.done ]; then
+    mkdir -p data/fbank
+    python3 ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare BPE train data and set of words"
+  lang_dir=data/lang
+  mkdir -p $lang_dir
+
+  if [ ! -f $lang_dir/train.txt ]; then
+    gunzip -c $dl_dir/tedlium3/LM/*.en.gz | sed 's: <\/s>::g' > $lang_dir/train_orig.txt
+
+    ./local/prepare_transcripts.py \
+      --input-text-path $lang_dir/train_orig.txt \
+      --output-text-path $lang_dir/train.txt
+  fi
+
+  if [ ! -f $lang_dir/words.txt ]; then
+
+    awk '{print $1}' $dl_dir/tedlium3/TEDLIUM.152k.dic |
+    sed 's:([0-9])::g' | sort | uniq > $lang_dir/words_orig.txt
+
+    ./local/prepare_words.py --lang-dir $lang_dir
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang/words.txt $lang_dir
+
+    ./local/train_bpe_model.py \
+      --lang-dir $lang_dir \
+      --vocab-size $vocab_size \
+      --transcript data/lang/train.txt
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir --oov "<unk>"
+    fi
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_4_gram_small.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      --max-arpa-warnings=-1 \
+      $dl_dir/lm/4gram_small.arpa > data/lm/G_4_gram_small.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram_big.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      --max-arpa-warnings=-1 \
+      $dl_dir/lm/4gram_big.arpa > data/lm/G_4_gram_big.fst.txt
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compile HLG"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/HLG.pt ]; then
+      ./local/compile_hlg.py \
+        --lang-dir $lang_dir \
+        --lm G_4_gram_small
+    fi
+  done
+fi
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/__init__.py b/egs/tedlium3/ASR/pruned_transducer_stateless/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/asr_datamodule.py b/egs/tedlium3/ASR/pruned_transducer_stateless/asr_datamodule.py
new file mode 120000
index 000000000..49b2ee483
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/beam_search.py b/egs/tedlium3/ASR/pruned_transducer_stateless/beam_search.py
new file mode 120000
index 000000000..7f9f6263f
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless/beam_search.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/conformer.py b/egs/tedlium3/ASR/pruned_transducer_stateless/conformer.py
new file mode 120000
index 000000000..8be0dc864
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/decode.py b/egs/tedlium3/ASR/pruned_transducer_stateless/decode.py
new file mode 100755
index 000000000..abba9d403
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/decode.py
@@ -0,0 +1,519 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            decoding_graph=decoding_graph,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+    dev_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    dev_dl = tedlium.valid_dataloaders(dev_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/decoder.py b/egs/tedlium3/ASR/pruned_transducer_stateless/decoder.py
new file mode 120000
index 000000000..206384eaa
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless/decoder.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/encoder_interface.py b/egs/tedlium3/ASR/pruned_transducer_stateless/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/export.py b/egs/tedlium3/ASR/pruned_transducer_stateless/export.py
new file mode 100644
index 000000000..aa22f82ec
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/export.py
@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+# 					                         Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless/export.py \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 29 \
+  --avg 13
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tedlium3/ASR
+    ./pruned_transducer_stateless/decode.py \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/joiner.py b/egs/tedlium3/ASR/pruned_transducer_stateless/joiner.py
new file mode 120000
index 000000000..b3d677eb5
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/local b/egs/tedlium3/ASR/pruned_transducer_stateless/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/model.py b/egs/tedlium3/ASR/pruned_transducer_stateless/model.py
new file mode 120000
index 000000000..6b78aed54
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless/model.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/pretrained.py b/egs/tedlium3/ASR/pruned_transducer_stateless/pretrained.py
new file mode 100644
index 000000000..8a89c3578
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/pretrained.py
@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		 2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless/exp/pretrained.pt is generated by
+./pruned_transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and modified_beam_search ",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/subsampling.py b/egs/tedlium3/ASR/pruned_transducer_stateless/subsampling.py
new file mode 120000
index 000000000..fd7ca8b30
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/subsampling.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/test_decoder.py b/egs/tedlium3/ASR/pruned_transducer_stateless/test_decoder.py
new file mode 100755
index 000000000..b97bf6150
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/test_decoder.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						   Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/tedlium3/ASR
+    python ./pruned_transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    unk_id = 2
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        unk_id=unk_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, vocab_size)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, vocab_size)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/train.py b/egs/tedlium3/ASR/pruned_transducer_stateless/train.py
new file mode 100755
index 000000000..170f37767
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/train.py
@@ -0,0 +1,767 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12350,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    unk_id = params.unk_id
+    y = convert_texts_into_ids(texts, unk_id, sp=sp)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    tedlium = TedLiumAsrDataModule(args)
+
+    train_cuts = tedlium.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 17 seconds
+        return 1.0 <= c.duration <= 17.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = tedlium.train_dataloaders(train_cuts)
+    valid_cuts = tedlium.dev_cuts()
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/transformer.py b/egs/tedlium3/ASR/pruned_transducer_stateless/transformer.py
new file mode 120000
index 000000000..214afed39
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/shared b/egs/tedlium3/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/tedlium3/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/README.md b/egs/tedlium3/ASR/transducer_stateless/README.md
new file mode 100644
index 000000000..9b6ed62f1
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/README.md
@@ -0,0 +1,20 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+You can use the following command to start the training:
+
+```bash
+cd egs/tedlium3/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 300
+```
diff --git a/egs/tedlium3/ASR/transducer_stateless/__init__.py b/egs/tedlium3/ASR/transducer_stateless/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py b/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py
new file mode 100644
index 000000000..d4a9e4bc9
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py
@@ -0,0 +1,366 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2021  Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class TedLiumAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. TEDLium3 dev
+    and test).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset.",
+        )
+
+    def train_dataloaders(
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=10,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                    max_frames_mask_fraction=0.15,
+                    p=0.9,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to get Musan cuts")
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        logging.info("About to create train dataloader")
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts_test: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        if self.args.on_the_fly_feats:
+            test = K2SpeechRecognitionDataset(
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            test = K2SpeechRecognitionDataset(
+                return_cuts=self.args.return_cuts,
+            )
+
+        test_sampler = DynamicBucketingSampler(
+            cuts_test,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=test_sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tedlium_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_test.jsonl.gz")
diff --git a/egs/tedlium3/ASR/transducer_stateless/beam_search.py b/egs/tedlium3/ASR/transducer_stateless/beam_search.py
new file mode 100644
index 000000000..1f99edaf3
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/beam_search.py
@@ -0,0 +1,539 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import torch
+from model import Transducer
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    unk_id = model.decoder.unk_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size, device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out, encoder_out_len, decoder_out_len
+        )
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y != blank_id and y != unk_id:
+            hyp.append(y)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def run_decoder(
+    ys: List[int],
+    model: Transducer,
+    decoder_cache: Dict[str, torch.Tensor],
+) -> torch.Tensor:
+    """Run the neural decoder model for a given hypothesis.
+
+    Args:
+      ys:
+        The current hypothesis.
+      model:
+        The transducer model.
+      decoder_cache:
+        Cache to save computations.
+    Returns:
+      Return a 1-D tensor of shape (decoder_out_dim,) containing
+      output of `model.decoder`.
+    """
+    context_size = model.decoder.context_size
+    key = "_".join(map(str, ys[-context_size:]))
+    if key in decoder_cache:
+        return decoder_cache[key]
+
+    device = model.device
+
+    decoder_input = torch.tensor([ys[-context_size:]], device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_cache[key] = decoder_out
+
+    return decoder_out
+
+
+def run_joiner(
+    key: str,
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    decoder_out: torch.Tensor,
+    encoder_out_len: torch.Tensor,
+    decoder_out_len: torch.Tensor,
+    joint_cache: Dict[str, torch.Tensor],
+):
+    """Run the joint network given outputs from the encoder and decoder.
+
+    Args:
+      key:
+        A key into the `joint_cache`.
+      model:
+        The transducer model.
+      encoder_out:
+        A tensor of shape (1, 1, encoder_out_dim).
+      decoder_out:
+        A tensor of shape (1, 1, decoder_out_dim).
+      encoder_out_len:
+        A tensor with value [1].
+      decoder_out_len:
+        A tensor with value [1].
+      joint_cache:
+        A dict to save computations.
+    Returns:
+      Return a tensor from the output of log-softmax.
+      Its shape is (vocab_size,).
+    """
+    if key in joint_cache:
+        return joint_cache[key]
+
+    logits = model.joiner(
+        encoder_out,
+        decoder_out,
+        encoder_out_len,
+        decoder_out_len,
+    )
+
+    # TODO(fangjun): Scale the blank posterior
+    log_prob = logits.log_softmax(dim=-1)
+    # log_prob is (1, 1, 1, vocab_size)
+
+    log_prob = log_prob.squeeze()
+    # Now log_prob is (vocab_size,)
+
+    joint_cache[key] = log_prob
+
+    return log_prob
+
+
+def modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = model.decoder.unk_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # current_encoder_out is of shape (1, 1, encoder_out_dim)
+        # fmt: on
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        # decoder_output is of shape (num_hyps, 1, decoder_output_dim)
+
+        current_encoder_out = current_encoder_out.expand(decoder_out.size(0), 1, -1)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(decoder_out.size(0)),
+            decoder_out_len.expand(decoder_out.size(0)),
+        )
+        # logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        topk_hyp_indexes = topk_hyp_indexes.tolist()
+        topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
+            new_ys = hyp.ys[:]
+            new_token = topk_token_indexes[i]
+            if new_token != blank_id and new_token != unk_id:
+                new_ys.append(new_token)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+            B.add(new_hyp)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    return ys
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = model.decoder.unk_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            decoder_out = run_decoder(
+                ys=y_star.ys, model=model, decoder_cache=decoder_cache
+            )
+
+            key = "_".join(map(str, y_star.ys[-context_size:]))
+            key += f"-t-{t}"
+            log_prob = run_joiner(
+                key=key,
+                model=model,
+                encoder_out=current_encoder_out,
+                decoder_out=decoder_out,
+                encoder_out_len=encoder_out_len,
+                decoder_out_len=decoder_out_len,
+                joint_cache=joint_cache,
+            )
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for idx in range(values.size(0)):
+                i = indices[idx].item()
+                if i == blank_id or i == unk_id:
+                    continue
+
+                new_ys = y_star.ys + [i]
+
+                new_log_prob = y_star.log_prob + values[idx]
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys
diff --git a/egs/tedlium3/ASR/transducer_stateless/conformer.py b/egs/tedlium3/ASR/transducer_stateless/conformer.py
new file mode 120000
index 000000000..8be0dc864
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/conformer.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/decode.py b/egs/tedlium3/ASR/transducer_stateless/decode.py
new file mode 100755
index 000000000..fb0e3116b
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/decode.py
@@ -0,0 +1,486 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 11 \
+        --exp-dir ./transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 11 \
+        --exp-dir ./transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 11 \
+        --exp-dir ./transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from beam_search import beam_search, greedy_search, modified_beam_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        beam_search or modified_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=3,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict):
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict):
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict):
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.decoding_method == "greedy_search":
+            hyp = greedy_search(
+                model=model,
+                encoder_out=encoder_out_i,
+                max_sym_per_frame=params.max_sym_per_frame,
+            )
+        elif params.decoding_method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        elif params.decoding_method == "modified_beam_search":
+            hyp = modified_beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+        hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+    dev_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    dev_dl = tedlium.valid_dataloaders(dev_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/transducer_stateless/decoder.py b/egs/tedlium3/ASR/transducer_stateless/decoder.py
new file mode 100644
index 000000000..f9a3814c6
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/decoder.py
@@ -0,0 +1,100 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        unk_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          unk_id:
+            The ID of the unk symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+        self.unk_id = unk_id
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, embedding_dim).
+        """
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        return embedding_out
diff --git a/egs/tedlium3/ASR/transducer_stateless/encoder_interface.py b/egs/tedlium3/ASR/transducer_stateless/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/export.py b/egs/tedlium3/ASR/transducer_stateless/export.py
new file mode 100644
index 000000000..48dcdc736
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/export.py
@@ -0,0 +1,252 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+# 					                         Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless/export.py \
+  --exp-dir ./transducer_stateless/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 29 \
+  --avg 11
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless/decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tedlium3/ASR
+    ./transducer_stateless/decode.py \
+        --exp-dir ./transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tedlium3/ASR/transducer_stateless/joiner.py b/egs/tedlium3/ASR/transducer_stateless/joiner.py
new file mode 120000
index 000000000..1aec6bfaf
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/joiner.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/local b/egs/tedlium3/ASR/transducer_stateless/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/model.py b/egs/tedlium3/ASR/transducer_stateless/model.py
new file mode 120000
index 000000000..16ddd93f0
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/model.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/pretrained.py b/egs/tedlium3/ASR/transducer_stateless/pretrained.py
new file mode 100644
index 000000000..81afd6a4e
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/pretrained.py
@@ -0,0 +1,337 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		 2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(3) modified beam search
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+You can also use `./transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./transducer_stateless/exp/pretrained.pt is generated by
+./transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+import torchaudio
+from beam_search import beam_search, greedy_search, modified_beam_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and modified_beam_search ",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=3,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+    for i in range(num_waves):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.method == "greedy_search":
+            hyp = greedy_search(
+                model=model,
+                encoder_out=encoder_out_i,
+                max_sym_per_frame=params.max_sym_per_frame,
+            )
+        elif params.method == "beam_search":
+            hyp = beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        elif params.method == "modified_beam_search":
+            hyp = modified_beam_search(
+                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+            )
+        else:
+            raise ValueError(f"Unsupported method: {params.method}")
+
+        hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/tedlium3/ASR/transducer_stateless/subsampling.py b/egs/tedlium3/ASR/transducer_stateless/subsampling.py
new file mode 120000
index 000000000..fd7ca8b30
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/subsampling.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/test_decoder.py b/egs/tedlium3/ASR/transducer_stateless/test_decoder.py
new file mode 100755
index 000000000..cc5f64951
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/test_decoder.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						   Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/tedlium3/ASR
+    python ./transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    unk_id = 2
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        unk_id=unk_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, embedding_dim)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, embedding_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/transducer_stateless/train.py b/egs/tedlium3/ASR/transducer_stateless/train.py
new file mode 100755
index 000000000..6fed32e81
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/train.py
@@ -0,0 +1,737 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--modified-transducer-prob",
+        type=float,
+        default=0.25,
+        help="""The probability to use modified transducer loss.
+        In modified transduer, it limits the maximum number of symbols
+        per frame to 1. See also the option --max-sym-per-frame in
+        transducer_stateless/decode.py
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "encoder_out_dim": 512,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.encoder_out_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.encoder_out_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    unk_id = params.unk_id
+    y = convert_texts_into_ids(texts, unk_id, sp=sp)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            modified_transducer_prob=params.modified_transducer_prob,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    tedlium = TedLiumAsrDataModule(args)
+
+    train_cuts = tedlium.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 17 seconds
+        return 1.0 <= c.duration <= 17.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = tedlium.train_dataloaders(train_cuts)
+    valid_cuts = tedlium.dev_cuts()
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/transducer_stateless/transformer.py b/egs/tedlium3/ASR/transducer_stateless/transformer.py
new file mode 120000
index 000000000..214afed39
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/transformer.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/__init__.py b/egs/tedlium3/ASR/zipformer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/tedlium3/ASR/zipformer/asr_datamodule.py b/egs/tedlium3/ASR/zipformer/asr_datamodule.py
new file mode 120000
index 000000000..49b2ee483
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/beam_search.py b/egs/tedlium3/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/decode.py b/egs/tedlium3/ASR/zipformer/decode.py
new file mode 100755
index 000000000..2c4123c20
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/decode.py
@@ -0,0 +1,833 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 9 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    x, x_lens = model.encoder_embed(feature, feature_lens)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+    encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    hyps = []
+    unk = sp.decode(sp.unk_id()).strip()
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            allow_partial=True,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyp = [w for w in hyp.split() if w != unk]
+            hyps.append(hyp)
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            allow_partial=True,
+        )
+        for hyp in hyp_tokens:
+            hyp = [word_table[i] for i in hyp if word_table[i] != unk]
+            hyps.append(hyp)
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+            allow_partial=True,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyp = [w for w in hyp.split() if w != unk]
+            hyps.append(hyp)
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            allow_partial=True,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyp = [w for w in hyp.split() if w != unk]
+            hyps.append(hyp)
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyp = [w for w in hyp.split() if w != unk]
+            hyps.append(hyp)
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyp = [w for w in hyp.split() if w != unk]
+            hyps.append(hyp)
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp = [w for w in sp.decode(hyp).split() if w != unk]
+            hyps.append(hyp)
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+
+    dev_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    dev_dl = tedlium.test_dataloaders(dev_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for name, dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=name,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/zipformer/decoder.py b/egs/tedlium3/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/encoder_interface.py b/egs/tedlium3/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/export.py b/egs/tedlium3/ASR/zipformer/export.py
new file mode 120000
index 000000000..dfc1bec08
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/joiner.py b/egs/tedlium3/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/local b/egs/tedlium3/ASR/zipformer/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/model.py b/egs/tedlium3/ASR/zipformer/model.py
new file mode 100644
index 000000000..90ec7e7aa
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/model.py
@@ -0,0 +1,223 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos, make_pad_mask
+from scaling import ScaledLinear
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dim) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+        self.simple_lm_proj = ScaledLinear(
+            decoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        rnnt_type: str = "regular",
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          rnnt_type:
+            The type of label topology to use for the transducer loss. One of "regular",
+            "modified", or "constrained".
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        # logging.info(f"Memory allocated at entry: {torch.cuda.memory_allocated() // 1000000}M")
+        x, x_lens = self.encoder_embed(x, x_lens)
+        # logging.info(f"Memory allocated after encoder_embed: {torch.cuda.memory_allocated() // 1000000}M")
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, x_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        # if self.training and random.random() < 0.25:
+        #    lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04)
+        # if self.training and random.random() < 0.25:
+        #    am = penalize_abs_values_gt(am, 30.0, 1.0e-04)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+                rnnt_type=rnnt_type,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+                rnnt_type=rnnt_type,
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/tedlium3/ASR/zipformer/optim.py b/egs/tedlium3/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/pretrained.py b/egs/tedlium3/ASR/zipformer/pretrained.py
new file mode 120000
index 000000000..0bd71dde4
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/profile.py b/egs/tedlium3/ASR/zipformer/profile.py
new file mode 120000
index 000000000..c93adbd14
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/profile.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/profile.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/scaling.py b/egs/tedlium3/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/scaling_converter.py b/egs/tedlium3/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/subsampling.py b/egs/tedlium3/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/train.py b/egs/tedlium3/ASR/zipformer/train.py
new file mode 100755
index 000000000..5ad01df27
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/train.py
@@ -0,0 +1,1308 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --full-libri 1 \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --full-libri 1 \
+  --max-duration 1000
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
+from model import Transducer
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=50,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.04, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--rnnt-type",
+        type=str,
+        default="regular",
+        choices=["regular", "modified", "constrained"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=1,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNNT loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = convert_texts_into_ids(texts, sp)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            rnnt_type=params.rnnt_type,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    tedlium = TedLiumAsrDataModule(args)
+
+    train_cuts = tedlium.train_cuts()
+    train_cuts = train_cuts.filter(lambda c: 1.0 <= c.duration <= 20.0)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = tedlium.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = tedlium.dev_cuts()
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tedlium3/ASR/zipformer/zipformer.py b/egs/tedlium3/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/timit/ASR/README.md b/egs/timit/ASR/README.md
new file mode 100644
index 000000000..d493fc479
--- /dev/null
+++ b/egs/timit/ASR/README.md
@@ -0,0 +1,3 @@
+
+Please refer to <https://icefall.readthedocs.io/en/latest/recipes/Non-streaming-ASR/timit/index.html>
+for how to run models in this recipe.
diff --git a/egs/timit/ASR/RESULTS.md b/egs/timit/ASR/RESULTS.md
new file mode 100644
index 000000000..d8ceb82b6
--- /dev/null
+++ b/egs/timit/ASR/RESULTS.md
@@ -0,0 +1,74 @@
+## Results
+
+### TIMIT training results (Tdnn_LSTM_CTC)
+#### 2021-11-16
+(Mingshuang Luo): Result of https://github.com/k2-fsa/icefall/pull/114
+
+TensorBoard log is available at https://tensorboard.dev/experiment/qhA1o025Q322kO34SlhWzg/#scalars
+
+Pretrained model is available at https://huggingface.co/luomingshuang/icefall_asr_timit_tdnn_lstm_ctc
+
+The best decoding results (PER) are listed below, we got this results by averaging models from epoch 16 to 25, and using `whole-lattice-rescoring` with lm_scale equals to 0.08.
+
+||TEST|
+|--|--|
+|PER| 19.71% |
+
+You can use the following commands to reproduce our results:
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+
+cd egs/timit/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0"
+python tdnn_lstm_ctc/train.py --bucketing-sampler True \
+                              --concatenate-cuts False \
+                              --max-duration 200 \
+                              --world-size 1 \
+                              --lang-dir data/lang_phone
+
+python tdnn_lstm_ctc/decode.py --epoch 25 \
+                               --avg 10 \
+                               --max-duration 20 \
+                               --lang-dir data/lang_phone
+```
+
+### TIMIT training results (Tdnn_LiGRU_CTC)
+#### 2021-11-16
+
+(Mingshuang Luo): Result of phone based Tdnn_LiGRU_CTC model, https://github.com/k2-fsa/icefall/pull/114
+
+TensorBoard log is available at https://tensorboard.dev/experiment/IlQxeq5vQJ2SEVP94Y5fyg/#scalars
+
+Pretrained model is available at https://huggingface.co/luomingshuang/icefall_asr_timit_tdnn_ligru_ctc
+
+The best decoding results (PER) are listed below, we got this results by averaging models from epoch 9 to 25, and using `whole-lattice-rescoring` decoding method with lm_scale equals to 0.1.
+
+||TEST|
+|--|--|
+|PER| 17.66% |
+
+You can use the following commands to reproduce our results:
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+
+cd egs/timit/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0"
+python tdnn_ligru_ctc/train.py --bucketing-sampler True \
+                              --concatenate-cuts False \
+                              --max-duration 200 \
+                              --world-size 1 \
+                              --lang-dir data/lang_phone
+
+python tdnn_ligru_ctc/decode.py --epoch 25 \
+                               --avg 17 \
+                               --max-duration 20 \
+                               --lang-dir data/lang_phone
+```
diff --git a/egs/timit/ASR/local/__init__.py b/egs/timit/ASR/local/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/timit/ASR/local/compile_hlg.py b/egs/timit/ASR/local/compile_hlg.py
new file mode 100755
index 000000000..c8562f4fb
--- /dev/null
+++ b/egs/timit/ASR/local/compile_hlg.py
@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_3_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str) -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path("data/lm/G.pt").is_file():
+        logging.info("Loading pre-compiled G")
+        d = torch.load("data/lm/G.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info("Loading G_3_gram.fst.txt")
+        with open("data/lm/G_3_gram.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), "data/lm/G.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    # LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # see https://github.com/k2-fsa/k2/pull/1140
+    labels = LG.labels
+    labels[labels >= first_token_disambig_id] = 0
+    LG.labels = labels
+
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lang_dir)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/timit/ASR/local/compute_fbank_musan.py b/egs/timit/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/timit/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/timit/ASR/local/compute_fbank_timit.py b/egs/timit/ASR/local/compute_fbank_timit.py
new file mode 100755
index 000000000..ecdf10ba9
--- /dev/null
+++ b/egs/timit/ASR/local/compute_fbank_timit.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the TIMIT dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_timit():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "TRAIN",
+        "DEV",
+        "TEST",
+    )
+    prefix = "timit"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_file = output_dir / f"{prefix}_cuts_{partition}.{suffix}"
+            if cuts_file.is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if partition == "TRAIN":
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(cuts_file)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_timit()
diff --git a/egs/timit/ASR/local/prepare_lang.py b/egs/timit/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..e9f283274
--- /dev/null
+++ b/egs/timit/ASR/local/prepare_lang.py
@@ -0,0 +1,386 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+from icefall.utils import str2bool
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+
+    sorted_ans = list(ans)
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    pronprob = 1.0
+    score = -math.log(pronprob)
+
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, score])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        tokens[i] = tokens[i] if i >= 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    lexicon_filename = lang_dir / "lexicon.txt"
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+
+    words = get_words(lexicon)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(lang_dir / "tokens.txt", token2id)
+    write_mapping(lang_dir / "words.txt", word2id)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(lang_dir / "L.png", title="L")
+        L_disambig.draw(lang_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/timit/ASR/local/prepare_lexicon.py b/egs/timit/ASR/local/prepare_lexicon.py
new file mode 100755
index 000000000..0cf0f0deb
--- /dev/null
+++ b/egs/timit/ASR/local/prepare_lexicon.py
@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input supervisions json dir "data/manifests"
+consisting of supervisions_TRAIN.json and does the following:
+
+1. Generate lexicon.txt.
+
+"""
+import argparse
+import json
+import logging
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--manifests-dir",
+        type=str,
+        help="""Input directory.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def prepare_lexicon(manifests_dir: str, lang_dir: str):
+    """
+    Args:
+      manifests_dir:
+        The manifests directory, e.g., data/manifests.
+      lang_dir:
+        The language directory, e.g., data/lang_phone.
+
+    Return:
+      The lexicon.txt file and the train.text in lang_dir.
+    """
+    import gzip
+
+    phones = set()
+
+    supervisions_train = Path(manifests_dir) / "timit_supervisions_TRAIN.jsonl.gz"
+    lexicon = Path(lang_dir) / "lexicon.txt"
+
+    logging.info(f"Loading {supervisions_train}!")
+    with gzip.open(supervisions_train, "r") as load_f:
+        for line in load_f.readlines():
+            load_dict = json.loads(line)
+            text = load_dict["text"]
+            # list the phone units and filter the empty item
+            phones_list = list(filter(None, text.split()))
+
+            for phone in phones_list:
+                if phone not in phones:
+                    phones.add(phone)
+
+    with open(lexicon, "w") as f:
+        for phone in sorted(phones):
+            f.write(phone + "  " + phone)
+            f.write("\n")
+        f.write("<UNK>  <UNK>")
+        f.write("\n")
+
+
+def main():
+    args = get_args()
+    manifests_dir = Path(args.manifests_dir)
+    lang_dir = Path(args.lang_dir)
+
+    logging.info("Generating lexicon.txt")
+    prepare_lexicon(manifests_dir, lang_dir)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/timit/ASR/prepare.sh b/egs/timit/ASR/prepare.sh
new file mode 100644
index 000000000..f25fe5add
--- /dev/null
+++ b/egs/timit/ASR/prepare.sh
@@ -0,0 +1,159 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+num_phones=39
+# Here we use num_phones=39 for modeling
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/timit
+#      You can find data, train_data.csv, test_data.csv, etc, inside it.
+#      You can download them from https://data.deepai.org/timit.zip
+#
+#  - $dl_dir/lm
+#      This directory contains the language model(LM) downloaded from
+#      https://huggingface.co/luomingshuang/timit_lm, and the LM is based
+#	     on 39 phones. About how to get these LM files, you can know it
+#      from https://github.com/luomingshuang/Train_LM_with_kaldilm.
+#
+#	    - lm_3_gram.arpa
+#     - lm_4_gram.arpa
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+splits_dir=$PWD/splits_dir
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download LM"
+  # We assume that you have installed the git-lfs, if not, you could install it
+  # using: `sudo apt-get install git-lfs && git-lfs install`
+  [ ! -e $dl_dir/lm ] && mkdir -p $dl_dir/lm
+  git clone https://huggingface.co/luomingshuang/timit_lm $dl_dir/lm
+  pushd $dl_dir/lm
+  git lfs pull
+  popd
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/timit,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/timit $dl_dir/timit
+  #
+  if [ ! -d $dl_dir/timit ]; then
+    lhotse download timit $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare timit manifest"
+  # We assume that you have downloaded the timit corpus
+  # to $dl_dir/timit
+  mkdir -p data/manifests
+  lhotse prepare timit -p $num_phones -j $nj $dl_dir/timit/data data/manifests
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for timit"
+  mkdir -p data/fbank
+  ./local/compute_fbank_timit.py
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  ./local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+
+  ./local/prepare_lexicon.py \
+   --manifests-dir data/manifests \
+   --lang-dir $lang_dir
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $dl_dir/lm/lm_3_gram.arpa > data/lm/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      $dl_dir/lm/lm_4_gram.arpa > data/lm/G_4_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+fi
diff --git a/egs/timit/ASR/shared b/egs/timit/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/timit/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/timit/ASR/tdnn_ligru_ctc/__init__.py b/egs/timit/ASR/tdnn_ligru_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/timit/ASR/tdnn_ligru_ctc/asr_datamodule.py b/egs/timit/ASR/tdnn_ligru_ctc/asr_datamodule.py
new file mode 120000
index 000000000..fa1b8cca3
--- /dev/null
+++ b/egs/timit/ASR/tdnn_ligru_ctc/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/timit/ASR/tdnn_ligru_ctc/decode.py b/egs/timit/ASR/tdnn_ligru_ctc/decode.py
new file mode 100644
index 000000000..4beeed18c
--- /dev/null
+++ b/egs/timit/ASR/tdnn_ligru_ctc/decode.py
@@ -0,0 +1,488 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import TimitAsrDataModule
+from model import TdnnLiGRU
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=19,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="whole-lattice-rescoring",
+        help="""Decoding method.
+        Supported values are:
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_ligru_ctc/exp/"),
+            "lang_dir": Path("data/lang_phone"),
+            "lm_dir": Path("data/lm"),
+            "feature_dim": 80,
+            "subsampling_factor": 2,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    feature = feature.permute(0, 2, 1)  # now feature is (N, C, T)
+
+    nnet_output = model(feature)
+    # nnet_output is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-{params.num_paths}"
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in ["nbest-rescoring", "whole-lattice-rescoring"]
+
+    lm_scale_list = [0.0, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09]
+    lm_scale_list += [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+
+    ans = dict()
+    for lm_scale_str, best_path in best_path_dict.items():
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            lexicon=lexicon,
+            G=G,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out PERs, per-phone error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"per-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tPER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, PER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TimitAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu"))
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in ["nbest-rescoring", "whole-lattice-rescoring"]:
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location="cpu")
+            G = k2.Fsa.from_dict(d).to(device)
+
+        if params.method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = TdnnLiGRU(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    timit = TimitAsrDataModule(args)
+    test_set = "TEST"
+    test_dl = timit.test_dataloaders()
+
+    results_dict = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+        HLG=HLG,
+        lexicon=lexicon,
+        G=G,
+    )
+
+    save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/timit/ASR/tdnn_ligru_ctc/model.py b/egs/timit/ASR/tdnn_ligru_ctc/model.py
new file mode 100644
index 000000000..9a594a969
--- /dev/null
+++ b/egs/timit/ASR/tdnn_ligru_ctc/model.py
@@ -0,0 +1,477 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+
+class TdnnLiGRU(nn.Module):
+    def __init__(
+        self, num_features: int, num_classes: int, subsampling_factor: int = 3
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            It reduces the number of output frames by this factor.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=self.subsampling_factor,  # stride: subsampling_factor!
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+        )
+        self.ligrus = nn.ModuleList(
+            [
+                LiGRU(
+                    input_shape=[None, None, 512],
+                    hidden_size=512,
+                    num_layers=1,
+                    bidirectional=True,
+                )
+                for _ in range(4)
+            ]
+        )
+        self.linears = nn.ModuleList(
+            [nn.Linear(in_features=1024, out_features=512) for _ in range(4)]
+        )
+        self.bnorms = nn.ModuleList(
+            [nn.BatchNorm1d(num_features=512, affine=False) for _ in range(4)]
+        )
+        self.dropout = nn.Dropout(0.2)
+        self.linear = nn.Linear(in_features=512, out_features=self.num_classes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Its shape is [N, C, T]
+
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x = self.tdnn(x)
+        x = x.permute(0, 2, 1)
+        for ligru, linear, bnorm in zip(self.ligrus, self.linears, self.bnorms):
+            x_new, _ = ligru(x)
+            x_new = linear(x_new)
+            x_new = bnorm(x_new.permute(0, 2, 1)).permute(0, 2, 1)
+            # (N, T, C) -> (N, C, T) -> (N, T, C)
+            x_new = self.dropout(x_new)
+            x = x_new + x  # skip connections
+
+        x = self.linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x
+
+
+class LiGRU(torch.nn.Module):
+    """This function implements a Light GRU (liGRU).
+    This LiGRU model is from speechbrain, please see
+    https://github.com/speechbrain/speechbrain/blob/develop/speechbrain/nnet/RNN.py
+
+    LiGRU is single-gate GRU model based on batch-norm + relu
+    activations + recurrent dropout. For more info see:
+
+    "M. Ravanelli, P. Brakel, M. Omologo, Y. Bengio,
+    Light Gated Recurrent Units for Speech Recognition,
+    in IEEE Transactions on Emerging Topics in Computational Intelligence,
+    2018" (https://arxiv.org/abs/1803.10225)
+
+    This is a custm RNN and to speed it up it must be compiled with
+    the torch just-in-time compiler (jit) right before using it.
+    You can compile it with:
+    compiled_model = torch.jit.script(model)
+
+    It accepts in input tensors formatted as (batch, time, fea).
+    In the case of 4d inputs like (batch, time, fea, channel) the tensor is
+    flattened as (batch, time, fea*channel).
+
+    Arguments
+    ---------
+    hidden_size : int
+        Number of output neurons (i.e, the dimensionality of the output).
+        values (i.e, time and frequency kernel sizes respectively).
+    input_shape : tuple
+        The shape of an example input.
+    nonlinearity : str
+        Type of nonlinearity (tanh, relu).
+    normalization : str
+        Type of normalization for the ligru model (batchnorm, layernorm).
+        Every string different from batchnorm and layernorm will result
+        in no normalization.
+    num_layers : int
+        Number of layers to employ in the RNN architecture.
+    bias : bool
+        If True, the additive bias b is adopted.
+    dropout : float
+        It is the dropout factor (must be between 0 and 1).
+    bidirectional : bool
+        If True, a bidirectional model that scans the sequence both
+        right-to-left and left-to-right is used.
+
+    Example
+    -------
+    >>> inp_tensor = torch.rand([4, 10, 20])
+    >>> net = LiGRU(input_shape=inp_tensor.shape, hidden_size=5)
+    >>> out_tensor, _ = net(inp_tensor)
+    >>>
+    torch.Size([4, 10, 5])
+    """
+
+    def __init__(
+        self,
+        hidden_size,
+        input_shape,
+        nonlinearity="relu",
+        normalization="batchnorm",
+        num_layers=1,
+        bias=True,
+        dropout=0.0,
+        bidirectional=False,
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.nonlinearity = nonlinearity
+        self.num_layers = num_layers
+        self.normalization = normalization
+        self.bias = bias
+        self.dropout = dropout
+        self.bidirectional = bidirectional
+        self.reshape = False
+
+        # Computing the feature dimensionality
+        if len(input_shape) > 3:
+            self.reshape = True
+        self.fea_dim = float(torch.prod(torch.tensor(input_shape[2:])))
+        self.batch_size = input_shape[0]
+        self.rnn = self._init_layers()
+
+    def _init_layers(self):
+        """Initializes the layers of the liGRU."""
+        rnn = torch.nn.ModuleList([])
+        current_dim = self.fea_dim
+
+        for i in range(self.num_layers):
+            rnn_lay = LiGRU_Layer(
+                current_dim,
+                self.hidden_size,
+                self.num_layers,
+                self.batch_size,
+                dropout=self.dropout,
+                nonlinearity=self.nonlinearity,
+                normalization=self.normalization,
+                bidirectional=self.bidirectional,
+            )
+            rnn.append(rnn_lay)
+
+            if self.bidirectional:
+                current_dim = self.hidden_size * 2
+            else:
+                current_dim = self.hidden_size
+        return rnn
+
+    def forward(self, x, hx: Optional[Tensor] = None):
+        """Returns the output of the liGRU.
+
+        Arguments
+        ---------
+        x : torch.Tensor
+            The input tensor.
+        hx : torch.Tensor
+            Starting hidden state.
+        """
+        # Reshaping input tensors for 4d inputs
+        if self.reshape:
+            if x.ndim == 4:
+                x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3])
+
+        # run ligru
+        output, hh = self._forward_ligru(x, hx=hx)
+
+        return output, hh
+
+    def _forward_ligru(self, x, hx: Optional[Tensor]):
+        """Returns the output of the vanilla liGRU.
+
+        Arguments
+        ---------
+        x : torch.Tensor
+            Input tensor.
+        hx : torch.Tensor
+        """
+        h = []
+        if hx is not None:
+            if self.bidirectional:
+                hx = hx.reshape(self.num_layers, self.batch_size * 2, self.hidden_size)
+        # Processing the different layers
+        for i, ligru_lay in enumerate(self.rnn):
+            if hx is not None:
+                x = ligru_lay(x, hx=hx[i])
+            else:
+                x = ligru_lay(x, hx=None)
+            h.append(x[:, -1, :])
+        h = torch.stack(h, dim=1)
+
+        if self.bidirectional:
+            h = h.reshape(h.shape[1] * 2, h.shape[0], self.hidden_size)
+        else:
+            h = h.transpose(0, 1)
+
+        return x, h
+
+
+class LiGRU_Layer(torch.nn.Module):
+    """This function implements Light-Gated Recurrent Units (ligru) layer.
+
+    Arguments
+    ---------
+    input_size : int
+        Feature dimensionality of the input tensors.
+    batch_size : int
+        Batch size of the input tensors.
+    hidden_size : int
+        Number of output neurons.
+    num_layers : int
+        Number of layers to employ in the RNN architecture.
+    nonlinearity : str
+        Type of nonlinearity (tanh, relu).
+    normalization : str
+        Type of normalization (batchnorm, layernorm).
+        Every string different from batchnorm and layernorm will result
+        in no normalization.
+    dropout : float
+        It is the dropout factor (must be between 0 and 1).
+    bidirectional : bool
+        if True, a bidirectional model that scans the sequence both
+        right-to-left and left-to-right is used.
+    """
+
+    def __init__(
+        self,
+        input_size,
+        hidden_size,
+        num_layers,
+        batch_size,
+        dropout=0.0,
+        nonlinearity="relu",
+        normalization="batchnorm",
+        bidirectional=False,
+    ):
+
+        super(LiGRU_Layer, self).__init__()
+        self.hidden_size = int(hidden_size)
+        self.input_size = int(input_size)
+        self.batch_size = batch_size
+        self.bidirectional = bidirectional
+        self.dropout = dropout
+        self.drop = torch.nn.Dropout(p=self.dropout, inplace=False)
+        self.N_drop_masks = 16000
+        self.drop_mask_cnt = 0
+        self.drop_mask_te = torch.tensor([1.0]).float()
+        self.w = nn.Linear(self.input_size, 2 * self.hidden_size, bias=False)
+        self.u = nn.Linear(self.hidden_size, 2 * self.hidden_size, bias=False)
+
+        # Initializing batch norm
+        self.normalize = False
+
+        if normalization == "batchnorm":
+            self.norm = nn.BatchNorm1d(2 * self.hidden_size, momentum=0.05)
+            self.normalize = True
+
+        elif normalization == "layernorm":
+            self.norm = torch.nn.LayerNorm(2 * self.hidden_size)
+            self.normalize = True
+        else:
+            # Normalization is disabled here. self.norm is only  formally
+            # initialized to avoid jit issues.
+            self.norm = torch.nn.LayerNorm(2 * self.hidden_size)
+            self.normalize = True
+
+        # Initial state
+        self.register_buffer("h_init", torch.zeros(1, self.hidden_size))
+
+        # Setting the activation function
+        if nonlinearity == "tanh":
+            self.act = torch.nn.Tanh()
+        elif nonlinearity == "sin":
+            self.act = torch.sin
+        elif nonlinearity == "leaky_relu":
+            self.act = torch.nn.LeakyReLU()
+        else:
+            self.act = torch.nn.ReLU()
+
+    def forward(self, x, hx: Optional[Tensor] = None):
+        # type: (Tensor, Optional[Tensor]) -> Tensor # noqa F821
+        """Returns the output of the liGRU layer.
+
+        Arguments
+        ---------
+        x : torch.Tensor
+            Input tensor.
+        """
+        if self.bidirectional:
+            x_flip = x.flip(1)
+            x = torch.cat([x, x_flip], dim=0)
+
+        # Change batch size if needed
+        self._change_batch_size(x)
+
+        # Feed-forward affine transformations (all steps in parallel)
+        w = self.w(x)
+
+        # Apply batch normalization
+        if self.normalize:
+            w_bn = self.norm(w.reshape(w.shape[0] * w.shape[1], w.shape[2]))
+            w = w_bn.reshape(w.shape[0], w.shape[1], w.shape[2])
+
+        # Processing time steps
+        if hx is not None:
+            h = self._ligru_cell(w, hx)
+        else:
+            h = self._ligru_cell(w, self.h_init)
+
+        if self.bidirectional:
+            h_f, h_b = h.chunk(2, dim=0)
+            h_b = h_b.flip(1)
+            h = torch.cat([h_f, h_b], dim=2)
+
+        return h
+
+    def _ligru_cell(self, w, ht):
+        """Returns the hidden states for each time step.
+
+        Arguments
+        ---------
+        wx : torch.Tensor
+            Linearly transformed input.
+        """
+        hiddens = []
+
+        # Sampling dropout mask
+        drop_mask = self._sample_drop_mask(w)
+
+        # Loop over time axis
+        for k in range(w.shape[1]):
+            gates = w[:, k] + self.u(ht)
+            at, zt = gates.chunk(2, 1)
+            zt = torch.sigmoid(zt)
+            hcand = self.act(at) * drop_mask
+            ht = zt * ht + (1 - zt) * hcand
+            hiddens.append(ht)
+
+        # Stacking hidden states
+        h = torch.stack(hiddens, dim=1)
+        return h
+
+    def _init_drop(self, batch_size):
+        """Initializes the recurrent dropout operation. To speed it up,
+        the dropout masks are sampled in advance.
+        """
+        self.N_drop_masks = 16000
+        self.drop_mask_cnt = 0
+
+        self.register_buffer(
+            "drop_masks",
+            self.drop(torch.ones(self.N_drop_masks, self.hidden_size)).data,
+        )
+        self.register_buffer("drop_mask_te", torch.tensor([1.0]).float())
+
+    def _sample_drop_mask(self, w):
+        """Selects one of the pre-defined dropout masks"""
+        if self.training:
+
+            # Sample new masks when needed
+            if self.drop_mask_cnt + self.batch_size > self.N_drop_masks:
+                self.drop_mask_cnt = 0
+                self.drop_masks = self.drop(
+                    torch.ones(self.N_drop_masks, self.hidden_size, device=w.device)
+                ).data
+
+            # Sampling the mask
+            left_boundary = self.drop_mask_cnt
+            right_boundary = self.drop_mask_cnt + self.batch_size
+            drop_mask = self.drop_masks[left_boundary:right_boundary]
+            self.drop_mask_cnt = self.drop_mask_cnt + self.batch_size
+
+        else:
+            self.drop_mask_te = self.drop_mask_te.to(w.device)
+            drop_mask = self.drop_mask_te
+
+        return drop_mask
+
+    def _change_batch_size(self, x):
+        """This function changes the batch size when it is different from
+        the one detected in the initialization method. This might happen in
+        the case of multi-gpu or when we have different batch sizes in train
+        and test. We also update the h_int and drop masks.
+        """
+        if self.batch_size != x.shape[0]:
+            self.batch_size = x.shape[0]
+
+            if self.training:
+                self.drop_masks = self.drop(
+                    torch.ones(
+                        self.N_drop_masks,
+                        self.hidden_size,
+                        device=x.device,
+                    )
+                ).data
diff --git a/egs/timit/ASR/tdnn_ligru_ctc/pretrained.py b/egs/timit/ASR/tdnn_ligru_ctc/pretrained.py
new file mode 100644
index 000000000..3fdf3b855
--- /dev/null
+++ b/egs/timit/ASR/tdnn_ligru_ctc/pretrained.py
@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from model import TdnnLiGRU
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding, rescore_with_whole_lattice
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.1,
+        help="""
+        Used only when method is whole-lattice-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 2,
+            "num_classes": 41,
+            "sample_rate": 16000,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = TdnnLiGRU(
+        num_features=params.feature_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+    if not hasattr(HLG, "lm_scores"):
+        # For whole-lattice-rescoring and attention-decoder
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method == "whole-lattice-rescoring":
+        logging.info(f"Loading G from {params.G}")
+        G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+        # Add epsilon self-loops to G as we will compose
+        # it with the whole lattice later
+        G = G.to(device)
+        G = k2.add_epsilon_self_loops(G)
+        G = k2.arc_sort(G)
+        G.lm_scores = G.scores.clone()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    features = features.permute(0, 2, 1)  # now features is (N, C, T)
+
+    with torch.no_grad():
+        nnet_output = model(features)
+        # nnet_output is (N, T, C)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "1best":
+        logging.info("Use HLG decoding")
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+    elif params.method == "whole-lattice-rescoring":
+        logging.info("Use HLG decoding + LM rescoring")
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=[params.ngram_lm_scale],
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/timit/ASR/tdnn_ligru_ctc/train.py b/egs/timit/ASR/tdnn_ligru_ctc/train.py
new file mode 100644
index 000000000..48b7feda0
--- /dev/null
+++ b/egs/timit/ASR/tdnn_ligru_ctc/train.py
@@ -0,0 +1,601 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import TimitAsrDataModule
+from lhotse.utils import fix_random_seed
+from model import TdnnLiGRU
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=25,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn_lstm_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - exp_dir: It specifies the directory where all training related
+                   files, e.g., checkpoints, log, etc, are saved
+
+        - lang_dir: It contains language related input files such as
+                    "lexicon.txt"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_ligru_ctc/exp"),
+            "lang_dir": Path("data/lang_phone"),
+            "lr": 1e-3,
+            "feature_dim": 80,
+            "weight_decay": 5e-4,
+            "subsampling_factor": 2,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of TdnnLstm in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    feature = feature.permute(0, 2, 1)  # now feature is (N, C, T)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(feature)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervisions = batch["supervisions"]
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+    decoding_graph = graph_compiler.compile(texts)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = TdnnLiGRU(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.AdamW(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    scheduler = StepLR(optimizer, step_size=2, gamma=0.8)
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    timit = TimitAsrDataModule(args)
+    train_dl = timit.train_dataloaders()
+    valid_dl = timit.valid_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if epoch > params.start_epoch:
+            logging.info(f"epoch {epoch}, lr: {scheduler.get_last_lr()[0]}")
+
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/lr",
+                scheduler.get_last_lr()[0],
+                params.batch_idx_train,
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        scheduler.step()
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    TimitAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/__init__.py b/egs/timit/ASR/tdnn_lstm_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py
new file mode 100644
index 000000000..5d1b3c367
--- /dev/null
+++ b/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py
@@ -0,0 +1,336 @@
+# Copyright      2021     Piotr Żelasko
+#                2022     Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import List, Union
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.dataset.datamodule import DataModule
+from icefall.utils import str2bool
+
+
+class TimitAsrDataModule(DataModule):
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        super().add_arguments(parser)
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--feature-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+    def train_dataloaders(self) -> DataLoader:
+        logging.info("About to get train cuts")
+        cuts_train = self.train_cuts()
+
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.feature_dir / "musan_cuts.jsonl.gz")
+
+        logging.info("About to create train dataset")
+        transforms = [CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20))]
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        # Set the value of num_frame_masks according to Lhotse's version.
+        # In different Lhotse's versions, the default of num_frame_masks is
+        # different.
+        num_frame_masks = 10
+        num_frame_masks_parameter = inspect.signature(SpecAugment.__init__).parameters[
+            "num_frame_masks"
+        ]
+        if num_frame_masks_parameter.default == 1:
+            num_frame_masks = 2
+        logging.info(f"Num frame mask: {num_frame_masks}")
+        input_transforms = [
+            SpecAugment(
+                num_frame_masks=num_frame_masks,
+                features_mask_size=27,
+                num_feature_masks=2,
+                frames_mask_size=100,
+            )
+        ]
+
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self) -> DataLoader:
+        logging.info("About to get dev cuts")
+        cuts_valid = self.valid_cuts()
+
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = SimpleCutSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self) -> Union[DataLoader, List[DataLoader]]:
+        cuts = self.test_cuts()
+        is_list = isinstance(cuts, list)
+        test_loaders = []
+        if not is_list:
+            cuts = [cuts]
+
+        for cuts_test in cuts:
+            logging.debug("About to create test dataset")
+            test = K2SpeechRecognitionDataset(
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+                if self.args.on_the_fly_feats
+                else PrecomputedFeatures(),
+                return_cuts=self.args.return_cuts,
+            )
+            sampler = SimpleCutSampler(cuts_test, max_duration=self.args.max_duration)
+            logging.debug("About to create test dataloader")
+            test_dl = DataLoader(test, batch_size=None, sampler=sampler, num_workers=1)
+            test_loaders.append(test_dl)
+
+        if is_list:
+            return test_loaders
+        else:
+            return test_loaders[0]
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest_lazy(
+            self.args.feature_dir / "timit_cuts_TRAIN.jsonl.gz"
+        )
+
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.feature_dir / "timit_cuts_DEV.jsonl.gz"
+        )
+
+        return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.debug("About to get test cuts")
+        cuts_test = load_manifest_lazy(
+            self.args.feature_dir / "timit_cuts_TEST.jsonl.gz"
+        )
+
+        return cuts_test
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/decode.py b/egs/timit/ASR/tdnn_lstm_ctc/decode.py
new file mode 100644
index 000000000..502a48def
--- /dev/null
+++ b/egs/timit/ASR/tdnn_lstm_ctc/decode.py
@@ -0,0 +1,486 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import TimitAsrDataModule
+from model import TdnnLstm
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=25,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="whole-lattice-rescoring",
+        help="""Decoding method.
+        Supported values are:
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp/"),
+            "lang_dir": Path("data/lang_phone"),
+            "lm_dir": Path("data/lm"),
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    feature = feature.permute(0, 2, 1)  # now feature is (N, C, T)
+
+    nnet_output = model(feature)
+    # nnet_output is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-{params.num_paths}"
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in ["nbest-rescoring", "whole-lattice-rescoring"]
+
+    lm_scale_list = [0.0, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09]
+    lm_scale_list += [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+
+    ans = dict()
+    for lm_scale_str, best_path in best_path_dict.items():
+        hyps = get_texts(best_path)
+        hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+        ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    lexicon: Lexicon,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      lexicon:
+        It contains word symbol table.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            lexicon=lexicon,
+            G=G,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out PERs, per-phone error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"per-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tPER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, PER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TimitAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu"))
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in ["nbest-rescoring", "whole-lattice-rescoring"]:
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location="cpu")
+            G = k2.Fsa.from_dict(d).to(device)
+
+        if params.method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    timit = TimitAsrDataModule(args)
+    test_set = "TEST"
+    test_dl = timit.test_dataloaders()
+    results_dict = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+        HLG=HLG,
+        lexicon=lexicon,
+        G=G,
+    )
+
+    save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/model.py b/egs/timit/ASR/tdnn_lstm_ctc/model.py
new file mode 100644
index 000000000..e211ad80d
--- /dev/null
+++ b/egs/timit/ASR/tdnn_lstm_ctc/model.py
@@ -0,0 +1,107 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+
+
+class TdnnLstm(nn.Module):
+    def __init__(
+        self, num_features: int, num_classes: int, subsampling_factor: int = 3
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            It reduces the number of output frames by this factor.
+        """
+        super().__init__()
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+            nn.Conv1d(
+                in_channels=512,
+                out_channels=512,
+                kernel_size=3,
+                stride=self.subsampling_factor,  # stride: subsampling_factor!
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=512, affine=False),
+        )
+        self.lstms = nn.ModuleList(
+            [nn.LSTM(input_size=512, hidden_size=512, num_layers=1) for _ in range(4)]
+        )
+        self.lstm_bnorms = nn.ModuleList(
+            [nn.BatchNorm1d(num_features=512, affine=False) for _ in range(5)]
+        )
+        self.dropout = nn.Dropout(0.2)
+        self.linear = nn.Linear(in_features=512, out_features=self.num_classes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Its shape is [N, C, T]
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x = self.tdnn(x)
+        x = x.permute(2, 0, 1)  # (N, C, T) -> (T, N, C) -> how LSTM expects it
+        for lstm, bnorm in zip(self.lstms, self.lstm_bnorms):
+            x_new, _ = lstm(x)
+            x_new = bnorm(x_new.permute(1, 2, 0)).permute(
+                2, 0, 1
+            )  # (T, N, C) -> (N, C, T) -> (T, N, C)
+            x_new = self.dropout(x_new)
+            x = x_new + x  # skip connections
+        x = x.transpose(
+            1, 0
+        )  # (T, N, C) -> (N, T, C) -> linear expects "features" in the last dim
+        x = self.linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/pretrained.py b/egs/timit/ASR/tdnn_lstm_ctc/pretrained.py
new file mode 100644
index 000000000..98c746ce5
--- /dev/null
+++ b/egs/timit/ASR/tdnn_lstm_ctc/pretrained.py
@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from model import TdnnLstm
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding, rescore_with_whole_lattice
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) whole-lattice-rescoring - Use an LM to rescore the
+            decoding lattice and then use 1best to decode the
+            rescored lattice.
+            We call it HLG decoding + n-gram LM rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--G",
+        type=str,
+        help="""An LM for rescoring.
+        Used only when method is
+        whole-lattice-rescoring.
+        It's usually a 4-gram LM.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.8,
+        help="""
+        Used only when method is whole-lattice-rescoring.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "num_classes": 41,
+            "sample_rate": 16000,
+            "search_beam": 20,
+            "output_beam": 5,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+    if not hasattr(HLG, "lm_scores"):
+        # For whole-lattice-rescoring and attention-decoder
+        HLG.lm_scores = HLG.scores.clone()
+
+    if params.method == "whole-lattice-rescoring":
+        logging.info(f"Loading G from {params.G}")
+        G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
+        # Add epsilon self-loops to G as we will compose
+        # it with the whole lattice later
+        G = G.to(device)
+        G = k2.add_epsilon_self_loops(G)
+        G = k2.arc_sort(G)
+        G.lm_scores = G.scores.clone()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    features = features.permute(0, 2, 1)  # now features is (N, C, T)
+
+    with torch.no_grad():
+        nnet_output = model(features)
+        # nnet_output is (N, T, C)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "1best":
+        logging.info("Use HLG decoding")
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+    elif params.method == "whole-lattice-rescoring":
+        logging.info("Use HLG decoding + LM rescoring")
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=[params.ngram_lm_scale],
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/train.py b/egs/timit/ASR/tdnn_lstm_ctc/train.py
new file mode 100644
index 000000000..be1ecffaa
--- /dev/null
+++ b/egs/timit/ASR/tdnn_lstm_ctc/train.py
@@ -0,0 +1,601 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import TimitAsrDataModule
+from lhotse.utils import fix_random_seed
+from model import TdnnLstm
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn_lstm_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - exp_dir: It specifies the directory where all training related
+                   files, e.g., checkpoints, log, etc, are saved
+
+        - lang_dir: It contains language related input files such as
+                    "lexicon.txt"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp"),
+            "lang_dir": Path("data/lang_phone"),
+            "lr": 1e-3,
+            "feature_dim": 80,
+            "weight_decay": 5e-4,
+            "subsampling_factor": 3,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of TdnnLstm in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    feature = feature.permute(0, 2, 1)  # now feature is (N, C, T)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(feature)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervisions = batch["supervisions"]
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+    decoding_graph = graph_compiler.compile(texts)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.AdamW(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    scheduler = StepLR(optimizer, step_size=8, gamma=0.8)
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    timit = TimitAsrDataModule(args)
+    train_dl = timit.train_dataloaders()
+    valid_dl = timit.valid_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if epoch > params.start_epoch:
+            logging.info(f"epoch {epoch}, lr: {scheduler.get_last_lr()[0]}")
+
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/lr",
+                scheduler.get_last_lr()[0],
+                params.batch_idx_train,
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        scheduler.step()
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    TimitAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/vctk/TTS/local/compute_spectrogram_vctk.py b/egs/vctk/TTS/local/compute_spectrogram_vctk.py
new file mode 100755
index 000000000..440ac1245
--- /dev/null
+++ b/egs/vctk/TTS/local/compute_spectrogram_vctk.py
@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Zengwei Yao,
+#                                                       Zengrui Jin,)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the VCTK dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/spectrogram.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    LilcomChunkyWriter,
+    Spectrogram,
+    SpectrogramConfig,
+    load_manifest,
+)
+from lhotse.audio import RecordingSet
+from lhotse.supervision import SupervisionSet
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_spectrogram_vctk():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/spectrogram")
+    num_jobs = min(32, os.cpu_count())
+
+    sampling_rate = 22050
+    frame_length = 1024 / sampling_rate  # (in second)
+    frame_shift = 256 / sampling_rate  # (in second)
+    use_fft_mag = True
+
+    prefix = "vctk"
+    suffix = "jsonl.gz"
+    partition = "all"
+
+    recordings = load_manifest(
+        src_dir / f"{prefix}_recordings_{partition}.jsonl.gz", RecordingSet
+    ).resample(sampling_rate=sampling_rate)
+    supervisions = load_manifest(
+        src_dir / f"{prefix}_supervisions_{partition}.jsonl.gz", SupervisionSet
+    )
+
+    config = SpectrogramConfig(
+        sampling_rate=sampling_rate,
+        frame_length=frame_length,
+        frame_shift=frame_shift,
+        use_fft_mag=use_fft_mag,
+    )
+    extractor = Spectrogram(config)
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{partition} already exists - skipping.")
+            return
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=recordings, supervisions=supervisions
+        )
+
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+        cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_spectrogram_vctk()
diff --git a/egs/vctk/TTS/local/display_manifest_statistics.py b/egs/vctk/TTS/local/display_manifest_statistics.py
new file mode 100755
index 000000000..0472e2cea
--- /dev/null
+++ b/egs/vctk/TTS/local/display_manifest_statistics.py
@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Zengwei Yao,
+#                                                  Zengrui Jin,)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in vits/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    path = "./data/spectrogram/vctk_cuts_all.jsonl.gz"
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 43873    │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 41:02:18 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 3.4      │
+├───────────────────────────┼──────────┤
+│ std                       │ 1.2      │
+├───────────────────────────┼──────────┤
+│ min                       │ 1.2      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 2.6      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 3.1      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 3.8      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 8.0      │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 9.1      │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 12.1     │
+├───────────────────────────┼──────────┤
+│ max                       │ 16.6     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 43873    │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 43873    │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 43873    │
+╘═══════════════════════════╧══════════╛
+SUPERVISION custom fields:
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 41:02:18 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 41:02:18 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:01 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+"""
diff --git a/egs/vctk/TTS/local/prepare_token_file.py b/egs/vctk/TTS/local/prepare_token_file.py
new file mode 100755
index 000000000..c6636c3ad
--- /dev/null
+++ b/egs/vctk/TTS/local/prepare_token_file.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file reads the texts in given manifest and generates the file that maps tokens to IDs.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict
+
+from lhotse import load_manifest
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest-file",
+        type=Path,
+        default=Path("data/spectrogram/vctk_cuts_all.jsonl.gz"),
+        help="Path to the manifest file",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=Path,
+        default=Path("data/tokens.txt"),
+        help="Path to the tokens",
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_token2id(manifest_file: Path) -> Dict[str, int]:
+    """Return a dict that maps token to IDs."""
+    extra_tokens = [
+        "<blk>",  # 0 for blank
+        "<sos/eos>",  # 1 for sos and eos symbols.
+        "<unk>",  # 2 for OOV
+    ]
+    all_tokens = set()
+
+    cut_set = load_manifest(manifest_file)
+
+    for cut in cut_set:
+        # Each cut only contain one supervision
+        assert len(cut.supervisions) == 1, len(cut.supervisions)
+        for t in cut.tokens:
+            all_tokens.add(t)
+
+    all_tokens = extra_tokens + list(all_tokens)
+
+    token2id: Dict[str, int] = {token: i for i, token in enumerate(all_tokens)}
+    return token2id
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    manifest_file = Path(args.manifest_file)
+    out_file = Path(args.tokens)
+
+    token2id = get_token2id(manifest_file)
+    write_mapping(out_file, token2id)
diff --git a/egs/vctk/TTS/local/prepare_tokens_vctk.py b/egs/vctk/TTS/local/prepare_tokens_vctk.py
new file mode 100755
index 000000000..32e1c7dfa
--- /dev/null
+++ b/egs/vctk/TTS/local/prepare_tokens_vctk.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao,
+#                                                       Zengrui Jin,)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file reads the texts in given manifest and save the new cuts with phoneme tokens.
+"""
+
+import logging
+from pathlib import Path
+
+import g2p_en
+import tacotron_cleaner.cleaners
+from lhotse import CutSet, load_manifest
+from tqdm.auto import tqdm
+
+
+def prepare_tokens_vctk():
+    output_dir = Path("data/spectrogram")
+    prefix = "vctk"
+    suffix = "jsonl.gz"
+    partition = "all"
+
+    cut_set = load_manifest(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+    g2p = g2p_en.G2p()
+
+    new_cuts = []
+    for cut in tqdm(cut_set):
+        # Each cut only contains one supervision
+        assert len(cut.supervisions) == 1, len(cut.supervisions)
+        text = cut.supervisions[0].text
+        # Text normalization
+        text = tacotron_cleaner.cleaners.custom_english_cleaners(text)
+        # Convert to phonemes
+        cut.tokens = g2p(text)
+        new_cuts.append(cut)
+
+    new_cut_set = CutSet.from_cuts(new_cuts)
+    new_cut_set.to_file(output_dir / f"{prefix}_cuts_with_tokens_{partition}.{suffix}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    prepare_tokens_vctk()
diff --git a/egs/vctk/TTS/local/validate_manifest.py b/egs/vctk/TTS/local/validate_manifest.py
new file mode 100755
index 000000000..cd466303e
--- /dev/null
+++ b/egs/vctk/TTS/local/validate_manifest.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/spectrogram/ljspeech_cuts_all.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.dataset.speech_synthesis import validate_for_tts
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    manifest = args.manifest
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    validate_for_tts(cut_set)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/vctk/TTS/prepare.sh b/egs/vctk/TTS/prepare.sh
new file mode 100755
index 000000000..87150ad31
--- /dev/null
+++ b/egs/vctk/TTS/prepare.sh
@@ -0,0 +1,131 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=0
+stop_stage=100
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: build monotonic_align lib"
+  if [ ! -d vits/monotonic_align/build ]; then
+    cd vits/monotonic_align
+    python setup.py build_ext --inplace
+    cd ../../
+  else 
+    log "monotonic_align lib already built"
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/VCTK,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/VCTK $dl_dir/VCTK
+  #
+  if [ ! -d $dl_dir/VCTK ]; then
+    lhotse download vctk $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare VCTK manifest"
+  # We assume that you have downloaded the VCTK corpus
+  # to $dl_dir/VCTK
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.vctk.done ]; then
+    lhotse prepare vctk --use-edinburgh-vctk-url true $dl_dir/VCTK data/manifests
+    touch data/manifests/.vctk.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute spectrogram for VCTK"
+  mkdir -p data/spectrogram
+  if [ ! -e data/spectrogram/.vctk.done ]; then
+    ./local/compute_spectrogram_vctk.py
+    touch data/spectrogram/.vctk.done
+  fi
+
+  if [ ! -e data/spectrogram/.vctk-validated.done ]; then
+    log "Validating data/fbank for VCTK"
+    ./local/validate_manifest.py \
+      data/spectrogram/vctk_cuts_all.jsonl.gz
+    touch data/spectrogram/.vctk-validated.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare phoneme tokens for VCTK"
+  if [ ! -e data/spectrogram/.vctk_with_token.done ]; then
+    ./local/prepare_tokens_vctk.py
+    mv data/spectrogram/vctk_cuts_with_tokens_all.jsonl.gz \
+      data/spectrogram/vctk_cuts_all.jsonl.gz
+    touch data/spectrogram/.vctk_with_token.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Split the VCTK cuts into train, valid and test sets"
+  if [ ! -e data/spectrogram/.vctk_split.done ]; then
+    lhotse subset --last 600 \
+      data/spectrogram/vctk_cuts_all.jsonl.gz \
+      data/spectrogram/vctk_cuts_validtest.jsonl.gz
+    lhotse subset --first 100 \
+      data/spectrogram/vctk_cuts_validtest.jsonl.gz \
+      data/spectrogram/vctk_cuts_valid.jsonl.gz
+    lhotse subset --last 500 \
+      data/spectrogram/vctk_cuts_validtest.jsonl.gz \
+      data/spectrogram/vctk_cuts_test.jsonl.gz
+
+    rm data/spectrogram/vctk_cuts_validtest.jsonl.gz
+
+    n=$(( $(gunzip -c data/spectrogram/vctk_cuts_all.jsonl.gz | wc -l) - 600 ))
+    lhotse subset --first $n  \
+      data/spectrogram/vctk_cuts_all.jsonl.gz \
+      data/spectrogram/vctk_cuts_train.jsonl.gz
+      touch data/spectrogram/.vctk_split.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Generate token file"
+  # We assume you have installed g2p_en and espnet_tts_frontend.
+  # If not, please install them with:
+  #   - g2p_en: `pip install g2p_en`, refer to https://github.com/Kyubyong/g2p
+  #   - espnet_tts_frontend, `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/
+  if [ ! -e data/tokens.txt ]; then
+    ./local/prepare_token_file.py \
+      --manifest-file data/spectrogram/vctk_cuts_train.jsonl.gz \
+      --tokens data/tokens.txt
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Generate speakers file"
+  if [ ! -e data/speakers.txt ]; then
+    gunzip -c data/manifests/vctk_supervisions_all.jsonl.gz \
+      | jq '.speaker' | sed 's/"//g' \
+      | sort | uniq > data/speakers.txt
+  fi
+fi
diff --git a/egs/vctk/TTS/shared b/egs/vctk/TTS/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/vctk/TTS/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/duration_predictor.py b/egs/vctk/TTS/vits/duration_predictor.py
new file mode 120000
index 000000000..9972b476f
--- /dev/null
+++ b/egs/vctk/TTS/vits/duration_predictor.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/duration_predictor.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/export-onnx.py b/egs/vctk/TTS/vits/export-onnx.py
new file mode 100755
index 000000000..667ac284b
--- /dev/null
+++ b/egs/vctk/TTS/vits/export-onnx.py
@@ -0,0 +1,284 @@
+#!/usr/bin/env python3
+#
+# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script exports a VITS model from PyTorch to ONNX.
+
+Export the model to ONNX:
+./vits/export-onnx.py \
+  --epoch 1000 \
+  --exp-dir vits/exp \
+  --tokens data/tokens.txt
+
+It will generate two files inside vits/exp:
+  - vits-epoch-1000.onnx
+  - vits-epoch-1000.int8.onnx (quantizated model)
+
+See ./test_onnx.py for how to use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import torch
+import torch.nn as nn
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from tokenizer import Tokenizer
+from train import get_model, get_params
+
+from icefall.checkpoint import load_checkpoint
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=1000,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="vits/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--speakers",
+        type=Path,
+        default=Path("data/speakers.txt"),
+        help="Path to speakers.txt file.",
+    )
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxModel(nn.Module):
+    """A wrapper for VITS generator."""
+
+    def __init__(self, model: nn.Module):
+        """
+        Args:
+          model:
+            A VITS generator.
+          frame_shift:
+            The frame shift in samples.
+        """
+        super().__init__()
+        self.model = model
+
+    def forward(
+        self,
+        tokens: torch.Tensor,
+        tokens_lens: torch.Tensor,
+        noise_scale: float = 0.667,
+        alpha: float = 1.0,
+        noise_scale_dur: float = 0.8,
+        speaker: int = 20,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of VITS.inference_batch
+
+        Args:
+          tokens:
+            Input text token indexes (1, T_text)
+          tokens_lens:
+            Number of tokens of shape (1,)
+          noise_scale (float):
+            Noise scale parameter for flow.
+          noise_scale_dur (float):
+            Noise scale parameter for duration predictor.
+          speaker (int):
+            Speaker ID.
+          alpha (float):
+            Alpha parameter to control the speed of generated speech.
+
+        Returns:
+          Return a tuple containing:
+            - audio, generated wavform tensor, (B, T_wav)
+        """
+        audio, _, _ = self.model.inference(
+            text=tokens,
+            text_lengths=tokens_lens,
+            noise_scale=noise_scale,
+            noise_scale_dur=noise_scale_dur,
+            sids=speaker,
+            alpha=alpha,
+        )
+        return audio
+
+
+def export_model_onnx(
+    model: nn.Module,
+    model_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given generator model to ONNX format.
+    The exported model has one input:
+
+        - tokens, a tensor of shape (1, T_text); dtype is torch.int64
+
+    and it has one output:
+
+        - audio, a tensor of shape (1, T'); dtype is torch.float32
+
+    Args:
+      model:
+        The VITS generator.
+      model_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    tokens = torch.randint(low=0, high=79, size=(1, 13), dtype=torch.int64)
+    tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64)
+    noise_scale = torch.tensor([1], dtype=torch.float32)
+    noise_scale_dur = torch.tensor([1], dtype=torch.float32)
+    alpha = torch.tensor([1], dtype=torch.float32)
+    speaker = torch.tensor([1], dtype=torch.int64)
+
+    torch.onnx.export(
+        model,
+        (tokens, tokens_lens, noise_scale, alpha, noise_scale_dur, speaker),
+        model_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "tokens",
+            "tokens_lens",
+            "noise_scale",
+            "alpha",
+            "noise_scale_dur",
+            "speaker",
+        ],
+        output_names=["audio"],
+        dynamic_axes={
+            "tokens": {0: "N", 1: "T"},
+            "tokens_lens": {0: "N"},
+            "audio": {0: "N", 1: "T"},
+            "speaker": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "VITS",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "VITS generator",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=model_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.blank_id
+    params.oov_id = tokenizer.oov_id
+    params.vocab_size = tokenizer.vocab_size
+
+    with open(args.speakers) as f:
+        speaker_map = {line.strip(): i for i, line in enumerate(f)}
+    params.num_spks = len(speaker_map)
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+
+    model = model.generator
+    model.to("cpu")
+    model.eval()
+
+    model = OnnxModel(model=model)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"generator parameters: {num_param}")
+
+    suffix = f"epoch-{params.epoch}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    model_filename = params.exp_dir / f"vits-{suffix}.onnx"
+    export_model_onnx(
+        model,
+        model_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported generator to {model_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    model_filename_int8 = params.exp_dir / f"vits-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=model_filename,
+        model_output=model_filename_int8,
+        weight_type=QuantType.QUInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/vctk/TTS/vits/flow.py b/egs/vctk/TTS/vits/flow.py
new file mode 120000
index 000000000..e65d91ea7
--- /dev/null
+++ b/egs/vctk/TTS/vits/flow.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/flow.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/generator.py b/egs/vctk/TTS/vits/generator.py
new file mode 120000
index 000000000..611679bfa
--- /dev/null
+++ b/egs/vctk/TTS/vits/generator.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/generator.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/hifigan.py b/egs/vctk/TTS/vits/hifigan.py
new file mode 120000
index 000000000..5ac025de7
--- /dev/null
+++ b/egs/vctk/TTS/vits/hifigan.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/hifigan.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/infer.py b/egs/vctk/TTS/vits/infer.py
new file mode 100755
index 000000000..06c25f02e
--- /dev/null
+++ b/egs/vctk/TTS/vits/infer.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+#
+# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao,
+#                                                     Zengrui Jin,)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script performs model inference on test set.
+
+Usage:
+./vits/infer.py \
+    --epoch 1000 \
+    --exp-dir ./vits/exp \
+    --max-duration 500
+"""
+
+
+import argparse
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+import torch.nn as nn
+import torchaudio
+from tokenizer import Tokenizer
+from train import get_model, get_params
+from tts_datamodule import VctkTtsDataModule
+
+from icefall.checkpoint import load_checkpoint
+from icefall.utils import AttributeDict, setup_logger
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=1000,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="vits/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    return parser
+
+
+def infer_dataset(
+    dl: torch.utils.data.DataLoader,
+    subset: str,
+    params: AttributeDict,
+    model: nn.Module,
+    tokenizer: Tokenizer,
+    speaker_map: Dict[str, int],
+) -> None:
+    """Decode dataset.
+    The ground-truth and generated audio pairs will be saved to `params.save_wav_dir`.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      tokenizer:
+        Used to convert text to phonemes.
+    """
+
+    #  Background worker save audios to disk.
+    def _save_worker(
+        subset: str,
+        batch_size: int,
+        cut_ids: List[str],
+        audio: torch.Tensor,
+        audio_pred: torch.Tensor,
+        audio_lens: List[int],
+        audio_lens_pred: List[int],
+    ):
+        for i in range(batch_size):
+            torchaudio.save(
+                str(params.save_wav_dir / subset / f"{cut_ids[i]}_gt.wav"),
+                audio[i : i + 1, : audio_lens[i]],
+                sample_rate=params.sampling_rate,
+            )
+            torchaudio.save(
+                str(params.save_wav_dir / subset / f"{cut_ids[i]}_pred.wav"),
+                audio_pred[i : i + 1, : audio_lens_pred[i]],
+                sample_rate=params.sampling_rate,
+            )
+
+    device = next(model.parameters()).device
+    num_cuts = 0
+    log_interval = 5
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    futures = []
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        for batch_idx, batch in enumerate(dl):
+            batch_size = len(batch["tokens"])
+
+            tokens = batch["tokens"]
+            tokens = tokenizer.tokens_to_token_ids(tokens)
+            tokens = k2.RaggedTensor(tokens)
+            row_splits = tokens.shape.row_splits(1)
+            tokens_lens = row_splits[1:] - row_splits[:-1]
+            tokens = tokens.to(device)
+            tokens_lens = tokens_lens.to(device)
+            # tensor of shape (B, T)
+            tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id)
+            speakers = (
+                torch.Tensor([speaker_map[sid] for sid in batch["speakers"]])
+                .int()
+                .to(device)
+            )
+
+            audio = batch["audio"]
+            audio_lens = batch["audio_lens"].tolist()
+            cut_ids = [cut.id for cut in batch["cut"]]
+
+            audio_pred, _, durations = model.inference_batch(
+                text=tokens,
+                text_lengths=tokens_lens,
+                sids=speakers,
+            )
+            audio_pred = audio_pred.detach().cpu()
+            # convert to samples
+            audio_lens_pred = (
+                (durations.sum(1) * params.frame_shift).to(dtype=torch.int64).tolist()
+            )
+
+            futures.append(
+                executor.submit(
+                    _save_worker,
+                    subset,
+                    batch_size,
+                    cut_ids,
+                    audio,
+                    audio_pred,
+                    audio_lens,
+                    audio_lens_pred,
+                )
+            )
+
+            num_cuts += batch_size
+
+            if batch_idx % log_interval == 0:
+                batch_str = f"{batch_idx}/{num_batches}"
+
+                logging.info(
+                    f"batch {batch_str}, cuts processed until now is {num_cuts}"
+                )
+        # return results
+        for f in futures:
+            f.result()
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    VctkTtsDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.suffix = f"epoch-{params.epoch}"
+
+    params.res_dir = params.exp_dir / "infer" / params.suffix
+    params.save_wav_dir = params.res_dir / "wav"
+    params.save_wav_dir.mkdir(parents=True, exist_ok=True)
+
+    setup_logger(f"{params.res_dir}/log-infer-{params.suffix}")
+    logging.info("Infer started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.blank_id
+    params.oov_id = tokenizer.oov_id
+    params.vocab_size = tokenizer.vocab_size
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    vctk = VctkTtsDataModule(args)
+    speaker_map = vctk.speakers()
+    params.num_spks = len(speaker_map)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+
+    model.to(device)
+    model.eval()
+
+    num_param_g = sum([p.numel() for p in model.generator.parameters()])
+    logging.info(f"Number of parameters in generator: {num_param_g}")
+    num_param_d = sum([p.numel() for p in model.discriminator.parameters()])
+    logging.info(f"Number of parameters in discriminator: {num_param_d}")
+    logging.info(f"Total number of parameters: {num_param_g + num_param_d}")
+
+    test_cuts = vctk.test_cuts()
+    test_dl = vctk.test_dataloaders(test_cuts)
+
+    valid_cuts = vctk.valid_cuts()
+    valid_dl = vctk.valid_dataloaders(valid_cuts)
+
+    infer_sets = {"test": test_dl, "valid": valid_dl}
+
+    for subset, dl in infer_sets.items():
+        save_wav_dir = params.res_dir / "wav" / subset
+        save_wav_dir.mkdir(parents=True, exist_ok=True)
+
+        logging.info(f"Processing {subset} set, saving to {save_wav_dir}")
+
+        infer_dataset(
+            dl=dl,
+            subset=subset,
+            params=params,
+            model=model,
+            tokenizer=tokenizer,
+            speaker_map=speaker_map,
+        )
+
+    logging.info(f"Wav files are saved to {params.save_wav_dir}")
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/vctk/TTS/vits/loss.py b/egs/vctk/TTS/vits/loss.py
new file mode 120000
index 000000000..672e5ff68
--- /dev/null
+++ b/egs/vctk/TTS/vits/loss.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/loss.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/monotonic_align b/egs/vctk/TTS/vits/monotonic_align
new file mode 120000
index 000000000..71934e7cc
--- /dev/null
+++ b/egs/vctk/TTS/vits/monotonic_align
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/monotonic_align
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/posterior_encoder.py b/egs/vctk/TTS/vits/posterior_encoder.py
new file mode 120000
index 000000000..41d64a3a6
--- /dev/null
+++ b/egs/vctk/TTS/vits/posterior_encoder.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/posterior_encoder.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/residual_coupling.py b/egs/vctk/TTS/vits/residual_coupling.py
new file mode 120000
index 000000000..f979adbf0
--- /dev/null
+++ b/egs/vctk/TTS/vits/residual_coupling.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/residual_coupling.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/test_onnx.py b/egs/vctk/TTS/vits/test_onnx.py
new file mode 100755
index 000000000..d85c0a27b
--- /dev/null
+++ b/egs/vctk/TTS/vits/test_onnx.py
@@ -0,0 +1,138 @@
+#!/usr/bin/env python3
+#
+# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script is used to test the exported onnx model by vits/export-onnx.py
+
+Use the onnx model to generate a wav:
+./vits/test_onnx.py \
+  --model-filename vits/exp/vits-epoch-1000.onnx \
+  --tokens data/tokens.txt
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+
+import onnxruntime as ort
+import torch
+import torchaudio
+from tokenizer import Tokenizer
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model-filename",
+        type=str,
+        required=True,
+        help="Path to the onnx model.",
+    )
+
+    parser.add_argument(
+        "--speakers",
+        type=Path,
+        default=Path("data/speakers.txt"),
+        help="Path to speakers.txt file.",
+    )
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(self, model_filename: str):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 4
+
+        self.session_opts = session_opts
+
+        self.model = ort.InferenceSession(
+            model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        logging.info(f"{self.model.get_modelmeta().custom_metadata_map}")
+
+    def __call__(
+        self, tokens: torch.Tensor, tokens_lens: torch.Tensor, speaker: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          tokens:
+            A 1-D tensor of shape (1, T)
+        Returns:
+            A tensor of shape (1, T')
+        """
+        noise_scale = torch.tensor([0.667], dtype=torch.float32)
+        noise_scale_dur = torch.tensor([0.8], dtype=torch.float32)
+        alpha = torch.tensor([1.0], dtype=torch.float32)
+
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: tokens.numpy(),
+                self.model.get_inputs()[1].name: tokens_lens.numpy(),
+                self.model.get_inputs()[2].name: noise_scale.numpy(),
+                self.model.get_inputs()[3].name: alpha.numpy(),
+                self.model.get_inputs()[4].name: noise_scale_dur.numpy(),
+                self.model.get_inputs()[5].name: speaker.numpy(),
+            },
+        )[0]
+        return torch.from_numpy(out)
+
+
+def main():
+    args = get_parser().parse_args()
+
+    tokenizer = Tokenizer(args.tokens)
+
+    with open(args.speakers) as f:
+        speaker_map = {line.strip(): i for i, line in enumerate(f)}
+    args.num_spks = len(speaker_map)
+
+    logging.info("About to create onnx model")
+    model = OnnxModel(args.model_filename)
+
+    text = "I went there to see the land, the people and how their system works, end quote."
+    tokens = tokenizer.texts_to_token_ids([text])
+    tokens = torch.tensor(tokens)  # (1, T)
+    tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64)  # (1, T)
+    speaker = torch.tensor([1], dtype=torch.int64)  # (1, )
+    audio = model(tokens, tokens_lens, speaker)  # (1, T')
+
+    torchaudio.save(str("test_onnx.wav"), audio, sample_rate=22050)
+    logging.info("Saved to test_onnx.wav")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/vctk/TTS/vits/text_encoder.py b/egs/vctk/TTS/vits/text_encoder.py
new file mode 120000
index 000000000..0efba277e
--- /dev/null
+++ b/egs/vctk/TTS/vits/text_encoder.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/text_encoder.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/tokenizer.py b/egs/vctk/TTS/vits/tokenizer.py
new file mode 120000
index 000000000..057b0dc4b
--- /dev/null
+++ b/egs/vctk/TTS/vits/tokenizer.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/tokenizer.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/train.py b/egs/vctk/TTS/vits/train.py
new file mode 100755
index 000000000..56f167a17
--- /dev/null
+++ b/egs/vctk/TTS/vits/train.py
@@ -0,0 +1,1000 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import numpy as np
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from tokenizer import Tokenizer
+from torch.cuda.amp import GradScaler, autocast
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import Optimizer
+from torch.utils.tensorboard import SummaryWriter
+from tts_datamodule import VctkTtsDataModule
+from utils import MetricsTracker, plot_feature, save_checkpoint
+from vits import VITS
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+LRSchedulerType = torch.optim.lr_scheduler._LRScheduler
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=1000,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="vits/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/tokens.txt",
+        help="""Path to vocabulary.""",
+    )
+
+    parser.add_argument(
+        "--lr", type=float, default=2.0e-4, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=20,
+        help="""Save checkpoint after processing this number of epochs"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.cur_epoch % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/epoch-{params.cur_epoch}.pt'.
+        Since it will take around 1000 epochs, we suggest using a large
+        save_every_n to save disk space.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            # training params
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": -1,  # 0
+            "log_interval": 50,
+            "valid_interval": 200,
+            "env_info": get_env_info(),
+            "sampling_rate": 22050,
+            "frame_shift": 256,
+            "frame_length": 1024,
+            "feature_dim": 513,  # 1024 // 2 + 1, 1024 is fft_length
+            "n_mels": 80,
+            "lambda_adv": 1.0,  # loss scaling coefficient for adversarial loss
+            "lambda_mel": 45.0,  # loss scaling coefficient for Mel loss
+            "lambda_feat_match": 2.0,  # loss scaling coefficient for feat match loss
+            "lambda_dur": 1.0,  # loss scaling coefficient for duration loss
+            "lambda_kl": 1.0,  # loss scaling coefficient for KL divergence loss
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict, model: nn.Module
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(filename, model=model)
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    mel_loss_params = {
+        "n_mels": params.n_mels,
+        "frame_length": params.frame_length,
+        "frame_shift": params.frame_shift,
+    }
+    generator_params = {
+        "hidden_channels": 192,
+        "spks": params.num_spks,
+        "langs": None,
+        "spk_embed_dim": None,
+        "global_channels": 256,
+        "segment_size": 32,
+        "text_encoder_attention_heads": 2,
+        "text_encoder_ffn_expand": 4,
+        "text_encoder_cnn_module_kernel": 5,
+        "text_encoder_blocks": 6,
+        "text_encoder_dropout_rate": 0.1,
+        "decoder_kernel_size": 7,
+        "decoder_channels": 512,
+        "decoder_upsample_scales": [8, 8, 2, 2],
+        "decoder_upsample_kernel_sizes": [16, 16, 4, 4],
+        "decoder_resblock_kernel_sizes": [3, 7, 11],
+        "decoder_resblock_dilations": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+        "use_weight_norm_in_decoder": True,
+        "posterior_encoder_kernel_size": 5,
+        "posterior_encoder_layers": 16,
+        "posterior_encoder_stacks": 1,
+        "posterior_encoder_base_dilation": 1,
+        "posterior_encoder_dropout_rate": 0.0,
+        "use_weight_norm_in_posterior_encoder": True,
+        "flow_flows": 4,
+        "flow_kernel_size": 5,
+        "flow_base_dilation": 1,
+        "flow_layers": 4,
+        "flow_dropout_rate": 0.0,
+        "use_weight_norm_in_flow": True,
+        "use_only_mean_in_flow": True,
+        "stochastic_duration_predictor_kernel_size": 3,
+        "stochastic_duration_predictor_dropout_rate": 0.5,
+        "stochastic_duration_predictor_flows": 4,
+        "stochastic_duration_predictor_dds_conv_layers": 3,
+    }
+    model = VITS(
+        vocab_size=params.vocab_size,
+        feature_dim=params.feature_dim,
+        sampling_rate=params.sampling_rate,
+        generator_params=generator_params,
+        mel_loss_params=mel_loss_params,
+        lambda_adv=params.lambda_adv,
+        lambda_mel=params.lambda_mel,
+        lambda_feat_match=params.lambda_feat_match,
+        lambda_dur=params.lambda_dur,
+        lambda_kl=params.lambda_kl,
+    )
+    return model
+
+
+def prepare_input(
+    batch: dict,
+    tokenizer: Tokenizer,
+    device: torch.device,
+    speaker_map: Dict[str, int],
+):
+    """Parse batch data"""
+    audio = batch["audio"].to(device)
+    features = batch["features"].to(device)
+    audio_lens = batch["audio_lens"].to(device)
+    features_lens = batch["features_lens"].to(device)
+    tokens = batch["tokens"]
+    speakers = (
+        torch.Tensor([speaker_map[sid] for sid in batch["speakers"]]).int().to(device)
+    )
+
+    tokens = tokenizer.tokens_to_token_ids(tokens)
+    tokens = k2.RaggedTensor(tokens)
+    row_splits = tokens.shape.row_splits(1)
+    tokens_lens = row_splits[1:] - row_splits[:-1]
+    tokens = tokens.to(device)
+    tokens_lens = tokens_lens.to(device)
+    # a tensor of shape (B, T)
+    tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id)
+
+    return audio, audio_lens, features, features_lens, tokens, tokens_lens, speakers
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    tokenizer: Tokenizer,
+    optimizer_g: Optimizer,
+    optimizer_d: Optimizer,
+    scheduler_g: LRSchedulerType,
+    scheduler_d: LRSchedulerType,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    speaker_map: Dict[str, int],
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      tokenizer:
+        Used to convert text to phonemes.
+      optimizer_g:
+        The optimizer for generator.
+      optimizer_d:
+        The optimizer for discriminator.
+      scheduler_g:
+        The learning rate scheduler for generator, we call step() every epoch.
+      scheduler_d:
+        The learning rate scheduler for discriminator, we call step() every epoch.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+
+    # used to summary the stats over iterations in one epoch
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            params=params,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            scheduler_g=scheduler_g,
+            scheduler_d=scheduler_d,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+
+        batch_size = len(batch["tokens"])
+        (
+            audio,
+            audio_lens,
+            features,
+            features_lens,
+            tokens,
+            tokens_lens,
+            speakers,
+        ) = prepare_input(batch, tokenizer, device, speaker_map)
+
+        loss_info = MetricsTracker()
+        loss_info["samples"] = batch_size
+
+        try:
+            with autocast(enabled=params.use_fp16):
+                # forward discriminator
+                loss_d, stats_d = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    sids=speakers,
+                    forward_generator=False,
+                )
+            for k, v in stats_d.items():
+                loss_info[k] = v * batch_size
+            # update discriminator
+            optimizer_d.zero_grad()
+            scaler.scale(loss_d).backward()
+            scaler.step(optimizer_d)
+
+            with autocast(enabled=params.use_fp16):
+                # forward generator
+                loss_g, stats_g = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    sids=speakers,
+                    forward_generator=True,
+                    return_sample=params.batch_idx_train % params.log_interval == 0,
+                )
+            for k, v in stats_g.items():
+                if "returned_sample" not in k:
+                    loss_info[k] = v * batch_size
+            # update generator
+            optimizer_g.zero_grad()
+            scaler.scale(loss_g).backward()
+            scaler.step(optimizer_g)
+            scaler.update()
+
+            # summary stats
+            tot_loss = tot_loss + loss_info
+        except:  # noqa
+            save_bad_model()
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if params.batch_idx_train % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (
+                cur_grad_scale < 32.0 and params.batch_idx_train % 400 == 0
+            ):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if params.batch_idx_train % params.log_interval == 0:
+            cur_lr_g = max(scheduler_g.get_last_lr())
+            cur_lr_d = max(scheduler_d.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"global_batch_idx: {params.batch_idx_train}, batch size: {batch_size}, "
+                f"loss[{loss_info}], tot_loss[{tot_loss}], "
+                f"cur_lr_g: {cur_lr_g:.2e}, cur_lr_d: {cur_lr_d:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate_g", cur_lr_g, params.batch_idx_train
+                )
+                tb_writer.add_scalar(
+                    "train/learning_rate_d", cur_lr_d, params.batch_idx_train
+                )
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+                if "returned_sample" in stats_g:
+                    speech_hat_, speech_, mel_hat_, mel_ = stats_g["returned_sample"]
+                    tb_writer.add_audio(
+                        "train/speech_hat_",
+                        speech_hat_,
+                        params.batch_idx_train,
+                        params.sampling_rate,
+                    )
+                    tb_writer.add_audio(
+                        "train/speech_",
+                        speech_,
+                        params.batch_idx_train,
+                        params.sampling_rate,
+                    )
+                    tb_writer.add_image(
+                        "train/mel_hat_",
+                        plot_feature(mel_hat_),
+                        params.batch_idx_train,
+                        dataformats="HWC",
+                    )
+                    tb_writer.add_image(
+                        "train/mel_",
+                        plot_feature(mel_),
+                        params.batch_idx_train,
+                        dataformats="HWC",
+                    )
+
+        if (
+            params.batch_idx_train % params.valid_interval == 0
+            and not params.print_diagnostics
+        ):
+            logging.info("Computing validation loss")
+            valid_info, (speech_hat, speech) = compute_validation_loss(
+                params=params,
+                model=model,
+                tokenizer=tokenizer,
+                valid_dl=valid_dl,
+                speaker_map=speaker_map,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+                tb_writer.add_audio(
+                    "train/valdi_speech_hat",
+                    speech_hat,
+                    params.batch_idx_train,
+                    params.sampling_rate,
+                )
+                tb_writer.add_audio(
+                    "train/valdi_speech",
+                    speech,
+                    params.batch_idx_train,
+                    params.sampling_rate,
+                )
+
+    loss_value = tot_loss["generator_loss"] / tot_loss["samples"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    tokenizer: Tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    speaker_map: Dict[str, int],
+    world_size: int = 1,
+    rank: int = 0,
+) -> Tuple[MetricsTracker, Tuple[np.ndarray, np.ndarray]]:
+    """Run the validation process."""
+    model.eval()
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+
+    # used to summary the stats over iterations
+    tot_loss = MetricsTracker()
+    returned_sample = None
+
+    with torch.no_grad():
+        for batch_idx, batch in enumerate(valid_dl):
+            batch_size = len(batch["tokens"])
+            (
+                audio,
+                audio_lens,
+                features,
+                features_lens,
+                tokens,
+                tokens_lens,
+                speakers,
+            ) = prepare_input(batch, tokenizer, device, speaker_map)
+
+            loss_info = MetricsTracker()
+            loss_info["samples"] = batch_size
+
+            # forward discriminator
+            loss_d, stats_d = model(
+                text=tokens,
+                text_lengths=tokens_lens,
+                feats=features,
+                feats_lengths=features_lens,
+                speech=audio,
+                speech_lengths=audio_lens,
+                sids=speakers,
+                forward_generator=False,
+            )
+            assert loss_d.requires_grad is False
+            for k, v in stats_d.items():
+                loss_info[k] = v * batch_size
+
+            # forward generator
+            loss_g, stats_g = model(
+                text=tokens,
+                text_lengths=tokens_lens,
+                feats=features,
+                feats_lengths=features_lens,
+                speech=audio,
+                speech_lengths=audio_lens,
+                sids=speakers,
+                forward_generator=True,
+            )
+            assert loss_g.requires_grad is False
+            for k, v in stats_g.items():
+                loss_info[k] = v * batch_size
+
+            # summary stats
+            tot_loss = tot_loss + loss_info
+
+            # infer for first batch:
+            if batch_idx == 0 and rank == 0:
+                inner_model = model.module if isinstance(model, DDP) else model
+                audio_pred, _, duration = inner_model.inference(
+                    text=tokens[0, : tokens_lens[0].item()],
+                    sids=speakers[0],
+                )
+                audio_pred = audio_pred.data.cpu().numpy()
+                audio_len_pred = (
+                    (duration.sum(0) * params.frame_shift).to(dtype=torch.int64).item()
+                )
+                assert audio_len_pred == len(audio_pred), (
+                    audio_len_pred,
+                    len(audio_pred),
+                )
+                audio_gt = audio[0, : audio_lens[0].item()].data.cpu().numpy()
+                returned_sample = (audio_pred, audio_gt)
+
+    if world_size > 1:
+        tot_loss.reduce(device)
+
+    loss_value = tot_loss["generator_loss"] / tot_loss["samples"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss, returned_sample
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    tokenizer: Tokenizer,
+    optimizer_g: torch.optim.Optimizer,
+    optimizer_d: torch.optim.Optimizer,
+    speaker_map: Dict[str, int],
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        (
+            audio,
+            audio_lens,
+            features,
+            features_lens,
+            tokens,
+            tokens_lens,
+            speakers,
+        ) = prepare_input(batch, tokenizer, device, speaker_map)
+        try:
+            # for discriminator
+            with autocast(enabled=params.use_fp16):
+                loss_d, stats_d = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    sids=speakers,
+                    forward_generator=False,
+                )
+            optimizer_d.zero_grad()
+            loss_d.backward()
+            # for generator
+            with autocast(enabled=params.use_fp16):
+                loss_g, stats_g = model(
+                    text=tokens,
+                    text_lengths=tokens_lens,
+                    feats=features,
+                    feats_lengths=features_lens,
+                    speech=audio,
+                    speech_lengths=audio_lens,
+                    sids=speakers,
+                    forward_generator=True,
+                )
+            optimizer_g.zero_grad()
+            loss_g.backward()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    tokenizer = Tokenizer(params.tokens)
+    params.blank_id = tokenizer.blank_id
+    params.oov_id = tokenizer.oov_id
+    params.vocab_size = tokenizer.vocab_size
+
+    vctk = VctkTtsDataModule(args)
+
+    train_cuts = vctk.train_cuts()
+    speaker_map = vctk.speakers()
+    params.num_spks = len(speaker_map)
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+    generator = model.generator
+    discriminator = model.discriminator
+
+    num_param_g = sum([p.numel() for p in generator.parameters()])
+    logging.info(f"Number of parameters in generator: {num_param_g}")
+    num_param_d = sum([p.numel() for p in discriminator.parameters()])
+    logging.info(f"Number of parameters in discriminator: {num_param_d}")
+    logging.info(f"Total number of parameters: {num_param_g + num_param_d}")
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer_g = torch.optim.AdamW(
+        generator.parameters(), lr=params.lr, betas=(0.8, 0.99), eps=1e-9
+    )
+    optimizer_d = torch.optim.AdamW(
+        discriminator.parameters(), lr=params.lr, betas=(0.8, 0.99), eps=1e-9
+    )
+
+    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optimizer_g, gamma=0.999875)
+    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optimizer_d, gamma=0.999875)
+
+    if checkpoints is not None:
+        # load state_dict for optimizers
+        if "optimizer_g" in checkpoints:
+            logging.info("Loading optimizer_g state dict")
+            optimizer_g.load_state_dict(checkpoints["optimizer_g"])
+        if "optimizer_d" in checkpoints:
+            logging.info("Loading optimizer_d state dict")
+            optimizer_d.load_state_dict(checkpoints["optimizer_d"])
+
+        # load state_dict for schedulers
+        if "scheduler_g" in checkpoints:
+            logging.info("Loading scheduler_g state dict")
+            scheduler_g.load_state_dict(checkpoints["scheduler_g"])
+        if "scheduler_d" in checkpoints:
+            logging.info("Loading scheduler_d state dict")
+            scheduler_d.load_state_dict(checkpoints["scheduler_d"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            # logging.warning(
+            #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_dl = vctk.train_dataloaders(train_cuts)
+
+    valid_cuts = vctk.valid_cuts()
+    valid_dl = vctk.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            tokenizer=tokenizer,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            speaker_map=speaker_map,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        logging.info(f"Start epoch {epoch}")
+
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        params.cur_epoch = epoch
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            tokenizer=tokenizer,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            scheduler_g=scheduler_g,
+            scheduler_d=scheduler_d,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            speaker_map=speaker_map,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        if epoch % params.save_every_n == 0 or epoch == params.num_epochs:
+            filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+            save_checkpoint(
+                filename=filename,
+                params=params,
+                model=model,
+                optimizer_g=optimizer_g,
+                optimizer_d=optimizer_d,
+                scheduler_g=scheduler_g,
+                scheduler_d=scheduler_d,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            if rank == 0:
+                if params.best_train_epoch == params.cur_epoch:
+                    best_train_filename = params.exp_dir / "best-train-loss.pt"
+                    copyfile(src=filename, dst=best_train_filename)
+
+                if params.best_valid_epoch == params.cur_epoch:
+                    best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+                    copyfile(src=filename, dst=best_valid_filename)
+
+        # step per epoch
+        scheduler_g.step()
+        scheduler_d.step()
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    VctkTtsDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/vctk/TTS/vits/transform.py b/egs/vctk/TTS/vits/transform.py
new file mode 120000
index 000000000..962647408
--- /dev/null
+++ b/egs/vctk/TTS/vits/transform.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/transform.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/tts_datamodule.py b/egs/vctk/TTS/vits/tts_datamodule.py
new file mode 100644
index 000000000..8b2a96b09
--- /dev/null
+++ b/egs/vctk/TTS/vits/tts_datamodule.py
@@ -0,0 +1,338 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022-2023  Xiaomi Corporation     (Authors: Mingshuang Luo,
+#                                                            Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Spectrogram, SpectrogramConfig, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+    SpeechSynthesisDataset,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class VctkTtsDataModule:
+    """
+    DataModule for tts experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="TTS data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/spectrogram"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--speakers",
+            type=Path,
+            default=Path("data/speakers.txt"),
+            help="Path to speakers.txt file.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, each batch will have the "
+            "field: batch['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=8,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to create train dataset")
+        train = SpeechSynthesisDataset(
+            return_text=False,
+            return_tokens=True,
+            return_spk_ids=True,
+            feature_input_strategy=eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            sampling_rate = 22050
+            config = SpectrogramConfig(
+                sampling_rate=sampling_rate,
+                frame_length=1024 / sampling_rate,  # (in second),
+                frame_shift=256 / sampling_rate,  # (in second)
+                use_fft_mag=True,
+            )
+            train = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                return_spk_ids=True,
+                feature_input_strategy=OnTheFlyFeatures(Spectrogram(config)),
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            sampling_rate = 22050
+            config = SpectrogramConfig(
+                sampling_rate=sampling_rate,
+                frame_length=1024 / sampling_rate,  # (in second),
+                frame_shift=256 / sampling_rate,  # (in second)
+                use_fft_mag=True,
+            )
+            validate = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                return_spk_ids=True,
+                feature_input_strategy=OnTheFlyFeatures(Spectrogram(config)),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                return_spk_ids=True,
+                feature_input_strategy=eval(self.args.input_strategy)(),
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create valid dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.info("About to create test dataset")
+        if self.args.on_the_fly_feats:
+            sampling_rate = 22050
+            config = SpectrogramConfig(
+                sampling_rate=sampling_rate,
+                frame_length=1024 / sampling_rate,  # (in second),
+                frame_shift=256 / sampling_rate,  # (in second)
+                use_fft_mag=True,
+            )
+            test = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                return_spk_ids=True,
+                feature_input_strategy=OnTheFlyFeatures(Spectrogram(config)),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            test = SpeechSynthesisDataset(
+                return_text=False,
+                return_tokens=True,
+                return_spk_ids=True,
+                feature_input_strategy=eval(self.args.input_strategy)(),
+                return_cuts=self.args.return_cuts,
+            )
+        test_sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=test_sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "vctk_cuts_train.jsonl.gz")
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get validation cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "vctk_cuts_valid.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "vctk_cuts_test.jsonl.gz")
+
+    @lru_cache()
+    def speakers(self) -> Dict[str, int]:
+        logging.info("About to get speakers")
+        with open(self.args.speakers) as f:
+            speakers = {line.strip(): i for i, line in enumerate(f)}
+        return speakers
diff --git a/egs/vctk/TTS/vits/utils.py b/egs/vctk/TTS/vits/utils.py
new file mode 120000
index 000000000..085e764b4
--- /dev/null
+++ b/egs/vctk/TTS/vits/utils.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/utils.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/vits.py b/egs/vctk/TTS/vits/vits.py
new file mode 120000
index 000000000..1f58cf6fe
--- /dev/null
+++ b/egs/vctk/TTS/vits/vits.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/vits.py
\ No newline at end of file
diff --git a/egs/vctk/TTS/vits/wavenet.py b/egs/vctk/TTS/vits/wavenet.py
new file mode 120000
index 000000000..28f0a78ee
--- /dev/null
+++ b/egs/vctk/TTS/vits/wavenet.py
@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/wavenet.py
\ No newline at end of file
diff --git a/egs/voxpopuli/ASR/README.md b/egs/voxpopuli/ASR/README.md
new file mode 100644
index 000000000..92aa26464
--- /dev/null
+++ b/egs/voxpopuli/ASR/README.md
@@ -0,0 +1,38 @@
+# Readme
+
+This recipe contains data preparation for the
+[VoxPopuli](https://github.com/facebookresearch/voxpopuli) dataset
+[(pdf)](https://aclanthology.org/2021.acl-long.80.pdf).
+At the moment, without model training.
+
+
+## audio per language
+
+| language |  Size  | Hrs. untranscribed | Hrs. transcribed |
+|----------|--------|--------------------|------------------|
+|  bg      |  295G  |  17.6K             |    -             |
+|  cs      |  308G  |  18.7K             |   62             |
+|  da      |  233G  |  13.6K             |    -             |
+|  de      |  379G  |  23.2K             |  282             |
+|  el      |  305G  |  17.7K             |    -             |
+|  en      |  382G  |  24.1K             |  543             |
+|  es      |  362G  |  21.4K             |  166             |
+|  et      |  179G  |  10.6K             |    3             |
+|  fi      |  236G  |  14.2K             |   27             |
+|  fr      |  376G  |  22.8K             |  211             |
+|  hr      |  132G  |   8.1K             |   43             |
+|  hu      |  297G  |  17.7K             |   63             |
+|  it      |  361G  |  21.9K             |   91             |
+|  lt      |  243G  |  14.4K             |    2             |
+|  lv      |  217G  |  13.1K             |    -             |
+|  mt      |  147G  |   9.1K             |    -             |
+|  nl      |  322G  |  19.0K             |   53             |
+|  pl      |  348G  |  21.2K             |  111             |
+|  pt      |  300G  |  17.5K             |    -             |
+|  ro      |  296G  |  17.9K             |   89             |
+|  sk      |  201G  |  12.1K             |   35             |
+|  sl      |  190G  |  11.3K             |   10             |
+|  sv      |  272G  |  16.3K             |    -             |
+|          |        |                    |                  |
+|  total   |  6.3T  |   384K             | 1791             |
+
diff --git a/egs/voxpopuli/ASR/local/compute_fbank.py b/egs/voxpopuli/ASR/local/compute_fbank.py
new file mode 100755
index 000000000..b63e51f29
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/compute_fbank.py
@@ -0,0 +1,248 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of VoxPopuli dataset.
+
+Usage example:
+
+  python3 ./local/compute_fbank.py \
+      --src-dir data/fbank --output-dir data/fbank \
+      --num-jobs 100 --num-workers 25 \
+      --prefix "voxpopuli-${task}-${lang}" \
+      --dataset train \
+      --trim-to-supervisions True \
+      --speed-perturb True
+
+It looks for raw CutSet in the directory data/fbank
+located at: `{src_dir}/{prefix}_cuts_{dataset}_raw.jsonl.gz`.
+
+The generated fbank features are saved in `data/fbank/{prefix}-{dataset}_feats`
+and CutSet manifest stored in `data/fbank/{prefix}_cuts_{dataset}.jsonl.gz`.
+
+Typically, the number of workers is smaller than number of jobs
+(see --num-jobs 100 --num-workers 25 in the example).
+And, the number of jobs should be at least the number of workers (it's checked).
+"""
+
+import argparse
+import logging
+import multiprocessing
+import os
+from concurrent.futures import ProcessPoolExecutor
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    LilcomChunkyWriter,
+    is_caching_enabled,
+    set_caching_enabled,
+)
+
+from icefall.utils import str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+    parser.add_argument(
+        "--src-dir",
+        type=str,
+        help="""Folder with the input manifest files.""",
+        default="data/manifests",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        help="""Folder with the output manifests (cuts) and feature files.""",
+        default="data/fbank",
+    )
+
+    parser.add_argument(
+        "--prefix",
+        type=str,
+        help="""Prefix of the manifest files.""",
+        default="",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank (train,test,dev).""",
+        default=None,
+    )
+
+    parser.add_argument(
+        "--num-jobs",
+        type=int,
+        help="""Number of jobs (i.e. files with extracted features)""",
+        default=50,
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        help="""Number of parallel workers""",
+        default=10,
+    )
+    parser.add_argument(
+        "--speed-perturb",
+        type=str2bool,
+        default=False,
+        help="""Enable speed perturbation for the set.""",
+    )
+    parser.add_argument(
+        "--trim-to-supervisions",
+        type=str2bool,
+        default=False,
+        help="""Apply `trim-to-supervision` to cut set.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_features(args: argparse.Namespace):
+    set_caching_enabled(True)  # lhotse
+
+    src_dir = Path(args.src_dir)
+    output_dir = Path(args.output_dir)
+    num_jobs = args.num_jobs
+    num_workers = min(args.num_workers, os.cpu_count())
+    num_mel_bins = 80
+
+    bpe_model = args.bpe_model
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    prefix = args.prefix  # "ELEF_TRAIN"
+    dataset = args.dataset
+    suffix = "jsonl.gz"
+
+    cuts_raw_filename = Path(f"{src_dir}/{prefix}_cuts_{dataset}_raw.{suffix}")
+    cuts_raw = CutSet.from_file(cuts_raw_filename)
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    cuts_filename = Path(f"{prefix}_cuts_{dataset}.{suffix}")
+    if (output_dir / cuts_filename).is_file():
+        logging.info(f"{output_dir/cuts_filename} already exists - skipping.")
+        return
+
+    logging.info(f"Processing {output_dir/cuts_filename}")
+    cut_set = cuts_raw
+
+    if bpe_model:
+        cut_set = filter_cuts(cut_set, sp)
+
+    if args.speed_perturb:
+        cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+
+    if args.trim_to_supervisions:
+        logging.info(f"About to `trim_to_supervisions()` {output_dir / cuts_filename}")
+        cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+    else:
+        logging.info(
+            "Not doing `trim_to_supervisions()`, "
+            "to enable use --trim-to-supervision=True"
+        )
+
+    cut_set = cut_set.to_eager()  # disallow lazy evaluation (sorting requires it)
+    cut_set = cut_set.sort_by_recording_id()  # enhances AudioCache hit rate
+
+    # We typically use `num_jobs=100, num_workers=20`
+    # - this is helpful for large databases
+    # - both values are configurable externally
+    assert num_jobs >= num_workers, (num_jobs, num_workers)
+    executor = ProcessPoolExecutor(
+        max_workers=num_workers,
+        mp_context=multiprocessing.get_context("spawn"),
+        initializer=set_caching_enabled,
+        initargs=(is_caching_enabled(),),
+    )
+
+    logging.info(
+        f"executor {executor} : num_workers {num_workers}, num_jobs {num_jobs}"
+    )
+
+    cut_set = cut_set.compute_and_store_features(
+        extractor=extractor,
+        storage_path=f"{output_dir / prefix}-{dataset}_feats",
+        num_jobs=num_jobs,
+        executor=executor,
+        storage_type=LilcomChunkyWriter,
+    )
+
+    # correct small deviations of duration, caused by speed-perturbation
+    for cut in cut_set:
+        assert len(cut.supervisions) == 1, (len(cut.supervisions), cut.id)
+        duration_difference = abs(cut.supervisions[0].duration - cut.duration)
+        tolerance = 0.02  # 20ms
+        if duration_difference == 0.0:
+            pass
+        elif duration_difference <= tolerance:
+            logging.info(
+                "small mismatch of the supervision duration "
+                f"(Δt = {duration_difference*1000}ms), "
+                f"correcting : cut.duration {cut.duration} -> "
+                f"supervision {cut.supervisions[0].duration}"
+            )
+            cut.supervisions[0].duration = cut.duration
+        else:
+            logging.error(
+                "mismatch of cut/supervision duration "
+                f"(Δt = {duration_difference*1000}ms) : "
+                f"cut.duration {cut.duration}, "
+                f"supervision {cut.supervisions[0].duration}"
+            )
+            raise ValueError(
+                "mismatch of cut/supervision duration "
+                f"(Δt = {duration_difference*1000}ms)"
+            )
+
+    # store the cutset
+    logging.info(f"storing CutSet to : `{output_dir / cuts_filename}`")
+    cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    logging.info(vars(args))
+
+    compute_fbank_features(args)
diff --git a/egs/voxpopuli/ASR/local/compute_fbank_musan.py b/egs/voxpopuli/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/voxpopuli/ASR/local/display_manifest_statistics.py b/egs/voxpopuli/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..36c99e126
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+Usage example:
+    python3 ./local/display_manifest_statistics.py data/fbank/*_cuts*.jsonl.gz
+
+See the function `remove_short_and_long_utt()` in transducer/train.py
+for usage.
+
+"""
+
+import argparse
+
+from lhotse import load_manifest_lazy
+
+
+def get_args():
+    parser = argparse.ArgumentParser("Compute statistics for 'cuts' .jsonl.gz")
+
+    parser.add_argument(
+        "filename",
+        help="data/fbank/imported_cuts_bison-train_trim.jsonl.gz",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    cuts = load_manifest_lazy(args.filename)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/voxpopuli/ASR/local/duration_from_supervision_manifest.py b/egs/voxpopuli/ASR/local/duration_from_supervision_manifest.py
new file mode 100755
index 000000000..957267fe8
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/duration_from_supervision_manifest.py
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script computes durations of datasets from
+the SupervisionSet manifests.
+
+Usage example:
+
+  python3 ./local/duration_from_supervision_manifest.py \
+    data/manifest/*_superivions*.jsonl.gz
+"""
+
+import argparse
+import gzip
+import json
+import logging
+import re
+import sys
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        "Read the raw text from the 'supervisions.jsonl.gz'"
+    )
+
+    parser.add_argument(
+        "filename",
+        help="supervisions.jsonl.gz",
+        nargs="+",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    total_duration = 0.0
+    total_n_utts = 0
+
+    for fname in args.filename:
+        if fname == "-":
+            fd = sys.stdin
+        elif re.match(r".*\.jsonl\.gz$", fname):
+            fd = gzip.open(fname, mode="r")
+        else:
+            fd = open(fname, mode="r")
+
+        fname_duration = 0.0
+        n_utts = 0
+        for line in fd:
+            js = json.loads(line)
+            fname_duration += js["duration"]
+            n_utts += 1
+
+        print(
+            f"Duration: {fname_duration/3600:7.2f} hours "
+            f"(eq. {fname_duration:7.0f} seconds, {n_utts} utts): {fname}"
+        )
+
+        if fd != sys.stdin:
+            fd.close()
+
+        total_duration += fname_duration
+        total_n_utts += n_utts
+
+    print(
+        f"Total duration: {total_duration/3600:7.2f} hours "
+        f"(eq. {total_duration:7.0f} seconds)"
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/voxpopuli/ASR/local/filter_cuts.py b/egs/voxpopuli/ASR/local/filter_cuts.py
new file mode 120000
index 000000000..27aca1729
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/filter_cuts.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/filter_cuts.py
\ No newline at end of file
diff --git a/egs/voxpopuli/ASR/local/prepare_lang_bpe.py b/egs/voxpopuli/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/voxpopuli/ASR/local/preprocess_voxpopuli.py b/egs/voxpopuli/ASR/local/preprocess_voxpopuli.py
new file mode 100755
index 000000000..4032537db
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/preprocess_voxpopuli.py
@@ -0,0 +1,178 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan Yang)
+#              2023  Brno University of Technology  (author: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Preprocess the database.
+- Convert RecordingSet and SupervisionSet to CutSet.
+- Apply text normalization to the transcripts.
+   - We take renormalized `orig_text` as `text` transcripts.
+   - The text normalization is separating punctuation from words.
+   - Also we put capital letter to the beginning of a sentence.
+
+The script is inspired in:
+    `egs/commonvoice/ASR/local/preprocess_commonvoice.py`
+
+Usage example:
+    python3 ./local/preprocess_voxpopuli.py \
+        --task asr --lang en
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Optional
+
+from lhotse import CutSet
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# from local/
+from separate_punctuation import separate_punctuation
+from uppercase_begin_of_sentence import UpperCaseBeginOfSentence
+
+from icefall.utils import str2bool
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+        default=None,
+    )
+
+    parser.add_argument(
+        "--task",
+        type=str,
+        help="""Task of VoxPopuli""",
+        default="asr",
+    )
+
+    parser.add_argument(
+        "--lang",
+        type=str,
+        help="""Language of VoxPopuli""",
+        required=True,
+    )
+
+    parser.add_argument(
+        "--use-original-text",
+        type=str2bool,
+        help="""Use 'original_text' from the annoattaion file,
+                otherwise 'normed_text' will be used
+                (see `data/manifests/${task}_${lang}.tsv.gz`).
+             """,
+        default=False,
+    )
+
+    return parser.parse_args()
+
+
+def normalize_text(utt: str) -> str:
+    utt = UpperCaseBeginOfSentence().process_line_text(separate_punctuation(utt))
+    return utt
+
+
+def preprocess_voxpopuli(
+    task: str,
+    language: str,
+    dataset: Optional[str] = None,
+    use_original_text: bool = False,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    if dataset is None:
+        dataset_parts = (
+            "dev",
+            "test",
+            "train",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    logging.info("Loading manifest")
+    prefix = f"voxpopuli-{task}-{language}"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix=suffix,
+        prefix=prefix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"{prefix}_cuts_{partition}_raw.{suffix}"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        if use_original_text:
+            logging.info("Using 'original_text' from the annotation file.")
+            logging.info(f"Normalizing text in {partition}")
+            for sup in m["supervisions"]:
+                # `orig_text` includes punctuation and true-case
+                orig_text = str(sup.custom["orig_text"])
+                # we replace `text` by normalized `orig_text`
+                sup.text = normalize_text(orig_text)
+        else:
+            logging.info("Using 'normed_text' from the annotation file.")
+
+        # remove supervisions with empty 'text'
+        m["supervisions"] = m["supervisions"].filter(lambda sup: len(sup.text) > 0)
+
+        # Create cut manifest with long-recordings.
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        ).resample(16000)
+
+        # Store the cut set incl. the resampling.
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    preprocess_voxpopuli(
+        task=args.task,
+        language=args.lang,
+        dataset=args.dataset,
+        use_original_text=args.use_original_text,
+    )
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/voxpopuli/ASR/local/separate_punctuation.py b/egs/voxpopuli/ASR/local/separate_punctuation.py
new file mode 100755
index 000000000..706d6fcd5
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/separate_punctuation.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+# Copyright    2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script chops the punctuation as standalone tokens.
+Example:
+    input: "This is fine. Yes, you are right."
+    output: "This is fine . Yes , you are right ."
+
+The script also handles exceptions in a hard-coded fashion.
+
+(same functionality could be done with `nltk.tokenize.word_tokenize()`,
+ but that would be an extra dependency)
+
+It can be used as a module, or as an executable script.
+
+Usage example #1:
+  `from separate_punctuation import separate_punctuation`
+
+Usage example #2:
+```
+  python3 ./local/separate_punctuation.py \
+    --ignore-columns 1 \
+    < ${kaldi_data}/text
+```
+"""
+
+import re
+import sys
+from argparse import ArgumentParser
+
+
+def separate_punctuation(text: str) -> str:
+    """
+    Text filtering function for separating punctuation.
+
+    Example:
+        input: "This is fine. Yes, you are right."
+        output: "This is fine . Yes , you are right ."
+
+    The exceptions for which the punctuation is
+    not splitted are hard-coded.
+    """
+
+    # remove non-desired punctuation symbols
+    text = re.sub('["„“«»]', "", text)
+
+    # separate [,.!?;] punctuation from words by space
+    text = re.sub(r"(\w)([,.!?;])", r"\1 \2", text)
+    text = re.sub(r"([,.!?;])(\w)", r"\1 \2", text)
+
+    # split to tokens
+    tokens = text.split()
+    tokens_out = []
+
+    # re-join the special cases of punctuation
+    for ii, tok in enumerate(tokens):
+        # no rewriting for 1st and last token
+        if ii > 0 and ii < len(tokens) - 1:
+            # **RULES ADDED FOR CZECH COMMON VOICE**
+
+            # fix "27 . dubna" -> "27. dubna", but keep punctuation separate,
+            if tok == "." and tokens[ii - 1].isdigit() and tokens[ii + 1].islower():
+                tokens_out[-1] = tokens_out[-1] + "."
+                continue
+
+            # fix "resp . pak" -> "resp. pak"
+            if tok == "." and tokens[ii - 1].isalpha() and tokens[ii + 1].islower():
+                tokens_out[-1] = tokens_out[-1] + "."
+                continue
+
+            # **RULES ADDED FOR ENGLISH COMMON VOICE**
+
+            # fix "A ." -> "A."
+            if tok == "." and re.match(r"^[A-Z]S", tokens[ii - 1]):
+                tokens_out[-1] = tokens_out[-1] + "."
+                continue
+
+            # fix "Mr ." -> "Mr."
+            exceptions = set(["Mr", "Mrs", "Ms"])
+            if tok == "." and tokens[ii - 1] in exceptions:
+                tokens_out[-1] = tokens_out[-1] + "."
+                continue
+
+        tokens_out.append(tok)
+
+    return " ".join(tokens_out)
+
+
+def get_args():
+    parser = ArgumentParser(
+        description="Separate punctuation from words: 'hello.' -> 'hello .'"
+    )
+    parser.add_argument(
+        "--ignore-columns", type=int, default=1, help="skip number of initial columns"
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    max_split = args.ignore_columns
+
+    while True:
+        line = sys.stdin.readline()
+        if not line:
+            break
+
+        *key, text = line.strip().split(maxsplit=max_split)
+        text_norm = separate_punctuation(text)
+
+        print(" ".join(key), text_norm)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/voxpopuli/ASR/local/text_from_manifest.py b/egs/voxpopuli/ASR/local/text_from_manifest.py
new file mode 100755
index 000000000..d9ab53b5a
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/text_from_manifest.py
@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+# Copyright    2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Print the text contained in `supervisions.jsonl.gz` or `cuts.jsonl.gz`.
+
+Usage example:
+    python3 ./local/text_from_manifest.py \
+        data/manifests/voxpopuli-asr-en_supervisions_dev.jsonl.gz
+"""
+
+import argparse
+import gzip
+import json
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        "Read the raw text from the 'supervisions.jsonl.gz'"
+    )
+    parser.add_argument("filename", help="supervisions.jsonl.gz")
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    with gzip.open(args.filename, mode="r") as fd:
+        for line in fd:
+            js = json.loads(line)
+            if "text" in js:
+                print(js["text"])  # supervisions.jsonl.gz
+            elif "supervisions" in js:
+                for s in js["supervisions"]:
+                    print(s["text"])  # cuts.jsonl.gz
+            else:
+                raise Exception(f"Unknown jsonl format of {args.filename}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/voxpopuli/ASR/local/train_bpe_model.py b/egs/voxpopuli/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/voxpopuli/ASR/local/uppercase_begin_of_sentence.py b/egs/voxpopuli/ASR/local/uppercase_begin_of_sentence.py
new file mode 100755
index 000000000..8e9de905f
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/uppercase_begin_of_sentence.py
@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+# Copyright    2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script introduces initial capital letter at the beginning of a sentence.
+It can be used as a module, or as an executable script.
+
+Usage example #1:
+  `from uppercase_begin_of_sentence import UpperCaseBeginOfSentence`
+
+Usage example #2:
+```
+  python3 ./local/uppercase_begin_of_sentence.py \
+    --ignore-columns 1 \
+    < ${kaldi_data}/text
+```
+"""
+
+import re
+import sys
+from argparse import ArgumentParser
+
+
+class UpperCaseBeginOfSentence:
+    """
+    This class introduces initial capital letter at the beginning of a sentence.
+    Capital letter is used, if previous symbol was punctuation token from
+    `set([".", "!", "?"])`.
+
+    The punctuation as previous token is memorized also across
+    `process_line_text()` calls.
+    """
+
+    def __init__(self):
+        # The 1st word will have Title-case
+        # This variable transfers context from previous line
+        self.prev_token_is_punct = True
+
+    def process_line_text(self, line_text: str) -> str:
+        """
+        It is assumed that punctuation in `line_text` was already separated,
+        example: "This is fine . Yes , you are right ."
+        """
+
+        words = line_text.split()
+        punct_set = set([".", "!", "?"])
+
+        for ii, w in enumerate(words):
+            # punctuation ?
+            if w in punct_set:
+                self.prev_token_is_punct = True
+                continue
+
+            # change case of word...
+            if self.prev_token_is_punct:
+                if re.match("<", w):
+                    continue  # skip <symbols>
+                # apply Title-case only on lowercase words.
+                if w.islower():
+                    words[ii] = w.title()
+                # change state
+                self.prev_token_is_punct = False
+
+        line_text_uc = " ".join(words)
+
+        return line_text_uc
+
+
+def get_args():
+    parser = ArgumentParser(
+        description="Put upper-case at the beginning of a sentence."
+    )
+    parser.add_argument(
+        "--ignore-columns", type=int, default=4, help="skip number of initial columns"
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    uc_bos = UpperCaseBeginOfSentence()
+    max_split = args.ignore_columns
+
+    while True:
+        line = sys.stdin.readline()
+        if not line:
+            break
+        line = line.strip()
+
+        if len(line.split()) > 1:
+            *key, text = line.strip().split(maxsplit=max_split)  # parse,
+            text_uc = uc_bos.process_line_text(text)  # process,
+            print(" ".join(key), text_uc)  # print,
+        else:
+            print(line)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/voxpopuli/ASR/local/validate_bpe_lexicon.py b/egs/voxpopuli/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/voxpopuli/ASR/local/validate_cutset_manifest.py b/egs/voxpopuli/ASR/local/validate_cutset_manifest.py
new file mode 100755
index 000000000..4659aa9cd
--- /dev/null
+++ b/egs/voxpopuli/ASR/local/validate_cutset_manifest.py
@@ -0,0 +1,123 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Brno University of Technology  (authors: Karel Veselý)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+- Supervision time bounds are within Cut time bounds
+- Duration of Cut and Superivion are equal
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz
+
+(Based on: `librispeech/ASR/local/validate_manifest.py`)
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+from lhotse.dataset.speech_recognition import validate_for_asr
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "cutset_manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def validate_one_supervision_per_cut(c: Cut):
+    if len(c.supervisions) != 1:
+        raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
+
+
+def validate_supervision_and_cut_time_bounds(c: Cut):
+    tol = 2e-3  # same tolerance as in 'validate_for_asr()'
+    s = c.supervisions[0]
+
+    # Supervision start time is relative to Cut ...
+    # https://lhotse.readthedocs.io/en/v0.10_e/cuts.html
+    if s.start < -tol:
+        raise ValueError(
+            f"{c.id}: Supervision start time {s.start} must not be negative."
+        )
+    if s.start > tol:
+        raise ValueError(
+            f"{c.id}: Supervision start time {s.start} "
+            "is not at the beginning of the Cut. "
+            "Please apply `lhotse cut trim-to-supervisions`."
+        )
+    if c.start + s.end > c.end + tol:
+        raise ValueError(
+            f"{c.id}: Supervision end time {c.start+s.end} is larger "
+            f"than cut end time {c.end}"
+        )
+
+    if s.duration != c.duration:
+        raise ValueError(
+            f"{c.id}: Cut duration {c.duration} and supervision duration "
+            f"{s.duration} must be the same.\n"
+            f"The difference causes problems in the training code : "
+            f"+/- 1 frame in `x`, `x_lens` in `Zipformer::forward()`.\n"
+            f"Did you forget to apply `trim_to_supervisions()` ?"
+        )
+
+
+def main():
+    args = get_args()
+
+    manifest = args.cutset_manifest
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    try:
+        for c in cut_set:
+            validate_one_supervision_per_cut(c)
+            validate_supervision_and_cut_time_bounds(c)
+
+        # Validation from K2 training
+        # - checks supervision start is 0
+        # - checks supervision.duration is not longer than cut.duration
+        # - there is tolerance 2ms
+        validate_for_asr(cut_set)
+    except BaseException as e:
+        logging.error(str(e))
+        raise
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/voxpopuli/ASR/prepare.sh b/egs/voxpopuli/ASR/prepare.sh
new file mode 100755
index 000000000..7cddad756
--- /dev/null
+++ b/egs/voxpopuli/ASR/prepare.sh
@@ -0,0 +1,257 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -euxo pipefail
+
+nj=20
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/voxpopuli/raw_audios/$lang/$year
+#      This directory contains *.ogg files with audio downloaded and extracted from archives:
+#       https://dl.fbaipublicfiles.com/voxpopuli/audios/${lang}_${year}.tar
+#
+#  - Note: the voxpopuli transcripts are downloaded to a ${tmp} folder
+#    as part of `lhotse prepare voxpopuli` from:
+#       https://dl.fbaipublicfiles.com/voxpopuli/annotations/asr/asr_${lang}.tsv.gz
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+#dl_dir=/mnt/matylda6/szoke/EU-ASR/DATA  # BUT
+
+musan_dir=${dl_dir}/musan
+#musan_dir=/mnt/matylda2/data/MUSAN  # BUT
+
+# Choose value from ASR_LANGUAGES:
+#
+# [ "en", "de", "fr", "es", "pl", "it", "ro", "hu", "cs", "nl", "fi", "hr",
+# "sk", "sl", "et", "lt" ]
+#
+# See ASR_LANGUAGES in:
+# https://github.com/lhotse-speech/lhotse/blob/c5f26afd100885b86e4244eeb33ca1986f3fa923/lhotse/recipes/voxpopuli.py#L54C4-L54C4
+lang=en
+
+task=asr
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/${lang}/lang_bpe_xxx,
+# data/${lang}/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data/${lang}".
+# You can safely remove "data/${lang}" and rerun this script to regenerate it.
+mkdir -p data/${lang}
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+log "musan_dir: $musan_dir"
+log "task: $task, lang: $lang"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/$release,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/$release $dl_dir/$release
+  #
+  if [ ! -d $dl_dir/voxpopuli/raw_audios/${lang} ]; then
+    lhotse download voxpopuli --subset $lang $dl_dir/voxpopuli
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $musan_dir/musan ]; then
+    lhotse download musan $musan_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare VoxPopuli manifest"
+  # We assume that you have downloaded the VoxPopuli corpus
+  # to $dl_dir/voxpopuli
+  if [ ! -e data/manifests/.voxpopuli-${task}-${lang}.done ]; then
+    # Warning : it requires Internet connection (it downloads transcripts to ${tmpdir})
+    lhotse prepare voxpopuli --task asr --lang $lang -j $nj $dl_dir/voxpopuli data/manifests
+    touch data/manifests/.voxpopuli-${task}-${lang}.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    #lhotse prepare musan $dl_dir/musan data/manifests
+    lhotse prepare musan $musan_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Preprocess VoxPopuli manifest"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.voxpopuli-${task}-${lang}-preprocess_complete ]; then
+    # recordings + supervisions -> cutset
+    ./local/preprocess_voxpopuli.py --task $task --lang $lang \
+        --use-original-text True
+    touch data/fbank/.voxpopuli-${task}-${lang}-preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for dev and test subsets of VoxPopuli"
+  mkdir -p data/fbank
+  for dataset in "dev" "test"; do
+    if [ ! -e data/fbank/.voxpopuli-${task}-${lang}-${dataset}.done ]; then
+      ./local/compute_fbank.py --src-dir data/fbank --output-dir data/fbank \
+          --num-jobs 50 --num-workers ${nj} \
+          --prefix "voxpopuli-${task}-${lang}" \
+          --dataset ${dataset} \
+          --trim-to-supervisions True
+      touch data/fbank/.voxpopuli-${task}-${lang}-${dataset}.done
+    fi
+  done
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for train set of VoxPopuli"
+  if [ ! -e data/fbank/.voxpopuli-${task}-${lang}-train.done ]; then
+    ./local/compute_fbank.py --src-dir data/fbank --output-dir data/fbank \
+        --num-jobs 100 --num-workers ${nj} \
+        --prefix "voxpopuli-${task}-${lang}" \
+        --dataset train \
+        --trim-to-supervisions True \
+        --speed-perturb True
+    touch data/fbank/.voxpopuli-${task}-${lang}-train.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Validate fbank manifests for VoxPopuli"
+  for dataset in "dev" "test" "train"; do
+    mkdir -p data/fbank/log/
+    ./local/validate_cutset_manifest.py \
+      data/fbank/voxpopuli-asr-en_cuts_${dataset}.jsonl.gz \
+      2>&1 | tee data/fbank/log/validate_voxpopuli-asr-en_cuts_${dataset}.log
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}_${lang}
+    mkdir -p $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      file=$(
+        find "data/fbank/voxpopuli-${task}-${lang}_cuts_train.jsonl.gz"
+      )
+      local/text_from_manifest.py $file >$lang_dir/transcript_words.txt
+      # gunzip -c ${file} | awk -F '"' '{print $30}' > $lang_dir/transcript_words.txt
+
+      # Ensure space only appears once
+      #sed -i 's/\t/ /g' $lang_dir/transcript_words.txt
+      #sed -i 's/[ ][ ]*/ /g' $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/words.txt ]; then
+      cat $lang_dir/transcript_words.txt | sed 's/ /\n/g' \
+        | sort -u | sed '/^$/d' > $lang_dir/words.txt
+      (echo '!SIL'; echo '<SPOKEN_NOISE>'; echo '<UNK>'; ) |
+        cat - $lang_dir/words.txt | sort | uniq | awk '
+        BEGIN {
+          print "<eps> 0";
+        }
+        {
+          if ($1 == "<s>") {
+            print "<s> is in the vocabulary!" | "cat 1>&2"
+            exit 1;
+          }
+          if ($1 == "</s>") {
+            print "</s> is in the vocabulary!" | "cat 1>&2"
+            exit 1;
+          }
+          printf("%s %d\n", $1, NR);
+        }
+        END {
+          printf("#0 %d\n", NR+1);
+          printf("<s> %d\n", NR+2);
+          printf("</s> %d\n", NR+3);
+        }' > $lang_dir/words || exit 1;
+      mv $lang_dir/words $lang_dir/words.txt
+    fi
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+      log "Converting L.pt to L.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L.pt \
+        $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+      log "Converting L_disambig.pt to L_disambig.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L_disambig.pt \
+        $lang_dir/L_disambig.fst
+    fi
+  done
+fi
diff --git a/egs/voxpopuli/ASR/shared b/egs/voxpopuli/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/voxpopuli/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/README.md b/egs/wenetspeech/ASR/README.md
new file mode 100644
index 000000000..44e631b4a
--- /dev/null
+++ b/egs/wenetspeech/ASR/README.md
@@ -0,0 +1,20 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with WenetSpeech.
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless2`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+| `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.
diff --git a/egs/wenetspeech/ASR/RESULTS.md b/egs/wenetspeech/ASR/RESULTS.md
new file mode 100644
index 000000000..1a0e0681f
--- /dev/null
+++ b/egs/wenetspeech/ASR/RESULTS.md
@@ -0,0 +1,294 @@
+## Results
+
+### WenetSpeech char-based training results (Non-streaming and streaming) on zipformer model
+
+This is the [pull request](https://github.com/k2-fsa/icefall/pull/1130) in icefall.
+
+#### Non-streaming
+
+Best results (num of params : ~76M):
+
+Type | Greedy(dev & net & meeting) | Beam search(dev & net & meeting) |  
+-- | -- | -- | --
+Non-streaming | 7.36 & 7.65 & 12.43 | 7.32 & 7.61 & 12.35 | --epoch=12
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 6 \
+  --num-epochs 12 \
+  --use-fp16 1 \
+  --max-duration 450 \
+  --training-subset L \
+  --lr-epochs 1.5 \
+  --context-size 2 \
+  --exp-dir zipformer/exp_L_context_2 \
+  --causal 0 \
+  --num-workers 8
+```
+
+Listed best results for each epoch below:
+
+Epoch | Greedy search(dev & net & meeting) | Modified beam search(dev & net & meeting) |  
+-- | -- | -- | --
+4 | 7.83 & 8.86 &13.73 | 7.75 & 8.81 & 13.67 | avg=1;blank-penalty=2
+5 | 7.75 & 8.46 & 13.38 | 7.68 & 8.41 & 13.27 | avg=1;blank-penalty=2
+6 | 7.72 & 8.19 & 13.16 | 7.62 & 8.14 & 13.06 | avg=1;blank-penalty=2
+7 | 7.59 & 8.08 & 12.97 | 7.53 & 8.01 & 12.87 | avg=2;blank-penalty=2
+8 | 7.68 & 7.87 & 12.96 | 7.61 & 7.81 & 12.88 | avg=1;blank-penalty=2
+9 | 7.57 & 7.77 & 12.87 | 7.5 & 7.71 & 12.77 | avg=1;blank-penalty=2
+10 | 7.45 & 7.7 & 12.69 | 7.39 & 7.63 & 12.59 | avg=2;blank-penalty=2
+11 | 7.35 & 7.67 & 12.46 | 7.31 & 7.63 & 12.43 | avg=3;blank-penalty=2
+12 | 7.36 & 7.65 & 12.43 | 7.32 & 7.61 & 12.35 | avg=4;blank-penalty=2
+
+The pre-trained model is available here : https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615
+
+
+#### Streaming
+
+Best results (num of params : ~76M):
+
+Type | Greedy(dev & net & meeting) | Beam search(dev & net & meeting) |  
+-- | -- | -- | --
+Streaming | 8.45 & 9.89 & 16.46 | 8.21 & 9.77 & 16.07 | --epoch=12; --chunk-size=16; --left-context-frames=256
+Streaming | 8.0 & 9.0 & 15.11 | 7.84 & 8.94 & 14.92 | --epoch=12; --chunk-size=32; --left-context-frames=256
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 12 \
+  --use-fp16 1 \
+  --max-duration 450 \
+  --training-subset L \
+  --lr-epochs 1.5 \
+  --context-size 2 \
+  --exp-dir zipformer/exp_L_causal_context_2 \
+  --causal 1 \
+  --num-workers 8
+```
+
+Best results for each epoch (--chunk-size=16; --left-context-frames=128)
+
+Epoch | Greedy search(dev & net & meeting) | Modified beam search(dev & net & meeting) |  
+-- | -- | -- | --
+6 | 9.14 & 10.75 & 18.15 | 8.79 & 10.54 & 17.64 | avg=1;blank-penalty=1.5
+7 | 9.11 & 10.61 & 17.86 | 8.8 & 10.42 & 17.29 | avg=1;blank-penalty=1.5
+8 | 8.89 & 10.32 & 17.44 | 8.59 & 10.09 & 16.9 | avg=1;blank-penalty=1.5
+9 | 8.86 & 10.11 & 17.35 | 8.55 & 9.87 & 16.76 | avg=1;blank-penalty=1.5
+10 | 8.66 & 10.0 & 16.94 | 8.39 & 9.83 & 16.47 | avg=2;blank-penalty=1.5
+11 | 8.58 & 9.92 & 16.67 | 8.32 & 9.77 & 16.27 | avg=3;blank-penalty=1.5
+12 | 8.45 & 9.89 & 16.46 | 8.21 & 9.77 & 16.07 | avg=4;blank-penalty=1.5
+
+The pre-trained model is available here: https://huggingface.co/pkufool/icefall-asr-zipformer-streaming-wenetspeech-20230615
+
+
+### WenetSpeech char-based training results (offline and streaming) (Pruned Transducer 5)
+
+#### 2022-07-22
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/447.
+
+When training with the L subset, the CERs are
+
+**Offline**:
+|decoding-method| epoch | avg | use-averaged-model | DEV | TEST-NET | TEST-MEETING|
+|-- | -- | -- | -- | -- | -- | --|
+|greedy_search | 4 | 1 | True | 8.22 | 9.03 | 14.54|
+|modified_beam_search | 4 | 1 | True | **8.17** | **9.04** | **14.44**|
+|fast_beam_search | 4 | 1 | True | 8.29 | 9.00 | 14.93|
+
+The offline training command for reproducing is given below:
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless5/exp_L_offline \
+  --world-size 8 \
+  --num-epochs 15 \
+  --start-epoch 2 \
+  --max-duration 120 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --average-period 1000 \
+  --training-subset L
+```
+
+The tensorboard training log can be found at https://tensorboard.dev/experiment/SvnN2jfyTB2Hjqu22Z7ZoQ/#scalars .
+
+
+A pre-trained offline model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline>
+
+**Streaming**:
+|decoding-method| epoch | avg | use-averaged-model | DEV | TEST-NET | TEST-MEETING|
+|--|--|--|--|--|--|--|
+| greedy_search | 7| 1| True | 8.78 | 10.12 | 16.16 |
+| modified_beam_search | 7| 1| True| **8.53**| **9.95** | **15.81** |
+| fast_beam_search | 7 | 1| True | 9.01 | 10.47 | 16.28 |
+
+The streaming training command for reproducing is given below:
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless5/exp_L_streaming \
+  --world-size 8 \
+  --num-epochs 15 \
+  --start-epoch 1 \
+  --max-duration 140 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --average-period 1000 \
+  --training-subset L \
+  --dynamic-chunk-training True \
+  --causal-convolution True \
+  --short-chunk-size 25 \
+  --num-left-chunks 4
+```
+
+The tensorboard training log can be found at https://tensorboard.dev/experiment/E2NXPVflSOKWepzJ1a1uDQ/#scalars .
+
+
+A pre-trained offline model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_streaming>
+
+### WenetSpeech char-based training results (Pruned Transducer 2)
+
+#### 2022-05-19
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/349.
+
+When training with the L subset, the CERs are
+
+|                                    |  dev  | test-net | test-meeting | comment                                  |
+|------------------------------------|-------|----------|--------------|------------------------------------------|
+|          greedy search             | 7.80  | 8.75     | 13.49        | --epoch 10, --avg 2, --max-duration 100  |
+| modified beam search (beam size 4) | 7.76  | 8.71     | 13.41        | --epoch 10, --avg 2, --max-duration 100  |
+| fast beam search (1best)  | 7.94  | 8.74     | 13.80        | --epoch 10, --avg 2, --max-duration 1500 |
+| fast beam search (nbest)  | 9.82  | 10.98    |     16.37   | --epoch 10, --avg 2, --max-duration 600 |
+| fast beam search (nbest oracle)  | 6.88 | 7.18    |     11.77   | --epoch 10, --avg 2, --max-duration 600 |
+| fast beam search (nbest LG, ngram_lm_scale=0.35)  | 8.83 | 9.88    |   15.47  | --epoch 10, --avg 2, --max-duration 600 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless2/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 15 \
+  --start-epoch 0 \
+  --max-duration 180 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --training-subset L
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/wM4ZUNtASRavJx79EOYYcg/#scalars
+
+The decoding command is:
+```
+epoch=10
+avg=2
+
+## greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+## modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+## fast beam search (1best)
+./pruned_transducer_stateless2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+
+## fast beam search (nbest)
+./pruned_transducer_stateless2/decode.py \
+        --epoch 10 \
+        --avg 2 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 600 \
+        --decoding-method fast_beam_search_nbest \
+        --beam 20.0 \
+        --max-contexts 8 \
+        --max-states 64 \
+        --num-paths 200 \
+        --nbest-scale 0.5
+
+## fast beam search (nbest oracle WER)
+./pruned_transducer_stateless2/decode.py \
+        --epoch 10 \
+        --avg 2 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 600 \
+        --decoding-method fast_beam_search_nbest_oracle \
+        --beam 20.0 \
+        --max-contexts 8 \
+        --max-states 64 \
+        --num-paths 200 \
+        --nbest-scale 0.5
+
+## fast beam search (with LG)
+./pruned_transducer_stateless2/decode.py \
+        --epoch 10 \
+        --avg 2 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 600 \
+        --decoding-method fast_beam_search_nbest_LG \
+        --ngram-lm-scale 0.35 \
+        --beam 20.0 \
+        --max-contexts 8 \
+        --max-states 64
+```
+
+When training with the M subset, the CERs are
+
+|                                    |   dev  | test-net  | test-meeting  | comment                                   |
+|------------------------------------|--------|-----------|---------------|-------------------------------------------|
+|          greedy search             | 10.40  | 11.31     | 19.64         | --epoch 29, --avg 11, --max-duration 100  |
+| modified beam search (beam size 4) |  9.85  | 11.04     | 18.20         | --epoch 29, --avg 11, --max-duration 100  |
+| fast beam search (set as default)  | 10.18  | 11.10     | 19.32         | --epoch 29, --avg 11, --max-duration 1500 |
+
+
+When training with the S subset, the CERs are
+
+|                                    |  dev   | test-net  | test-meeting  | comment                                   |
+|------------------------------------|--------|-----------|---------------|-------------------------------------------|
+|          greedy search             | 19.92  | 25.20     | 35.35         | --epoch 29, --avg 24, --max-duration 100  |
+| modified beam search (beam size 4) | 18.62  | 23.88     | 33.80         | --epoch 29, --avg 24, --max-duration 100  |
+| fast beam search (set as default)  | 19.31  | 24.41     | 34.87         | --epoch 29, --avg 24, --max-duration 1500 |
+
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2>
diff --git a/egs/wenetspeech/ASR/finetune.sh b/egs/wenetspeech/ASR/finetune.sh
new file mode 100755
index 000000000..8559780e9
--- /dev/null
+++ b/egs/wenetspeech/ASR/finetune.sh
@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# This is an example script for fine-tuning. Here, we fine-tune a model trained
+# on WenetSpeech on Aishell. The model used for fine-tuning is
+# pruned_transducer_stateless2 (zipformer). If you want to fine-tune model
+# from another recipe, you can adapt ./pruned_transducer_stateless2/finetune.py
+# for that recipe. If you have any problem, please open up an issue in https://github.com/k2-fsa/icefall/issues.
+
+# We assume that you have already prepared the Aishell manfiest&features under ./data.
+# If you haven't done that, please see https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/prepare.sh.
+
+. shared/parse_options.sh || exit 1
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download Pre-trained model"
+
+  # clone from huggingface
+  git lfs install
+  git clone https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2
+
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Start fine-tuning"
+
+  # The following configuration of lr schedule should work well
+  # You may also tune the following parameters to adjust learning rate schedule
+  initial_lr=0.0001
+  lr_epochs=100
+  lr_batches=100000
+
+  # We recommend to start from an averaged model
+  finetune_ckpt=icefall_asr_wenetspeech_pruned_transducer_stateless2/exp/pretrained_epoch_10_avg_2.pt
+  lang_dir=icefall_asr_wenetspeech_pruned_transducer_stateless2/data/lang_char
+  export CUDA_VISIBLE_DEVICES="0,1"
+
+  ./pruned_transducer_stateless2/finetune.py \
+    --world-size 2 \
+    --master-port 18180 \
+    --num-epochs 15 \
+    --context-size 2 \
+    --exp-dir pruned_transducer_stateless2/exp_aishell_finetune \
+    --initial-lr $initial_lr \
+    --lr-epochs $lr_epochs \
+    --lr-batches $lr_batches \
+    --lang-dir $lang_dir \
+    --do-finetune True \
+    --finetune-ckpt $finetune_ckpt \
+    --max-duration 200
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Decoding"
+
+  epoch=4
+  avg=4
+
+  for m in greedy_search modified_beam_search; do
+    python pruned_transducer_stateless2/decode_aishell.py \
+    --epoch $epoch \
+    --avg $avg \
+    --context-size 2 \
+    --beam-size 4 \
+    --exp-dir pruned_transducer_stateless2/exp_aishell_finetune \
+    --max-duration 400 \
+    --decoding-method $m
+  done
+fi
diff --git a/egs/wenetspeech/ASR/local/compile_lg.py b/egs/wenetspeech/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/local/compute_fbank_musan.py b/egs/wenetspeech/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py
new file mode 100755
index 000000000..1af08fee2
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def compute_fbank_wenetspeech_dev_test():
+    in_out_dir = Path("data/fbank")
+    # number of workers in dataloader
+    num_workers = 42
+
+    # number of seconds in a batch
+    batch_duration = 600
+
+    subsets = ("DEV", "TEST_NET", "TEST_MEETING")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    for partition in subsets:
+        cuts_path = in_out_dir / f"cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = in_out_dir / f"cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{in_out_dir}/feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def main():
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_wenetspeech_dev_test()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py
new file mode 100755
index 000000000..99d39bbdc
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py
@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+    set_audio_duration_mismatch_tolerance,
+    set_caching_enabled,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--training-subset",
+        type=str,
+        default="L",
+        help="The training subset for computing fbank feature.",
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the L subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+    return parser
+
+
+def compute_fbank_wenetspeech_splits(args):
+    subset = args.training_subset
+    subset = str(subset)
+    num_splits = args.num_splits
+    output_dir = f"data/fbank/{subset}_split_{num_splits}"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    set_audio_duration_mismatch_tolerance(0.01)  # 10ms tolerance
+    set_caching_enabled(False)
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"cuts_{subset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"cuts_{subset}_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/feats_{subset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def main():
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+
+    log_filename = "log-compute_fbank_wenetspeech_splits"
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    log_filename = f"{log_filename}-{date_time}"
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=logging.INFO,
+        filemode="w",
+    )
+
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter(formatter))
+    logging.getLogger("").addHandler(console)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    compute_fbank_wenetspeech_splits(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/local/display_manifest_statistics.py b/egs/wenetspeech/ASR/local/display_manifest_statistics.py
new file mode 100644
index 000000000..36e4ac5c3
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    paths = [
+        "./data/fbank/cuts_S.jsonl.gz",
+        "./data/fbank/cuts_M.jsonl.gz",
+        "./data/fbank/cuts_L.jsonl.gz",
+        "./data/fbank/cuts_DEV.jsonl.gz",
+        "./data/fbank/cuts_TEST_NET.jsonl.gz",
+        "./data/fbank/cuts_TEST_MEETING.jsonl.gz",
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest_lazy(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/cuts_L.jsonl.gz
+
+Cuts count: 43874235
+Total duration (hours): 30217.3
+Speech duration (hours): 30217.3 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.5
+std     1.7
+min     0.2
+25%     1.4
+50%     2.0
+75%     3.0
+99%     8.4
+99.5%   9.1
+99.9%   15.4
+max     405.1
+
+Starting display the statistics for ./data/fbank/cuts_S.jsonl.gz
+Duration statistics (seconds):
+mean    2.4
+std     1.8
+min     0.2
+25%     1.4
+50%     2.0
+75%     2.9
+99%     8.0
+99.5%   8.7
+99.9%   11.9
+max     405.1
+
+Starting display the statistics for ./data/fbank/cuts_M.jsonl.gz
+Cuts count: 4543341
+Total duration (hours): 3021.1
+Speech duration (hours): 3021.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.4
+std     1.6
+min     0.2
+25%     1.4
+50%     2.0
+75%     2.9
+99%     8.0
+99.5%   8.8
+99.9%   12.1
+max     405.1
+
+Starting display the statistics for ./data/fbank/cuts_DEV.jsonl.gz
+Cuts count: 13825
+Total duration (hours): 20.0
+Speech duration (hours): 20.0 (100.0%)
+***
+Duration statistics (seconds):
+mean    5.2
+std     2.2
+min     1.0
+25%     3.3
+50%     4.9
+75%     7.0
+99%     9.6
+99.5%   9.8
+99.9%   10.0
+max     10.0
+
+Starting display the statistics for ./data/fbank/cuts_TEST_NET.jsonl.gz
+Cuts count: 24774
+Total duration (hours): 23.1
+Speech duration (hours): 23.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.4
+std     2.6
+min     0.1
+25%     1.4
+50%     2.4
+75%     4.8
+99%     13.1
+99.5%   14.5
+99.9%   18.5
+max     33.3
+
+Starting display the statistics for ./data/fbank/cuts_TEST_MEETING.jsonl.gz
+Cuts count: 8370
+Total duration (hours): 15.2
+Speech duration (hours): 15.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.5
+std     3.5
+min     0.8
+25%     3.7
+50%     5.8
+75%     8.8
+99%     15.2
+99.5%   16.0
+99.9%   18.8
+max     24.6
+
+"""
diff --git a/egs/wenetspeech/ASR/local/prepare_char.py b/egs/wenetspeech/ASR/local/prepare_char.py
new file mode 100755
index 000000000..d8622842f
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/prepare_char.py
@@ -0,0 +1,242 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input `lang_dir`, which should contain::
+    - lang_dir/text,
+    - lang_dir/words.txt
+and generates the following files in the directory `lang_dir`:
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            tokens_list = list(line)
+            for token in tokens_list:
+                if token not in tokens:
+                    tokens[token] = len(tokens)
+    return tokens
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang directory.")
+    args = parser.parse_args()
+
+    lang_dir = Path(args.lang_dir)
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/local/prepare_char_lm_training_data.py b/egs/wenetspeech/ASR/local/prepare_char_lm_training_data.py
new file mode 120000
index 000000000..2374cafdd
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/prepare_char_lm_training_data.py
@@ -0,0 +1 @@
+../../../aishell/ASR/local/prepare_char_lm_training_data.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/local/prepare_lang.py b/egs/wenetspeech/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/local/prepare_words.py b/egs/wenetspeech/ASR/local/prepare_words.py
new file mode 100644
index 000000000..d5f833db1
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/prepare_words.py
@@ -0,0 +1,94 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input words.txt without ids:
+    - words_no_ids.txt
+and generates the new words.txt with related ids.
+    - words.txt
+"""
+
+
+import argparse
+import logging
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare words.txt",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/words_no_ids.txt",
+        type=str,
+        help="the words file without ids for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/words.txt",
+        type=str,
+        help="the words file with ids for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    add_words = ["<eps> 0", "!SIL 1", "<SPOKEN_NOISE> 2", "<UNK> 3"]
+    new_lines.extend(add_words)
+
+    logging.info("Starting reading the input file")
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        idx = 4 + i
+        new_line = str(x.strip("\n")) + " " + str(idx)
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the words.txt")
+    f_out = open(output_file, "w", encoding="utf-8")
+
+    # LG decoding needs below symbols.
+    id1, id2, id3 = (
+        str(len(new_lines)),
+        str(len(new_lines) + 1),
+        str(len(new_lines) + 2),
+    )
+    add_words = ["#0 " + id1, "<s> " + id2, "</s> " + id3]
+    new_lines.extend(add_words)
+
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/local/preprocess_wenetspeech.py b/egs/wenetspeech/ASR/local/preprocess_wenetspeech.py
new file mode 100755
index 000000000..5de3c23a9
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/preprocess_wenetspeech.py
@@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import re
+from pathlib import Path
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall import setup_logger
+from icefall.utils import str2bool
+
+# Similar text filtering and normalization procedure as in:
+# https://github.com/SpeechColab/WenetSpeech/blob/main/toolkits/kaldi/wenetspeech_data_prep.sh
+
+
+def normalize_text(
+    utt: str,
+    # punct_pattern=re.compile(r"<(COMMA|PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    punct_pattern=re.compile(r"<(PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    whitespace_pattern=re.compile(r"\s\s+"),
+) -> str:
+    return whitespace_pattern.sub(" ", punct_pattern.sub("", utt))
+
+
+def has_no_oov(
+    sup: SupervisionSegment,
+    oov_pattern=re.compile(r"<(SIL|MUSIC|NOISE|OTHER)>"),
+) -> bool:
+    return oov_pattern.search(sup.text) is None
+
+
+def preprocess_wenet_speech(perturb_speed: bool = False):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    # Note: By default, we preprocess all sub-parts.
+    # You can delete those that you don't need.
+    # For instance, if you don't want to use the L subpart, just remove
+    # the line below containing "L"
+    dataset_parts = (
+        "DEV",
+        "TEST_NET",
+        "TEST_MEETING",
+        "S",
+        "M",
+        "L",
+    )
+
+    logging.info("Loading manifest (may take 10 minutes)")
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix="jsonl.gz",
+        prefix="wenetspeech",
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"cuts_{partition}_raw.jsonl.gz"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        # Note this step makes the recipe different than LibriSpeech:
+        # We must filter out some utterances and remove punctuation
+        # to be consistent with Kaldi.
+        logging.info("Filtering OOV utterances from supervisions")
+        m["supervisions"] = m["supervisions"].filter(has_no_oov)
+        logging.info(f"Normalizing text in {partition}")
+        for sup in m["supervisions"]:
+            text = str(sup.text)
+            orig_text = text
+            sup.text = normalize_text(sup.text)
+            text = str(sup.text)
+            if len(orig_text) != len(text):
+                logging.info(
+                    f"\nOriginal text vs normalized text:\n{orig_text}\n{text}"
+                )
+
+        # Create long-recording cut manifests.
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        )
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        if partition not in ["DEV", "TEST_NET", "TEST_MEETING"] and perturb_speed:
+            logging.info(
+                f"Speed perturb for {partition} with factors 0.9 and 1.1 "
+                "(Perturbing may take 8 minutes and saving may take 20 minutes)"
+            )
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.",
+    )
+    return parser.parse_args()
+
+
+def main():
+    setup_logger(log_filename="./log-preprocess-wenetspeech")
+
+    args = get_args()
+    preprocess_wenet_speech(perturb_speed=args.perturb_speed)
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/local/sort_lm_training_data.py b/egs/wenetspeech/ASR/local/sort_lm_training_data.py
new file mode 120000
index 000000000..efef2c445
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/sort_lm_training_data.py
@@ -0,0 +1 @@
+../../../aishell/ASR/local/sort_lm_training_data.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/local/text2segments.py b/egs/wenetspeech/ASR/local/text2segments.py
new file mode 100644
index 000000000..bdf5a3984
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/text2segments.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#              2022  Xiaomi Corp.        (authors: Weiji Zhuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input "text", which refers to the transcript file for
+WenetSpeech:
+    - text
+and generates the output file text_word_segmentation which is implemented
+with word segmenting:
+    - text_words_segmentation
+"""
+
+
+import argparse
+from multiprocessing import Pool
+
+import jieba
+import paddle
+from tqdm import tqdm
+
+# In PaddlePaddle 2.x, dynamic graph mode is turned on by default,
+# and 'data()' is only supported in static graph mode. So if you
+# want to use this api, should call 'paddle.enable_static()' before
+# this api to enter static graph mode.
+# paddle.enable_static()
+# paddle.disable_signal_handler()
+jieba.enable_paddle()
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Chinese Word Segmentation for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--num-process",
+        "-n",
+        default=20,
+        type=int,
+        help="the number of processes",
+    )
+    parser.add_argument(
+        "--input-file",
+        "-i",
+        default="data/lang_char/text",
+        type=str,
+        help="the input text file for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        "-o",
+        default="data/lang_char/text_words_segmentation",
+        type=str,
+        help="the text implemented with words segmenting for WenetSpeech",
+    )
+
+    return parser
+
+
+def cut(lines):
+    if lines is not None:
+        cut_lines = jieba.cut(lines, use_paddle=True)
+        return [i for i in cut_lines]
+    else:
+        return None
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    num_process = args.num_process
+    input_file = args.input_file
+    output_file = args.output_file
+    # parallel mode does not support use_paddle
+    # jieba.enable_parallel(num_process)
+
+    with open(input_file, "r", encoding="utf-8") as fr:
+        lines = fr.readlines()
+
+    with Pool(processes=num_process) as p:
+        new_lines = list(tqdm(p.imap(cut, lines), total=len(lines)))
+
+    with open(output_file, "w", encoding="utf-8") as fw:
+        for line in new_lines:
+            fw.write(" ".join(line) + "\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/local/text2token.py b/egs/wenetspeech/ASR/local/text2token.py
new file mode 100755
index 000000000..d1d237a68
--- /dev/null
+++ b/egs/wenetspeech/ASR/local/text2token.py
@@ -0,0 +1,189 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument("--space", default="<space>", type=str, help="space symbol")
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument("text", type=str, default=False, nargs="?", help="input text")
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer)
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = [z.replace(" ", args.space) for z in a_flat]
+
+        print("".join(a_chars))
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/prepare.sh b/egs/wenetspeech/ASR/prepare.sh
new file mode 100755
index 000000000..b0525de60
--- /dev/null
+++ b/egs/wenetspeech/ASR/prepare.sh
@@ -0,0 +1,367 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=0
+stop_stage=100
+
+# Split L subset to this number of pieces
+# This is to avoid OOM during feature extraction.
+num_splits=1000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/WenetSpeech
+#      You can find audio, WenetSpeech.json inside it.
+#      You can apply for the download credentials by following
+#      https://github.com/wenet-e2e/WenetSpeech#download
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+lang_char_dir=data/lang_char
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  [ ! -e $dl_dir/WenetSpeech ] && mkdir -p $dl_dir/WenetSpeech
+
+  # If you have pre-downloaded it to /path/to/WenetSpeech,
+  # you can create a symlink
+  #
+  # ln -sfv /path/to/WenetSpeech $dl_dir/WenetSpeech
+  #
+  if [ ! -d $dl_dir/WenetSpeech/wenet_speech ] && [ ! -f $dl_dir/WenetSpeech/metadata/v1.list ]; then
+    log "Stage 0: You should download WenetSpeech first"
+    exit 1;
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #ln -sfv /path/to/musan $dl_dir/musan
+
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare WenetSpeech manifest"
+  # We assume that you have downloaded the WenetSpeech corpus
+  # to $dl_dir/WenetSpeech
+  mkdir -p data/manifests
+  lhotse prepare wenet-speech $dl_dir/WenetSpeech data/manifests -j $nj
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Preprocess WenetSpeech manifest"
+  if [ ! -f data/fbank/.preprocess_complete ]; then
+    python3 ./local/preprocess_wenetspeech.py --perturb-speed True
+    touch data/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute features for DEV and TEST subsets of WenetSpeech (may take 2 minutes)"
+  python3 ./local/compute_fbank_wenetspeech_dev_test.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Split S subset into ${num_splits} pieces"
+  split_dir=data/fbank/S_split_${num_splits}
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split $num_splits ./data/fbank/cuts_S_raw.jsonl.gz $split_dir
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Split M subset into ${num_splits} piece"
+  split_dir=data/fbank/M_split_${num_splits}
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split $num_splits ./data/fbank/cuts_M_raw.jsonl.gz $split_dir
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Split L subset into ${num_splits} pieces"
+  split_dir=data/fbank/L_split_${num_splits}
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split $num_splits ./data/fbank/cuts_L_raw.jsonl.gz $split_dir
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compute features for S"
+  python3 ./local/compute_fbank_wenetspeech_splits.py \
+    --training-subset S \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --start 0 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compute features for M"
+  python3 ./local/compute_fbank_wenetspeech_splits.py \
+    --training-subset M \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --start 0 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Compute features for L"
+  python3 ./local/compute_fbank_wenetspeech_splits.py \
+    --training-subset L \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --start 0 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Combine features for S"
+  if [ ! -f data/fbank/cuts_S.jsonl.gz ]; then
+    pieces=$(find data/fbank/S_split_1000 -name "cuts_S.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_S.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Combine features for M"
+  if [ ! -f data/fbank/cuts_M.jsonl.gz ]; then
+    pieces=$(find data/fbank/M_split_1000 -name "cuts_M.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_M.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: Combine features for L"
+  if [ ! -f data/fbank/cuts_L.jsonl.gz ]; then
+    pieces=$(find data/fbank/L_split_1000 -name "cuts_L.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_L.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+  log "Stage 14: Compute fbank for musan"
+  mkdir -p data/fbank
+  ./local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 15 ] && [ $stop_stage -ge 15 ]; then
+  log "Stage 15: Prepare char based lang"
+  mkdir -p $lang_char_dir
+
+  if ! which jq; then
+      echo "This script is intended to be used with jq but you have not installed jq
+      Note: in Linux, you can install jq with the following command:
+      1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+      2. chmod +x ./jq
+      3. cp jq /usr/bin" && exit 1
+  fi
+  if [ ! -f $lang_char_dir/text ] || [ ! -s $lang_char_dir/text ]; then
+    log "Prepare text."
+    gunzip -c data/manifests/wenetspeech_supervisions_L.jsonl.gz \
+      | jq '.text' | sed 's/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text
+  fi
+
+  # The implementation of chinese word segmentation for text,
+  # and it will take about 15 minutes.
+  if [ ! -f $lang_char_dir/text_words_segmentation ]; then
+    python3 ./local/text2segments.py \
+      --num-process $nj \
+      --input-file $lang_char_dir/text \
+      --output-file $lang_char_dir/text_words_segmentation
+  fi
+
+  cat $lang_char_dir/text_words_segmentation | sed 's/ /\n/g' \
+    | sort -u | sed '/^$/d' | uniq > $lang_char_dir/words_no_ids.txt
+
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    python3 ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+fi
+
+if [ $stage -le 16 ] && [ $stop_stage -ge 16 ]; then
+  log "Stage 16: Prepare char based L_disambig.pt"
+  if [ ! -f data/lang_char/L_disambig.pt ]; then
+    python3 ./local/prepare_char.py \
+      --lang-dir data/lang_char
+  fi
+fi
+
+# If you don't want to use LG for decoding, the following steps are not necessary.
+if [ $stage -le 17 ] && [ $stop_stage -ge 17 ]; then
+  log "Stage 17: Prepare G"
+  # It will take about 20 minutes.
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+  if [ ! -f $lang_char_dir/3-gram.unpruned.arpa ]; then
+    python3 ./shared/make_kn_lm.py \
+      -ngram-order 3 \
+      -text $lang_char_dir/text_words_segmentation \
+      -lm $lang_char_dir/3-gram.unpruned.arpa
+  fi
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building LG
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_char_dir/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $lang_char_dir/3-gram.unpruned.arpa > data/lm/G_3_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 18 ] && [ $stop_stage -ge 18 ]; then
+  log "Stage 18: Compile LG"
+  python ./local/compile_lg.py --lang-dir $lang_char_dir
+fi
+
+# prepare RNNLM data
+if [ $stage -le 19 ] && [ $stop_stage -ge 19 ]; then
+  log "Stage 19: Prepare LM training data"
+
+  log "Processing char based data"
+  text_out_dir=data/lm_char
+
+  mkdir -p $text_out_dir
+
+  log "Genearating training text data"
+  
+  if [ ! -f $text_out_dir/lm_data.pt ]; then
+    ./local/prepare_char_lm_training_data.py \
+      --lang-char data/lang_char \
+      --lm-data $lang_char_dir/text_words_segmentation \
+      --lm-archive $text_out_dir/lm_data.pt
+  fi
+
+  log "Generating DEV text data"
+  # prepare validation text data 
+  if [ ! -f $text_out_dir/valid_text_words_segmentation ]; then
+    valid_text=${text_out_dir}/
+
+    gunzip -c data/manifests/wenetspeech_supervisions_DEV.jsonl.gz \
+      | jq '.text' | sed 's/"//g' \
+      | ./local/text2token.py -t "char" > $text_out_dir/valid_text
+    
+    python3 ./local/text2segments.py \
+      --num-process $nj \
+      --input-file $text_out_dir/valid_text \
+      --output-file $text_out_dir/valid_text_words_segmentation
+  fi
+
+  ./local/prepare_char_lm_training_data.py \
+    --lang-char data/lang_char \
+    --lm-data $text_out_dir/valid_text_words_segmentation \
+    --lm-archive $text_out_dir/lm_data_valid.pt
+
+  # prepare TEST text data 
+  if [ ! -f $text_out_dir/TEST_text_words_segmentation ]; then
+    log "Prepare text for test set."
+    for test_set in TEST_MEETING TEST_NET; do
+        gunzip -c data/manifests/wenetspeech_supervisions_${test_set}.jsonl.gz \
+          | jq '.text' | sed 's/"//g' \
+          | ./local/text2token.py -t "char" > $text_out_dir/${test_set}_text
+
+        python3 ./local/text2segments.py \
+          --num-process $nj \
+          --input-file $text_out_dir/${test_set}_text \
+          --output-file $text_out_dir/${test_set}_text_words_segmentation
+    done
+    
+    cat $text_out_dir/TEST_*_text_words_segmentation > $text_out_dir/test_text_words_segmentation
+  fi
+
+  ./local/prepare_char_lm_training_data.py \
+    --lang-char data/lang_char \
+    --lm-data $text_out_dir/test_text_words_segmentation \
+    --lm-archive $text_out_dir/lm_data_test.pt
+
+fi
+
+# sort RNNLM data
+if [ $stage -le 20 ] && [ $stop_stage -ge 20 ]; then
+  text_out_dir=data/lm_char
+
+  log "Sort lm data"
+
+  ./local/sort_lm_training_data.py \
+    --in-lm-data $text_out_dir/lm_data.pt \
+    --out-lm-data $text_out_dir/sorted_lm_data.pt \
+    --out-statistics $text_out_dir/statistics.txt
+
+  ./local/sort_lm_training_data.py \
+    --in-lm-data $text_out_dir/lm_data_valid.pt \
+    --out-lm-data $text_out_dir/sorted_lm_data-valid.pt \
+    --out-statistics $text_out_dir/statistics-valid.txt
+
+  ./local/sort_lm_training_data.py \
+    --in-lm-data $text_out_dir/lm_data_test.pt \
+    --out-lm-data $text_out_dir/sorted_lm_data-test.pt \
+    --out-statistics $text_out_dir/statistics-test.txt
+fi
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+if [ $stage -le 21 ] && [ $stop_stage -ge 21 ]; then
+  log "Stage 21: Train RNN LM model"
+  python ../../../icefall/rnn_lm/train.py \
+    --start-epoch 0 \
+    --world-size 2 \
+    --num-epochs 20 \
+    --use-fp16 0 \
+    --embedding-dim 2048 \
+    --hidden-dim 2048 \
+    --num-layers 2 \
+    --batch-size 400 \
+    --exp-dir rnnlm_char/exp \
+    --lm-data data/lm_char/sorted_lm_data.pt \
+    --lm-data-valid data/lm_char/sorted_lm_data-valid.pt \
+    --vocab-size 5537 \
+    --master-port 12340
+fi
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/__init__.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/aishell.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/aishell.py
new file mode 120000
index 000000000..f7321272b
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/aishell.py
@@ -0,0 +1 @@
+../../../aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
new file mode 100644
index 000000000..1dbfb9709
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -0,0 +1,409 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    load_manifest,
+    load_manifest_lazy,
+    set_caching_enabled,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class WenetSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--training-subset",
+            type=str,
+            default="L",
+            help="The training subset for using",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                buffer_size=300000,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+
+        valid_dl = DataLoader(
+            validate,
+            batch_size=None,
+            sampler=valid_sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.info("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest_lazy(
+            self.args.manifest_dir / f"cuts_{self.args.training_subset}.jsonl.gz"
+        )
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_DEV.jsonl.gz")
+
+    @lru_cache()
+    def test_net_cuts(self) -> List[CutSet]:
+        logging.info("About to get TEST_NET cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST_NET.jsonl.gz")
+
+    @lru_cache()
+    def test_meeting_cuts(self) -> List[CutSet]:
+        logging.info("About to get TEST_MEETING cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST_MEETING.jsonl.gz")
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/conformer.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/conformer.py
new file mode 120000
index 000000000..a65957180
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py
new file mode 100755
index 000000000..2bafe25d6
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py
@@ -0,0 +1,689 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When training with the L subset, usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 10 \
+        --avg 2 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 10 \
+        --avg 2 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search (1best)
+./pruned_transducer_stateless2/decode.py \
+        --epoch 10 \
+        --avg 2 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+
+(4) fast beam search (nbest)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 10 \
+    --avg 2 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(5) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 10 \
+    --avg 2 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (with LG)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 10 \
+    --avg 2 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--batch",
+        type=int,
+        default=None,
+        help="It specifies the batch checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to
+        specify `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.35,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            sentence = "".join([lexicon.word_table[i] for i in hyp])
+            hyps.append(list(sentence))
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=graph_compiler.texts_to_ids(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text)) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+        if (
+            params.decoding_method == "fast_beam_search_nbest"
+            or params.decoding_method == "fast_beam_search_nbest_oracle"
+        ):
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    elif params.batch is not None:
+        filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints([filenames], device=device))
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lg_filename = params.lang_dir + "/LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    dev_cuts = wenetspeech.valid_cuts()
+    dev_dl = wenetspeech.valid_dataloaders(dev_cuts)
+
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_net_dl = wenetspeech.test_dataloaders(test_net_cuts)
+
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+    test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts)
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_dl = [dev_dl, test_net_dl, test_meeting_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode_aishell.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode_aishell.py
new file mode 100755
index 000000000..2e644ec2f
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode_aishell.py
@@ -0,0 +1,547 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless2/decode.py \
+    --epoch 84 \
+    --avg 25 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from aishell import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from finetune import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if i >= 1:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+    test_dl = aishell.test_dataloaders(test_cuts)
+    dev_dl = aishell.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/decoder.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/export-onnx.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/export-onnx.py
new file mode 100755
index 000000000..140b1d37f
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/export-onnx.py
@@ -0,0 +1,518 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=icefall_asr_wenetspeech_pruned_transducer_stateless2
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/Linv.pt"
+git lfs pull --include "exp/pretrained_epoch_10_avg_2.pt"
+
+cd exp
+ln -s pretrained_epoch_10_avg_2.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless2/export-onnx.py \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless5",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py
new file mode 100755
index 000000000..5d25daf5e
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py
@@ -0,0 +1,389 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 10 \
+  --avg 2 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Please refer to
+https://k2-fsa.github.io/sherpa/python/offline_asr/conformer/index.html
+for how to use `cpu_jit.pt` for speech recognition.
+
+It will also generate 3 other files: `encoder_jit_script.pt`,
+`decoder_jit_script.pt`, and `joiner_jit_script.pt`. Check ./jit_pretrained.py
+for how to use them.
+
+(2) Export to torchscript model using torch.jit.trace()
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 10 \
+  --avg 2 \
+  --jit-trace 1
+
+It will generate the following files:
+    - encoder_jit_trace.pt
+    - decoder_jit_trace.pt
+    - joiner_jit_trace.pt
+
+Check ./jit_pretrained.py for usage.
+
+(3) Export `model.state_dict()`
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 10 \
+  --avg 2
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/wenetspeech/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+
+You can find pretrained models at
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2/tree/main/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate 4 files:
+         - encoder_jit_script.pt
+         - decoder_jit_script.pt
+         - joiner_jit_script.pt
+         - cpu_jit.pt (which combines the above 3 files)
+
+        Check ./jit_pretrained.py for how to use xxx_jit_script.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace.
+        It will generate 3 files:
+         - encoder_jit_trace.pt
+         - decoder_jit_trace.pt
+         - joiner_jit_trace.pt
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def export_encoder_model_jit_script(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.script()
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    script_model = torch.jit.script(encoder_model)
+    script_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_script(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.script()
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    script_model = torch.jit.script(decoder_model)
+    script_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_script(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+    """
+    script_model = torch.jit.script(joiner_model)
+    script_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script")
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+
+        # Also export encoder/decoder/joiner separately
+        encoder_filename = params.exp_dir / "encoder_jit_script.pt"
+        export_encoder_model_jit_script(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_script.pt"
+        export_decoder_model_jit_script(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_script.pt"
+        export_joiner_model_jit_script(model.joiner, joiner_filename)
+    elif params.jit_trace is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py
new file mode 100755
index 000000000..c34f1593d
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py
@@ -0,0 +1,1054 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Xiaoyu Yang,
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/finetune.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --do-finetune 1 \
+  --max-duration 100
+
+"""
+
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from aishell import AishellAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_finetune_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument("--do-finetune", type=str2bool, default=False)
+
+    parser.add_argument(
+        "--init-modules",
+        type=str,
+        default=None,
+        help="""
+        Modules to be initialized. It matches all parameters starting with
+        a specific key. The keys are given with Comma separated. If None,
+        all modules will be initialised. For example, if you only want to
+        initialise all parameters staring with "encoder", use "encoder";
+        if you want to initialise parameters starting with encoder or decoder,
+        use "encoder,joiner".
+        """,
+    )
+
+    parser.add_argument(
+        "--finetune-ckpt",
+        type=str,
+        default=None,
+        help="Fine-tuning from which checkpoint (a path to a .pt file)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        pruned_transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.0001,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=100000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. During fine-tuning, we set this very large so that the
+        learning rate slowly decays with number of batches. You may tune
+        its value by yourself.
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=100,
+        help="""Number of epochs that affects how rapidly the learning rate
+        decreases. During fine-tuning, we set this very large so that the
+        learning rate slowly decays with number of batches. You may tune
+        its value by yourself.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--valid-interval",
+        type=int,
+        default=3000,
+        help="""When training_subset is L, set the valid_interval to 3000.
+        When training_subset is M, set the valid_interval to 1000.
+        When training_subset is S, set the valid_interval to 400.
+        """,
+    )
+
+    parser.add_argument(
+        "--model-warm-step",
+        type=int,
+        default=3000,
+        help="""When training_subset is L, set the model_warm_step to 3000.
+        When training_subset is M, set the model_warm_step to 500.
+        When training_subset is S, set the model_warm_step to 100.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+    Explanation of options saved in `params`:
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+        - best_train_epoch: It is the epoch that has the best training loss.
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+        - subsampling_factor:  The subsampling factor for the model.
+        - encoder_dim: Hidden dim for multi-head attention model.
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def load_model_params(
+    ckpt: str, model: nn.Module, init_modules: List[str] = None, strict: bool = True
+):
+    """Load model params from checkpoint
+
+    Args:
+        ckpt (str): Path to the checkpoint
+        model (nn.Module): model to be loaded
+
+    """
+    logging.info(f"Loading checkpoint from {ckpt}")
+    checkpoint = torch.load(ckpt, map_location="cpu")
+
+    # if module list is empty, load the whole model from ckpt
+    if not init_modules:
+        if next(iter(checkpoint["model"])).startswith("module."):
+            logging.info("Loading checkpoint saved by DDP")
+
+            dst_state_dict = model.state_dict()
+            src_state_dict = checkpoint["model"]
+            for key in dst_state_dict.keys():
+                src_key = "{}.{}".format("module", key)
+                dst_state_dict[key] = src_state_dict.pop(src_key)
+            assert len(src_state_dict) == 0
+            model.load_state_dict(dst_state_dict, strict=strict)
+        else:
+            model.load_state_dict(checkpoint["model"], strict=strict)
+    else:
+        src_state_dict = checkpoint["model"]
+        dst_state_dict = model.state_dict()
+        for module in init_modules:
+            logging.info(f"Loading parameters starting with prefix {module}")
+            src_keys = [
+                k for k in src_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
+            dst_keys = [
+                k for k in dst_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
+            assert set(src_keys) == set(dst_keys)  # two sets should match exactly
+            for key in src_keys:
+                dst_state_dict[key] = src_state_dict.pop(key)
+
+        model.load_state_dict(dst_state_dict, strict=strict)
+
+    return None
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if isinstance(y, list):
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # load model parameters for model fine-tuning
+    if params.do_finetune:
+        modules = params.init_modules.split(",") if params.init_modules else None
+        checkpoints = load_model_params(
+            ckpt=params.finetune_ckpt, model=model, init_modules=modules
+        )
+    else:
+        assert params.start_epoch > 0, params.start_epoch
+        checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aishell = AishellAsrDataModule(args)
+    train_dl = aishell.train_dataloaders(aishell.train_cuts())
+    valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0 if params.start_epoch == 1 else 1.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(
+        parser
+    )  # you may replace this with your own dataset
+    add_finetune_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/jit_pretrained.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/jit_pretrained.py
new file mode 100755
index 000000000..f90dd2b43
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/jit_pretrained.py
@@ -0,0 +1,336 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads torchscript models, either exported by `torch.jit.trace()`
+or by `torch.jit.script()`, and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --tokens data/lang_char/tokens.txt \
+  --epoch 10 \
+  --avg 2 \
+  --jit-trace 1
+
+or
+
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --tokens data/lang_char/tokens.txt \
+  --epoch 10 \
+  --avg 2 \
+  --jit 1
+
+Usage of this script:
+
+./pruned_transducer_stateless2/jit_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless2/exp/encoder_jit_trace.pt \
+  --decoder-model-filename ./pruned_transducer_stateless2/exp/decoder_jit_trace.pt \
+  --joiner-model-filename ./pruned_transducer_stateless2/exp/joiner_jit_trace.pt \
+  --tokens data/lang_char/tokens.txt \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+or
+
+./pruned_transducer_stateless2/jit_pretrained.py \
+  --encoder-model-filename ./pruned_transducer_stateless2/exp/encoder_jit_script.pt \
+  --decoder-model-filename ./pruned_transducer_stateless2/exp/decoder_jit_script.pt \
+  --joiner-model-filename ./pruned_transducer_stateless2/exp/joiner_jit_script.pt \
+  --tokens data/lang_char/tokens.txt \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+You can find pretrained models at
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2/tree/main/exp
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner torchscript model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    decoder: torch.jit.ScriptModule,
+    joiner: torch.jit.ScriptModule,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_size: int,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      decoder:
+        The decoder model.
+      joiner:
+        The joiner model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, C)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+      context_size:
+        The context size of the decoder model.
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    device = encoder_out.device
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = decoder(
+        decoder_input,
+        need_pad=torch.tensor([False]),
+    ).squeeze(1)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out
+        # current_encoder_out's shape: (batch_size, encoder_out_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+
+        logits = joiner(
+            current_encoder_out,
+            decoder_out,
+        )
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = decoder(
+                decoder_input,
+                need_pad=torch.tensor([False]),
+            )
+            decoder_out = decoder_out.squeeze(1)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    encoder = torch.jit.load(args.encoder_model_filename)
+    decoder = torch.jit.load(args.decoder_model_filename)
+    joiner = torch.jit.load(args.joiner_model_filename)
+
+    encoder.eval()
+    decoder.eval()
+    joiner.eval()
+
+    encoder.to(device)
+    decoder.to(device)
+    joiner.to(device)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = encoder(
+        x=features,
+        x_lens=feature_lengths,
+    )
+
+    hyps = greedy_search(
+        decoder=decoder,
+        joiner=joiner,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        context_size=args.context_size,
+    )
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = "".join([symbol_table[i] for i in hyp])
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/joiner.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/lstmp.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/lstmp.py
new file mode 120000
index 000000000..b82e115fc
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/lstmp.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/model.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_check.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_check.py
new file mode 100755
index 000000000..2d46eede1
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_check.py
@@ -0,0 +1,303 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script checks that exported onnx models produce the same output
+with the given torchscript model for the same input.
+
+Usage:
+
+./pruned_transducer_stateless2/onnx_check.py \
+  --jit-filename ./t/cpu_jit.pt \
+  --onnx-encoder-filename ./t/encoder.onnx \
+  --onnx-decoder-filename ./t/decoder.onnx \
+  --onnx-joiner-filename ./t/joiner.onnx \
+  --onnx-joiner-encoder-proj-filename ./t/joiner_encoder_proj.onnx \
+  --onnx-joiner-decoder-proj-filename ./t/joiner_decoder_proj.onnx
+
+You can generate cpu_jit.pt, encoder.onnx, decoder.onnx, and other
+xxx.onnx files using ./export.py
+
+We provide pretrained models at:
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2/tree/main/exp
+"""
+
+import argparse
+import logging
+
+from icefall import is_module_available
+
+if not is_module_available("onnxruntime"):
+    raise ValueError("Please 'pip install onnxruntime' first.")
+
+import onnxruntime as ort
+import torch
+
+ort.set_default_logger_severity(3)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript model exported by torch.jit.script",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-encoder-proj-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner encoder projection model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-decoder-proj-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner decoder projection model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    model: torch.jit.ScriptModule,
+    encoder_session: ort.InferenceSession,
+):
+    inputs = encoder_session.get_inputs()
+    outputs = encoder_session.get_outputs()
+    input_names = [n.name for n in inputs]
+    output_names = [n.name for n in outputs]
+
+    assert inputs[0].shape == ["N", "T", 80]
+    assert inputs[1].shape == ["N"]
+
+    for N in [1, 5]:
+        for T in [12, 25]:
+            print("N, T", N, T)
+            x = torch.rand(N, T, 80, dtype=torch.float32)
+            x_lens = torch.randint(low=10, high=T + 1, size=(N,))
+            x_lens[0] = T
+
+            encoder_inputs = {
+                input_names[0]: x.numpy(),
+                input_names[1]: x_lens.numpy(),
+            }
+            encoder_out, encoder_out_lens = encoder_session.run(
+                output_names,
+                encoder_inputs,
+            )
+
+            torch_encoder_out, torch_encoder_out_lens = model.encoder(x, x_lens)
+
+            encoder_out = torch.from_numpy(encoder_out)
+            assert torch.allclose(encoder_out, torch_encoder_out, atol=1e-05), (
+                (encoder_out - torch_encoder_out).abs().max(),
+                encoder_out.shape,
+                torch_encoder_out.shape,
+            )
+
+
+def test_decoder(
+    model: torch.jit.ScriptModule,
+    decoder_session: ort.InferenceSession,
+):
+    inputs = decoder_session.get_inputs()
+    outputs = decoder_session.get_outputs()
+    input_names = [n.name for n in inputs]
+    output_names = [n.name for n in outputs]
+
+    assert inputs[0].shape == ["N", 2]
+    for N in [1, 5, 10]:
+        y = torch.randint(low=1, high=500, size=(10, 2))
+
+        decoder_inputs = {input_names[0]: y.numpy()}
+        decoder_out = decoder_session.run(
+            output_names,
+            decoder_inputs,
+        )[0]
+        decoder_out = torch.from_numpy(decoder_out)
+
+        torch_decoder_out = model.decoder(y, need_pad=False)
+        assert torch.allclose(decoder_out, torch_decoder_out, atol=1e-5), (
+            (decoder_out - torch_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    model: torch.jit.ScriptModule,
+    joiner_session: ort.InferenceSession,
+    joiner_encoder_proj_session: ort.InferenceSession,
+    joiner_decoder_proj_session: ort.InferenceSession,
+):
+    joiner_inputs = joiner_session.get_inputs()
+    joiner_outputs = joiner_session.get_outputs()
+    joiner_input_names = [n.name for n in joiner_inputs]
+    joiner_output_names = [n.name for n in joiner_outputs]
+
+    assert joiner_inputs[0].shape == ["N", 512]
+    assert joiner_inputs[1].shape == ["N", 512]
+
+    joiner_encoder_proj_inputs = joiner_encoder_proj_session.get_inputs()
+    encoder_proj_input_name = joiner_encoder_proj_inputs[0].name
+
+    assert joiner_encoder_proj_inputs[0].shape == ["N", 512]
+
+    joiner_encoder_proj_outputs = joiner_encoder_proj_session.get_outputs()
+    encoder_proj_output_name = joiner_encoder_proj_outputs[0].name
+
+    joiner_decoder_proj_inputs = joiner_decoder_proj_session.get_inputs()
+    decoder_proj_input_name = joiner_decoder_proj_inputs[0].name
+
+    assert joiner_decoder_proj_inputs[0].shape == ["N", 512]
+
+    joiner_decoder_proj_outputs = joiner_decoder_proj_session.get_outputs()
+    decoder_proj_output_name = joiner_decoder_proj_outputs[0].name
+
+    for N in [1, 5, 10]:
+        encoder_out = torch.rand(N, 512)
+        decoder_out = torch.rand(N, 512)
+
+        projected_encoder_out = torch.rand(N, 512)
+        projected_decoder_out = torch.rand(N, 512)
+
+        joiner_inputs = {
+            joiner_input_names[0]: projected_encoder_out.numpy(),
+            joiner_input_names[1]: projected_decoder_out.numpy(),
+        }
+        joiner_out = joiner_session.run(joiner_output_names, joiner_inputs)[0]
+        joiner_out = torch.from_numpy(joiner_out)
+
+        torch_joiner_out = model.joiner(
+            projected_encoder_out,
+            projected_decoder_out,
+            project_input=False,
+        )
+        assert torch.allclose(joiner_out, torch_joiner_out, atol=1e-5), (
+            (joiner_out - torch_joiner_out).abs().max()
+        )
+
+        # Now test encoder_proj
+        joiner_encoder_proj_inputs = {encoder_proj_input_name: encoder_out.numpy()}
+        joiner_encoder_proj_out = joiner_encoder_proj_session.run(
+            [encoder_proj_output_name], joiner_encoder_proj_inputs
+        )[0]
+        joiner_encoder_proj_out = torch.from_numpy(joiner_encoder_proj_out)
+
+        torch_joiner_encoder_proj_out = model.joiner.encoder_proj(encoder_out)
+        assert torch.allclose(
+            joiner_encoder_proj_out, torch_joiner_encoder_proj_out, atol=1e-5
+        ), ((joiner_encoder_proj_out - torch_joiner_encoder_proj_out).abs().max())
+
+        # Now test decoder_proj
+        joiner_decoder_proj_inputs = {decoder_proj_input_name: decoder_out.numpy()}
+        joiner_decoder_proj_out = joiner_decoder_proj_session.run(
+            [decoder_proj_output_name], joiner_decoder_proj_inputs
+        )[0]
+        joiner_decoder_proj_out = torch.from_numpy(joiner_decoder_proj_out)
+
+        torch_joiner_decoder_proj_out = model.joiner.decoder_proj(decoder_out)
+        assert torch.allclose(
+            joiner_decoder_proj_out, torch_joiner_decoder_proj_out, atol=1e-5
+        ), ((joiner_decoder_proj_out - torch_joiner_decoder_proj_out).abs().max())
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    model = torch.jit.load(args.jit_filename)
+
+    options = ort.SessionOptions()
+    options.inter_op_num_threads = 1
+    options.intra_op_num_threads = 1
+
+    logging.info("Test encoder")
+    encoder_session = ort.InferenceSession(
+        args.onnx_encoder_filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+    test_encoder(model, encoder_session)
+
+    logging.info("Test decoder")
+    decoder_session = ort.InferenceSession(
+        args.onnx_decoder_filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+    test_decoder(model, decoder_session)
+
+    logging.info("Test joiner")
+    joiner_session = ort.InferenceSession(
+        args.onnx_joiner_filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+    joiner_encoder_proj_session = ort.InferenceSession(
+        args.onnx_joiner_encoder_proj_filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+    joiner_decoder_proj_session = ort.InferenceSession(
+        args.onnx_joiner_decoder_proj_filename,
+        sess_options=options,
+        providers=["CPUExecutionProvider"],
+    )
+    test_joiner(
+        model,
+        joiner_session,
+        joiner_encoder_proj_session,
+        joiner_decoder_proj_session,
+    )
+    logging.info("Finished checking ONNX models")
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220727)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_pretrained.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_pretrained.py
new file mode 120000
index 000000000..f1bfbee49
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/onnx_pretrained.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless5/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/optim.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/pretrained.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/pretrained.py
new file mode 100755
index 000000000..c3d67ad92
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/pretrained.py
@@ -0,0 +1,334 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		         2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+(2) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+(3) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
+Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
+./pruned_transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is beam_search
+        and modified_beam_search """,
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --decoding-method is greedy_search.
+        """,
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = "".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling_converter.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling_converter.py
new file mode 120000
index 000000000..db93d155b
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
new file mode 100644
index 000000000..49977e01b
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
@@ -0,0 +1,1066 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+For training with the L subset:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless2/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 15 \
+  --start-epoch 0 \
+  --max-duration 180 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --training-subset L
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 10 \
+  --start-epoch 0 \
+  --max-duration 180 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --use-fp16 True \
+  --training-subset L
+
+For training with the M subset:
+
+./pruned_transducer_stateless2/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 29 \
+  --start-epoch 0 \
+  --max-duration 180 \
+  --valid-interval 1000 \
+  --model-warm-step 500 \
+  --save-every-n 1000 \
+  --training-subset M
+
+For training with the S subset:
+
+./pruned_transducer_stateless2/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 29 \
+  --start-epoch 0 \
+  --max-duration 180 \
+  --valid-interval 400 \
+  --model-warm-step 100 \
+  --save-every-n 1000 \
+  --training-subset S
+"""
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        pruned_transducer_stateless2/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=20,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--valid-interval",
+        type=int,
+        default=3000,
+        help="""When training_subset is L, set the valid_interval to 3000.
+        When training_subset is M, set the valid_interval to 1000.
+        When training_subset is S, set the valid_interval to 400.
+        """,
+    )
+
+    parser.add_argument(
+        "--model-warm-step",
+        type=int,
+        default=3000,
+        help="""When training_subset is L, set the model_warm_step to 3000.
+        When training_subset is M, set the model_warm_step to 500.
+        When training_subset is S, set the model_warm_step to 100.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+    Explanation of options saved in `params`:
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+        - best_train_epoch: It is the epoch that has the best training loss.
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+        - subsampling_factor:  The subsampling factor for the model.
+        - encoder_dim: Hidden dim for multi-head attention model.
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`.
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 0:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if isinstance(y, list):
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    train_cuts = wenetspeech.train_cuts()
+    valid_cuts = wenetspeech.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 10 seconds
+        #
+        # Caution: There is a reason to select 10.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 10.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = c.supervisions[0].text.replace(" ", "")
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    valid_dl = wenetspeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = wenetspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if not params.print_diagnostics and params.start_batch == 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    texts = batch["supervisions"]["text"]
+    num_tokens = sum(len(i) for i in texts)
+
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params)
+            raise
+
+
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/__init__.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/beam_search.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..02d01b343
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/beam_search.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/conformer.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/conformer.py
new file mode 100644
index 000000000..23a877b2f
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1,1531 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask, subsequent_chunk_mask
+
+
+class Conformer(EncoderInterface):
+    """
+    Args:
+        num_features (int): Number of input features
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+        dynamic_chunk_training (bool): whether to use dynamic chunk training, if
+            you want to train a streaming model, this is expected to be True.
+            When setting True, it will use a masking strategy to make the attention
+            see only limited left and right context.
+        short_chunk_threshold (float): a threshold to determinize the chunk size
+            to be used in masking training, if the randomly generated chunk size
+            is greater than ``max_len * short_chunk_threshold`` (max_len is the
+            max sequence length of current batch) then it will use
+            full context in training (i.e. with chunk size equals to max_len).
+            This will be used only when dynamic_chunk_training is True.
+        short_chunk_size (int): see docs above, if the randomly generated chunk
+            size equals to or less than ``max_len * short_chunk_threshold``, the
+            chunk size will be sampled uniformly from 1 to short_chunk_size.
+            This also will be used only when dynamic_chunk_training is True.
+        num_left_chunks (int): the left context (in chunks) attention can see, the
+            chunk size is decided by short_chunk_threshold and short_chunk_size.
+            A minus value means seeing full left context.
+            This also will be used only when dynamic_chunk_training is True.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        dynamic_chunk_training: bool = False,
+        short_chunk_threshold: float = 0.75,
+        short_chunk_size: int = 25,
+        num_left_chunks: int = -1,
+        causal: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__()
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_layers = num_encoder_layers
+        self.d_model = d_model
+        self.cnn_module_kernel = cnn_module_kernel
+        self.causal = causal
+        self.dynamic_chunk_training = dynamic_chunk_training
+        self.short_chunk_threshold = short_chunk_threshold
+        self.short_chunk_size = short_chunk_size
+        self.num_left_chunks = num_left_chunks
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+            causal,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self._init_state: List[torch.Tensor] = [torch.empty(0)]
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, d_model)
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        assert x.size(0) == lengths.max().item()
+
+        src_key_padding_mask = make_pad_mask(lengths)
+
+        if self.dynamic_chunk_training:
+            assert (
+                self.causal
+            ), "Causal convolution is required for streaming conformer."
+            max_len = x.size(0)
+            chunk_size = torch.randint(1, max_len, (1,)).item()
+            if chunk_size > (max_len * self.short_chunk_threshold):
+                chunk_size = max_len
+            else:
+                chunk_size = chunk_size % self.short_chunk_size + 1
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=self.num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+        else:
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=None,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        return x, lengths
+
+    @torch.jit.export
+    def get_init_state(
+        self, left_context: int, device: torch.device
+    ) -> List[torch.Tensor]:
+        """Return the initial cache state of the model.
+        Args:
+          left_context: The left context size (in frames after subsampling).
+        Returns:
+          Return the initial state of the model, it is a list containing two
+          tensors, the first one is the cache for attentions which has a shape
+          of (num_encoder_layers, left_context, encoder_dim), the second one
+          is the cache of conv_modules which has a shape of
+          (num_encoder_layers, cnn_module_kernel - 1, encoder_dim).
+          NOTE: the returned tensors are on the given device.
+        """
+        if len(self._init_state) == 2 and self._init_state[0].size(1) == left_context:
+            # Note: It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        init_states: List[torch.Tensor] = [
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    left_context,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+            torch.zeros(
+                (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    self.d_model,
+                ),
+                device=device,
+            ),
+        ]
+
+        self._init_state = init_states
+
+        return init_states
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[Tensor]] = None,
+        processed_lens: Optional[Tensor] = None,
+        left_context: int = 64,
+        right_context: int = 4,
+        chunk_size: int = 16,
+        simulate_streaming: bool = False,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states:
+            The decode states for previous frames which contains the cached data.
+            It has two elements, the first element is the attn_cache which has
+            a shape of (encoder_layers, left_context, batch, attention_dim),
+            the second element is the conv_cache which has a shape of
+            (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+            Note: states will be modified in this function.
+          processed_lens:
+            How many frames (after subsampling) have been processed for each sequence.
+          left_context:
+            How many previous frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `left_context` frames
+            of left context, some have more.
+          right_context:
+            How many future frames the attention can see in current chunk.
+            Note: It's not that each individual frame has `right_context` frames
+            of right context, some have more.
+          chunk_size:
+            The chunk size for decoding, this will be used to simulate streaming
+            decoding using masking.
+          simulate_streaming:
+            If setting True, it will use a masking strategy to simulate streaming
+            fashion (i.e. every chunk data only see limited left context and
+            right context). The whole sequence is supposed to be send at a time
+            When using simulate_streaming.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+            - decode_states, the updated states including the information
+              of current chunk.
+        """
+
+        # x: [N, T, C]
+        # Caution: We assume the subsampling factor is 4!
+
+        #  lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning
+        #
+        # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
+
+        if not simulate_streaming:
+            assert states is not None
+            assert processed_lens is not None
+            assert (
+                len(states) == 2
+                and states[0].shape
+                == (self.encoder_layers, left_context, x.size(0), self.d_model)
+                and states[1].shape
+                == (
+                    self.encoder_layers,
+                    self.cnn_module_kernel - 1,
+                    x.size(0),
+                    self.d_model,
+                )
+            ), f"""The length of states MUST be equal to 2, and the shape of
+             first element should be {(self.encoder_layers, left_context, x.size(0), self.d_model)},
+             given {states[0].shape}. the shape of second element should be
+             {(self.encoder_layers, self.cnn_module_kernel - 1, x.size(0), self.d_model)},
+             given {states[1].shape}."""
+
+            lengths -= 2  # we will cut off 1 frame on each side of encoder_embed output
+
+            src_key_padding_mask = make_pad_mask(lengths)
+
+            processed_mask = torch.arange(left_context, device=x.device).expand(
+                x.size(0), left_context
+            )
+            processed_lens = processed_lens.view(x.size(0), 1)
+            processed_mask = (processed_lens <= processed_mask).flip(1)
+
+            src_key_padding_mask = torch.cat(
+                [processed_mask, src_key_padding_mask], dim=1
+            )
+
+            embed = self.encoder_embed(x)
+
+            # cut off 1 frame on each size of embed as they see the padding
+            # value which causes a training and decoding mismatch.
+            embed = embed[:, 1:-1, :]
+
+            embed, pos_enc = self.encoder_pos(embed, left_context)
+            embed = embed.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+
+            x, states = self.encoder.chunk_forward(
+                embed,
+                pos_enc,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+                states=states,
+                left_context=left_context,
+                right_context=right_context,
+            )  # (T, B, F)
+            if right_context > 0:
+                x = x[0:-right_context, ...]
+                lengths -= right_context
+        else:
+            assert states is None
+            states = []  # just to make torch.script.jit happy
+            # this branch simulates streaming decoding using mask as we are
+            # using in training time.
+            src_key_padding_mask = make_pad_mask(lengths)
+            x = self.encoder_embed(x)
+            x, pos_emb = self.encoder_pos(x)
+            x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+            assert x.size(0) == lengths.max().item()
+
+            num_left_chunks = -1
+            if left_context >= 0:
+                assert left_context % chunk_size == 0
+                num_left_chunks = left_context // chunk_size
+
+            mask = ~subsequent_chunk_mask(
+                size=x.size(0),
+                chunk_size=chunk_size,
+                num_left_chunks=num_left_chunks,
+                device=x.device,
+            )
+            x = self.encoder(
+                x,
+                pos_emb,
+                mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return x, lengths, states
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        causal (bool): Whether to use causal convolution in conformer encoder
+            layer. This MUST be True when using dynamic_chunk_training and streaming decoding.
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+        causal: bool = False,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel, causal=causal)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        conv, _ = self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        src = src + self.dropout(conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        """
+        Pass the input through the encoder layer.
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        assert not self.training
+        assert len(states) == 2
+        assert states[0].shape == (left_context, src.size(1), src.size(2))
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # We put the attention cache this level (i.e. before linear transformation)
+        # to save memory consumption, when decoding in streaming fashion, the
+        # batch size would be thousands (for 32GB machine), if we cache key & val
+        # separately, it needs extra several GB memory.
+        # TODO(WeiKang): Move cache to self_attn level (i.e. cache key & val
+        # separately) if needed.
+        key = torch.cat([states[0], src], dim=0)
+        val = key
+        if right_context > 0:
+            states[0] = key[
+                -(left_context + right_context) : -right_context, ...  # noqa
+            ]
+        else:
+            states[0] = key[-left_context:, ...]
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            key,
+            val,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+            left_context=left_context,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        conv, conv_cache = self.conv_module(src, states[1], right_context)
+        states[1] = conv_cache
+
+        src = src + self.dropout(conv)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        return src, states
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+        """
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+    @torch.jit.export
+    def chunk_forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        states: List[Tensor],
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+        left_context: int = 0,
+        right_context: int = 0,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        r"""Pass the input through the encoder layers in turn.
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            states:
+              The decode states for previous frames which contains the cached data.
+              It has two elements, the first element is the attn_cache which has
+              a shape of (encoder_layers, left_context, batch, attention_dim),
+              the second element is the conv_cache which has a shape of
+              (encoder_layers, cnn_module_kernel-1, batch, conv_dim).
+              Note: states will be modified in this function.
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+            left_context:
+              How many previous frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `left_context` frames
+              of left context, some have more.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*(S+left_context)-1, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+        """
+        assert not self.training
+        assert len(states) == 2
+        assert states[0].shape == (
+            self.num_layers,
+            left_context,
+            src.size(1),
+            src.size(2),
+        )
+        assert states[1].size(0) == self.num_layers
+
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            cache = [states[0][layer_index], states[1][layer_index]]
+            output, cache = mod.chunk_forward(
+                output,
+                pos_emb,
+                states=cache,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+                left_context=left_context,
+                right_context=right_context,
+            )
+            states[0][layer_index] = cache[0]
+            states[1][layer_index] = cache[1]
+
+        return output, states
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor, left_context: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_1 = x.size(1) + left_context
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_1 * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_1, self.d_model)
+        pe_negative = torch.zeros(x_size_1, self.d_model)
+        position = torch.arange(0, x_size_1, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+        """
+        self.extend_pe(x, left_context)
+        x_size_1 = x.size(1) + left_context
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_1
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+    Examples::
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            left_context=left_context,
+        )
+
+    def rel_shift(self, x: Tensor, left_context: int = 0) -> Tensor:
+        """Compute relative positional encoding.
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+
+        time2 = time1 + left_context
+        if not torch.jit.is_tracing():
+            assert (
+                n == left_context + 2 * time1 - 1
+            ), f"{n} == {left_context} + 2 * {time1} - 1"
+
+        if torch.jit.is_tracing():
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(time2)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, time1, time2)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time1_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, time1, time2),
+                (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+                storage_offset=n_stride * (time1 - 1),
+            )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1)
+        p = p.permute(0, 2, 3, 1)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(q_with_bias_v, p)  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd, left_context)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+
+        # If we are using dynamic_chunk_training and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`, at this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax), so, we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, tgt_len, src_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+        causal (bool): Whether to use causal convolution.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        bias: bool = True,
+        causal: bool = False,
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+        self.causal = causal
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.lorder = kernel_size - 1
+        padding = (kernel_size - 1) // 2
+        if self.causal:
+            padding = 0
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        cache: Optional[Tensor] = None,
+        right_context: int = 0,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module.
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: The cache of depthwise_conv, only used in real streaming
+                decoding.
+            right_context:
+              How many future frames the attention can see in current chunk.
+              Note: It's not that each individual frame has `right_context` frames
+              of right context, some have more.
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+        Returns:
+            If cache is None return the output tensor (#time, batch, channels).
+            If cache is not None, return a tuple of Tensor, the first one is
+            the output tensor (#time, batch, channels), the second one is the
+            new cache for next chunk (#kernel_size - 1, batch, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        if self.causal and self.lorder > 0:
+            if cache is None:
+                # Make depthwise_conv causal by
+                # manualy padding self.lorder zeros to the left
+                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+            else:
+                assert not self.training, "Cache should be None in training time"
+                assert cache.size(0) == self.lorder
+                x = torch.cat([cache.permute(1, 2, 0), x], dim=2)
+                if right_context > 0:
+                    cache = x.permute(2, 0, 1)[
+                        -(self.lorder + right_context) : (-right_context),  # noqa
+                        ...,
+                    ]
+                else:
+                    cache = x.permute(2, 0, 1)[-self.lorder :, ...]  # noqa
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        # torch.jit.script requires return types be the same as annotated above
+        if cache is None:
+            cache = torch.empty(0)
+
+        return x.permute(2, 0, 1), cache
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+        Args:
+          x:
+            Its shape is (N, T, idim).
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..d665f3364
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,918 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo)
+# Copyright 2022 Xiaomi Corporation (Author: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When training with the L subset, the offline usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 4 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless5/exp_L_offline \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 4 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless5/exp_L_offline \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 4 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless5/exp_L_offline \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+
+When training with the L subset, the streaming usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+        --lang-dir data/lang_char \
+        --exp-dir pruned_transducer_stateless5/exp_L_streaming \
+        --use-averaged-model True \
+        --max-duration 600 \
+        --epoch 7 \
+        --avg 1 \
+        --decoding-method greedy_search \
+        --simulate-streaming 1 \
+        --causal-convolution 1 \
+        --decode-chunk-size 16 \
+        --left-context 64
+
+(2) modified beam search
+./pruned_transducer_stateless5/decode.py \
+        --lang-dir data/lang_char \
+        --exp-dir pruned_transducer_stateless5/exp_L_streaming \
+        --use-averaged-model True \
+        --max-duration 600 \
+        --epoch 7 \
+        --avg 1 \
+        --decoding-method modified_beam_search \
+        --simulate-streaming 1 \
+        --causal-convolution 1 \
+        --decode-chunk-size 16 \
+        --left-context 64
+
+(3) fast beam search
+./pruned_transducer_stateless5/decode.py \
+        --lang-dir data/lang_char \
+        --exp-dir pruned_transducer_stateless5/exp_L_streaming \
+        --use-averaged-model True \
+        --max-duration 600 \
+        --epoch 7 \
+        --avg 1 \
+        --decoding-method fast_beam_search \
+        --simulate-streaming 1 \
+        --causal-convolution 1 \
+        --decode-chunk-size 16 \
+        --left-context 64
+
+(4) modified beam search with RNNLM shallow fusion
+./pruned_transducer_stateless5/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam-size 4 \
+    --lm-type rnn \
+    --lm-scale 0.3 \
+    --lm-exp-dir /path/to/LM \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+"""
+
+
+import argparse
+import glob
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import ContextGraph, LmScorer, NgramLm
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=3,
+        help="""Token Ngram used for rescoring.
+            Used only when the decoding method is
+            modified_beam_search_ngram_rescoring, or LODR
+            """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="""ID of the backoff symbol.
+                Used only when the decoding method is
+                modified_beam_search_ngram_rescoring""",
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+            encoder_out_lens=encoder_out_lens,
+            context_graph=context_graph,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        key = f"beam_size_{params.beam_size}"
+        if params.has_contexts:
+            key += f"-context-score-{params.context_score}"
+        else:
+            key += "-no-context-words"
+        return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    ngram_lm: Optional[NgramLm] = None,
+    ngram_lm_scale: float = 1.0,
+    LM: Optional[LmScorer] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list("".join(text.split())) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+            context_graph=context_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    )
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+        if params.has_contexts:
+            params.suffix += f"-context-score-{params.context_score}"
+        else:
+            params.suffix += "-no-contexts-words"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    if params.use_shallow_fusion:
+        if params.lm_type == "rnn":
+            params.suffix += f"-rnnlm-lm-scale-{params.lm_scale}"
+        elif params.lm_type == "transformer":
+            params.suffix += f"-transformer-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+    # only load N-gram LM when needed
+    if "ngram" in params.decoding_method or "LODR" in params.decoding_method:
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"lm filename: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    # import pdb; pdb.set_trace()
+    # only load the neural network LM if doing shallow fusion
+    if params.use_shallow_fusion:
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+
+        num_param = sum([p.numel() for p in LM.parameters()])
+        logging.info(f"Number of model parameters: {num_param}")
+    else:
+        LM = None
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build([(c, 0.0) for c in contexts])
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    dev_cuts = wenetspeech.valid_cuts()
+    dev_dl = wenetspeech.valid_dataloaders(dev_cuts)
+
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_net_dl = wenetspeech.test_dataloaders(test_net_cuts)
+
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+    test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts)
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_dls = [dev_dl, test_net_dl, test_meeting_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+            LM=LM,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode_stream.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode_stream.py
new file mode 100644
index 000000000..e522943c0
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode_stream.py
@@ -0,0 +1,146 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class DecodeStream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        initial_states: List[torch.Tensor],
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+    ) -> None:
+        """
+        Args:
+          initial_states:
+            Initial decode states of the model, e.g. the return value of
+            `get_init_state` in conformer.py
+          decoding_graph:
+            Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+            Used only when decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+        """
+        if params.decoding_method == "fast_beam_search":
+            assert decoding_graph is not None
+            assert device == decoding_graph.device
+
+        self.params = params
+        self.cut_id = cut_id
+        self.LOG_EPS = math.log(1e-10)
+
+        self.states = initial_states
+
+        # It contains a 2-D tensors representing the feature frames.
+        self.features: torch.Tensor = None
+
+        self.num_frames: int = 0
+        # how many frames have been processed. (before subsampling).
+        # we only modify this value in `func:get_feature_frames`.
+        self.num_processed_frames: int = 0
+
+        self._done: bool = False
+
+        # The transcript of current utterance.
+        self.ground_truth: str = ""
+
+        # The decoding result (partial or final) of current utterance.
+        self.hyp: List = []
+
+        # how many frames have been processed, after subsampling (i.e. a
+        # cumulative sum of the second return value of
+        # encoder.streaming_forward
+        self.done_frames: int = 0
+
+        self.pad_length = (params.right_context + 2) * params.subsampling_factor + 3
+
+        if params.decoding_method == "greedy_search":
+            self.hyp = [params.blank_id] * params.context_size
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[params.blank_id] * params.context_size,
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    @property
+    def done(self) -> bool:
+        """Return True if all the features are processed."""
+        return self._done
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    def set_features(
+        self,
+        features: torch.Tensor,
+    ) -> None:
+        """Set features tensor of current utterance."""
+        assert features.dim() == 2, features.dim()
+        self.features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, self.pad_length),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+        self.num_frames = self.features.size(0)
+
+    def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]:
+        """Consume chunk_size frames of features"""
+        chunk_length = chunk_size + self.pad_length
+
+        ret_length = min(self.num_frames - self.num_processed_frames, chunk_length)
+
+        ret_features = self.features[
+            self.num_processed_frames : self.num_processed_frames + ret_length  # noqa
+        ]
+
+        self.num_processed_frames += chunk_size
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_features, ret_length
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.params.decoding_method == "greedy_search":
+            return self.hyp[self.params.context_size :]  # noqa
+        elif self.params.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.params.context_size :]  # noqa
+        else:
+            assert self.params.decoding_method == "fast_beam_search"
+            return self.hyp
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decoder.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..6775ee67e
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/decoder.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..972e44ca4
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py
new file mode 100755
index 000000000..921766ad4
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/export-onnx-streaming.py
@@ -0,0 +1,679 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_streaming
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_streaming
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/Linv.pt"
+git lfs pull --include "exp/pretrained_epoch_7_avg_1.pt"
+
+cd exp
+ln -s pretrained_epoch_7_avg_1.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx-streaming.py \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained-streaming.py for how to
+use the exported ONNX models.
+
+You can find the exported models in
+https://huggingface.co/csukuangfj/sherpa-onnx-streaming-conformer-zh-2023-05-23
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+from icefall.lexicon import Lexicon
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from decoder import Decoder
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+        self.num_encoder_layers = encoder.encoder_layers
+        self.encoder_dim = encoder.d_model
+        self.cnn_module_kernel = encoder.cnn_module_kernel
+
+        # Note you can tune these values
+        self.left_context = 64  # after subsampling
+        self.chunk_size = 16  # after subsampling
+        self.right_context = 0  # after subsampling
+
+        subsampling_factor = 4
+        self.pad_length = (self.right_context + 2) * subsampling_factor + 3
+
+        self.T = (self.chunk_size * subsampling_factor) + self.pad_length
+        self.decode_chunk_len = self.chunk_size * subsampling_factor
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cached_attn: torch.Tensor,
+        cached_conv: torch.Tensor,
+        processed_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, self.T, C)
+          cached_attn:
+            A 3-D tensor of shape
+            (num_encoder_layers, self.left_context, N, self.encoder_dim)
+          cached_conv:
+            A 3-D tensor of shape
+            (num_encoder_layers, self.cnn_module_kernel-1, N, self.encoder_dim)
+          processed_lens:
+            A 1-D tensor of shape (N,). It contains number of processed frames
+            after subsampling. Its dtype is torch.int64.
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, self.chunk_size, joiner_dim)
+            - new_cached_attn, it has the same shape as cached_attn
+            - new_cached_conv, it has the same shape as cached_conv
+        """
+        assert x.size(1) == self.T, (x.shape, self.T)
+        N = x.size(0)
+        x_lens = torch.full((N,), fill_value=self.T, device=x.device, dtype=torch.int64)
+
+        (
+            encoder_out,
+            _,
+            [new_cached_attn, new_cached_conv],
+        ) = self.encoder.streaming_forward(
+            x,
+            x_lens,
+            states=[cached_attn, cached_conv],
+            processed_lens=processed_lens,
+            left_context=self.left_context,
+            right_context=self.right_context,
+            chunk_size=self.chunk_size,
+        )
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, new_cached_attn, new_cached_conv
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    N = 1
+    x = torch.zeros(N, encoder_model.T, 80, dtype=torch.float32)
+    cached_attn = torch.zeros(
+        encoder_model.num_encoder_layers,
+        encoder_model.left_context,
+        N,
+        encoder_model.encoder_dim,
+    )
+    cached_conv = torch.zeros(
+        encoder_model.num_encoder_layers,
+        encoder_model.cnn_module_kernel - 1,
+        N,
+        encoder_model.encoder_dim,
+    )
+    processed_lens = torch.zeros((N,), dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, cached_attn, cached_conv, processed_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "cached_attn", "cached_conv", "processed_lens"],
+        output_names=["encoder_out", "new_cached_attn", "new_cached_conv"],
+        dynamic_axes={
+            "x": {0: "N"},
+            "cached_attn": {2: "N"},
+            "cached_conv": {2: "N"},
+            "processed_lens": {0: "N"},
+            "encoder_out": {0: "N"},
+            "new_cached_attn": {2: "N"},
+            "new_cached_conv": {2: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless5",
+        "pad_length": str(encoder_model.pad_length),
+        "decode_chunk_len": str(encoder_model.decode_chunk_len),
+        "encoder_dim": str(encoder_model.encoder_dim),
+        "num_encoder_layers": str(encoder_model.num_encoder_layers),
+        "cnn_module_kernel": str(encoder_model.cnn_module_kernel),
+        "left_context": str(encoder_model.left_context),
+        "right_context": str(encoder_model.right_context),
+        "chunk_size": str(encoder_model.chunk_size),
+        "T": str(encoder_model.T),
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    if not params.causal_convolution:
+        logging.info("Seting causal_convolution to True for exporting streaming models")
+        params.causal_convolution = True
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum(p.numel() for p in model.parameters())
+    logging.info(f"Number of model parameters: {num_param}")
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/export-onnx.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/export-onnx.py
new file mode 100755
index 000000000..037c7adf1
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/export-onnx.py
@@ -0,0 +1,603 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/Linv.pt"
+git lfs pull --include "exp/pretrained_epoch_9_avg_1.pt"
+
+cd exp
+ln -s pretrained_epoch_9_avg_1.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "conformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless5",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/export.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..cb541070e
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,208 @@
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage for offline:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp_L_offline \
+  --lang-dir data/lang_char \
+  --epoch 4 \
+  --avg 1
+
+It will generate a file exp_dir/pretrained.pt for offline ASR.
+
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp_L_offline \
+  --lang-dir data/lang_char \
+  --epoch 4 \
+  --avg 1 \
+  --jit True
+
+It will generate a file exp_dir/cpu_jit.pt for offline ASR.
+
+Usage for streaming:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp_L_streaming \
+  --lang-dir data/lang_char \
+  --epoch 7 \
+  --avg 1
+
+It will generate a file exp_dir/pretrained.pt for streaming ASR.
+
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp_L_streaming \
+  --lang-dir data/lang_char \
+  --epoch 7 \
+  --avg 1 \
+  --jit True
+
+It will generate a file exp_dir/cpu_jit.pt for streaming ASR.
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/wenetspeech/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 4 \
+        --avg 1 \
+        --decoding-method greedy_search \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        convert_scaled_to_non_scaled(model, inplace=True)
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/joiner.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..f5279e151
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/joiner.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/lstmp.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/lstmp.py
new file mode 120000
index 000000000..d13a1e063
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/lstmp.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/lstmp.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/model.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..7b417fd89
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/model.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py
new file mode 100755
index 000000000..ee8252a90
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py
@@ -0,0 +1,235 @@
+#!/usr/bin/env python3
+#
+# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script checks that exported onnx models produce the same output
+with the given torchscript model for the same input.
+
+We use the pre-trained model from
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/Linv.pt"
+git lfs pull --include "exp/pretrained_epoch_4_avg_1.pt"
+git lfs pull --include "exp/cpu_jit_epoch_4_avg_1_torch.1.7.1.pt"
+
+cd exp
+ln -s pretrained_epoch_9_avg_1_torch.1.7.1.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+4. Run this file
+
+./pruned_transducer_stateless5/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit_epoch_4_avg_1_torch.1.7.1.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+"""
+
+import argparse
+import logging
+
+from icefall import is_module_available
+from onnx_pretrained import OnnxModel
+
+import torch
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit-filename",
+        required=True,
+        type=str,
+        help="Path to the torchscript model",
+    )
+
+    parser.add_argument(
+        "--onnx-encoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx encoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-decoder-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx decoder model",
+    )
+
+    parser.add_argument(
+        "--onnx-joiner-filename",
+        required=True,
+        type=str,
+        help="Path to the onnx joiner model",
+    )
+
+    return parser
+
+
+def test_encoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    C = 80
+    for i in range(3):
+        N = torch.randint(low=1, high=20, size=(1,)).item()
+        T = torch.randint(low=30, high=50, size=(1,)).item()
+        logging.info(f"test_encoder: iter {i}, N={N}, T={T}")
+
+        x = torch.rand(N, T, C)
+        x_lens = torch.randint(low=30, high=T + 1, size=(N,))
+        x_lens[0] = T
+
+        torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens)
+        torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out)
+
+        onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens)
+
+        assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), (
+            (torch_encoder_out - onnx_encoder_out).abs().max()
+        )
+
+
+def test_decoder(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    context_size = onnx_model.context_size
+    vocab_size = onnx_model.vocab_size
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_decoder: iter {i}, N={N}")
+        x = torch.randint(
+            low=1,
+            high=vocab_size,
+            size=(N, context_size),
+            dtype=torch.int64,
+        )
+        torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False]))
+        torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out)
+        torch_decoder_out = torch_decoder_out.squeeze(1)
+
+        onnx_decoder_out = onnx_model.run_decoder(x)
+        assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), (
+            (torch_decoder_out - onnx_decoder_out).abs().max()
+        )
+
+
+def test_joiner(
+    torch_model: torch.jit.ScriptModule,
+    onnx_model: OnnxModel,
+):
+    encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1]
+    decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1]
+    for i in range(10):
+        N = torch.randint(1, 100, size=(1,)).item()
+        logging.info(f"test_joiner: iter {i}, N={N}")
+        encoder_out = torch.rand(N, encoder_dim)
+        decoder_out = torch.rand(N, decoder_dim)
+
+        projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out)
+        projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out)
+
+        torch_joiner_out = torch_model.joiner(encoder_out, decoder_out)
+        onnx_joiner_out = onnx_model.run_joiner(
+            projected_encoder_out, projected_decoder_out
+        )
+
+        assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), (
+            (torch_joiner_out - onnx_joiner_out).abs().max()
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_model = torch.jit.load(args.jit_filename)
+
+    onnx_model = OnnxModel(
+        encoder_model_filename=args.onnx_encoder_filename,
+        decoder_model_filename=args.onnx_decoder_filename,
+        joiner_model_filename=args.onnx_joiner_filename,
+    )
+
+    logging.info("Test encoder")
+    test_encoder(torch_model, onnx_model)
+
+    logging.info("Test decoder")
+    test_decoder(torch_model, onnx_model)
+
+    logging.info("Test joiner")
+    test_joiner(torch_model, onnx_model)
+    logging.info("Finished checking ONNX models")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+# See https://github.com/pytorch/pytorch/issues/38342
+# and https://github.com/pytorch/pytorch/issues/33354
+#
+# If we don't do this, the delay increases whenever there is
+# a new request that changes the actual batch size.
+# If you use `py-spy dump --pid <server-pid> --native`, you will
+# see a lot of time is spent in re-compiling the torch script model.
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_set_profiling_mode(False)
+torch._C._set_graph_executor_optimize(False)
+if __name__ == "__main__":
+    torch.manual_seed(20220727)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_pretrained-streaming.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_pretrained-streaming.py
new file mode 100755
index 000000000..c31db6859
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_pretrained-streaming.py
@@ -0,0 +1,460 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This script loads ONNX models exported by ./export-onnx.py
+and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_streaming
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_streaming
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/Linv.pt"
+git lfs pull --include "exp/pretrained_epoch_7_avg_1.pt"
+
+cd exp
+ln -s pretrained_epoch_7_avg_1.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx-streaming.py \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file with the exported ONNX models
+
+./pruned_transducer_stateless5/onnx_pretrained-streaming.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_char/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav
+
+Note: Even though this script only supports decoding a single file,
+the exported ONNX models do support batch processing.
+
+You can find the exported models in
+https://huggingface.co/csukuangfj/sherpa-onnx-streaming-conformer-zh-2023-05-23
+"""
+
+import argparse
+import logging
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_file",
+        type=str,
+        help="The input sound file to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.init_encoder_states()
+
+    def init_encoder_states(self, batch_size: int = 1):
+        encoder_meta = self.encoder.get_modelmeta().custom_metadata_map
+        print(encoder_meta)
+
+        model_type = encoder_meta["model_type"]
+        assert model_type == "conformer", model_type
+
+        decode_chunk_len = int(encoder_meta["decode_chunk_len"])
+        T = int(encoder_meta["T"])
+        pad_length = int(encoder_meta["pad_length"])
+
+        encoder_dim = int(encoder_meta["encoder_dim"])
+        cnn_module_kernel = int(encoder_meta["cnn_module_kernel"])
+        left_context = int(encoder_meta["left_context"])
+        num_encoder_layers = int(encoder_meta["num_encoder_layers"])
+
+        self.cached_attn = torch.zeros(
+            num_encoder_layers,
+            left_context,
+            batch_size,
+            encoder_dim,
+        ).numpy()
+        self.cached_conv = torch.zeros(
+            num_encoder_layers,
+            cnn_module_kernel - 1,
+            batch_size,
+            encoder_dim,
+        ).numpy()
+
+        logging.info(f"decode_chunk_len: {decode_chunk_len}")
+        logging.info(f"T: {T}")
+        logging.info(f"pad_length: {pad_length}")
+        logging.info(f"encoder_dim: {encoder_dim}")
+        logging.info(f"cnn_module_kernel: {cnn_module_kernel}")
+        logging.info(f"left_context: {left_context}")
+        logging.info(f"num_encoder_layers: {num_encoder_layers}")
+
+        self.segment = T
+        self.offset = decode_chunk_len
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def _build_encoder_input_output(
+        self, x: torch.Tensor, processed_lens: int
+    ) -> Tuple[Dict[str, np.ndarray], List[str]]:
+        assert x.size(0) == 1
+        encoder_input = {
+            "x": x.numpy(),
+            "cached_attn": self.cached_attn,
+            "cached_conv": self.cached_conv,
+            "processed_lens": torch.full(
+                (1,), fill_value=processed_lens, dtype=torch.int64
+            ).numpy(),
+        }
+        encoder_output = ["encoder_out", "new_cached_attn", "new_cached_conv"]
+
+        return encoder_input, encoder_output
+
+    def _update_states(self, states: List[np.ndarray]):
+        self.cached_attn = states[0]
+        self.cached_conv = states[1]
+
+    def run_encoder(self, x: torch.Tensor, num_processed_frames: int) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, self.T, C). It only implements N == 1
+          num_processed_frames:
+            Number of processed frames before subsampling.
+        Returns:
+          Return a 3-D tensor of shape (N, chunk_size, joiner_dim)
+        """
+        # assume subsampling_factor is 4
+        num_processed_frames = num_processed_frames // 4
+        encoder_input, encoder_output_names = self._build_encoder_input_output(
+            x, num_processed_frames
+        )
+        out = self.encoder.run(encoder_output_names, encoder_input)
+
+        self._update_states(out[1:])
+
+        return torch.from_numpy(out[0])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    context_size: int,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+) -> List[int]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (1, T, joiner_dim)
+      context_size:
+        The context size of the decoder model.
+      decoder_out:
+        Optional. Decoder output of the previous chunk.
+      hyp:
+        Decoding results for previous chunks.
+    Returns:
+      Return the decoded results so far.
+    """
+
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor([hyp], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert hyp is not None, hyp
+
+    encoder_out = encoder_out.squeeze(0)
+    T = encoder_out.size(0)
+    for t in range(T):
+        cur_encoder_out = encoder_out[t : t + 1]
+        joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0)
+        y = joiner_out.argmax(dim=0).item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+            decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {args.sound_file}")
+    waves = read_sound_files(
+        filenames=[args.sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    tail_padding = torch.zeros(int(1.0 * sample_rate), dtype=torch.float32)
+    wave_samples = torch.cat([waves, tail_padding])
+
+    num_processed_frames = 0
+    segment = model.segment
+    offset = model.offset
+
+    context_size = model.context_size
+    hyp = None
+    decoder_out = None
+
+    chunk = int(1 * sample_rate)  # 1 second
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0)
+            frames = frames.unsqueeze(0)
+            encoder_out = model.run_encoder(frames, num_processed_frames)
+            hyp, decoder_out = greedy_search(
+                model,
+                encoder_out,
+                context_size,
+                decoder_out,
+                hyp,
+            )
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    text = ""
+    for i in hyp[context_size:]:
+        text += symbol_table[i]
+    text = text.replace("▁", " ").strip()
+
+    logging.info(args.sound_file)
+    logging.info(text)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py
new file mode 100755
index 000000000..c784853ee
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py
@@ -0,0 +1,428 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+We use the pre-trained model from
+https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/Linv.pt"
+git lfs pull --include "exp/pretrained_epoch_4_avg_1.pt"
+git lfs pull --include "exp/cpu_jit_epoch_4_avg_1_torch.1.7.1.pt"
+
+cd exp
+ln -s pretrained_epoch_9_avg_1_torch.1.7.1.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --lang-dir $repo/data/lang_char \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 24 \
+  --dim-feedforward 1536 \
+  --nhead 8 \
+  --encoder-dim 384 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+3. Run this file
+
+./pruned_transducer_stateless5/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_char/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder_model_filename: str,
+        decoder_model_filename: str,
+        joiner_model_filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 4
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder_model_filename)
+        self.init_decoder(decoder_model_filename)
+        self.init_joiner(joiner_model_filename)
+
+    def init_encoder(self, encoder_model_filename: str):
+        self.encoder = ort.InferenceSession(
+            encoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+    def init_decoder(self, decoder_model_filename: str):
+        self.decoder = ort.InferenceSession(
+            decoder_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
+        self.context_size = int(decoder_meta["context_size"])
+        self.vocab_size = int(decoder_meta["vocab_size"])
+
+        logging.info(f"context_size: {self.context_size}")
+        logging.info(f"vocab_size: {self.vocab_size}")
+
+    def init_joiner(self, joiner_model_filename: str):
+        self.joiner = ort.InferenceSession(
+            joiner_model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
+        self.joiner_dim = int(joiner_meta["joiner_dim"])
+
+        logging.info(f"joiner_dim: {self.joiner_dim}")
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 2-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, its shape is (N, T', joiner_dim)
+            - encoder_out_lens, its shape is (N,)
+        """
+        out = self.encoder.run(
+            [
+                self.encoder.get_outputs()[0].name,
+                self.encoder.get_outputs()[1].name,
+            ],
+            {
+                self.encoder.get_inputs()[0].name: x.numpy(),
+                self.encoder.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A 2-D tensor of shape (N, context_size)
+        Returns:
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        out = self.decoder.run(
+            [self.decoder.get_outputs()[0].name],
+            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
+        )[0]
+
+        return torch.from_numpy(out)
+
+    def run_joiner(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        out = self.joiner.run(
+            [self.joiner.get_outputs()[0].name],
+            {
+                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
+                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(out)
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(
+    model: OnnxModel,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, joiner_dim)
+      encoder_out_lens:
+        A 1-D tensor of shape (N,).
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = 0  # hard-code to 0
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    context_size = model.context_size
+    hyps = [[blank_id] * context_size for _ in range(N)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        dtype=torch.int64,
+    )  # (N, context_size)
+
+    decoder_out = model.run_decoder(decoder_input)
+
+    offset = 0
+    for batch_size in batch_size_list:
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = packed_encoder_out.data[start:end]
+        # current_encoder_out's shape: (batch_size, joiner_dim)
+        offset = end
+
+        decoder_out = decoder_out[:batch_size]
+        logits = model.run_joiner(current_encoder_out, decoder_out)
+
+        # logits'shape (batch_size, vocab_size)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
+            decoder_input = torch.tensor(
+                decoder_input,
+                dtype=torch.int64,
+            )
+            decoder_out = model.run_decoder(decoder_input)
+
+    sorted_ans = [h[context_size:] for h in hyps]
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)
+
+    hyps = greedy_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+    )
+    s = "\n"
+
+    symbol_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += symbol_table[i]
+        return text.replace("▁", " ").strip()
+
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = token_ids_to_words(hyp)
+        s += f"{filename}:\n{words}\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/optim.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..210374f22
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/optim.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/pretrained.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100644
index 000000000..1cac20435
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,334 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		         2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Offline Usage:
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless5/exp_L_offline/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+(2) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless5/exp_L_offline/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+(3) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp_L_offline/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+You can also use `./pruned_transducer_stateless5/exp_L_offline/epoch-xx.pt`.
+Note: ./pruned_transducer_stateless5/exp_L_offline/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=48000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and modified_beam_search ",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/scaling.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..ff7bfeda9
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/scaling.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/scaling_converter.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/scaling_converter.py
new file mode 120000
index 000000000..e58473a04
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/scaling_converter.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_beam_search.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_beam_search.py
new file mode 100644
index 000000000..810d94135
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_beam_search.py
@@ -0,0 +1,283 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from decode_stream import DecodeStream
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    T = encoder_out.size(1)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (N, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+        # print(current_encoder_out.shape)
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+    num_active_paths: int = 4,
+) -> None:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      num_active_paths:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(num_active_paths)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    streams: List[DecodeStream],
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+    A lattice is first generated by Fsa-based beam search, then we get the
+    recognition by applying shortest path on the lattice.
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      streams:
+        A list of stream objects.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+    best_path = one_best_decoding(lattice)
+    hyp_tokens = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyp_tokens[i]
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_decode.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_decode.py
new file mode 100644
index 000000000..3a4dc3cb8
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_decode.py
@@ -0,0 +1,665 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+(1) greedy search
+python pruned_transducer_stateless5/streaming_decode.py \
+        --epoch 7 \
+        --avg 1 \
+        --decode-chunk-size 16 \
+        --left-context 64 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless5/exp_L_streaming \
+        --decoding-method greedy_search \
+        --num-decode-streams 2000
+
+(2) modified beam search
+python pruned_transducer_stateless5/streaming_decode.py \
+        --epoch 7 \
+        --avg 1 \
+        --decode-chunk-size 16 \
+        --left-context 64 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless5/exp_L_streaming \
+        --decoding-method modified_beam_search \
+        --num-decode-streams 2000
+
+(3) fast beam search
+python pruned_transducer_stateless5/streaming_decode.py \
+        --epoch 7 \
+        --avg 1 \
+        --decode-chunk-size 16 \
+        --left-context 64 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless5/exp_L_streaming \
+        --decoding-method fast_beam_search \
+        --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num-active-paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        decode_stream.set_features(fbank(samples.to(device)))
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                hyp = decode_streams[i].decoding_result()
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        list(decode_streams[i].ground_truth),
+                        [lexicon.token_table[idx] for idx in hyp],
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            hyp = decode_streams[i].decoding_result()
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    list(decode_streams[i].ground_truth),
+                    [lexicon.token_table[idx] for idx in hyp],
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        # sort results so we can easily compare the difference between two
+        # recognition results
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    dev_cuts = wenetspeech.valid_cuts()
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_cuts = [dev_cuts, test_net_cuts, test_meeting_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..931e699d9
--- /dev/null
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1205 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage for offline ASR:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless5/exp_L_offline \
+  --world-size 8 \
+  --num-epochs 15 \
+  --start-epoch 2 \
+  --max-duration 120 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --average-period 1000 \
+  --training-subset L
+
+Usage for streaming ASR:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless5/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless5/exp_L_streaming \
+  --world-size 8 \
+  --num-epochs 15 \
+  --start-epoch 1 \
+  --max-duration 140 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --average-period 1000 \
+  --training-subset L \
+  --dynamic-chunk-training True \
+  --causal-convolution True \
+  --short-chunk-size 25 \
+  --num-left-chunks 4
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--valid-interval",
+        type=int,
+        default=3000,
+        help="""When training_subset is L, set the valid_interval to 3000.
+        When training_subset is M, set the valid_interval to 1000.
+        When training_subset is S, set the valid_interval to 400.
+        """,
+    )
+
+    parser.add_argument(
+        "--model-warm-step",
+        type=int,
+        default=3000,
+        help="""When training_subset is L, set the model_warm_step to 3000.
+        When training_subset is M, set the model_warm_step to 500.
+        When training_subset is S, set the model_warm_step to 100.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if isinstance(y, list):
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    train_cuts = wenetspeech.train_cuts()
+    valid_cuts = wenetspeech.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 10 seconds
+        #
+        # Caution: There is a reason to select 10.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 10.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = c.supervisions[0].text.replace(" ", "")
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    valid_dl = wenetspeech.valid_dataloaders(valid_cuts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = wenetspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if not params.print_diagnostics and params.start_batch == 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    texts = batch["supervisions"]["text"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y)
+
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/shared b/egs/wenetspeech/ASR/shared
new file mode 120000
index 000000000..e9461a6d7
--- /dev/null
+++ b/egs/wenetspeech/ASR/shared
@@ -0,0 +1 @@
+../../librispeech/ASR/shared
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/__init__.py b/egs/wenetspeech/ASR/zipformer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/wenetspeech/ASR/zipformer/asr_datamodule.py b/egs/wenetspeech/ASR/zipformer/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/beam_search.py b/egs/wenetspeech/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/decode.py b/egs/wenetspeech/ASR/zipformer/decode.py
new file mode 100755
index 000000000..0fbc8244b
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/decode.py
@@ -0,0 +1,818 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao
+#                                                 Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(3) fast beam search (trivial_graph)
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(4) fast beam search (LG)
+./zipformer/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from train import add_model_arguments, get_model, get_params
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest_oracle
+        If you use fast_beam_search_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--ilme-scale",
+        type=float,
+        default=0.2,
+        help="""
+        Used only when --decoding_method is fast_beam_search_LG.
+        It specifies the scale for the internal language model estimation.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search, fast_beam_search_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    x, x_lens = model.encoder_embed(feature, feature_lens)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+    encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_LG":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            blank_penalty=params.blank_penalty,
+            ilme_scale=params.ilme_scale,
+        )
+        for hyp in hyp_tokens:
+            sentence = "".join([lexicon.word_table[i] for i in hyp])
+            hyps.append(list(sentence))
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=graph_compiler.texts_to_ids(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            blank_penalty=params.blank_penalty,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            blank_penalty=params.blank_penalty,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                    blank_penalty=params.blank_penalty,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                    blank_penalty=params.blank_penalty,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    key = f"blank_penalty_{params.blank_penalty}"
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search_" + key: hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key += f"_beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ilme_scale_{params.ilme_scale}"
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}_" + key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list("".join(text.split())) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "modified_beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest_oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"_ilme_scale_{params.ilme_scale}"
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+    params.suffix += f"-blank-penalty-{params.blank_penalty}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
+            )
+        return T > 0
+
+    dev_cuts = wenetspeech.valid_cuts()
+    dev_cuts = dev_cuts.filter(remove_short_utt)
+    dev_dl = wenetspeech.valid_dataloaders(dev_cuts)
+
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_net_cuts = test_net_cuts.filter(remove_short_utt)
+    test_net_dl = wenetspeech.test_dataloaders(test_net_cuts)
+
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+    test_meeting_cuts = test_meeting_cuts.filter(remove_short_utt)
+    test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts)
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_dls = [dev_dl, test_net_dl, test_meeting_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            graph_compiler=graph_compiler,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/zipformer/decode_stream.py b/egs/wenetspeech/ASR/zipformer/decode_stream.py
new file mode 120000
index 000000000..b8d8ddfc4
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decode_stream.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/decoder.py b/egs/wenetspeech/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/encoder_interface.py b/egs/wenetspeech/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/export-onnx-streaming.py b/egs/wenetspeech/ASR/zipformer/export-onnx-streaming.py
new file mode 120000
index 000000000..2962eb784
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/export-onnx.py b/egs/wenetspeech/ASR/zipformer/export-onnx.py
new file mode 120000
index 000000000..70a15683c
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/export.py b/egs/wenetspeech/ASR/zipformer/export.py
new file mode 120000
index 000000000..dfc1bec08
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/jit_pretrained.py b/egs/wenetspeech/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/jit_pretrained_streaming.py b/egs/wenetspeech/ASR/zipformer/jit_pretrained_streaming.py
new file mode 120000
index 000000000..1962351e9
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/joiner.py b/egs/wenetspeech/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/model.py b/egs/wenetspeech/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/onnx_check.py b/egs/wenetspeech/ASR/zipformer/onnx_check.py
new file mode 120000
index 000000000..f3dd42004
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_check.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/onnx_decode.py b/egs/wenetspeech/ASR/zipformer/onnx_decode.py
new file mode 100755
index 000000000..ed5f6db08
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/onnx_decode.py
@@ -0,0 +1,334 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads ONNX exported models and uses them to decode the test sets.
+
+We use the pre-trained model from
+https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/wenetspeech/ASR
+
+repo_url=https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "data/lang_char/tokens.txt"
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-9999.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --epoch 9999 \
+  --avg 1 \
+  --exp-dir $repo/exp/
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-9999-avg-1.onnx
+  - decoder-epoch-9999-avg-1.onnx
+  - joiner-epoch-9999-avg-1.onnx
+
+2. Run this file
+
+./zipformer/onnx_decode.py \
+  --exp-dir ./zipformer/exp \
+  --max-duration 600 \
+  --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \
+"""
+
+
+import argparse
+import logging
+import time
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from lhotse.cut import Cut
+from onnx_pretrained import OnnxModel, greedy_search
+
+from icefall.utils import setup_logger, store_transcripts, write_error_stats
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_char/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="Valid values are greedy_search and modified_beam_search",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    model: OnnxModel, token_table: k2.SymbolTable, batch: dict
+) -> List[List[str]]:
+    """Decode one batch and return the result.
+    Currently it only greedy_search is supported.
+
+    Args:
+      model:
+        The neural model.
+      token_table:
+        Mapping ids to tokens.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+
+    Returns:
+      Return the decoded results for each utterance.
+    """
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(dtype=torch.int64)
+
+    encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens)
+
+    hyps = greedy_search(
+        model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens
+    )
+
+    hyps = [[token_table[h] for h in hyp] for hyp in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+) -> Tuple[List[Tuple[str, List[str], List[str]]], float]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+      token_table:
+        Mapping ids to tokens.
+
+    Returns:
+      - A list of tuples. Each tuple contains three elements:
+         - cut_id,
+         - reference transcript,
+         - predicted result.
+      - The total duration (in seconds) of the dataset.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 10
+    total_duration = 0
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]])
+
+        hyps = decode_one_batch(model=model, token_table=token_table, batch=batch)
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = list(ref_text)
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+
+    return results, total_duration
+
+
+def save_results(
+    res_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+):
+    recog_path = res_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = res_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    errs_info = res_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("WER", file=f)
+        print(wer, file=f)
+
+    s = "\nFor {}, WER is {}:\n".format(test_set_name, wer)
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    assert (
+        args.decoding_method == "greedy_search"
+    ), "Only supports greedy_search currently."
+    res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}"
+
+    setup_logger(f"{res_dir}/log-decode")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    logging.info(f"Device: {device}")
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+    assert token_table[0] == "<blk>"
+
+    logging.info(vars(args))
+
+    logging.info("About to create model")
+    model = OnnxModel(
+        encoder_model_filename=args.encoder_model_filename,
+        decoder_model_filename=args.decoder_model_filename,
+        joiner_model_filename=args.joiner_model_filename,
+    )
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
+            )
+        return T > 0
+
+    dev_cuts = wenetspeech.valid_cuts()
+    dev_cuts = dev_cuts.filter(remove_short_utt)
+    dev_dl = wenetspeech.valid_dataloaders(dev_cuts)
+
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_net_cuts = test_net_cuts.filter(remove_short_utt)
+    test_net_dl = wenetspeech.test_dataloaders(test_net_cuts)
+
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+    test_meeting_cuts = test_meeting_cuts.filter(remove_short_utt)
+    test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts)
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_dl = [dev_dl, test_net_dl, test_meeting_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        start_time = time.time()
+        results, total_duration = decode_dataset(
+            dl=test_dl, model=model, token_table=token_table
+        )
+        end_time = time.time()
+        elapsed_seconds = end_time - start_time
+        rtf = elapsed_seconds / total_duration
+
+        logging.info(f"Elapsed time: {elapsed_seconds:.3f} s")
+        logging.info(f"Wave duration: {total_duration:.3f} s")
+        logging.info(
+            f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}"
+        )
+
+        save_results(res_dir=res_dir, test_set_name=test_set, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/zipformer/onnx_pretrained-streaming.py b/egs/wenetspeech/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 120000
index 000000000..cfea104c2
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/onnx_pretrained.py b/egs/wenetspeech/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/optim.py b/egs/wenetspeech/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/pretrained.py b/egs/wenetspeech/ASR/zipformer/pretrained.py
new file mode 120000
index 000000000..0bd71dde4
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/scaling.py b/egs/wenetspeech/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/scaling_converter.py b/egs/wenetspeech/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/streaming_beam_search.py b/egs/wenetspeech/ASR/zipformer/streaming_beam_search.py
new file mode 120000
index 000000000..b1ed54557
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/streaming_decode.py b/egs/wenetspeech/ASR/zipformer/streaming_decode.py
new file mode 100755
index 000000000..94c5fae5f
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/streaming_decode.py
@@ -0,0 +1,881 @@
+#!/usr/bin/env python3
+# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
+#                                                  Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+from asr_datamodule import WenetSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch import Tensor, nn
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="Path to the lang dir(containing lexicon, tokens, etc.)",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_init_states(
+    model: nn.Module,
+    batch_size: int = 1,
+    device: torch.device = torch.device("cpu"),
+) -> List[torch.Tensor]:
+    """
+    Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+    is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+    states[-2] is the cached left padding for ConvNeXt module,
+    of shape (batch_size, num_channels, left_pad, num_freqs)
+    states[-1] is processed_lens of shape (batch,), which records the number
+    of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+    """
+    states = model.encoder.get_init_states(batch_size, device)
+
+    embed_states = model.encoder_embed.get_init_states(batch_size, device)
+    states.append(embed_states)
+
+    processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+    states.append(processed_lens)
+
+    return states
+
+
+def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance. For element-n,
+        state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
+        state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
+        cached_val2, cached_conv1, cached_conv2).
+        state_list[n][-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+        state_list[n][-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+    """
+    batch_size = len(state_list)
+    assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
+    tot_num_layers = (len(state_list[0]) - 2) // 6
+
+    batch_states = []
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key = torch.cat(
+            [state_list[i][layer_offset] for i in range(batch_size)], dim=1
+        )
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn = torch.cat(
+            [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1 = torch.cat(
+            [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2 = torch.cat(
+            [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1 = torch.cat(
+            [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2 = torch.cat(
+            [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
+        )
+        batch_states += [
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ]
+
+    cached_embed_left_pad = torch.cat(
+        [state_list[i][-2] for i in range(batch_size)], dim=0
+    )
+    batch_states.append(cached_embed_left_pad)
+
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
+    batch_states.append(processed_lens)
+
+    return batch_states
+
+
+def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+        batch_states: A list of cached tensors of all encoder layers. For layer-i,
+          states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+          cached_conv1, cached_conv2).
+          state_list[-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+          states[-1] is processed_lens of shape (batch,), which records the number
+          of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Returns:
+        state_list: A list of list. Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+    """
+    assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
+    tot_num_layers = (len(batch_states) - 2) // 6
+
+    processed_lens = batch_states[-1]
+    batch_size = processed_lens.shape[0]
+
+    state_list = [[] for _ in range(batch_size)]
+
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1_list = batch_states[layer_offset + 2].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2_list = batch_states[layer_offset + 3].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1_list = batch_states[layer_offset + 4].chunk(
+            chunks=batch_size, dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2_list = batch_states[layer_offset + 5].chunk(
+            chunks=batch_size, dim=0
+        )
+        for i in range(batch_size):
+            state_list[i] += [
+                cached_key_list[i],
+                cached_nonlin_attn_list[i],
+                cached_val1_list[i],
+                cached_val2_list[i],
+                cached_conv1_list[i],
+                cached_conv2_list[i],
+            ]
+
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(cached_embed_left_pad_list[i])
+
+    processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(processed_lens_list[i])
+
+    return state_list
+
+
+def streaming_forward(
+    features: Tensor,
+    feature_lens: Tensor,
+    model: nn.Module,
+    states: List[Tensor],
+    chunk_size: int,
+    left_context_len: int,
+) -> Tuple[Tensor, Tensor, List[Tensor]]:
+    """
+    Returns encoder outputs, output lengths, and updated states.
+    """
+    cached_embed_left_pad = states[-2]
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        cached_left_pad=cached_embed_left_pad,
+    )
+    assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    # processed_mask is used to mask out initial states
+    processed_mask = torch.arange(left_context_len, device=x.device).expand(
+        x.size(0), left_context_len
+    )
+    processed_lens = states[-1]  # (batch,)
+    # (batch, left_context_size)
+    processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+    # Update processed lengths
+    new_processed_lens = processed_lens + x_lens
+
+    # (batch, left_context_size + chunk_size)
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    encoder_states = states[:-2]
+    (
+        encoder_out,
+        encoder_out_lens,
+        new_encoder_states,
+    ) = model.encoder.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=encoder_states,
+        src_key_padding_mask=src_key_padding_mask,
+    )
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    new_states = new_encoder_states + [
+        new_cached_embed_left_pad,
+        new_processed_lens,
+    ]
+    return encoder_out, encoder_out_lens, new_states
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+    chunk_size = int(params.chunk_size)
+    left_context_len = int(params.left_context_frames)
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []  # Used in fast-beam-search
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(chunk_size * 2)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # Make sure the length after encoder_embed is at least 1.
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    tail_length = chunk_size * 2 + 7 + 2 * 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+
+    encoder_out, encoder_out_lens, new_states = streaming_forward(
+        features=features,
+        feature_lens=feature_lens,
+        model=model,
+        states=states,
+        chunk_size=chunk_size,
+        left_context_len=left_context_len,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(
+            model=model,
+            encoder_out=encoder_out,
+            streams=decode_streams,
+            blank_penalty=params.blank_penalty,
+        )
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = torch.tensor(processed_lens, device=device)
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+            blank_penalty=params.blank_penalty,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+            blank_penalty=params.blank_penalty,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      lexicon:
+        The Lexicon.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = get_init_states(model=model, batch_size=1, device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        if audio.max() > 1:
+            logging.warning(
+                f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
+                f"Clipping to [-1, 1]."
+            )
+            audio = np.clip(audio, -1, 1)
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=30)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        list(decode_streams[i].ground_truth.strip()),
+                        [
+                            lexicon.token_table[idx]
+                            for idx in decode_streams[i].decoding_result()
+                        ],
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    [
+                        lexicon.token_table[idx]
+                        for idx in decode_streams[i].decoding_result()
+                    ],
+                )
+            )
+            del decode_streams[i]
+
+    key = f"blank_penalty_{params.blank_penalty}"
+    if params.decoding_method == "greedy_search":
+        key = f"greedy_search_{key}"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}_{key}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}_{key}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    assert params.causal, params.causal
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
+    assert (
+        "," not in params.left_context_frames
+    ), "left_context_frames should be one value in decoding."
+    params.suffix += f"-chunk-{params.chunk_size}"
+    params.suffix += f"-left-context-{params.left_context_frames}"
+    params.suffix += f"-blank-penalty-{params.blank_penalty}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    dev_cuts = wenetspeech.valid_cuts()
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_meeting_cuts = wenetspeech.test_meeting_cuts()
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_cuts = [dev_cuts, test_net_cuts, test_meeting_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/zipformer/subsampling.py b/egs/wenetspeech/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/wenetspeech/ASR/zipformer/train.py b/egs/wenetspeech/ASR/zipformer/train.py
new file mode 100755
index 000000000..b1557dedb
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/train.py
@@ -0,0 +1,1350 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 12 \
+  --start-epoch 1 \
+  --exp-dir zipformer/exp \
+  --training-subset L
+  --lr-epochs 1.5 \
+  --max-duration 350
+
+# For mix precision training:
+
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 12 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --training-subset L \
+  --lr-epochs 1.5 \
+  --max-duration 750
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="""Feedforward dimension of the zipformer encoder layers, per stack, comma separated.""",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="""Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="""Embedding dimension in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="""Query/key dimension per head in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="""Value dimension per head in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="""Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="""Unmasked dimensions in the encoders, relates to augmentation during training. A single int or comma-separated list.  Must be <= each corresponding encoder_dim.""",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="""Sizes of convolutional kernels in convolution modules in each encoder stack: a single int or comma-separated list.""",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="""Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. Must be just -1 if --causal=False""",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="""Maximum left-contexts for causal training, measured in frames which will
+        be converted to a number of chunks.  If splitting into chunks,
+        chunk left-context frames will be chosen randomly from this list; else not relevant.""",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="""Reference batch duration for purposes of adjusting batch counts for setting various schedules inside the model""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="""The context size in the decoder. 1 means bigram; 2 means tri-gram""",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="""The prune range for rnnt loss, it means how many symbols(context)
+        we are using to compute the loss""",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="""The scale to smooth the loss with lm
+        (output of prediction network) part.""",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="""The scale to smooth the loss with am (output of encoder network) part.""",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="""To get pruning ranges, we will calculate a simple version
+        loss(joiner is just addition), this simple loss also uses for
+        training (as a regularization item). We will scale the simple loss
+        with this parameter before adding to the final loss.""",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, _ = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    train_cuts = wenetspeech.train_cuts()
+    valid_cuts = wenetspeech.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 15 seconds
+        #
+        # Caution: There is a reason to select 15.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 15.0:
+            # logging.warning(
+            #    f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0]
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = wenetspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_dl = wenetspeech.valid_dataloaders(valid_cuts)
+
+    if False and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      graph_compiler:
+        The compiler to encode texts to ids.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    texts = supervisions["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/wenetspeech/ASR/zipformer/zipformer.py b/egs/wenetspeech/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/wenetspeech/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/README.md b/egs/xbmu_amdo31/ASR/README.md
new file mode 100644
index 000000000..0a441d070
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/README.md
@@ -0,0 +1,16 @@
+# Introduction
+About the XBMU-AMDO31 corpus
+XBMU-AMDO31 is an open-source Amdo Tibetan speech corpus published by Northwest Minzu University.
+publicly available on https://huggingface.co/datasets/syzym/xbmu_amdo31
+
+XBMU-AMDO31 dataset is a speech recognition corpus of Amdo Tibetan dialect. 
+The open source corpus contains 31 hours of speech data and resources related 
+to build speech recognition systems,including transcribed texts and a Tibetan 
+pronunciation lexicon.
+(The lexicon is a Tibetan lexicon of the Lhasa dialect, which has been reused 
+for the Amdo dialect because of the uniformity of the Tibetan language)
+The dataset can be used to train a model for Amdo Tibetan Automatic Speech Recognition (ASR). 
+
+This recipe includes some different ASR models trained with XBMU-AMDO31.
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/RESULTS.md b/egs/xbmu_amdo31/ASR/RESULTS.md
new file mode 100644
index 000000000..1bd9b2e2b
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/RESULTS.md
@@ -0,0 +1,92 @@
+## Results
+
+### XBMU-AMDO31 BPE training result (Stateless Transducer)
+
+#### Pruned transducer stateless 5
+
+[./pruned_transducer_stateless5](./pruned_transducer_stateless5)
+
+It uses pruned RNN-T.
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/syzym/icefall-asr-xbmu-amdo31-pruned-transducer-stateless5-2022-11-29>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 87801200, i.e., 87.8 M
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 11.06| 11.73| --epoch 28 --avg 23 --max-duration 600|
+| beam search            | 10.64| 11.42| --epoch 28 --avg 23 --max-duration 600|
+| modified beam search   | 10.57| 11.24| --epoch 28 --avg 23 --max-duration 600|
+
+
+Training command is:
+
+```bash
+cd egs/xbmu_amdo31/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0"
+
+./pruned_transducer_stateless5/train.py
+```
+
+**Caution**: It uses `--context-size=1`.
+
+
+The decoding command is:
+```bash
+for method in greedy_search beam_search modified_beam_search;
+do
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 23 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method $method
+done
+```
+
+### pruned_transducer_stateless7 (zipformer)
+
+See <https://github.com/k2-fsa/icefall/pull/672> for more details.
+
+[pruned_transducer_stateless7](./pruned_transducer_stateless7)
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/syzym/icefall-asr-xbmu-amdo31-pruned-transducer-stateless7-2022-12-02>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+Number of model parameters: 70369391, i.e., 70.37 M
+
+|                      | test | dev  | comment                                |
+|----------------------|------|------|----------------------------------------|
+| greedy search        | 10.06| 10.59| --epoch 23 --avg 11 --max-duration 600 |
+| beam search          | 9.77 | 10.11| --epoch 23 --avg 11 --max-duration 600 |
+| modified beam search | 9.7  | 10.12| --epoch 23 --avg 11 --max-duration 600 |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+
+./pruned_transducer_stateless7/train.py
+```
+
+The decoding commands are:
+```bash
+for m in greedy_search beam_search modified_beam_search; do
+  for epoch in 23; do
+    for avg in 11; do
+      ./pruned_transducer_stateless7/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --exp-dir ./pruned_transducer_stateless7/exp \
+          --max-duration 600 \
+          --decoding-method $m
+    done
+  done
+done
+```
diff --git a/egs/xbmu_amdo31/ASR/local/compile_hlg.py b/egs/xbmu_amdo31/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/compile_lg.py b/egs/xbmu_amdo31/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/compute_fbank_musan.py b/egs/xbmu_amdo31/ASR/local/compute_fbank_musan.py
new file mode 120000
index 000000000..5833f2484
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/compute_fbank_musan.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/compute_fbank_xbmu_amdo31.py b/egs/xbmu_amdo31/ASR/local/compute_fbank_xbmu_amdo31.py
new file mode 100755
index 000000000..a593e7be3
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/compute_fbank_xbmu_amdo31.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the XBMU-AMDO31 dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+    return parser.parse_args()
+
+
+def compute_fbank_xbmu_amdo31(bpe_model: Optional[str] = None):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+    prefix = "xbmu_amdo31"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if bpe_model:
+                cut_set = filter_cuts(cut_set, sp)
+
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_xbmu_amdo31(bpe_model=args.bpe_model)
diff --git a/egs/xbmu_amdo31/ASR/local/convert_transcript_words_to_tokens.py b/egs/xbmu_amdo31/ASR/local/convert_transcript_words_to_tokens.py
new file mode 120000
index 000000000..2ce13fd69
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/convert_transcript_words_to_tokens.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/filter_cuts.py b/egs/xbmu_amdo31/ASR/local/filter_cuts.py
new file mode 120000
index 000000000..27aca1729
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/filter_cuts.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/filter_cuts.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/generate_unique_lexicon.py b/egs/xbmu_amdo31/ASR/local/generate_unique_lexicon.py
new file mode 120000
index 000000000..c0aea1403
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/generate_unique_lexicon.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/prepare_lang.py b/egs/xbmu_amdo31/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/prepare_lang_bpe.py b/egs/xbmu_amdo31/ASR/local/prepare_lang_bpe.py
new file mode 120000
index 000000000..36b40e7fc
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/prepare_lm_training_data.py b/egs/xbmu_amdo31/ASR/local/prepare_lm_training_data.py
new file mode 120000
index 000000000..abc00d421
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/prepare_lm_training_data.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/sort_lm_training_data.py b/egs/xbmu_amdo31/ASR/local/sort_lm_training_data.py
new file mode 120000
index 000000000..1d6ccbe33
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/sort_lm_training_data.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/sort_lm_training_data.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/train_bpe_model.py b/egs/xbmu_amdo31/ASR/local/train_bpe_model.py
new file mode 120000
index 000000000..6fad36421
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/local/validate_bpe_lexicon.py b/egs/xbmu_amdo31/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/prepare.sh b/egs/xbmu_amdo31/ASR/prepare.sh
new file mode 100755
index 000000000..21836840c
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/prepare.sh
@@ -0,0 +1,357 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/xbmu_amdo31
+#      You can find data, resource, etc, inside it.
+#      You can download them from https://huggingface.co/datasets/syzym/xbmu_amdo31
+#
+#  - $dl_dir/lm
+#      This directory contains the following files downloaded from
+#       git lfs install
+#       https://huggingface.co/syzym/xbmu_amdo31_lm
+#
+#        - tibetan.3-gram.arpa
+#        - tibetan.4-gram.arpa
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "stage -1: Download LM"
+  # We assume that you have installed the git-lfs, if not, you could install it
+  # using: `sudo apt-get install git-lfs && git-lfs install`
+  git lfs 1>/dev/null 2>&1 || (echo "please install git-lfs, consider using: sudo apt-get install git-lfs && git-lfs install" && exit 1)
+
+  if [ ! -f $dl_dir/lm/3-gram.unpruned.arpa ]; then
+    git clone https://huggingface.co/syzym/xbmu_amdo31_lm $dl_dir/lm
+    pushd $dl_dir/lm
+    git lfs pull --include "tibetan.3-gram.arpa"
+    git lfs pull --include "tibetan.4-gram.arpa"
+    popd
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/xbmu_amdo31,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/xbmu_amdo31 $dl_dir/xbmu_amdo31
+  #
+  
+  if [ ! -f $dl_dir/xbmu_amdo31 ]; then
+    git lfs 1>/dev/null 2>&1 || (echo "please install git-lfs, consider using: sudo apt-get install git-lfs && git-lfs install" && exit 1)
+    lhotse download xbmu-amdo31 $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare xbmu_amdo31 manifest"
+  # We assume that you have downloaded the xbmu_amdo31 corpus
+  # to $dl_dir/xbmu_amdo31
+  if [ ! -f data/manifests/.xbmu_amdo31_manifests.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare xbmu-amdo31 $dl_dir/xbmu_amdo31 data/manifests
+    touch data/manifests/.xbmu_amdo31_manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for xbmu_amdo31"
+  if [ ! -f data/fbank/.xbmu_amdo31.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_xbmu_amdo31.py
+    touch data/fbank/.xbmu_amdo31.done
+  fi
+fi
+
+
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/xbmu_amdo31/resource/lexicon.txt |
+    sort | uniq > $lang_dir/lexicon.txt
+
+  ./local/generate_unique_lexicon.py --lang-dir $lang_dir
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+fi
+
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang_phone/words.txt $lang_dir
+
+  if [ ! -f $lang_dir/transcript_words.txt ]; then
+    log "Generate data to train phone based bigram P"
+    xbmu_amdo31_text=$dl_dir/xbmu_amdo31/data/transcript/transcript_clean.txt
+    xbmu_amdo31_train_uid=$dl_dir/xbmu_amdo31/data/transcript/xbmu_amdo31_train_uid
+    find $dl_dir/xbmu_amdo31/data/wav/train -name "*.wav" | sed 's/\.wav//g' | awk -F '-' '{print $NF}' > $xbmu_amdo31_train_uid
+    awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $xbmu_amdo31_train_uid $xbmu_amdo31_text |
+	    cut -d " " -f 2- > $lang_dir/transcript_words.txt
+  fi
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare bigram P"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+      ./local/convert_transcript_words_to_tokens.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --transcript $lang_dir/transcript_words.txt \
+        --oov "<UNK>" \
+        > $lang_dir/transcript_tokens.txt
+    fi
+
+    if [ ! -f $lang_dir/P.arpa ]; then
+      ./shared/make_kn_lm.py \
+        -ngram-order 2 \
+        -text $lang_dir/transcript_tokens.txt \
+        -lm $lang_dir/P.arpa
+    fi
+
+    if [ ! -f $lang_dir/P.fst.txt ]; then
+      python3 -m kaldilm \
+        --read-symbol-table="$lang_dir/tokens.txt" \
+        --disambig-symbol='#0' \
+        --max-order=2 \
+        $lang_dir/P.arpa > $lang_dir/P.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $dl_dir/lm/tibetan.3-gram.arpa > data/lm/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      $dl_dir/lm/tibetan.4-gram.arpa > data/lm/G_4_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+  done
+fi
+
+# Compile LG for RNN-T fast_beam_search decoding
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Compile LG"
+  ./local/compile_lg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_lg.py --lang-dir $lang_dir
+  done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Generate LM training data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    lang_dir=data/lang_bpe_${vocab_size}
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/lm/lm_train.txt \
+      --lm-archive $out_dir/lm_data.pt
+  done
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Generate LM validation data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/valid.txt ]; then
+      files=$dl_dir/xbmu_amdo31/data/transcript/dev_text
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > $out_dir/valid.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/valid.txt \
+      --lm-archive $out_dir/lm_data-valid.pt
+  done
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: Generate LM test data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/test.txt ]; then
+        files=$dl_dir/xbmu_amdo31/data/transcript/test_text
+        cat $f | cut -d " " -f 2- > $out_dir/test.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/test.txt \
+      --lm-archive $out_dir/lm_data-test.pt
+  done
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+  log "Stage 14: Sort LM training data"
+  # Sort LM training data by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of BPE tokens
+  # in a sentence.
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data.pt \
+      --out-lm-data $out_dir/sorted_lm_data.pt \
+      --out-statistics $out_dir/statistics.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-valid.pt \
+      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+      --out-statistics $out_dir/statistics-valid.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-test.pt \
+      --out-lm-data $out_dir/sorted_lm_data-test.pt \
+      --out-statistics $out_dir/statistics-test.txt
+  done
+fi
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/__init__.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py
new file mode 100644
index 000000000..7594fb28e
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -0,0 +1,408 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+# Copyright      2022  Northwest Minzu University     (Author: Senyan Li)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import CutConcatenate  # noqa F401 for PrecomputedFeatures
+from lhotse.dataset import (
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import AudioSamples  # noqa F401 For AudioSamples
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class Xbmu_AmdoAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        f = self.args.manifest_dir / "xbmu_amdo31_cuts_train.jsonl.gz"
+        logging.info(f"About to get train cuts from {f}")
+        cuts_train = load_manifest_lazy(f)
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        f = self.args.manifest_dir / "xbmu_amdo31_cuts_dev.jsonl.gz"
+        logging.info(f"About to get valid cuts from {f}")
+        cuts_valid = load_manifest_lazy(f)
+        return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        f = self.args.manifest_dir / "xbmu_amdo31_cuts_test.jsonl.gz"
+        logging.info(f"About to get test cuts from {f}")
+        cuts_test = load_manifest_lazy(f)
+        return cuts_test
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/beam_search.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/conformer.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/conformer.py
new file mode 120000
index 000000000..c7c1a4b6e
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decode.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decode.py
new file mode 100755
index 000000000..b77f734e3
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decode.py
@@ -0,0 +1,964 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+(2) beam search (not recommended)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+(4) fast beam search (one best)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+(5) fast beam search (nbest)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+(7) fast beam search (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search with RNNLM shallow fusion (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 4 \
+    --max-contexts 4 \
+    --rnn-lm-scale 0.4 \
+    --rnn-lm-exp-dir /path/to/RNNLM/exp \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import Xbmu_AmdoAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_rnnlm_shallow_fusion,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_rnnlm_shallow_fusion # for rnn lm shallow fusion
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-scale",
+        type=float,
+        default=0.0,
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    rnnlm: Optional[RnnLmModel] = None,
+    rnnlm_scale: float = 1.0,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_rnnlm_shallow_fusion":
+        hyp_tokens = modified_beam_search_rnnlm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            sp=sp,
+            rnnlm=rnnlm,
+            rnnlm_scale=rnnlm_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    rnnlm: Optional[RnnLmModel] = None,
+    rnnlm_scale: float = 1.0,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        logging.info(f"Decoding {batch_idx}-th batch")
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            rnnlm=rnnlm,
+            rnnlm_scale=rnnlm_scale,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Xbmu_AmdoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_rnnlm_shallow_fusion",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    params.suffix += f"-rnnlm-lm-scale-{params.rnn_lm_scale}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    rnn_lm_model = None
+    rnn_lm_scale = params.rnn_lm_scale
+    if params.decoding_method == "modified_beam_search_rnnlm_shallow_fusion":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.vocab_size,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        assert params.rnn_lm_avg == 1
+
+        load_checkpoint(
+            f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+            rnn_lm_model,
+        )
+        rnn_lm_model.to(device)
+        rnn_lm_model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    xbmu_amdo = Xbmu_AmdoAsrDataModule(args)
+
+    test_cuts = xbmu_amdo.test_cuts()
+
+    test_dl = xbmu_amdo.test_dataloaders(test_cuts)
+
+    test_sets = ["test"]
+    test_dl = [test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            rnnlm=rnn_lm_model,
+            rnnlm_scale=rnn_lm_scale,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decode_stream.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decode_stream.py
new file mode 120000
index 000000000..d59ef95f7
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decode_stream.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decoder.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/export.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/export.py
new file mode 100755
index 000000000..54f656859
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/export.py
@@ -0,0 +1,287 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        convert_scaled_to_non_scaled(model, inplace=True)
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/joiner.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/lstmp.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/lstmp.py
new file mode 120000
index 000000000..b82e115fc
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/lstmp.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/model.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/model.py
new file mode 120000
index 000000000..a99e74334
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/optim.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/pretrained.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/pretrained.py
new file mode 100755
index 000000000..74a2210c3
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/pretrained.py
@@ -0,0 +1,344 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/scaling.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/scaling_converter.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/scaling_converter.py
new file mode 120000
index 000000000..db93d155b
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/streaming_beam_search.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/streaming_beam_search.py
new file mode 120000
index 000000000..1199a61d6
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/streaming_decode.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/streaming_decode.py
new file mode 120000
index 000000000..f29284163
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/streaming_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/streaming_decode.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/test_model.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/test_model.py
new file mode 100755
index 000000000..9aad32014
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/test_model.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless4/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 18
+    params.dim_feedforward = 1024
+    params.encoder_dim = 256
+    params.nhead = 4
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py
new file mode 100755
index 000000000..a6fa46b17
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py
@@ -0,0 +1,1187 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless5/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless5/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import Xbmu_AmdoAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    display_and_save_batch,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=24,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=1536,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=384,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--dynamic-chunk-training",
+        type=str2bool,
+        default=False,
+        help="""Whether to use dynamic_chunk_training, if you want a streaming
+        model, this requires to be True.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal-convolution",
+        type=str2bool,
+        default=False,
+        help="""Whether to use causal convolution, this requires to be True when
+        using dynamic_chunk_training.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=25,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--delay-penalty",
+        type=float,
+        default=0.0,
+        help="""A constant value used to penalize symbol delay,
+        to encourage streaming models to emit symbols earlier.
+        See https://github.com/k2-fsa/k2/issues/955 and
+        https://arxiv.org/pdf/2211.00490.pdf for more details.""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        dynamic_chunk_training=params.dynamic_chunk_training,
+        short_chunk_size=params.short_chunk_size,
+        num_left_chunks=params.num_left_chunks,
+        causal=params.causal_convolution,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+            reduction="none",
+            delay_penalty=params.delay_penalty if warmup >= 2.0 else 0,
+        )
+        simple_loss_is_finite = torch.isfinite(simple_loss)
+        pruned_loss_is_finite = torch.isfinite(pruned_loss)
+        is_finite = simple_loss_is_finite & pruned_loss_is_finite
+        if not torch.all(is_finite):
+            logging.info(
+                "Not all losses are finite!\n"
+                f"simple_loss: {simple_loss}\n"
+                f"pruned_loss: {pruned_loss}"
+            )
+            display_and_save_batch(batch, params=params, sp=sp)
+            simple_loss = simple_loss[simple_loss_is_finite]
+            pruned_loss = pruned_loss[pruned_loss_is_finite]
+
+            # If the batch contains more than 10 utterances AND
+            # if either all simple_loss or pruned_loss is inf or nan,
+            # we stop the training process by raising an exception
+            if torch.all(~simple_loss_is_finite) or torch.all(~pruned_loss_is_finite):
+                raise ValueError(
+                    "There are too many utterances in this batch "
+                    "leading to inf or nan losses."
+                )
+
+        simple_loss = simple_loss.sum()
+        pruned_loss = pruned_loss.sum()
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0 if warmup < 1.0 else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        # info["frames"] is an approximate number for two reasons:
+        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
+        # (2) If some utterances in the batch lead to inf/nan loss, they
+        #     are filtered out.
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.dynamic_chunk_training:
+        assert (
+            params.causal_convolution
+        ), "dynamic_chunk_training requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    xbmu_amdo = Xbmu_AmdoAsrDataModule(args)
+
+    train_cuts = xbmu_amdo.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./conformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 1) // 2 - 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = xbmu_amdo.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = xbmu_amdo.valid_cuts()
+    valid_dl = xbmu_amdo.valid_dataloaders(valid_cuts)
+
+    if params.start_batch <= 0 and not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+    warmup: float,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                    warmup=warmup,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+
+def main():
+    parser = get_parser()
+    Xbmu_AmdoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/__init__.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/asr_datamodule.py
new file mode 120000
index 000000000..c473a600a
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless5/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/beam_search.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/decode.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/decode.py
new file mode 100755
index 000000000..e334e690a
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/decode.py
@@ -0,0 +1,837 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import Xbmu_AmdoAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    Xbmu_AmdoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    xbmu_amdo = Xbmu_AmdoAsrDataModule(args)
+
+    test_cuts = xbmu_amdo.test_cuts()
+
+    test_dl = xbmu_amdo.test_dataloaders(test_cuts)
+
+    test_sets = [
+        "test",
+    ]
+    test_dl = [
+        test_dl,
+    ]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/decoder.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/decoder.py
new file mode 120000
index 000000000..8283d8c5a
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/encoder_interface.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/export.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/export.py
new file mode 120000
index 000000000..2713792e6
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/export.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/jit_pretrained.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/jit_pretrained.py
new file mode 120000
index 000000000..a44034e34
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/joiner.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/joiner.py
new file mode 120000
index 000000000..0f0c3c90a
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/model.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/model.py
new file mode 120000
index 000000000..0d8bc665b
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/optim.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/optim.py
new file mode 120000
index 000000000..8a05abb5f
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/pretrained.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/pretrained.py
new file mode 100755
index 000000000..d05bafcfb
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/pretrained.py
@@ -0,0 +1,355 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./pruned_transducer_stateless7/export.py \
+  --exp-dir ./pruned_transducer_stateless7/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) greedy search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless7/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless7/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless7/exp/pretrained.pt is generated by
+./pruned_transducer_stateless7/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/scaling.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/scaling.py
new file mode 120000
index 000000000..5f9be9fe0
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/scaling_converter.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/scaling_converter.py
new file mode 120000
index 000000000..f9960e5c6
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/test_model.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/test_model.py
new file mode 120000
index 000000000..7ceac5d10
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/test_model.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py
new file mode 100755
index 000000000..8c53972fd
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py
@@ -0,0 +1,1213 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7/exp \
+  --full-libri 1 \
+  --max-duration 550
+
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import Xbmu_AmdoAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=2.0)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    xbmu_amdo = Xbmu_AmdoAsrDataModule(args)
+
+    train_cuts = xbmu_amdo.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = xbmu_amdo.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = xbmu_amdo.valid_cuts()
+    valid_dl = xbmu_amdo.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    Xbmu_AmdoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/zipformer.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/zipformer.py
new file mode 120000
index 000000000..f2f66041e
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7/zipformer.py
\ No newline at end of file
diff --git a/egs/xbmu_amdo31/ASR/shared b/egs/xbmu_amdo31/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/xbmu_amdo31/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/yesno/ASR/README.md b/egs/yesno/ASR/README.md
new file mode 100644
index 000000000..38b491fc6
--- /dev/null
+++ b/egs/yesno/ASR/README.md
@@ -0,0 +1,14 @@
+## Yesno recipe
+
+This is the simplest ASR recipe in `icefall`.
+
+It can be run on CPU and takes less than 30 seconds to
+get the following WER:
+
+```
+[test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
+```
+
+Please refer to
+<https://icefall.readthedocs.io/en/latest/recipes/Non-streaming-ASR/yesno/index.html>
+for detailed instructions.
diff --git a/egs/yesno/ASR/local/compile_hlg.py b/egs/yesno/ASR/local/compile_hlg.py
new file mode 100755
index 000000000..e0a94bf08
--- /dev/null
+++ b/egs/yesno/ASR/local/compile_hlg.py
@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str) -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    logging.info("Loading G.fst.txt")
+    with open("data/lm/G.fst.txt") as f:
+        G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    # LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # see https://github.com/k2-fsa/k2/pull/1140
+    labels = LG.labels
+    labels[labels >= first_token_disambig_id] = 0
+    LG.labels = labels
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lang_dir)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/yesno/ASR/local/compute_fbank_yesno.py b/egs/yesno/ASR/local/compute_fbank_yesno.py
new file mode 100755
index 000000000..75d95df68
--- /dev/null
+++ b/egs/yesno/ASR/local/compute_fbank_yesno.py
@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+
+"""
+This file computes fbank features of the yesno dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or it wastes a
+# lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_yesno():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+
+    # This dataset is rather small, so we use only one job
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 23
+
+    dataset_parts = (
+        "train",
+        "test",
+    )
+    prefix = "yesno"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(sampling_rate=8000, num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_file = output_dir / f"{prefix}_cuts_{partition}.{suffix}"
+            if cuts_file.is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 1,  # use one job
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(cuts_file)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_yesno()
diff --git a/egs/yesno/ASR/local/prepare_lang.py b/egs/yesno/ASR/local/prepare_lang.py
new file mode 100755
index 000000000..f7fde7796
--- /dev/null
+++ b/egs/yesno/ASR/local/prepare_lang.py
@@ -0,0 +1,367 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    out_dir = Path("data/lang_phone")
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/local/prepare_lang_fst.py b/egs/yesno/ASR/local/prepare_lang_fst.py
new file mode 120000
index 000000000..c5787c534
--- /dev/null
+++ b/egs/yesno/ASR/local/prepare_lang_fst.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_fst.py
\ No newline at end of file
diff --git a/egs/yesno/ASR/prepare.sh b/egs/yesno/ASR/prepare.sh
new file mode 100755
index 000000000..41db0cf7c
--- /dev/null
+++ b/egs/yesno/ASR/prepare.sh
@@ -0,0 +1,97 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+dl_dir=$PWD/download
+
+lang_dir=data/lang_phone
+lm_dir=data/lm
+
+. shared/parse_options.sh || exit 1
+
+mkdir -p $lang_dir
+mkdir -p $lm_dir
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+  mkdir -p $dl_dir
+
+  if [ ! -f $dl_dir/waves_yesno/.completed ]; then
+    lhotse download yesno $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare yesno manifest"
+  mkdir -p data/manifests
+  lhotse prepare yesno $dl_dir/waves_yesno data/manifests
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute fbank for yesno"
+  mkdir -p data/fbank
+  ./local/compute_fbank_yesno.py
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare lang"
+  # NOTE: "<UNK> SIL" is added for implementation convenience
+  # as the graph compiler code requires that there is a OOV word
+  # in the lexicon.
+  (
+    echo "<SIL> SIL"
+    echo "YES Y"
+    echo "NO N"
+    echo "<UNK> SIL"
+  ) > $lang_dir/lexicon.txt
+
+  ./local/prepare_lang.py
+  ./local/prepare_lang_fst.py --lang-dir ./data/lang_phone --has-silence 1
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare G"
+  # We use a unigram G
+  cat <<EOF > $lm_dir/G.arpa
+
+\data\\
+ngram 1=4
+
+\1-grams:
+-1 NO
+-1 YES
+-99 <s>
+-1 </s>
+
+\end\\
+
+EOF
+
+  if [ ! -f $lm_dir/G.fst.txt ]; then
+    python3 -m kaldilm \
+      --read-symbol-table="$lang_dir/words.txt" \
+      --disambig-symbol='#0' \
+      $lm_dir/G.arpa > $lm_dir/G.fst.txt
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compile HLG"
+  if [ ! -f $lang_dir/HLG.pt ]; then
+    ./local/compile_hlg.py --lang-dir $lang_dir
+  fi
+fi
diff --git a/egs/yesno/ASR/shared b/egs/yesno/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/yesno/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/yesno/ASR/tdnn/README.md b/egs/yesno/ASR/tdnn/README.md
new file mode 100644
index 000000000..2b6116f0a
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/README.md
@@ -0,0 +1,8 @@
+
+## How to run this recipe
+
+You can find detailed instructions by visiting
+<https://icefall.readthedocs.io/en/latest/recipes/yesno.html>
+
+It describes how to run this recipe and how to use
+a pre-trained model with `./pretrained.py`.
diff --git a/egs/yesno/ASR/tdnn/asr_datamodule.py b/egs/yesno/ASR/tdnn/asr_datamodule.py
new file mode 100644
index 000000000..dc66b217d
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/asr_datamodule.py
@@ -0,0 +1,257 @@
+# Copyright      2021  Piotr Żelasko
+#                2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import List
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.dataset.datamodule import DataModule
+from icefall.utils import str2bool
+
+
+class YesNoAsrDataModule(DataModule):
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    """
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        super().add_arguments(parser)
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--feature-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=30.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=False,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=10,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=False,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+    def train_dataloaders(self) -> DataLoader:
+        logging.info("About to get train cuts")
+        cuts_train = self.train_cuts()
+
+        logging.info("About to create train dataset")
+        transforms = []
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(
+                    FbankConfig(sampling_rate=8000, num_mel_bins=23)
+                ),
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+        )
+
+        return train_dl
+
+    def test_dataloaders(self) -> DataLoader:
+        logging.info("About to get test cuts")
+        cuts_test = self.test_cuts()
+
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=23)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts_test,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest_lazy(
+            self.args.feature_dir / "yesno_cuts_train.jsonl.gz"
+        )
+        return cuts_train
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        cuts_test = load_manifest_lazy(
+            self.args.feature_dir / "yesno_cuts_test.jsonl.gz"
+        )
+        return cuts_test
diff --git a/egs/yesno/ASR/tdnn/decode.py b/egs/yesno/ASR/tdnn/decode.py
new file mode 100755
index 000000000..f520607af
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/decode.py
@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import YesNoAsrDataModule
+from model import Tdnn
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=14,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=2,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn/exp/"),
+            "lang_dir": Path("data/lang_phone"),
+            "feature_dim": 23,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    word_table: k2.SymbolTable,
+) -> List[List[int]]:
+    """Decode one batch and return the result in a list-of-list.
+    Each sub list contains the word IDs for an utterance in the batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding.
+        - params.method is "nbest", it uses nbest decoding.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+        (https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py)
+      word_table:
+        It is the word symbol table.
+    Returns:
+      Return the decoding result. `len(ans)` == batch size.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    nnet_output = model(feature)
+    # nnet_output is (N, T, C)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    best_path = one_best_decoding(
+        lattice=lattice, use_double_scores=params.use_double_scores
+    )
+    hyps = get_texts(best_path)
+    hyps = [[word_table[i] for i in ids] for ids in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    word_table: k2.SymbolTable,
+) -> List[Tuple[str, List[str], List[str]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      word_table:
+        It is word symbol table.
+    Returns:
+      Return a tuple contains two elements (ref_text, hyp_text):
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            word_table=word_table,
+        )
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    exp_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+) -> None:
+    """Save results to `exp_dir`.
+    Args:
+      exp_dir:
+        The output directory. This function create the following files inside
+        this directory:
+
+            - recogs-{test_set_name}.text
+
+                It contains the reference and hypothesis results, like below::
+
+                    ref=['NO', 'NO', 'NO', 'YES', 'NO', 'NO', 'NO', 'YES']
+                    hyp=['NO', 'NO', 'NO', 'YES', 'NO', 'NO', 'NO', 'YES']
+                    ref=['NO', 'NO', 'YES', 'NO', 'YES', 'NO', 'NO', 'YES']
+                    hyp=['NO', 'NO', 'YES', 'NO', 'YES', 'NO', 'NO', 'YES']
+
+            - errs-{test_set_name}.txt
+
+                It contains the detailed WER.
+      test_set_name:
+        The name of the test set, which will be part of the result filename.
+      results:
+        A list of tuples, each of which contains (ref_words, hyp_words).
+    Returns:
+      Return None.
+    """
+    recog_path = exp_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = exp_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        write_error_stats(f, f"{test_set_name}", results)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    YesNoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    params["env_info"] = get_env_info()
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu"))
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    model = Tdnn(
+        num_features=params.feature_dim,
+        num_classes=max_token_id + 1,  # +1 for the blank symbol
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")
+        return
+
+    model.to(device)
+    model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    yes_no = YesNoAsrDataModule(args)
+    test_dl = yes_no.test_dataloaders()
+    results = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+        HLG=HLG,
+        word_table=lexicon.word_table,
+    )
+
+    save_results(exp_dir=params.exp_dir, test_set_name="test_set", results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/tdnn/export.py b/egs/yesno/ASR/tdnn/export.py
new file mode 100755
index 000000000..c40cf8cd1
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/export.py
@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+
+"""
+This file is for exporting trained models to a checkpoint
+or to a torchscript model.
+
+(1) Generate the checkpoint tdnn/exp/pretrained.pt
+
+./tdnn/export.py \
+  --epoch 14 \
+  --avg 2
+
+See ./tdnn/pretrained.py for how to use the generated file.
+
+(2) Generate torchscript model tdnn/exp/cpu_jit.pt
+
+./tdnn/export.py \
+  --epoch 14 \
+  --avg 2 \
+  --jit 1
+
+See ./tdnn/jit_pretrained.py for how to use the generated file.
+"""
+
+import argparse
+import logging
+
+import torch
+from model import Tdnn
+from train import get_params
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=14,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=2,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    model = Tdnn(
+        num_features=params.feature_dim,
+        num_classes=max_token_id + 1,  # +1 for the blank symbol
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/export_onnx.py b/egs/yesno/ASR/tdnn/export_onnx.py
new file mode 100755
index 000000000..2436ca81b
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/export_onnx.py
@@ -0,0 +1,158 @@
+#!/usr/bin/env python3
+
+"""
+This file is for exporting trained models to onnx.
+
+Usage:
+
+    ./tdnn/export_onnx.py \
+      --epoch 14 \
+      --avg 2
+
+The above command generates the following two files:
+  - ./exp/model-epoch-14-avg-2.onnx
+  - ./exp/model-epoch-14-avg-2.int8.onnx
+
+See ./tdnn/onnx_pretrained.py for how to use them.
+"""
+
+import argparse
+import logging
+from typing import Dict
+
+import onnx
+import torch
+from model import Tdnn
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from train import get_params
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=14,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=2,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = str(value)
+
+    onnx.save(model, filename)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    model = Tdnn(
+        num_features=params.feature_dim,
+        num_classes=max_token_id + 1,  # +1 for the blank symbol
+    )
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to("cpu")
+    model.eval()
+
+    N = 1
+    T = 100
+    C = params.feature_dim
+    x = torch.rand(N, T, C)
+
+    opset_version = 13
+    onnx_filename = f"{params.exp_dir}/model-epoch-{params.epoch}-avg-{params.avg}.onnx"
+    torch.onnx.export(
+        model,
+        x,
+        onnx_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x"],
+        output_names=["log_prob"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "log_prob": {0: "N", 1: "T"},
+        },
+    )
+
+    logging.info(f"Saved to {onnx_filename}")
+    meta_data = {
+        "model_type": "tdnn",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "non-streaming tdnn for the yesno recipe",
+        "vocab_size": max_token_id + 1,
+    }
+
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=onnx_filename, meta_data=meta_data)
+
+    logging.info("Generate int8 quantization models")
+    onnx_filename_int8 = (
+        f"{params.exp_dir}/model-epoch-{params.epoch}-avg-{params.avg}.int8.onnx"
+    )
+
+    quantize_dynamic(
+        model_input=onnx_filename,
+        model_output=onnx_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+    logging.info(f"Saved to {onnx_filename_int8}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/jit_pretrained.py b/egs/yesno/ASR/tdnn/jit_pretrained.py
new file mode 100755
index 000000000..7581ecb83
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/jit_pretrained.py
@@ -0,0 +1,198 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use a torchscript model for decoding.
+
+Usage:
+
+  ./tdnn/jit_pretrained.py \
+    --nn-model ./tdnn/exp/cpu_jit.pt \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+Note that to generate ./tdnn/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+from typing import List
+import math
+
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./tdnn/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 23,
+            "num_classes": 4,  # [<blk>, N, SIL, Y]
+            "sample_rate": 8000,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output = model(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    best_path = one_best_decoding(
+        lattice=lattice, use_double_scores=params.use_double_scores
+    )
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py
new file mode 100755
index 000000000..ff8c742af
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py
@@ -0,0 +1,208 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with H
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+  ./tdnn/jit_pretrained_decode_with_H.py \
+    --nn-model ./tdnn/exp/cpu_jit.pt \
+    --H ./data/lang_phone/H.fst \
+    --tokens ./data/lang_phone/tokens.txt \
+    ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+    ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+Note that to generate ./tdnn/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./tdnn/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument("--H", type=str, required=True, help="Path to H.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_tokens(tokens_txt: str) -> Dict[int, str]:
+    id2token = dict()
+    with open(tokens_txt, encoding="utf-8") as f:
+        for line in f:
+            token, idx = line.strip().split()
+            id2token[int(idx)] = token
+
+    return id2token
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    H: kaldifst,
+    id2token: Dict[int, str],
+) -> List[str]:
+    decodable = DecodableCtc(nnet_output)
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(H, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        print(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        print(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2token[i - 1] for i in osymbols_out if id2token[i - 1] != "SIL"]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading H from {args.H}")
+    H = kaldifst.StdVectorFst.read(args.H)
+
+    sample_rate = 8000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 23
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output = model(features)
+
+    id2token = read_tokens(args.tokens)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[0],
+            nnet_output=nnet_output[i],
+            H=H,
+            id2token=id2token,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py
new file mode 100755
index 000000000..05ba74f9a
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py
@@ -0,0 +1,207 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with HL
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+  ./tdnn/jit_pretrained_decode_with_HL.py \
+    --nn-model ./tdnn/exp/cpu_jit.pt \
+    --HL ./data/lang_phone/HL.fst \
+    --words ./data/lang_phone/words.txt \
+    ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+    ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+Note that to generate ./tdnn/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./tdnn/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HL", type=str, required=True, help="Path to HL.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_words(words_txt: str) -> Dict[int, str]:
+    id2word = dict()
+    with open(words_txt, encoding="utf-8") as f:
+        for line in f:
+            word, idx = line.strip().split()
+            id2word[int(idx)] = word
+
+    return id2word
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    HL: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    decodable = DecodableCtc(nnet_output)
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HL, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        print(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        print(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    hyps = [id2word[i] for i in osymbols_out if id2word[i] != "<SIL>"]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HL from {args.HL}")
+    HL = kaldifst.StdVectorFst.read(args.HL)
+
+    sample_rate = 8000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 23
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output = model(features)
+
+    id2word = read_words(args.words)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[0],
+            nnet_output=nnet_output[i],
+            HL=HL,
+            id2word=id2word,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/model.py b/egs/yesno/ASR/tdnn/model.py
new file mode 100755
index 000000000..52cff37e0
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/model.py
@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corp.       (author: Fangjun Kuang)
+
+
+import torch
+import torch.nn as nn
+
+
+class Tdnn(nn.Module):
+    def __init__(self, num_features: int, num_classes: int):
+        """
+        Args:
+          num_features:
+            Model input dimension.
+          num_classes:
+            Model output dimension
+        """
+        super().__init__()
+
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=32,
+                kernel_size=3,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=32, affine=False),
+            nn.Conv1d(
+                in_channels=32,
+                out_channels=32,
+                kernel_size=5,
+                dilation=2,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=32, affine=False),
+            nn.Conv1d(
+                in_channels=32,
+                out_channels=32,
+                kernel_size=5,
+                dilation=4,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=32, affine=False),
+        )
+        self.output_linear = nn.Linear(in_features=32, out_features=num_classes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The input tensor with shape [N, T, C]
+
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x = x.permute(0, 2, 1)  # [N, T, C] -> [N, C, T]
+        x = self.tdnn(x)
+        x = x.permute(0, 2, 1)  # [N, C, T] -> [N, T, C]
+        x = self.output_linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x
+
+
+def test_tdnn():
+    num_features = 23
+    num_classes = 4
+    model = Tdnn(num_features=num_features, num_classes=num_classes)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    N = 2
+    T = 100
+    C = num_features
+    x = torch.randn(N, T, C)
+    y = model(x)
+    print(x.shape)
+    print(y.shape)
+
+
+if __name__ == "__main__":
+    test_tdnn()
diff --git a/egs/yesno/ASR/tdnn/onnx_pretrained.py b/egs/yesno/ASR/tdnn/onnx_pretrained.py
new file mode 100755
index 000000000..72a1d69c8
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/onnx_pretrained.py
@@ -0,0 +1,243 @@
+#!/usr/bin/env python3
+
+"""
+This file shows how to use an ONNX model for decoding with onnxruntime.
+
+Usage:
+
+(1) Use a not quantized ONNX model, i.e., a float32 model
+
+  ./tdnn/onnx_pretrained.py \
+    --nn-model ./tdnn/exp/model-epoch-14-avg-2.onnx \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+(2) Use a quantized ONNX model, i.e., an int8 model
+
+  ./tdnn/onnx_pretrained.py \
+    --nn-model ./tdnn/exp/model-epoch-14-avg-2.int8.onnx \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+Note that to generate ./tdnn/exp/model-epoch-14-avg-2.onnx,
+and ./tdnn/exp/model-epoch-14-avg-2.onnx,
+you can use ./export_onnx.py --epoch 14 --avg 2
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.utils import AttributeDict, get_texts
+
+
+class OnnxModel:
+    def __init__(self, nn_model: str):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        meta = self.model.get_modelmeta().custom_metadata_map
+        self.vocab_size = int(meta["vocab_size"])
+
+    def run(
+        self,
+        x: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+        Returns:
+          Return a 3-D tensor log_prob of shape (N, T, C)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0])
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./tdnn/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 23,
+            "sample_rate": 8000,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    logging.info(f"device: {device}")
+
+    logging.info(f"Loading onnx model {params.nn_model}")
+    model = OnnxModel(params.nn_model)
+
+    logging.info(f"Loading HLG from {args.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    # Note: We don't use key padding mask for attention during decoding
+    nnet_output = model.run(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    best_path = one_best_decoding(
+        lattice=lattice, use_double_scores=params.use_double_scores
+    )
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/pretrained.py b/egs/yesno/ASR/tdnn/pretrained.py
new file mode 100755
index 000000000..987c49de6
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/pretrained.py
@@ -0,0 +1,221 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file shows how to use a checkpoint for decoding.
+
+Usage:
+
+  ./tdnn/pretrained.py \
+    --checkpoint ./tdnn/exp/pretrained.pt \
+    --HLG ./data/lang_phone/HLG.pt \
+    --words-file ./data/lang_phone/words.txt \
+    download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    download/waves_yesno/0_0_1_0_0_0_1_0.wav
+
+Note that to generate ./tdnn/exp/pretrained.pt,
+you can use ./export.py
+"""
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from model import Tdnn
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.utils import AttributeDict, get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint(). "
+        "You can use ./tdnn/export.py to obtain it.",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.pt.")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 23,
+            "num_classes": 4,  # [<blk>, N, SIL, Y]
+            "sample_rate": 8000,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+
+    model = Tdnn(
+        num_features=params.feature_dim,
+        num_classes=params.num_classes,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    # Note: We don't use key padding mask for attention during decoding
+    nnet_output = model(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HLG,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+    )
+
+    best_path = one_best_decoding(
+        lattice=lattice, use_double_scores=params.use_double_scores
+    )
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/train.py b/egs/yesno/ASR/tdnn/train.py
new file mode 100755
index 000000000..335493491
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/train.py
@@ -0,0 +1,575 @@
+#!/usr/bin/env python3
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import YesNoAsrDataModule
+from lhotse.utils import fix_random_seed
+from model import Tdnn
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=15,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - exp_dir: It specifies the directory where all training related
+                   files, e.g., checkpoints, log, etc, are saved
+
+        - lang_dir: It contains language related input files such as
+                    "lexicon.txt"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - start_epoch:  If it is not zero, load checkpoint `start_epoch-1`
+                        and continue training from that checkpoint.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn/exp"),
+            "lang_dir": Path("data/lang_phone"),
+            "lr": 1e-2,
+            "feature_dim": 23,
+            "weight_decay": 1e-6,
+            "start_epoch": 0,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 20,
+            "valid_interval": 10,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Tdnn in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(feature)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervisions = batch["supervisions"]
+    texts = supervisions["text"]
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    decoding_graph = graph_compiler.compile(texts)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+    )
+
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    params["env_info"] = get_env_info()
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"device: {device}")
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = Tdnn(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.SGD(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    yes_no = YesNoAsrDataModule(args)
+    train_dl = yes_no.train_dataloaders()
+
+    # There are only 60 waves: 30 files are used for training
+    # and the remaining 30 files are used for testing.
+    # We use test data as validation.
+    valid_dl = yes_no.test_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=None,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    YesNoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/transducer/__init__.py b/egs/yesno/ASR/transducer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/yesno/ASR/transducer/asr_datamodule.py b/egs/yesno/ASR/transducer/asr_datamodule.py
new file mode 120000
index 000000000..c9c8adb57
--- /dev/null
+++ b/egs/yesno/ASR/transducer/asr_datamodule.py
@@ -0,0 +1 @@
+../tdnn/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/yesno/ASR/transducer/beam_search.py b/egs/yesno/ASR/transducer/beam_search.py
new file mode 100644
index 000000000..b98090636
--- /dev/null
+++ b/egs/yesno/ASR/transducer/beam_search.py
@@ -0,0 +1,69 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List
+
+import torch
+from transducer.model import Transducer
+
+
+def greedy_search(model: Transducer, encoder_out: torch.Tensor) -> List[str]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    device = model.device
+
+    sos = torch.tensor([blank_id], device=device).reshape(1, 1)
+    decoder_out, (h, c) = model.decoder(sos)
+    T = encoder_out.size(1)
+    t = 0
+    hyp = []
+    max_u = 1000  # terminate after this number of steps
+    u = 0
+
+    while t < T and u < max_u:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(current_encoder_out, decoder_out)
+
+        log_prob = logits.log_softmax(dim=-1)
+        # log_prob is (N, 1, 1)
+        # TODO: Use logits.argmax()
+        y = log_prob.argmax()
+        if y != blank_id:
+            hyp.append(y.item())
+            y = y.reshape(1, 1)
+            decoder_out, (h, c) = model.decoder(y, (h, c))
+            u += 1
+        else:
+            t += 1
+    id2word = {1: "YES", 2: "NO"}
+
+    hyp = [id2word[i] for i in hyp]
+
+    return hyp
diff --git a/egs/yesno/ASR/transducer/decode.py b/egs/yesno/ASR/transducer/decode.py
new file mode 100755
index 000000000..7f13e417a
--- /dev/null
+++ b/egs/yesno/ASR/transducer/decode.py
@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+from asr_datamodule import YesNoAsrDataModule
+from transducer.beam_search import greedy_search
+from transducer.decoder import Decoder
+from transducer.encoder import Tdnn
+from transducer.joiner import Joiner
+from transducer.model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=125,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="Directory from which to load the checkpoints",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 23,
+            # encoder/decoder params
+            "vocab_size": 3,  # blank, yes, no
+            "blank_id": 0,
+            "embedding_dim": 32,
+            "hidden_dim": 16,
+            "num_decoder_layers": 4,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+) -> List[List[int]]:
+    """Decode one batch and return the result in a list-of-list.
+    Each sub list contains the word IDs for an utterance in the batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding.
+        - params.method is "nbest", it uses nbest decoding.
+
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+        (https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py)
+    Returns:
+      Return the decoding result. `len(ans)` == batch size.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+    feature_lens = batch["supervisions"]["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        hyp = greedy_search(model=model, encoder_out=encoder_out_i)
+        hyps.append(hyp)
+
+    #  hyps = [[word_table[i] for i in ids] for ids in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+) -> List[Tuple[List[int], List[int]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+    Returns:
+      Return a tuple contains two elements (ref_text, hyp_text):
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps = decode_one_batch(
+            params=params,
+            model=model,
+            batch=batch,
+        )
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    exp_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[List[int], List[int]]],
+) -> None:
+    """Save results to `exp_dir`.
+    Args:
+      exp_dir:
+        The output directory. This function create the following files inside
+        this directory:
+
+            - recogs-{test_set_name}.text
+
+                It contains the reference and hypothesis results, like below::
+
+                    ref=['NO', 'NO', 'NO', 'YES', 'NO', 'NO', 'NO', 'YES']
+                    hyp=['NO', 'NO', 'NO', 'YES', 'NO', 'NO', 'NO', 'YES']
+                    ref=['NO', 'NO', 'YES', 'NO', 'YES', 'NO', 'NO', 'YES']
+                    hyp=['NO', 'NO', 'YES', 'NO', 'YES', 'NO', 'NO', 'YES']
+
+            - errs-{test_set_name}.txt
+
+                It contains the detailed WER.
+      test_set_name:
+        The name of the test set, which will be part of the result filename.
+      results:
+        A list of tuples, each of which contains (ref_words, hyp_words).
+    Returns:
+      Return None.
+    """
+    recog_path = exp_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = exp_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        write_error_stats(f, f"{test_set_name}", results)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = Tdnn(
+        num_features=params.feature_dim,
+        output_dim=params.hidden_dim,
+    )
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.hidden_dim,
+        embedding_dropout=0.4,
+        rnn_dropout=0.4,
+    )
+    joiner = Joiner(input_dim=params.hidden_dim, output_dim=params.vocab_size)
+    transducer = Transducer(encoder=encoder, decoder=decoder, joiner=joiner)
+    return transducer
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    YesNoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+    params["env_info"] = get_env_info()
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    yes_no = YesNoAsrDataModule(args)
+    test_dl = yes_no.test_dataloaders()
+    results = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+    )
+
+    save_results(exp_dir=params.exp_dir, test_set_name="test_set", results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/transducer/decoder.py b/egs/yesno/ASR/transducer/decoder.py
new file mode 100644
index 000000000..7ae540d03
--- /dev/null
+++ b/egs/yesno/ASR/transducer/decoder.py
@@ -0,0 +1,92 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        num_layers: int,
+        hidden_dim: int,
+        embedding_dropout: float = 0.0,
+        rnn_dropout: float = 0.0,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          num_layers:
+            Number of RNN layers.
+          hidden_dim:
+            Hidden dimension of RNN layers.
+          embedding_dropout:
+            Dropout rate for the embedding layer.
+          rnn_dropout:
+            Dropout for RNN layers.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.embedding_dropout = nn.Dropout(embedding_dropout)
+        self.rnn = nn.LSTM(
+            input_size=embedding_dim,
+            hidden_size=hidden_dim,
+            num_layers=num_layers,
+            batch_first=True,
+            dropout=rnn_dropout,
+        )
+        self.blank_id = blank_id
+        self.output_linear = nn.Linear(hidden_dim, hidden_dim)
+
+    def forward(
+        self,
+        y: torch.Tensor,
+        states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          states:
+            A tuple of two tensors containing the states information of
+            RNN layers in this decoder.
+        Returns:
+          Return a tuple containing:
+
+            - rnn_output, a tensor of shape (N, U, C)
+            - (h, c), which contain the state information for RNN layers.
+              Both are of shape (num_layers, N, C)
+        """
+        embedding_out = self.embedding(y)
+        embedding_out = self.embedding_dropout(embedding_out)
+        rnn_out, (h, c) = self.rnn(embedding_out, states)
+        out = self.output_linear(rnn_out)
+
+        return out, (h, c)
diff --git a/egs/yesno/ASR/transducer/encoder.py b/egs/yesno/ASR/transducer/encoder.py
new file mode 100644
index 000000000..8c50df293
--- /dev/null
+++ b/egs/yesno/ASR/transducer/encoder.py
@@ -0,0 +1,87 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+
+
+# We use a TDNN model as encoder, as it works very well with CTC training
+# for this tiny dataset.
+class Tdnn(nn.Module):
+    def __init__(self, num_features: int, output_dim: int):
+        """
+        Args:
+          num_features:
+            Model input dimension.
+          ouput_dim:
+            Model output dimension
+        """
+        super().__init__()
+
+        # Note: We don't use paddings inside conv layers
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=32,
+                kernel_size=3,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=32, affine=False),
+            nn.Conv1d(
+                in_channels=32,
+                out_channels=32,
+                kernel_size=5,
+                dilation=2,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=32, affine=False),
+            nn.Conv1d(
+                in_channels=32,
+                out_channels=32,
+                kernel_size=5,
+                dilation=4,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=32, affine=False),
+        )
+        self.output_linear = nn.Linear(in_features=32, out_features=output_dim)
+
+    def forward(self, x: torch.Tensor, x_lens: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The input tensor with shape (N, T, C)
+          x_lens:
+            It contains the number of frames in each utterance in x
+            before padding.
+
+        Returns:
+          Return a tuple with 2 tensors:
+
+            - logits, a tensor of shape (N, T, C)
+            - logit_lens, a tensor of shape (N,)
+        """
+        x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+        x = self.tdnn(x)
+        x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        logits = self.output_linear(x)
+
+        # the first conv layer reduces T by 3-1 frames
+        # the second layer reduces T by (5-1)*2 frames
+        # the second layer reduces T by (5-1)*4 frames
+        # Number of output frames is 2 + 4*2 + 4*4 = 2 + 8 + 16 = 26
+        x_lens = x_lens - 26
+        return logits, x_lens
diff --git a/egs/yesno/ASR/transducer/joiner.py b/egs/yesno/ASR/transducer/joiner.py
new file mode 100644
index 000000000..0422f8a6f
--- /dev/null
+++ b/egs/yesno/ASR/transducer/joiner.py
@@ -0,0 +1,55 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, C).
+        Returns:
+          Return a tensor of shape (N, T, U, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == decoder_out.size(2)
+
+        encoder_out = encoder_out.unsqueeze(2)
+        # Now encoder_out is (N, T, 1, C)
+
+        decoder_out = decoder_out.unsqueeze(1)
+        # Now decoder_out is (N, 1, U, C)
+
+        logit = encoder_out + decoder_out
+        logit = F.relu(logit)
+
+        output = self.output_linear(logit)
+
+        return output
diff --git a/egs/yesno/ASR/transducer/model.py b/egs/yesno/ASR/transducer/model.py
new file mode 100644
index 000000000..caf9bed37
--- /dev/null
+++ b/egs/yesno/ASR/transducer/model.py
@@ -0,0 +1,120 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Note we use `rnnt_loss` from torchaudio, which exists only in
+torchaudio >= v0.10.0. It also means you have to use torch >= v1.10.0
+"""
+import k2
+import torch
+import torch.nn as nn
+import torchaudio
+import torchaudio.functional
+
+from icefall.utils import add_sos
+
+assert hasattr(torchaudio.functional, "rnnt_loss"), (
+    f"Current torchaudio version: {torchaudio.__version__}\n"
+    "Please install a version >= 0.10.0"
+)
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: nn.Module,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        decoder_out, _ = self.decoder(sos_y_padded)
+
+        logits = self.joiner(encoder_out, decoder_out)
+
+        # rnnt_loss requires 0 padded targets
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        loss = torchaudio.functional.rnnt_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="mean",
+        )
+
+        return loss
diff --git a/egs/yesno/ASR/transducer/test_decoder.py b/egs/yesno/ASR/transducer/test_decoder.py
new file mode 100755
index 000000000..88c54f678
--- /dev/null
+++ b/egs/yesno/ASR/transducer/test_decoder.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/yesno/ASR
+    python ./transducer/test_decoder.py
+"""
+
+import torch
+from transducer.decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    num_layers = 2
+    hidden_dim = 6
+    N = 3
+    U = 5
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        num_layers=num_layers,
+        hidden_dim=hidden_dim,
+        embedding_dropout=0.0,
+        rnn_dropout=0.0,
+    )
+    x = torch.randint(1, vocab_size, (N, U))
+    rnn_out, (h, c) = decoder(x)
+
+    assert rnn_out.shape == (N, U, hidden_dim)
+    assert h.shape == (num_layers, N, hidden_dim)
+    assert c.shape == (num_layers, N, hidden_dim)
+
+    rnn_out, (h, c) = decoder(x, (h, c))
+    assert rnn_out.shape == (N, U, hidden_dim)
+    assert h.shape == (num_layers, N, hidden_dim)
+    assert c.shape == (num_layers, N, hidden_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/transducer/test_encoder.py b/egs/yesno/ASR/transducer/test_encoder.py
new file mode 100755
index 000000000..481fb558b
--- /dev/null
+++ b/egs/yesno/ASR/transducer/test_encoder.py
@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/yesno/ASR
+    python ./transducer/test_encoder.py
+"""
+
+import torch
+from transducer.encoder import Tdnn
+
+
+def test_encoder():
+    input_dim = 10
+    output_dim = 20
+    encoder = Tdnn(input_dim, output_dim)
+    N = 10
+    T = 85
+    x = torch.rand(N, T, input_dim)
+    x_lens = torch.randint(low=30, high=T, size=(N,), dtype=torch.int32)
+    logits, logit_lens = encoder(x, x_lens)
+    assert logits.shape == (N, T - 26, output_dim)
+    assert torch.all(torch.eq(x_lens - 26, logit_lens))
+
+
+def main():
+    test_encoder()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/transducer/test_joiner.py b/egs/yesno/ASR/transducer/test_joiner.py
new file mode 100755
index 000000000..2773ca319
--- /dev/null
+++ b/egs/yesno/ASR/transducer/test_joiner.py
@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/yesno/ASR
+    python ./transducer/test_joiner.py
+"""
+
+
+import torch
+from transducer.joiner import Joiner
+
+
+def test_joiner():
+    N = 2
+    T = 3
+    C = 4
+    U = 5
+
+    joiner = Joiner(C, 10)
+
+    encoder_out = torch.rand(N, T, C)
+    decoder_out = torch.rand(N, U, C)
+
+    joint = joiner(encoder_out, decoder_out)
+    assert joint.shape == (N, T, U, 10)
+
+
+def main():
+    test_joiner()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/transducer/test_transducer.py b/egs/yesno/ASR/transducer/test_transducer.py
new file mode 100755
index 000000000..db7bf9c68
--- /dev/null
+++ b/egs/yesno/ASR/transducer/test_transducer.py
@@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/yesno/ASR
+    python ./transducer/test_transducer.py
+"""
+
+
+import k2
+import torch
+from transducer.decoder import Decoder
+from transducer.encoder import Tdnn
+from transducer.joiner import Joiner
+from transducer.model import Transducer
+
+
+def test_transducer():
+    # encoder params
+    input_dim = 10
+    output_dim = 20
+
+    # decoder params
+    vocab_size = 3
+    blank_id = 0
+    embedding_dim = 128
+    num_layers = 2
+
+    encoder = Tdnn(input_dim, output_dim)
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        num_layers=num_layers,
+        hidden_dim=output_dim,
+        embedding_dropout=0.0,
+        rnn_dropout=0.0,
+    )
+
+    joiner = Joiner(output_dim, vocab_size)
+    transducer = Transducer(encoder=encoder, decoder=decoder, joiner=joiner)
+
+    y = k2.RaggedTensor([[1, 2, 1], [1, 1, 1, 2, 1]])
+    N = y.dim0
+    T = 50
+
+    x = torch.rand(N, T, input_dim)
+    x_lens = torch.randint(low=30, high=T, size=(N,), dtype=torch.int32)
+    x_lens[0] = T
+
+    loss = transducer(x, x_lens, y)
+    print(loss)
+
+
+def main():
+    test_transducer()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/yesno/ASR/transducer/train.py b/egs/yesno/ASR/transducer/train.py
new file mode 100755
index 000000000..88866ae81
--- /dev/null
+++ b/egs/yesno/ASR/transducer/train.py
@@ -0,0 +1,587 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import YesNoAsrDataModule
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transducer.decoder import Decoder
+from transducer.encoder import Tdnn
+from transducer.joiner import Joiner
+from transducer.model import Transducer
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_labels(texts: List[str]) -> k2.RaggedTensor:
+    """
+    Args:
+      texts:
+        A list of transcripts. Each transcript contains spaces separated
+        "NO" or "YES".
+    Returns:
+      Return a ragged tensor containing the corresponding word ID.
+    """
+    # blank is 0
+    word2id = {"YES": 1, "NO": 2}
+    word_ids = []
+    for t in texts:
+        words = t.split()
+        ids = [word2id[w] for w in words]
+        word_ids.append(ids)
+
+    return k2.RaggedTensor(word_ids)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=200,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="Directory to save results",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - start_epoch:  If it is not zero, load checkpoint `start_epoch-1`
+                        and continue training from that checkpoint.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+
+    """
+    params = AttributeDict(
+        {
+            "lr": 1e-3,
+            "feature_dim": 23,
+            "weight_decay": 1e-6,
+            "start_epoch": 0,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 20,
+            "valid_interval": 10,
+            # encoder/decoder params
+            "vocab_size": 3,  # blank, yes, no
+            "blank_id": 0,
+            "embedding_dim": 32,
+            "hidden_dim": 16,
+            "num_decoder_layers": 4,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Tdnn in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    feature_lens = batch["supervisions"]["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    labels = get_labels(texts).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(x=feature, x_lens=feature_lens, y=labels)
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = feature.size(0)
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = Tdnn(
+        num_features=params.feature_dim,
+        output_dim=params.hidden_dim,
+    )
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.hidden_dim,
+        embedding_dropout=0.4,
+        rnn_dropout=0.4,
+    )
+    joiner = Joiner(input_dim=params.hidden_dim, output_dim=params.vocab_size)
+    transducer = Transducer(encoder=encoder, decoder=decoder, joiner=joiner)
+
+    return transducer
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    params["env_info"] = get_env_info()
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"device: {device}")
+
+    model = get_transducer_model(params)
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    model.device = device
+
+    optimizer = optim.Adam(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    yes_no = YesNoAsrDataModule(args)
+    train_dl = yes_no.train_dataloaders()
+
+    # There are only 60 waves: 30 files are used for training
+    # and the remaining 30 files are used for testing.
+    # We use test data as validation.
+    valid_dl = yes_no.test_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=None,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    YesNoAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/__init__.py b/icefall/__init__.py
new file mode 100644
index 000000000..b1e4313e9
--- /dev/null
+++ b/icefall/__init__.py
@@ -0,0 +1,76 @@
+# isort:skip_file
+
+from . import checkpoint, decode, dist, env, utils
+
+from .byte_utils import (
+    byte_decode,
+    byte_encode,
+    smart_byte_decode,
+)
+
+from .checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+    remove_checkpoints,
+    save_checkpoint,
+    save_checkpoint_with_global_batch_idx,
+)
+
+from .context_graph import ContextGraph, ContextState
+
+from .decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+
+from .dist import (
+    cleanup_dist,
+    setup_dist,
+)
+
+from .env import (
+    get_env_info,
+    get_git_branch_name,
+    get_git_date,
+    get_git_sha1,
+)
+
+from .utils import (
+    AttributeDict,
+    MetricsTracker,
+    add_eos,
+    add_sos,
+    concat,
+    encode_supervisions,
+    get_alignments,
+    get_executor,
+    get_texts,
+    is_cjk,
+    is_jit_tracing,
+    is_module_available,
+    l1_norm,
+    l2_norm,
+    linf_norm,
+    load_alignments,
+    make_pad_mask,
+    measure_gradient_norms,
+    measure_weight_norms,
+    optim_step_and_measure_param_change,
+    save_alignments,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    subsequent_chunk_mask,
+    tokenize_by_CJK_char,
+    write_error_stats,
+)
+
+from .ngram_lm import NgramLm, NgramLmStateCost
+
+from .lm_wrapper import LmScorer
diff --git a/icefall/ali.py b/icefall/ali.py
new file mode 100644
index 000000000..c3e4b2662
--- /dev/null
+++ b/icefall/ali.py
@@ -0,0 +1,142 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict, List, Tuple
+
+import torch
+from torch.nn.utils.rnn import pad_sequence
+
+
+def save_alignments(
+    alignments: Dict[str, List[int]],
+    subsampling_factor: int,
+    filename: str,
+) -> None:
+    """Save alignments to a file.
+
+    Args:
+      alignments:
+        A dict containing alignments. Keys of the dict are utterances and
+        values are the corresponding framewise alignments after subsampling.
+      subsampling_factor:
+        The subsampling factor of the model.
+      filename:
+        Path to save the alignments.
+    Returns:
+      Return None.
+    """
+    ali_dict = {
+        "subsampling_factor": subsampling_factor,
+        "alignments": alignments,
+    }
+    torch.save(ali_dict, filename)
+
+
+def load_alignments(filename: str) -> Tuple[int, Dict[str, List[int]]]:
+    """Load alignments from a file.
+
+    Args:
+      filename:
+        Path to the file containing alignment information.
+        The file should be saved by :func:`save_alignments`.
+    Returns:
+      Return a tuple containing:
+        - subsampling_factor: The subsampling_factor used to compute
+          the alignments.
+        - alignments: A dict containing utterances and their corresponding
+          framewise alignment, after subsampling.
+    """
+    ali_dict = torch.load(filename)
+    subsampling_factor = ali_dict["subsampling_factor"]
+    alignments = ali_dict["alignments"]
+    return subsampling_factor, alignments
+
+
+def convert_alignments_to_tensor(
+    alignments: Dict[str, List[int]], device: torch.device
+) -> Dict[str, torch.Tensor]:
+    """Convert alignments from list of int to a 1-D torch.Tensor.
+
+    Args:
+      alignments:
+        A dict containing alignments. Keys are utterance IDs and
+        values are their corresponding frame-wise alignments.
+      device:
+        The device to move the alignments to.
+    Returns:
+      Return a dict using 1-D torch.Tensor to store the alignments.
+      The dtype of the tensor are `torch.int64`. We choose `torch.int64`
+      because `torch.nn.functional.one_hot` requires that.
+    """
+    ans = {}
+    for utt_id, ali in alignments.items():
+        ali = torch.tensor(ali, dtype=torch.int64, device=device)
+        ans[utt_id] = ali
+    return ans
+
+
+def lookup_alignments(
+    cut_ids: List[str],
+    alignments: Dict[str, torch.Tensor],
+    num_classes: int,
+    log_score: float = -10,
+) -> torch.Tensor:
+    """Return a mask constructed from alignments by a list of cut IDs.
+
+    The returned mask is a 3-D tensor of shape (N, T, C). For each frame,
+    i.e., each row, of the returned mask, positions not corresponding to
+    the alignments are filled with `log_score`, while the position
+    specified by the alignment is filled with 0. For instance, if the alignments
+    of two utterances are:
+
+        [ [1, 3, 2], [1, 0, 4, 2] ]
+    num_classes is 5 and log_score is -10,  then the returned mask is
+
+        [
+          [[-10, 0, -10, -10, -10],
+           [-10, -10, -10, 0, -10],
+           [-10, -10, 0, -10, -10],
+           [0, -10, -10, -10, -10]],
+          [[-10, 0, -10, -10, -10],
+           [0, -10, -10, -10, -10],
+           [-10, -10, -10, -10, 0],
+           [-10, -10, 0, -10, -10]]
+        ]
+    Note: We pad the alignment of the first utterance with 0.
+
+    Args:
+      cut_ids:
+        A list of utterance IDs.
+      alignments:
+        A dict containing alignments. The keys are utterance IDs and the values
+        are framewise alignments.
+      num_classes:
+        The max token ID + 1 that appears in the alignments.
+      log_score:
+        Positions in the returned tensor not corresponding to the alignments
+        are filled with this value.
+    Returns:
+      Return a 3-D torch.float32 tensor of shape (N, T, C).
+    """
+    # We assume all utterances have their alignments.
+    ali = [alignments[cut_id] for cut_id in cut_ids]
+    padded_ali = pad_sequence(ali, batch_first=True, padding_value=0)
+    padded_one_hot = torch.nn.functional.one_hot(
+        padded_ali,
+        num_classes=num_classes,
+    )
+    mask = (1 - padded_one_hot) * float(log_score)
+    return mask
diff --git a/icefall/bpe_graph_compiler.py b/icefall/bpe_graph_compiler.py
new file mode 100644
index 000000000..d9659c2dd
--- /dev/null
+++ b/icefall/bpe_graph_compiler.py
@@ -0,0 +1,94 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from pathlib import Path
+from typing import List, Union
+
+import k2
+import sentencepiece as spm
+import torch
+
+
+class BpeCtcTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lang_dir: Path,
+        device: Union[str, torch.device] = "cpu",
+        sos_token: str = "<sos/eos>",
+        eos_token: str = "<sos/eos>",
+    ) -> None:
+        """
+        Args:
+          lang_dir:
+            This directory is expected to contain the following files:
+
+                - bpe.model
+                - words.txt
+          device:
+            It indicates CPU or CUDA.
+          sos_token:
+            The word piece that represents sos.
+          eos_token:
+            The word piece that represents eos.
+        """
+        lang_dir = Path(lang_dir)
+        model_file = lang_dir / "bpe.model"
+        sp = spm.SentencePieceProcessor()
+        sp.load(str(model_file))
+        self.sp = sp
+        self.word_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+        self.device = device
+
+        self.sos_id = self.sp.piece_to_id(sos_token)
+        self.eos_id = self.sp.piece_to_id(eos_token)
+
+        assert self.sos_id != self.sp.unk_id()
+        assert self.eos_id != self.sp.unk_id()
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of piece IDs.
+
+        Args:
+          texts:
+            It is a list of strings. Each string consists of space(s)
+            separated words. An example containing two strings is given below:
+
+                ['HELLO ICEFALL', 'HELLO k2']
+        Returns:
+          Return a list-of-list of piece IDs.
+        """
+        return self.sp.encode(texts, out_type=int)
+
+    def compile(
+        self,
+        piece_ids: List[List[int]],
+        modified: bool = False,
+    ) -> k2.Fsa:
+        """Build a ctc graph from a list-of-list piece IDs.
+
+        Args:
+          piece_ids:
+            It is a list-of-list integer IDs.
+          modified:
+           See :func:`k2.ctc_graph` for its meaning.
+        Return:
+          Return an FsaVec, which is the result of composing a
+          CTC topology with linear FSAs constructed from the given
+          piece IDs.
+        """
+        graph = k2.ctc_graph(piece_ids, modified=modified, device=self.device)
+        return graph
diff --git a/icefall/byte_utils.py b/icefall/byte_utils.py
new file mode 100644
index 000000000..79c1c7545
--- /dev/null
+++ b/icefall/byte_utils.py
@@ -0,0 +1,313 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+#
+# This file was copied and modified from https://github.com/facebookresearch/fairseq/blob/main/fairseq/data/encoders/byte_utils.py
+
+import re
+import unicodedata
+
+
+WHITESPACE_NORMALIZER = re.compile(r"\s+")
+SPACE = chr(32)
+SPACE_ESCAPE = chr(9601)
+BPE_UNK = chr(8263)
+
+PRINTABLE_BASE_CHARS = [
+    256,
+    257,
+    258,
+    259,
+    260,
+    261,
+    262,
+    263,
+    264,
+    265,
+    266,
+    267,
+    268,
+    269,
+    270,
+    271,
+    272,
+    273,
+    274,
+    275,
+    276,
+    277,
+    278,
+    279,
+    280,
+    281,
+    282,
+    283,
+    284,
+    285,
+    286,
+    287,
+    32,
+    33,
+    34,
+    35,
+    36,
+    37,
+    38,
+    39,
+    40,
+    41,
+    42,
+    43,
+    44,
+    45,
+    46,
+    47,
+    48,
+    49,
+    50,
+    51,
+    52,
+    53,
+    54,
+    55,
+    56,
+    57,
+    58,
+    59,
+    60,
+    61,
+    62,
+    63,
+    64,
+    65,
+    66,
+    67,
+    68,
+    69,
+    70,
+    71,
+    72,
+    73,
+    74,
+    75,
+    76,
+    77,
+    78,
+    79,
+    80,
+    81,
+    82,
+    83,
+    84,
+    85,
+    86,
+    87,
+    88,
+    89,
+    90,
+    91,
+    92,
+    93,
+    94,
+    95,
+    96,
+    97,
+    98,
+    99,
+    100,
+    101,
+    102,
+    103,
+    104,
+    105,
+    106,
+    107,
+    108,
+    109,
+    110,
+    111,
+    112,
+    113,
+    114,
+    115,
+    116,
+    117,
+    118,
+    119,
+    120,
+    121,
+    122,
+    123,
+    124,
+    125,
+    126,
+    288,
+    289,
+    290,
+    291,
+    292,
+    293,
+    294,
+    295,
+    296,
+    297,
+    298,
+    299,
+    300,
+    301,
+    302,
+    303,
+    304,
+    305,
+    308,
+    309,
+    310,
+    311,
+    312,
+    313,
+    314,
+    315,
+    316,
+    317,
+    318,
+    321,
+    322,
+    323,
+    324,
+    325,
+    326,
+    327,
+    328,
+    330,
+    331,
+    332,
+    333,
+    334,
+    335,
+    336,
+    337,
+    338,
+    339,
+    340,
+    341,
+    342,
+    343,
+    344,
+    345,
+    346,
+    347,
+    348,
+    349,
+    350,
+    351,
+    352,
+    353,
+    354,
+    355,
+    356,
+    357,
+    358,
+    359,
+    360,
+    361,
+    362,
+    363,
+    364,
+    365,
+    366,
+    367,
+    368,
+    369,
+    370,
+    371,
+    372,
+    373,
+    374,
+    375,
+    376,
+    377,
+    378,
+    379,
+    380,
+    381,
+    382,
+    384,
+    385,
+    386,
+    387,
+    388,
+    389,
+    390,
+    391,
+    392,
+    393,
+    394,
+    395,
+    396,
+    397,
+    398,
+    399,
+    400,
+    401,
+    402,
+    403,
+    404,
+    405,
+    406,
+    407,
+    408,
+    409,
+    410,
+    411,
+    412,
+    413,
+    414,
+    415,
+    416,
+    417,
+    418,
+    419,
+    420,
+    421,
+    422,
+]
+
+for c in PRINTABLE_BASE_CHARS:
+    assert unicodedata.normalize("NFKC", chr(c)) == chr(c), c
+
+BYTE_TO_BCHAR = {b: chr(PRINTABLE_BASE_CHARS[b]) for b in range(256)}
+BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()}
+BCHAR_TO_BYTE[BPE_UNK] = 32  # map unk to space
+
+
+def byte_encode(x: str) -> str:
+    normalized = WHITESPACE_NORMALIZER.sub(SPACE, x)
+    return "".join([BYTE_TO_BCHAR[b] for b in normalized.encode("utf-8")])
+
+
+def byte_decode(x: str) -> str:
+    try:
+        return bytes([BCHAR_TO_BYTE[bc] for bc in x]).decode("utf-8")
+    except ValueError:
+        return ""
+
+
+def smart_byte_decode(x: str) -> str:
+    output = byte_decode(x)
+    if output == "":
+        # DP the best recovery (max valid chars) if it's broken
+        n_bytes = len(x)
+        f = [0 for _ in range(n_bytes + 1)]
+        pt = [0 for _ in range(n_bytes + 1)]
+        for i in range(1, n_bytes + 1):
+            f[i], pt[i] = f[i - 1], i - 1
+            for j in range(1, min(4, i) + 1):
+                if f[i - j] + 1 > f[i] and len(byte_decode(x[i - j : i])) > 0:
+                    f[i], pt[i] = f[i - j] + 1, i - j
+        cur_pt = n_bytes
+        while cur_pt > 0:
+            if f[cur_pt] == f[pt[cur_pt]] + 1:
+                output = byte_decode(x[pt[cur_pt] : cur_pt]) + output
+            cur_pt = pt[cur_pt]
+    return output
diff --git a/icefall/char_graph_compiler.py b/icefall/char_graph_compiler.py
new file mode 100644
index 000000000..5f9571d42
--- /dev/null
+++ b/icefall/char_graph_compiler.py
@@ -0,0 +1,121 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from typing import List
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+class CharCtcTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lexicon: Lexicon,
+        device: torch.device,
+        sos_token: str = "<sos/eos>",
+        eos_token: str = "<sos/eos>",
+        oov: str = "<unk>",
+    ):
+        """
+        Args:
+          lexicon:
+            It is built from `data/lang_char/lexicon.txt`.
+          device:
+            The device to use for operations compiling transcripts to FSAs.
+          oov:
+            Out of vocabulary token. When a word(token) in the transcript
+            does not exist in the token list, it is replaced with `oov`.
+        """
+
+        assert oov in lexicon.token_table
+
+        self.oov_id = lexicon.token_table[oov]
+        self.token_table = lexicon.token_table
+
+        self.device = device
+
+        self.sos_id = self.token_table[sos_token]
+        self.eos_id = self.token_table[eos_token]
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of token IDs.
+
+        Args:
+          texts:
+            It is a list of strings.
+            An example containing two strings is given below:
+
+                ['你好中国', '北京欢迎您']
+        Returns:
+          Return a list-of-list of token IDs.
+        """
+        ids: List[List[int]] = []
+        whitespace = re.compile(r"([ \t])")
+        for text in texts:
+            text = re.sub(whitespace, "", text)
+            sub_ids = [
+                self.token_table[txt] if txt in self.token_table else self.oov_id
+                for txt in text
+            ]
+            ids.append(sub_ids)
+        return ids
+
+    def texts_to_ids_with_bpe(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts (which include chars and bpes)
+           to a list-of-list of token IDs.
+
+        Args:
+          texts:
+            It is a list of strings.
+            An example containing two strings is given below:
+
+                [['你', '好', '▁C', 'hina'], ['北','京', '▁', 'welcome', '您']
+        Returns:
+          Return a list-of-list of token IDs.
+        """
+        ids: List[List[int]] = []
+        for text in texts:
+            text = text.split("/")
+            sub_ids = [
+                self.token_table[txt] if txt in self.token_table else self.oov_id
+                for txt in text
+            ]
+            ids.append(sub_ids)
+        return ids
+
+    def compile(
+        self,
+        token_ids: List[List[int]],
+        modified: bool = False,
+    ) -> k2.Fsa:
+        """Build a ctc graph from a list-of-list token IDs.
+
+        Args:
+          piece_ids:
+            It is a list-of-list integer IDs.
+         modified:
+           See :func:`k2.ctc_graph` for its meaning.
+        Return:
+          Return an FsaVec, which is the result of composing a
+          CTC topology with linear FSAs constructed from the given
+          piece IDs.
+        """
+        graph = k2.ctc_graph(token_ids, modified=modified, device=self.device)
+        return graph
diff --git a/icefall/checkpoint.py b/icefall/checkpoint.py
new file mode 100644
index 000000000..c83c56a53
--- /dev/null
+++ b/icefall/checkpoint.py
@@ -0,0 +1,477 @@
+# Copyright  2021-2022  Xiaomi Corporation  (authors: Fangjun Kuang,
+#                                                     Zengwei Yao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import glob
+import logging
+import os
+import re
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Union
+
+import torch
+import torch.nn as nn
+from lhotse.dataset.sampling.base import CutSampler
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import Optimizer
+
+# use duck typing for LRScheduler since we have different possibilities, see
+# our class LRScheduler.
+LRSchedulerType = object
+
+
+def save_checkpoint(
+    filename: Path,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    params: Optional[Dict[str, Any]] = None,
+    optimizer: Optional[Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+) -> None:
+    """Save training information to a file.
+
+    Args:
+      filename:
+        The checkpoint filename.
+      model:
+        The model to be saved. We only save its `state_dict()`.
+      model_avg:
+        The stored model averaged from the start of training.
+      params:
+        User defined parameters, e.g., epoch, loss.
+      optimizer:
+        The optimizer to be saved. We only save its `state_dict()`.
+      scheduler:
+        The scheduler to be saved. We only save its `state_dict()`.
+      scalar:
+        The GradScaler to be saved. We only save its `state_dict()`.
+      rank:
+        Used in DDP. We save checkpoint only for the node whose rank is 0.
+    Returns:
+      Return None.
+    """
+    if rank != 0:
+        return
+
+    logging.info(f"Saving checkpoint to {filename}")
+
+    if isinstance(model, DDP):
+        model = model.module
+
+    checkpoint = {
+        "model": model.state_dict(),
+        "optimizer": optimizer.state_dict() if optimizer is not None else None,
+        "scheduler": scheduler.state_dict() if scheduler is not None else None,
+        "grad_scaler": scaler.state_dict() if scaler is not None else None,
+        "sampler": sampler.state_dict() if sampler is not None else None,
+    }
+
+    if model_avg is not None:
+        checkpoint["model_avg"] = model_avg.to(torch.float32).state_dict()
+
+    if params:
+        for k, v in params.items():
+            assert k not in checkpoint
+            checkpoint[k] = v
+
+    torch.save(checkpoint, filename)
+
+
+def load_checkpoint(
+    filename: Path,
+    model: nn.Module,
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    strict: bool = False,
+) -> Dict[str, Any]:
+    """
+    TODO: document it
+    """
+    logging.info(f"Loading checkpoint from {filename}")
+    checkpoint = torch.load(filename, map_location="cpu")
+
+    if next(iter(checkpoint["model"])).startswith("module."):
+        logging.info("Loading checkpoint saved by DDP")
+
+        dst_state_dict = model.state_dict()
+        src_state_dict = checkpoint["model"]
+        for key in dst_state_dict.keys():
+            src_key = "{}.{}".format("module", key)
+            dst_state_dict[key] = src_state_dict.pop(src_key)
+        assert len(src_state_dict) == 0
+        model.load_state_dict(dst_state_dict, strict=strict)
+    else:
+        model.load_state_dict(checkpoint["model"], strict=strict)
+
+    checkpoint.pop("model")
+
+    if model_avg is not None and "model_avg" in checkpoint:
+        logging.info("Loading averaged model")
+        model_avg.load_state_dict(checkpoint["model_avg"], strict=strict)
+        checkpoint.pop("model_avg")
+
+    def load(name, obj):
+        s = checkpoint.get(name, None)
+        if obj and s:
+            obj.load_state_dict(s)
+            checkpoint.pop(name)
+
+    load("optimizer", optimizer)
+    load("scheduler", scheduler)
+    load("grad_scaler", scaler)
+    load("sampler", sampler)
+
+    return checkpoint
+
+
+def average_checkpoints(
+    filenames: List[Path], device: torch.device = torch.device("cpu")
+) -> dict:
+    """Average a list of checkpoints.
+
+    Args:
+      filenames:
+        Filenames of the checkpoints to be averaged. We assume all
+        checkpoints are saved by :func:`save_checkpoint`.
+      device:
+        Move checkpoints to this device before averaging.
+    Returns:
+      Return a dict (i.e., state_dict) which is the average of all
+      model state dicts contained in the checkpoints.
+    """
+    n = len(filenames)
+
+    avg = torch.load(filenames[0], map_location=device)["model"]
+
+    # Identify shared parameters. Two parameters are said to be shared
+    # if they have the same data_ptr
+    uniqued: Dict[int, str] = dict()
+
+    for k, v in avg.items():
+        v_data_ptr = v.data_ptr()
+        if v_data_ptr in uniqued:
+            continue
+        uniqued[v_data_ptr] = k
+
+    uniqued_names = list(uniqued.values())
+
+    for i in range(1, n):
+        state_dict = torch.load(filenames[i], map_location=device)["model"]
+        for k in uniqued_names:
+            avg[k] += state_dict[k]
+
+    for k in uniqued_names:
+        if avg[k].is_floating_point():
+            avg[k] /= n
+        else:
+            avg[k] //= n
+
+    return avg
+
+
+def save_checkpoint_with_global_batch_idx(
+    out_dir: Path,
+    global_batch_idx: int,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    params: Optional[Dict[str, Any]] = None,
+    optimizer: Optional[Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+):
+    """Save training info after processing given number of batches.
+
+    Args:
+      out_dir:
+        The directory to save the checkpoint.
+      global_batch_idx:
+        The number of batches processed so far from the very start of the
+        training. The saved checkpoint will have the following filename:
+
+            f'out_dir / checkpoint-{global_batch_idx}.pt'
+      model:
+        The neural network model whose `state_dict` will be saved in the
+        checkpoint.
+      model_avg:
+        The stored model averaged from the start of training.
+      params:
+        A dict of training configurations to be saved.
+      optimizer:
+        The optimizer used in the training. Its `state_dict` will be saved.
+      scheduler:
+        The learning rate scheduler used in the training. Its `state_dict` will
+        be saved.
+      scaler:
+        The scaler used for mix precision training. Its `state_dict` will
+        be saved.
+      sampler:
+        The sampler used in the training dataset.
+      rank:
+        The rank ID used in DDP training of the current node. Set it to 0
+        if DDP is not used.
+    """
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    filename = out_dir / f"checkpoint-{global_batch_idx}.pt"
+    save_checkpoint(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        scaler=scaler,
+        sampler=sampler,
+        rank=rank,
+    )
+
+
+def find_checkpoints(out_dir: Path, iteration: int = 0) -> List[str]:
+    """Find all available checkpoints in a directory.
+
+    The checkpoint filenames have the form: `checkpoint-xxx.pt`
+    where xxx is a numerical value.
+
+    Assume you have the following checkpoints in the folder `foo`:
+
+        - checkpoint-1.pt
+        - checkpoint-20.pt
+        - checkpoint-300.pt
+        - checkpoint-4000.pt
+
+    Case 1 (Return all checkpoints)::
+
+      find_checkpoints(out_dir='foo')
+
+    Case 2 (Return checkpoints newer than checkpoint-20.pt, i.e.,
+    checkpoint-4000.pt, checkpoint-300.pt, and checkpoint-20.pt)
+
+        find_checkpoints(out_dir='foo', iteration=20)
+
+    Case 3 (Return checkpoints older than checkpoint-20.pt, i.e.,
+    checkpoint-20.pt, checkpoint-1.pt)::
+
+        find_checkpoints(out_dir='foo', iteration=-20)
+
+    Args:
+      out_dir:
+        The directory where to search for checkpoints.
+      iteration:
+        If it is 0, return all available checkpoints.
+        If it is positive, return the checkpoints whose iteration number is
+        greater than or equal to `iteration`.
+        If it is negative, return the checkpoints whose iteration number is
+        less than or equal to `-iteration`.
+    Returns:
+      Return a list of checkpoint filenames, sorted in descending
+      order by the numerical value in the filename.
+    """
+    checkpoints = list(glob.glob(f"{out_dir}/checkpoint-[0-9]*.pt"))
+    pattern = re.compile(r"checkpoint-([0-9]+).pt")
+    iter_checkpoints = []
+    for c in checkpoints:
+        result = pattern.search(c)
+        if not result:
+            logging.warn(f"Invalid checkpoint filename {c}")
+            continue
+
+        iter_checkpoints.append((int(result.group(1)), c))
+
+    # iter_checkpoints is a list of tuples. Each tuple contains
+    # two elements: (iteration_number, checkpoint-iteration_number.pt)
+
+    iter_checkpoints = sorted(iter_checkpoints, reverse=True, key=lambda x: x[0])
+    if iteration >= 0:
+        ans = [ic[1] for ic in iter_checkpoints if ic[0] >= iteration]
+    else:
+        ans = [ic[1] for ic in iter_checkpoints if ic[0] <= -iteration]
+
+    return ans
+
+
+def remove_checkpoints(
+    out_dir: Path,
+    topk: int,
+    rank: int = 0,
+):
+    """Remove checkpoints from the given directory.
+
+    We assume that checkpoint filename has the form `checkpoint-xxx.pt`
+    where xxx is a number, representing the number of processed batches
+    when saving that checkpoint. We sort checkpoints by filename and keep
+    only the `topk` checkpoints with the highest `xxx`.
+
+    Args:
+      out_dir:
+        The directory containing checkpoints to be removed.
+      topk:
+        Number of checkpoints to keep.
+      rank:
+        If using DDP for training, it is the rank of the current node.
+        Use 0 if no DDP is used for training.
+    """
+    assert topk >= 1, topk
+    if rank != 0:
+        return
+    checkpoints = find_checkpoints(out_dir)
+
+    if len(checkpoints) == 0:
+        logging.warn(f"No checkpoints found in {out_dir}")
+        return
+
+    if len(checkpoints) <= topk:
+        return
+
+    to_remove = checkpoints[topk:]
+    for c in to_remove:
+        os.remove(c)
+
+
+def update_averaged_model(
+    params: Dict[str, Tensor],
+    model_cur: Union[nn.Module, DDP],
+    model_avg: nn.Module,
+) -> None:
+    """Update the averaged model:
+    model_avg = model_cur * (average_period / batch_idx_train)
+      + model_avg * ((batch_idx_train - average_period) / batch_idx_train)
+
+    Args:
+      params:
+        User defined parameters, e.g., epoch, loss.
+      model_cur:
+        The current model.
+      model_avg:
+        The averaged model to be updated.
+    """
+    weight_cur = params.average_period / params.batch_idx_train
+    weight_avg = 1 - weight_cur
+
+    if isinstance(model_cur, DDP):
+        model_cur = model_cur.module
+
+    cur = model_cur.state_dict()
+    avg = model_avg.state_dict()
+
+    average_state_dict(
+        state_dict_1=avg,
+        state_dict_2=cur,
+        weight_1=weight_avg,
+        weight_2=weight_cur,
+    )
+
+
+def average_checkpoints_with_averaged_model(
+    filename_start: str,
+    filename_end: str,
+    device: torch.device = torch.device("cpu"),
+) -> Dict[str, Tensor]:
+    """Average model parameters over the range with given
+    start model (excluded) and end model.
+
+    Let start = batch_idx_train of model-start;
+        end = batch_idx_train of model-end;
+        interval = end - start.
+    Then the average model over range from start (excluded) to end is
+    (1) avg = (model_end * end - model_start * start) / interval.
+    It can be written as
+    (2) avg = model_end * weight_end + model_start * weight_start,
+        where weight_end = end / interval,
+              weight_start = -start / interval = 1 - weight_end.
+    Since the terms `weight_end` and `weight_start` would be large
+    if the model has been trained for lots of batches, which would cause
+    overflow when multiplying the model parameters.
+    To avoid this, we rewrite (2) as:
+    (3) avg = (model_end + model_start * (weight_start / weight_end))
+              * weight_end
+
+    The model index could be epoch number or iteration number.
+
+    Args:
+      filename_start:
+        Checkpoint filename of the start model. We assume it
+        is saved by :func:`save_checkpoint`.
+      filename_end:
+        Checkpoint filename of the end model. We assume it
+        is saved by :func:`save_checkpoint`.
+      device:
+        Move checkpoints to this device before averaging.
+    """
+    state_dict_start = torch.load(filename_start, map_location=device)
+    state_dict_end = torch.load(filename_end, map_location=device)
+
+    batch_idx_train_start = state_dict_start["batch_idx_train"]
+    batch_idx_train_end = state_dict_end["batch_idx_train"]
+    interval = batch_idx_train_end - batch_idx_train_start
+    assert interval > 0, interval
+    weight_end = batch_idx_train_end / interval
+    weight_start = 1 - weight_end
+
+    model_end = state_dict_end["model_avg"]
+    model_start = state_dict_start["model_avg"]
+    avg = model_end
+
+    # scale the weight to avoid overflow
+    average_state_dict(
+        state_dict_1=avg,
+        state_dict_2=model_start,
+        weight_1=1.0,
+        weight_2=weight_start / weight_end,
+        scaling_factor=weight_end,
+    )
+
+    return avg
+
+
+def average_state_dict(
+    state_dict_1: Dict[str, Tensor],
+    state_dict_2: Dict[str, Tensor],
+    weight_1: float,
+    weight_2: float,
+    scaling_factor: float = 1.0,
+) -> Dict[str, Tensor]:
+    """Average two state_dict with given weights:
+    state_dict_1 = (state_dict_1 * weight_1 + state_dict_2 * weight_2)
+      * scaling_factor
+    It is an in-place operation on state_dict_1 itself.
+    """
+    # Identify shared parameters. Two parameters are said to be shared
+    # if they have the same data_ptr
+    uniqued: Dict[int, str] = dict()
+    for k, v in state_dict_1.items():
+        v_data_ptr = v.data_ptr()
+        if v_data_ptr in uniqued:
+            continue
+        uniqued[v_data_ptr] = k
+
+    uniqued_names = list(uniqued.values())
+    for k in uniqued_names:
+        v = state_dict_1[k]
+        if torch.is_floating_point(v):
+            v *= weight_1
+            v += state_dict_2[k].to(device=state_dict_1[k].device) * weight_2
+            v *= scaling_factor
diff --git a/icefall/context_graph.py b/icefall/context_graph.py
new file mode 100644
index 000000000..b3d7972a8
--- /dev/null
+++ b/icefall/context_graph.py
@@ -0,0 +1,453 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import shutil
+from collections import deque
+from typing import Dict, List, Optional, Tuple
+
+
+class ContextState:
+    """The state in ContextGraph"""
+
+    def __init__(
+        self,
+        id: int,
+        token: int,
+        token_score: float,
+        node_score: float,
+        output_score: float,
+        is_end: bool,
+    ):
+        """Create a ContextState.
+
+        Args:
+          id:
+            The node id, only for visualization now. A node is in [0, graph.num_nodes).
+            The id of the root node is always 0.
+          token:
+            The token id.
+          token_score:
+            The bonus for each token during decoding, which will hopefully
+            boost the token up to survive beam search.
+          node_score:
+            The accumulated bonus from root of graph to current node, it will be
+            used to calculate the score for fail arc.
+          output_score:
+            The total scores of matched phrases, sum of the node_score of all
+            the output node for current node.
+          is_end:
+            True if current token is the end of a context.
+        """
+        self.id = id
+        self.token = token
+        self.token_score = token_score
+        self.node_score = node_score
+        self.output_score = output_score
+        self.is_end = is_end
+        self.next = {}
+        self.fail = None
+        self.output = None
+
+
+class ContextGraph:
+    """The ContextGraph is modified from Aho-Corasick which is mainly
+    a Trie with a fail arc for each node.
+    See https://en.wikipedia.org/wiki/Aho%E2%80%93Corasick_algorithm for more details
+    of Aho-Corasick algorithm.
+
+    A ContextGraph contains some words / phrases that we expect to boost their
+    scores during decoding. If the substring of a decoded sequence matches the word / phrase
+    in the ContextGraph, we will give the decoded sequence a bonus to make it survive
+    beam search.
+    """
+
+    def __init__(self, context_score: float):
+        """Initialize a ContextGraph with the given ``context_score``.
+
+        A root node will be created (**NOTE:** the token of root is hardcoded to -1).
+
+        Args:
+          context_score:
+            The bonus score for each token(note: NOT for each word/phrase, it means longer
+            word/phrase will have larger bonus score, they have to be matched though).
+            Note: This is just the default score for each token, the users can manually
+            specify the context_score for each word/phrase (i.e. different phrase might
+            have different token score).
+        """
+        self.context_score = context_score
+        self.num_nodes = 0
+        self.root = ContextState(
+            id=self.num_nodes,
+            token=-1,
+            token_score=0,
+            node_score=0,
+            output_score=0,
+            is_end=False,
+        )
+        self.root.fail = self.root
+
+    def _fill_fail_output(self):
+        """This function fills the fail arc for each trie node, it can be computed
+        in linear time by performing a breadth-first search starting from the root.
+        See https://en.wikipedia.org/wiki/Aho%E2%80%93Corasick_algorithm for the
+        details of the algorithm.
+        """
+        queue = deque()
+        for token, node in self.root.next.items():
+            node.fail = self.root
+            queue.append(node)
+        while queue:
+            current_node = queue.popleft()
+            for token, node in current_node.next.items():
+                fail = current_node.fail
+                if token in fail.next:
+                    fail = fail.next[token]
+                else:
+                    fail = fail.fail
+                    while token not in fail.next:
+                        fail = fail.fail
+                        if fail.token == -1:  # root
+                            break
+                    if token in fail.next:
+                        fail = fail.next[token]
+                node.fail = fail
+                # fill the output arc
+                output = node.fail
+                while not output.is_end:
+                    output = output.fail
+                    if output.token == -1:  # root
+                        output = None
+                        break
+                node.output = output
+                node.output_score += 0 if output is None else output.output_score
+                queue.append(node)
+
+    def build(self, token_ids: List[Tuple[List[int], float]]):
+        """Build the ContextGraph from a list of token list.
+        It first build a trie from the given token lists, then fill the fail arc
+        for each trie node.
+
+        See https://en.wikipedia.org/wiki/Trie for how to build a trie.
+
+        Args:
+          token_ids:
+            The given token lists to build the ContextGraph, it is a list of tuple of
+            token list and its customized score, the token list contains the token ids
+            for a word/phrase. The token id could be an id of a char
+            (modeling with single Chinese char) or an id of a BPE
+            (modeling with BPEs). The score is the total score for current token list,
+            0 means using the default value (i.e. self.context_score).
+
+        Note: The phrases would have shared states, the score of the shared states is
+              the maximum value among all the tokens sharing this state.
+        """
+        for (tokens, score) in token_ids:
+            node = self.root
+            # If has customized score using the customized token score, otherwise
+            # using the default score
+            context_score = (
+                self.context_score if score == 0.0 else round(score / len(tokens), 2)
+            )
+            for i, token in enumerate(tokens):
+                node_next = {}
+                if token not in node.next:
+                    self.num_nodes += 1
+                    node_id = self.num_nodes
+                    token_score = context_score
+                    is_end = i == len(tokens) - 1
+                else:
+                    # node exists, get the score of shared state.
+                    token_score = max(context_score, node.next[token].token_score)
+                    node_id = node.next[token].id
+                    node_next = node.next[token].next
+                    is_end = i == len(tokens) - 1 or node.next[token].is_end
+                node_score = node.node_score + token_score
+                node.next[token] = ContextState(
+                    id=node_id,
+                    token=token,
+                    token_score=token_score,
+                    node_score=node_score,
+                    output_score=node_score if is_end else 0,
+                    is_end=is_end,
+                )
+                node.next[token].next = node_next
+                node = node.next[token]
+        self._fill_fail_output()
+
+    def forward_one_step(
+        self, state: ContextState, token: int
+    ) -> Tuple[float, ContextState]:
+        """Search the graph with given state and token.
+
+        Args:
+          state:
+            The given token containing trie node to start.
+          token:
+            The given token.
+
+        Returns:
+          Return a tuple of score and next state.
+        """
+        node = None
+        score = 0
+        # token matched
+        if token in state.next:
+            node = state.next[token]
+            score = node.token_score
+        else:
+            # token not matched
+            # We will trace along the fail arc until it matches the token or reaching
+            # root of the graph.
+            node = state.fail
+            while token not in node.next:
+                node = node.fail
+                if node.token == -1:  # root
+                    break
+
+            if token in node.next:
+                node = node.next[token]
+
+            # The score of the fail path
+            score = node.node_score - state.node_score
+        assert node is not None
+        return (score + node.output_score, node)
+
+    def finalize(self, state: ContextState) -> Tuple[float, ContextState]:
+        """When reaching the end of the decoded sequence, we need to finalize
+        the matching, the purpose is to subtract the added bonus score for the
+        state that is not the end of a word/phrase.
+
+        Args:
+          state:
+            The given state(trie node).
+
+        Returns:
+          Return a tuple of score and next state. If state is the end of a word/phrase
+          the score is zero, otherwise the score is the score of a implicit fail arc
+          to root. The next state is always root.
+        """
+        # The score of the fail arc
+        score = -state.node_score
+        return (score, self.root)
+
+    def draw(
+        self,
+        title: Optional[str] = None,
+        filename: Optional[str] = "",
+        symbol_table: Optional[Dict[int, str]] = None,
+    ) -> "Digraph":  # noqa
+        """Visualize a ContextGraph via graphviz.
+
+        Render ContextGraph as an image via graphviz, and return the Digraph object;
+        and optionally save to file `filename`.
+        `filename` must have a suffix that graphviz understands, such as
+        `pdf`, `svg` or `png`.
+
+        Note:
+          You need to install graphviz to use this function::
+
+            pip install graphviz
+
+        Args:
+           title:
+              Title to be displayed in image, e.g. 'A simple FSA example'
+           filename:
+              Filename to (optionally) save to, e.g. 'foo.png', 'foo.svg',
+              'foo.png'  (must have a suffix that graphviz understands).
+           symbol_table:
+              Map the token ids to symbols.
+        Returns:
+          A Diagraph from grahpviz.
+        """
+
+        try:
+            import graphviz
+        except Exception:
+            print("You cannot use `to_dot` unless the graphviz package is installed.")
+            raise
+
+        graph_attr = {
+            "rankdir": "LR",
+            "size": "8.5,11",
+            "center": "1",
+            "orientation": "Portrait",
+            "ranksep": "0.4",
+            "nodesep": "0.25",
+        }
+        if title is not None:
+            graph_attr["label"] = title
+
+        default_node_attr = {
+            "shape": "circle",
+            "style": "bold",
+            "fontsize": "14",
+        }
+
+        final_state_attr = {
+            "shape": "doublecircle",
+            "style": "bold",
+            "fontsize": "14",
+        }
+
+        final_state = -1
+        dot = graphviz.Digraph(name="Context Graph", graph_attr=graph_attr)
+
+        seen = set()
+        queue = deque()
+        queue.append(self.root)
+        # root id is always 0
+        dot.node("0", label="0", **default_node_attr)
+        dot.edge("0", "0", color="red")
+        seen.add(0)
+
+        while len(queue):
+            current_node = queue.popleft()
+            for token, node in current_node.next.items():
+                if node.id not in seen:
+                    node_score = f"{node.node_score:.2f}".rstrip("0").rstrip(".")
+                    output_score = f"{node.output_score:.2f}".rstrip("0").rstrip(".")
+                    label = f"{node.id}/({node_score}, {output_score})"
+                    if node.is_end:
+                        dot.node(str(node.id), label=label, **final_state_attr)
+                    else:
+                        dot.node(str(node.id), label=label, **default_node_attr)
+                    seen.add(node.id)
+                weight = f"{node.token_score:.2f}".rstrip("0").rstrip(".")
+                label = str(token) if symbol_table is None else symbol_table[token]
+                dot.edge(str(current_node.id), str(node.id), label=f"{label}/{weight}")
+                dot.edge(
+                    str(node.id),
+                    str(node.fail.id),
+                    color="red",
+                )
+                if node.output is not None:
+                    dot.edge(
+                        str(node.id),
+                        str(node.output.id),
+                        color="green",
+                    )
+                queue.append(node)
+
+        if filename:
+            _, extension = os.path.splitext(filename)
+            if extension == "" or extension[0] != ".":
+                raise ValueError(
+                    "Filename needs to have a suffix like .png, .pdf, .svg: {}".format(
+                        filename
+                    )
+                )
+
+            import tempfile
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                temp_fn = dot.render(
+                    filename="temp",
+                    directory=tmp_dir,
+                    format=extension[1:],
+                    cleanup=True,
+                )
+
+                shutil.move(temp_fn, filename)
+
+        return dot
+
+
+def _test(queries, score):
+    contexts_str = [
+        "S",
+        "HE",
+        "SHE",
+        "SHELL",
+        "HIS",
+        "HERS",
+        "HELLO",
+        "THIS",
+        "THEM",
+    ]
+
+    # test default score (1)
+    contexts = []
+    for s in contexts_str:
+        contexts.append(([ord(x) for x in s], score))
+
+    context_graph = ContextGraph(context_score=1)
+    context_graph.build(contexts)
+
+    symbol_table = {}
+    for contexts in contexts_str:
+        for s in contexts:
+            symbol_table[ord(s)] = s
+
+    context_graph.draw(
+        title="Graph for: " + " / ".join(contexts_str),
+        filename=f"context_graph_{score}.pdf",
+        symbol_table=symbol_table,
+    )
+
+    for query, expected_score in queries.items():
+        total_scores = 0
+        state = context_graph.root
+        for q in query:
+            score, state = context_graph.forward_one_step(state, ord(q))
+            total_scores += score
+        score, state = context_graph.finalize(state)
+        assert state.token == -1, state.token
+        total_scores += score
+        assert round(total_scores, 2) == expected_score, (
+            total_scores,
+            expected_score,
+            query,
+        )
+
+
+if __name__ == "__main__":
+    # test default score
+    queries = {
+        "HEHERSHE": 14,  # "HE", "HE", "HERS", "S", "SHE", "HE"
+        "HERSHE": 12,  # "HE", "HERS", "S", "SHE", "HE"
+        "HISHE": 9,  # "HIS", "S", "SHE", "HE"
+        "SHED": 6,  # "S", "SHE", "HE"
+        "SHELF": 6,  # "S", "SHE", "HE"
+        "HELL": 2,  # "HE"
+        "HELLO": 7,  # "HE", "HELLO"
+        "DHRHISQ": 4,  # "HIS", "S"
+        "THEN": 2,  # "HE"
+    }
+    _test(queries, 0)
+
+    # test custom score (5)
+    # S : 5
+    # HE : 5 (2.5 + 2.5)
+    # SHE : 8.34 (5 + 1.67 + 1.67)
+    # SHELL : 10.34 (5 + 1.67 + 1.67 + 1 + 1)
+    # HIS : 5.84 (2.5 + 1.67 + 1.67)
+    # HERS : 7.5 (2.5 + 2.5 + 1.25 + 1.25)
+    # HELLO : 8 (2.5 + 2.5 + 1 + 1 + 1)
+    # THIS : 5 (1.25 + 1.25 + 1.25 + 1.25)
+    queries = {
+        "HEHERSHE": 35.84,  # "HE", "HE", "HERS", "S", "SHE", "HE"
+        "HERSHE": 30.84,  # "HE", "HERS", "S", "SHE", "HE"
+        "HISHE": 24.18,  # "HIS", "S", "SHE", "HE"
+        "SHED": 18.34,  # "S", "SHE", "HE"
+        "SHELF": 18.34,  # "S", "SHE", "HE"
+        "HELL": 5,  # "HE"
+        "HELLO": 13,  # "HE", "HELLO"
+        "DHRHISQ": 10.84,  # "HIS", "S"
+        "THEN": 5,  # "HE"
+    }
+
+    _test(queries, 5)
diff --git a/icefall/ctc/.gitignore b/icefall/ctc/.gitignore
new file mode 100644
index 000000000..8154cb57f
--- /dev/null
+++ b/icefall/ctc/.gitignore
@@ -0,0 +1,2 @@
+*.pdf
+*.gv
diff --git a/icefall/ctc/README.md b/icefall/ctc/README.md
new file mode 100644
index 000000000..0096bc096
--- /dev/null
+++ b/icefall/ctc/README.md
@@ -0,0 +1,17 @@
+# Introduction
+
+This folder uses [kaldifst][kaldifst] for graph construction
+and decoders from [kaldi-decoder][kaldi-decoder] for CTC decoding.
+
+It supports only `CPU`.
+
+You can use
+
+```bash
+pip install kaldifst kaldi-decoder
+```
+to install the dependencies.
+
+[kaldi-decoder]: https://github.com/i2-fsa/kaldi-decoder
+[kaldifst]: https://github.com/k2-fsa/kaldifst
+[k2]: https://github.com/k2-fsa/k2
diff --git a/icefall/ctc/__init__.py b/icefall/ctc/__init__.py
new file mode 100644
index 000000000..b546b31af
--- /dev/null
+++ b/icefall/ctc/__init__.py
@@ -0,0 +1,6 @@
+from .prepare_lang import (
+    Lexicon,
+    make_lexicon_fst_no_silence,
+    make_lexicon_fst_with_silence,
+)
+from .topo import add_disambig_self_loops, add_one, build_standard_ctc_topo
diff --git a/icefall/ctc/prepare_lang.py b/icefall/ctc/prepare_lang.py
new file mode 100644
index 000000000..4801b1beb
--- /dev/null
+++ b/icefall/ctc/prepare_lang.py
@@ -0,0 +1,334 @@
+# Copyright    2023  Xiaomi Corp.        (author: Fangjun Kuang)
+
+"""
+The lang_dir should contain the following files:
+ - "lexicon_disambig.txt"
+ - "tokens.txt"
+ - "words.txt"
+"""
+
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import List, Tuple
+
+import kaldifst
+import re
+
+
+class Lexicon:
+    """Once constructed it is immutable"""
+
+    def __init__(
+        self,
+        lang_dir: str,
+        disambig_pattern: str = re.compile(r"^#\d+$"),
+    ):
+        """
+        Args:
+          lang_dir:
+            The path to the lang directory. We expect that it contains the
+            following files:
+              - lexicon_disambig.txt
+              - tokens.txt
+              - words.txt
+
+            The format of the above files is described below.
+
+            (1) lexicon_disambig.txt
+
+            Each line in the lexicon_disambig.txt has the following format:
+
+                word token1 token2 ... tokenN
+
+            That is, the first field is the word, the remaining fields are
+            pronunciations of this word. Fields are separated by space(s).
+
+            (2) tokens.txt
+
+            Each line in tokens.txt has two fields separated by space(s):
+
+                token ID
+
+            The first field is the token symbol and the second filed is the
+            integer ID of the token.
+
+            (3) words.txt
+
+            Each line in words.txt has two fields separated by space(s):
+
+                word ID
+
+            The first field is the word symbol and the second filed is the
+            integer ID of the word.
+          disambig_pattern:
+            It contains the pattern for disambiguation symbols.
+        """
+        lang_dir = Path(lang_dir)
+
+        lexicon_txt = lang_dir / "lexicon_disambig.txt"
+        tokens_txt = lang_dir / "tokens.txt"
+        words_txt = lang_dir / "words.txt"
+
+        assert lexicon_txt.is_file(), lexicon_txt
+        assert tokens_txt.is_file(), tokens_txt
+        assert words_txt.is_file(), words_txt
+
+        self._read_lexicon(lexicon_txt)
+        self._read_tokens(tokens_txt)
+        self._read_words(words_txt)
+
+        self.disambig_pattern = disambig_pattern
+
+        max_disambig_id = -1
+        for s, i in self.token2id.items():
+            if self.disambig_pattern.match(s) and i > max_disambig_id:
+                max_disambig_id = i
+
+        self.max_disambig_id = max_disambig_id
+
+    def _read_lexicon(self, lexicon_txt: str):
+        word2phones = defaultdict(list)
+        with open(lexicon_txt, encoding="utf-8") as f:
+            for line in f:
+                word_phones = line.strip().split()
+                assert len(word_phones) >= 2, (word_phones, line)
+                word = word_phones[0]
+                phones: str = " ".join(word_phones[1:])
+                word2phones[word].append(phones)
+                # We use a list here since a word may have multiple
+                # pronunciations
+
+        self.word2phones = word2phones
+
+    def _read_tokens(self, tokens_txt):
+        token2id = dict()
+        id2token = dict()
+        with open(tokens_txt, encoding="utf-8") as f:
+            for line in f:
+                token_id = line.strip().split()
+                assert len(token_id) == 2, token_id
+
+                token = token_id[0]
+                idx = int(token_id[1])
+
+                assert token not in token2id, f"Duplicate token {line}"
+                assert idx not in id2token, f"Duplicate ID {line}"
+
+                token2id[token] = idx
+                id2token[idx] = token
+        self.token2id = token2id
+        self.id2token = id2token
+
+    def _read_words(self, words_txt):
+        word2id = dict()
+        id2word = dict()
+        with open(words_txt, encoding="utf-8") as f:
+            for line in f:
+                word_id = line.strip().split()
+                assert len(word_id) == 2, word_id
+
+                word = word_id[0]
+                idx = int(word_id[1])
+
+                assert word not in word2id, f"Duplicate token {line}"
+                assert idx not in id2word, f"Duplicate ID {line}"
+
+                word2id[word] = idx
+                id2word[idx] = word
+
+        self.word2id = word2id
+        self.id2word = id2word
+
+    def __iter__(self) -> Tuple[str, List[str]]:
+        for word, phones_list in self.word2phones.items():
+            for phones in phones_list:
+                yield word, phones
+
+    def __str__(self):
+        return str(self.word2phones)
+
+    @property
+    def tokens(self) -> List[int]:
+        """Return a list of token IDs excluding those from
+        disambiguation symbols.
+
+        Caution:
+          0 is not a token ID so it is excluded from the return value.
+        """
+        ans = []
+        for s in self.token2id:
+            if not self.disambig_pattern.match(s):
+                ans.append(self.token2id[s])
+        if 0 in ans:
+            ans.remove(0)
+        ans.sort()
+        return ans
+
+
+# See also
+# http://vpanayotov.blogspot.com/2012/06/kaldi-decoding-graph-construction.html
+def make_lexicon_fst_with_silence(
+    lexicon: Lexicon,
+    sil_prob: float = 0.5,
+    sil_phone: str = "SIL",
+    attach_symbol_table: bool = True,
+) -> kaldifst.StdVectorFst:
+    phone2id = lexicon.token2id
+    word2id = lexicon.word2id
+
+    assert sil_phone in phone2id
+
+    assert sil_phone in phone2id, sil_phone
+
+    sil_cost = -1 * math.log(sil_prob)
+    no_sil_cost = -1 * math.log(1.0 - sil_prob)
+
+    fst = kaldifst.StdVectorFst()
+
+    start_state = fst.add_state()
+    loop_state = fst.add_state()
+    sil_state = fst.add_state()
+
+    fst.start = start_state
+    fst.set_final(state=loop_state, weight=0)
+
+    fst.add_arc(
+        state=start_state,
+        arc=kaldifst.StdArc(
+            ilabel=0,
+            olabel=0,
+            weight=no_sil_cost,
+            nextstate=loop_state,
+        ),
+    )
+
+    fst.add_arc(
+        state=start_state,
+        arc=kaldifst.StdArc(
+            ilabel=0,
+            olabel=0,
+            weight=sil_cost,
+            nextstate=sil_state,
+        ),
+    )
+
+    fst.add_arc(
+        state=sil_state,
+        arc=kaldifst.StdArc(
+            ilabel=phone2id[sil_phone],
+            olabel=0,
+            weight=0,
+            nextstate=loop_state,
+        ),
+    )
+
+    for word, phones in lexicon:
+        phoneseq = phones.split()
+        pron_cost = 0
+        cur_state = loop_state
+
+        for i in range(len(phoneseq) - 1):
+            next_state = fst.add_state()
+            fst.add_arc(
+                state=cur_state,
+                arc=kaldifst.StdArc(
+                    ilabel=phone2id[phoneseq[i]],
+                    olabel=word2id[word] if i == 0 else 0,
+                    weight=pron_cost if i == 0 else 0,
+                    nextstate=next_state,
+                ),
+            )
+            cur_state = next_state
+
+        i = len(phoneseq) - 1  # note: i == -1 if phoneseq is empty.
+
+        fst.add_arc(
+            state=cur_state,
+            arc=kaldifst.StdArc(
+                ilabel=phone2id[phoneseq[i]] if i >= 0 else 0,
+                olabel=word2id[word] if i <= 0 else 0,
+                weight=no_sil_cost + (pron_cost if i <= 0 else 0),
+                nextstate=loop_state,
+            ),
+        )
+
+        fst.add_arc(
+            state=cur_state,
+            arc=kaldifst.StdArc(
+                ilabel=phone2id[phoneseq[i]] if i >= 0 else 0,
+                olabel=word2id[word] if i <= 0 else 0,
+                weight=sil_cost + (pron_cost if i <= 0 else 0),
+                nextstate=sil_state,
+            ),
+        )
+
+    if attach_symbol_table:
+        isym = kaldifst.SymbolTable()
+        for p, i in phone2id.items():
+            isym.add_symbol(symbol=p, key=i)
+        fst.input_symbols = isym
+
+        osym = kaldifst.SymbolTable()
+        for w, i in word2id.items():
+            osym.add_symbol(symbol=w, key=i)
+        fst.output_symbols = osym
+
+    return fst
+
+
+def make_lexicon_fst_no_silence(
+    lexicon: Lexicon,
+    attach_symbol_table: bool = True,
+) -> kaldifst.StdVectorFst:
+    phone2id = lexicon.token2id
+    word2id = lexicon.word2id
+
+    fst = kaldifst.StdVectorFst()
+
+    start_state = fst.add_state()
+    fst.start = start_state
+    fst.set_final(state=start_state, weight=0)
+
+    for word, phones in lexicon:
+        phoneseq = phones.split()
+        pron_cost = 0
+        cur_state = start_state
+
+        for i in range(len(phoneseq) - 1):
+            next_state = fst.add_state()
+            fst.add_arc(
+                state=cur_state,
+                arc=kaldifst.StdArc(
+                    ilabel=phone2id[phoneseq[i]],
+                    olabel=word2id[word] if i == 0 else 0,
+                    weight=pron_cost if i == 0 else 0,
+                    nextstate=next_state,
+                ),
+            )
+            cur_state = next_state
+
+        i = len(phoneseq) - 1  # note: i == -1 if phoneseq is empty.
+
+        fst.add_arc(
+            state=cur_state,
+            arc=kaldifst.StdArc(
+                ilabel=phone2id[phoneseq[i]] if i >= 0 else 0,
+                olabel=word2id[word] if i <= 0 else 0,
+                weight=pron_cost if i <= 0 else 0,
+                nextstate=start_state,
+            ),
+        )
+
+    if attach_symbol_table:
+        isym = kaldifst.SymbolTable()
+        for p, i in phone2id.items():
+            isym.add_symbol(symbol=p, key=i)
+        fst.input_symbols = isym
+
+        osym = kaldifst.SymbolTable()
+        for w, i in word2id.items():
+            osym.add_symbol(symbol=w, key=i)
+        fst.output_symbols = osym
+
+    return fst
diff --git a/icefall/ctc/test_ctc_topo.py b/icefall/ctc/test_ctc_topo.py
new file mode 100755
index 000000000..4d4667209
--- /dev/null
+++ b/icefall/ctc/test_ctc_topo.py
@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from pathlib import Path
+
+import graphviz
+import kaldifst
+import sentencepiece as spm
+from prepare_lang import (
+    Lexicon,
+    make_lexicon_fst_no_silence,
+    make_lexicon_fst_with_silence,
+)
+from topo import add_disambig_self_loops, add_one, build_standard_ctc_topo
+
+
+def test_yesno():
+    lang_dir = "/Users/fangjun/open-source/icefall/egs/yesno/ASR/data/lang_phone"
+    if not Path(lang_dir).is_dir():
+        print(f"{lang_dir} does not exist! Skip testing")
+        return
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    H = build_standard_ctc_topo(max_token_id=max_token_id)
+
+    isym = kaldifst.SymbolTable()
+    isym.add_symbol(symbol="<blk>", key=0)
+    for i in range(1, max_token_id + 1):
+        isym.add_symbol(symbol=lexicon.id2token[i], key=i)
+
+    osym = kaldifst.SymbolTable()
+    osym.add_symbol(symbol="<eps>", key=0)
+    for i in range(1, max_token_id + 1):
+        osym.add_symbol(symbol=lexicon.id2token[i], key=i)
+
+    H.input_symbols = isym
+    H.output_symbols = osym
+
+    fst_dot = kaldifst.draw(H, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="standard_ctc_topo_yesno.pdf")
+    # See the link below to visualize the above PDF
+    # https://t.ly/7uXZ9
+
+    # Now test HL
+
+    # We need to add one to all tokens since we want to use ID 0
+    # for epsilon
+    add_one(H, treat_ilabel_zero_specially=False, update_olabel=True)
+
+    add_disambig_self_loops(
+        H,
+        start=lexicon.token2id["#0"] + 1,
+        end=lexicon.max_disambig_id,
+    )
+
+    fst_dot = kaldifst.draw(H, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="standard_ctc_topo_disambig_yesno.pdf")
+
+    L = make_lexicon_fst_with_silence(lexicon)
+
+    # We also need to change the input labels of L
+    add_one(L, treat_ilabel_zero_specially=True, update_olabel=False)
+
+    H.output_symbols = None
+
+    kaldifst.arcsort(H, sort_type="olabel")
+    kaldifst.arcsort(L, sort_type="ilabel")
+    HL = kaldifst.compose(H, L)
+
+    lexicon.id2token[0] = "<blk>"
+    lexicon.token2id["<blk>"] = 0
+
+    isym = kaldifst.SymbolTable()
+    isym.add_symbol(symbol="<eps>", key=0)
+    for i in range(0, lexicon.max_disambig_id + 1):
+        isym.add_symbol(symbol=lexicon.id2token[i], key=i + 1)
+
+    osym = kaldifst.SymbolTable()
+    for i, word in lexicon.id2word.items():
+        osym.add_symbol(symbol=word, key=i)
+
+    HL.input_symbols = isym
+    HL.output_symbols = osym
+
+    fst_dot = kaldifst.draw(HL, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="HL_yesno.pdf")
+
+
+def test_librispeech():
+    lang_dir = (
+        "/star-fj/fangjun/open-source/icefall-2/egs/librispeech/ASR/data/lang_bpe_500"
+    )
+
+    if not Path(lang_dir).is_dir():
+        print(f"{lang_dir} does not exist! Skip testing")
+        return
+
+    lexicon = Lexicon(lang_dir)
+    HL = kaldifst.StdVectorFst.read(lang_dir + "/HL.fst")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(lang_dir + "/bpe.model")
+
+    i = lexicon.word2id["HELLOA"]
+    k = lexicon.word2id["WORLD"]
+    print(i, k)
+    s = f"""
+        0 1 {i} {i}
+        1 2 {k} {k}
+        2
+    """
+    fst = kaldifst.compile(
+        s=s,
+        acceptor=False,
+    )
+
+    L = make_lexicon_fst_no_silence(lexicon, attach_symbol_table=False)
+    kaldifst.arcsort(L, sort_type="olabel")
+    with open("L.fst.txt", "w") as f:
+        print(L, file=f)
+
+    fst = kaldifst.compose(L, fst)
+    print(fst)
+    fst_dot = kaldifst.draw(fst, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="a.pdf")
+    print(sp.encode(["HELLOA", "WORLD"]))
+
+
+def main():
+    test_yesno()
+    test_librispeech()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/ctc/test_prepare_lang.py b/icefall/ctc/test_prepare_lang.py
new file mode 100755
index 000000000..6c4b9e510
--- /dev/null
+++ b/icefall/ctc/test_prepare_lang.py
@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from pathlib import Path
+
+import graphviz
+import kaldifst
+from prepare_lang import Lexicon, make_lexicon_fst_with_silence
+
+
+def test_yesno():
+    lang_dir = "/Users/fangjun/open-source/icefall/egs/yesno/ASR/data/lang_phone"
+    if not Path(lang_dir).is_dir():
+        print(f"{lang_dir} does not exist! Skip testing")
+        return
+
+    lexicon = Lexicon(lang_dir)
+
+    L = make_lexicon_fst_with_silence(lexicon)
+
+    isym = kaldifst.SymbolTable()
+    for i, token in lexicon.id2token.items():
+        isym.add_symbol(symbol=token, key=i)
+
+    osym = kaldifst.SymbolTable()
+    for i, word in lexicon.id2word.items():
+        osym.add_symbol(symbol=word, key=i)
+
+    L.input_symbols = isym
+    L.output_symbols = osym
+    fst_dot = kaldifst.draw(L, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="L_yesno.pdf")
+    # See the link below to visualize the above PDF
+    # https://t.ly/jMfXW
+
+
+def main():
+    test_yesno()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/ctc/topo.py b/icefall/ctc/topo.py
new file mode 100644
index 000000000..6a96dd038
--- /dev/null
+++ b/icefall/ctc/topo.py
@@ -0,0 +1,137 @@
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import kaldifst
+
+
+# Note the name contains `standard`; it means there will be non-standard
+# topologies.
+def build_standard_ctc_topo(max_token_id: int) -> kaldifst.StdVectorFst:
+    """Build a standard CTC topology.
+
+    Args:
+      Maximum valid token ID. We assume token IDs are contiguous
+      and starts from 0. In other words, the vocabulary size is
+      ``max_token_id + 1``. We assume the ID of the blank symbol is 0.
+    """
+    # Token ID starts from 0 and there are as many states as the
+    # number of tokens.
+    #
+    # Note that epsilon is not a token and the token with ID 0 in tokens.txt
+    # is not an epsilon. It means input label 0 of the resulting FST does
+    # not represent an epsilon.
+    #
+    # You can use the function `add_one()` to modify the input/output labels
+    # of the resulting FST
+
+    num_states = max_token_id + 1
+
+    # Step 1: Create as many states as the number of tokens.
+    # Each state is a final state
+    fst = kaldifst.StdVectorFst()
+    for i in range(num_states):
+        s = fst.add_state()
+        fst.set_final(state=s, weight=0)
+
+    # Step 2: Set state 0 as the start state.
+    # We assume the ID of the blank symbol is 0.
+    fst.start = 0
+
+    # Step 3: Build a fully connected graph.
+    for i in range(num_states):
+        for k in range(num_states):
+            fst.add_arc(
+                state=i,
+                arc=kaldifst.StdArc(
+                    ilabel=k,
+                    olabel=k if i != k else 0,  # if i==k, it is a self loop
+                    weight=0,
+                    nextstate=k,
+                ),
+            )
+    # Please see ./test_ctc_topo.py if you want to know what the resulting
+    # FST looks like
+
+    return fst
+
+
+def add_one(
+    fst: kaldifst.StdVectorFst,
+    treat_ilabel_zero_specially: bool,
+    update_olabel: bool,
+) -> None:
+    """Modify the input and output labels of the given FST in-place.
+
+    Args:
+      fst:
+        The FST to be modified. It is changed in-place.
+      treat_ilabel_zero_specially:
+        If True, then every non-zero input label is increased by one and the
+        zero input label is not changed.
+        If False, then every input label is increased by one.
+      update_olabel:
+        If False, the output label is not changed.
+        If True, then every non-zero output label is increased by one.
+        In either case, output label with 0 is not changed.
+    """
+    for state in kaldifst.StateIterator(fst):
+        for arc in kaldifst.ArcIterator(fst, state):
+            # If treat_ilabel_zero_specially is False, we always change it
+            # Otherwise, we only change non-zero input labels
+            if treat_ilabel_zero_specially is False or arc.ilabel != 0:
+                arc.ilabel += 1
+
+            if update_olabel and arc.olabel != 0:
+                arc.olabel += 1
+
+    if fst.input_symbols is not None:
+        input_symbols = kaldifst.SymbolTable()
+        input_symbols.add_symbol(symbol="<eps>", key=0)
+
+        for i in range(0, fst.input_symbols.num_symbols()):
+            s = fst.input_symbols.find(i)
+            input_symbols.add_symbol(symbol=s, key=i + 1)
+
+        fst.input_symbols = input_symbols
+
+    if update_olabel and fst.output_symbols is not None:
+        output_symbols = kaldifst.SymbolTable()
+        output_symbols.add_symbol(symbol="<eps>", key=0)
+
+        for i in range(0, fst.output_symbols.num_symbols()):
+            s = fst.output_symbols.find(i)
+            output_symbols.add_symbol(symbol=s, key=i + 1)
+
+        fst.output_symbols = output_symbols
+
+
+def add_disambig_self_loops(fst: kaldifst.StdVectorFst, start: int, end: int):
+    """Add self-loops to each state.
+
+    For each disambig symbol, we add a self-loop with input label disambig_id
+    and output label diambig_id of that disambig symbol.
+
+    Args:
+      fst:
+        It is changed in-place.
+      start:
+        The ID of #0
+      end:
+        The ID of the last disambig symbol. For instance if there are 3
+        disambig symbols ``#0``, ``#1``, and ``#2``, then ``end`` is the ID
+        of ``#2``.
+    """
+    for state in kaldifst.StateIterator(fst):
+        for i in range(start, end + 1):
+            fst.add_arc(
+                state=state,
+                arc=kaldifst.StdArc(
+                    ilabel=i,
+                    olabel=i,
+                    weight=0,
+                    nextstate=state,
+                ),
+            )
+
+    if fst.output_symbols:
+        for i in range(start, end + 1):
+            fst.output_symbols.add_symbol(symbol=f"#{i-start}", key=i)
diff --git a/icefall/dataset/__init__.py b/icefall/dataset/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/icefall/dataset/datamodule.py b/icefall/dataset/datamodule.py
new file mode 100644
index 000000000..97918ffd8
--- /dev/null
+++ b/icefall/dataset/datamodule.py
@@ -0,0 +1,60 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+from typing import List, Union
+
+from lhotse import CutSet
+from torch.utils.data import DataLoader
+
+
+class DataModule:
+    """
+    Contains dataset-related code. It is intended to read/construct Lhotse cuts,
+    and create Dataset/Sampler/DataLoader out of them.
+
+    There is a separate method to create each of train/valid/test DataLoader.
+    In principle, there might be multiple DataLoaders for each of
+    train/valid/test
+    (e.g. when a corpus has multiple test sets).
+    The API of this class allows to return lists of CutSets/DataLoaders.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        pass
+
+    def train_cuts(self) -> Union[CutSet, List[CutSet]]:
+        raise NotImplementedError()
+
+    def valid_cuts(self) -> Union[CutSet, List[CutSet]]:
+        raise NotImplementedError()
+
+    def test_cuts(self) -> Union[CutSet, List[CutSet]]:
+        raise NotImplementedError()
+
+    def train_dataloaders(self) -> Union[DataLoader, List[DataLoader]]:
+        raise NotImplementedError()
+
+    def valid_dataloaders(self) -> Union[DataLoader, List[DataLoader]]:
+        raise NotImplementedError()
+
+    def test_dataloaders(self) -> Union[DataLoader, List[DataLoader]]:
+        raise NotImplementedError()
diff --git a/icefall/decode.py b/icefall/decode.py
new file mode 100644
index 000000000..23f9fb9b3
--- /dev/null
+++ b/icefall/decode.py
@@ -0,0 +1,1243 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from typing import Dict, List, Optional, Union
+
+import k2
+import torch
+
+from icefall.utils import add_eos, add_sos, get_texts
+
+DEFAULT_LM_SCALE = [
+    0.01,
+    0.05,
+    0.08,
+    0.1,
+    0.3,
+    0.5,
+    0.6,
+    0.7,
+    0.9,
+    1.0,
+    1.1,
+    1.2,
+    1.3,
+    1.5,
+    1.7,
+    1.9,
+    2.0,
+    2.1,
+    2.2,
+    2.3,
+    2.5,
+    3.0,
+    4.0,
+    5.0,
+]
+
+
+def _intersect_device(
+    a_fsas: k2.Fsa,
+    b_fsas: k2.Fsa,
+    b_to_a_map: torch.Tensor,
+    sorted_match_a: bool,
+    batch_size: int = 50,
+) -> k2.Fsa:
+    """This is a wrapper of k2.intersect_device and its purpose is to split
+    b_fsas into several batches and process each batch separately to avoid
+    CUDA OOM error.
+
+    The arguments and return value of this function are the same as
+    :func:`k2.intersect_device`.
+    """
+    num_fsas = b_fsas.shape[0]
+    if num_fsas <= batch_size:
+        return k2.intersect_device(
+            a_fsas, b_fsas, b_to_a_map=b_to_a_map, sorted_match_a=sorted_match_a
+        )
+
+    num_batches = (num_fsas + batch_size - 1) // batch_size
+    splits = []
+    for i in range(num_batches):
+        start = i * batch_size
+        end = min(start + batch_size, num_fsas)
+        splits.append((start, end))
+
+    ans = []
+    for start, end in splits:
+        indexes = torch.arange(start, end).to(b_to_a_map)
+
+        fsas = k2.index_fsa(b_fsas, indexes)
+        b_to_a = k2.index_select(b_to_a_map, indexes)
+        path_lattice = k2.intersect_device(
+            a_fsas, fsas, b_to_a_map=b_to_a, sorted_match_a=sorted_match_a
+        )
+        ans.append(path_lattice)
+
+    return k2.cat(ans)
+
+
+def get_lattice(
+    nnet_output: torch.Tensor,
+    decoding_graph: k2.Fsa,
+    supervision_segments: torch.Tensor,
+    search_beam: float,
+    output_beam: float,
+    min_active_states: int,
+    max_active_states: int,
+    subsampling_factor: int = 1,
+) -> k2.Fsa:
+    """Get the decoding lattice from a decoding graph and neural
+    network output.
+    Args:
+      nnet_output:
+        It is the output of a neural model of shape `(N, T, C)`.
+      decoding_graph:
+        An Fsa, the decoding graph. It can be either an HLG
+        (see `compile_HLG.py`) or an H (see `k2.ctc_topo`).
+      supervision_segments:
+        A 2-D **CPU** tensor of dtype `torch.int32` with 3 columns.
+        Each row contains information for a supervision segment. Column 0
+        is the `sequence_index` indicating which sequence this segment
+        comes from; column 1 specifies the `start_frame` of this segment
+        within the sequence; column 2 contains the `duration` of this
+        segment.
+      search_beam:
+        Decoding beam, e.g. 20.  Smaller is faster, larger is more exact
+        (less pruning). This is the default value; it may be modified by
+        `min_active_states` and `max_active_states`.
+      output_beam:
+         Beam to prune output, similar to lattice-beam in Kaldi.  Relative
+         to best path of output.
+      min_active_states:
+        Minimum number of FSA states that are allowed to be active on any given
+        frame for any given intersection/composition task. This is advisory,
+        in that it will try not to have fewer than this number active.
+        Set it to zero if there is no constraint.
+      max_active_states:
+        Maximum number of FSA states that are allowed to be active on any given
+        frame for any given intersection/composition task. This is advisory,
+        in that it will try not to exceed that but may not always succeed.
+        You can use a very large number if no constraint is needed.
+      subsampling_factor:
+        The subsampling factor of the model.
+    Returns:
+      An FsaVec containing the decoding result. It has axes [utt][state][arc].
+    """
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=subsampling_factor - 1,
+    )
+
+    lattice = k2.intersect_dense_pruned(
+        decoding_graph,
+        dense_fsa_vec,
+        search_beam=search_beam,
+        output_beam=output_beam,
+        min_active_states=min_active_states,
+        max_active_states=max_active_states,
+    )
+
+    return lattice
+
+
+class Nbest(object):
+    """
+    An Nbest object contains two fields:
+
+        (1) fsa. It is an FsaVec containing a vector of **linear** FSAs.
+                 Its axes are [path][state][arc]
+        (2) shape. Its type is :class:`k2.RaggedShape`.
+                   Its axes are [utt][path]
+
+    The field `shape` has two axes [utt][path]. `shape.dim0` contains
+    the number of utterances, which is also the number of rows in the
+    supervision_segments. `shape.tot_size(1)` contains the number
+    of paths, which is also the number of FSAs in `fsa`.
+
+    Caution:
+      Don't be confused by the name `Nbest`. The best in the name `Nbest`
+      has nothing to do with `best scores`. The important part is
+      `N` in `Nbest`, not `best`.
+    """
+
+    def __init__(self, fsa: k2.Fsa, shape: k2.RaggedShape) -> None:
+        """
+        Args:
+          fsa:
+            An FsaVec with axes [path][state][arc]. It is expected to contain
+            a list of **linear** FSAs.
+          shape:
+            A ragged shape with two axes [utt][path].
+        """
+        assert len(fsa.shape) == 3, f"fsa.shape: {fsa.shape}"
+        assert shape.num_axes == 2, f"num_axes: {shape.num_axes}"
+
+        if fsa.shape[0] != shape.tot_size(1):
+            raise ValueError(
+                f"{fsa.shape[0]} vs {shape.tot_size(1)}\n"
+                "Number of FSAs in `fsa` does not match the given shape"
+            )
+
+        self.fsa = fsa
+        self.shape = shape
+
+    def __str__(self):
+        s = "Nbest("
+        s += f"Number of utterances:{self.shape.dim0}, "
+        s += f"Number of Paths:{self.fsa.shape[0]})"
+        return s
+
+    @staticmethod
+    def from_lattice(
+        lattice: k2.Fsa,
+        num_paths: int,
+        use_double_scores: bool = True,
+        nbest_scale: float = 0.5,
+    ) -> "Nbest":
+        """Construct an Nbest object by **sampling** `num_paths` from a lattice.
+
+        Each sampled path is a linear FSA.
+
+        We assume `lattice.labels` contains token IDs and `lattice.aux_labels`
+        contains word IDs.
+
+        Args:
+          lattice:
+            An FsaVec with axes [utt][state][arc].
+          num_paths:
+            Number of paths to **sample** from the lattice
+            using :func:`k2.random_paths`.
+          use_double_scores:
+            True to use double precision in :func:`k2.random_paths`.
+            False to use single precision.
+          scale:
+            Scale `lattice.score` before passing it to :func:`k2.random_paths`.
+            A smaller value leads to more unique paths at the risk of being not
+            to sample the path with the best score.
+        Returns:
+          Return an Nbest instance.
+        """
+        saved_scores = lattice.scores.clone()
+        lattice.scores *= nbest_scale
+        # path is a ragged tensor with dtype torch.int32.
+        # It has three axes [utt][path][arc_pos]
+        path = k2.random_paths(
+            lattice, num_paths=num_paths, use_double_scores=use_double_scores
+        )
+        lattice.scores = saved_scores
+
+        # word_seq is a k2.RaggedTensor sharing the same shape as `path`
+        # but it contains word IDs. Note that it also contains 0s and -1s.
+        # The last entry in each sublist is -1.
+        # It axes is [utt][path][word_id]
+        if isinstance(lattice.aux_labels, torch.Tensor):
+            word_seq = k2.ragged.index(lattice.aux_labels, path)
+        else:
+            word_seq = lattice.aux_labels.index(path)
+            word_seq = word_seq.remove_axis(word_seq.num_axes - 2)
+        word_seq = word_seq.remove_values_leq(0)
+
+        # Each utterance has `num_paths` paths but some of them transduces
+        # to the same word sequence, so we need to remove repeated word
+        # sequences within an utterance. After removing repeats, each utterance
+        # contains different number of paths
+        #
+        # `new2old` is a 1-D torch.Tensor mapping from the output path index
+        # to the input path index.
+        _, _, new2old = word_seq.unique(
+            need_num_repeats=False, need_new2old_indexes=True
+        )
+
+        # kept_path is a ragged tensor with dtype torch.int32.
+        # It has axes [utt][path][arc_pos]
+        kept_path, _ = path.index(new2old, axis=1, need_value_indexes=False)
+
+        # utt_to_path_shape has axes [utt][path]
+        utt_to_path_shape = kept_path.shape.get_layer(0)
+
+        # Remove the utterance axis.
+        # Now kept_path has only two axes [path][arc_pos]
+        kept_path = kept_path.remove_axis(0)
+
+        # labels is a ragged tensor with 2 axes [path][token_id]
+        # Note that it contains -1s.
+        labels = k2.ragged.index(lattice.labels.contiguous(), kept_path)
+
+        # Remove -1 from labels as we will use it to construct a linear FSA
+        labels = labels.remove_values_eq(-1)
+
+        if isinstance(lattice.aux_labels, k2.RaggedTensor):
+            # lattice.aux_labels is a ragged tensor with dtype torch.int32.
+            # It has 2 axes [arc][word], so aux_labels is also a ragged tensor
+            # with 2 axes [arc][word]
+            aux_labels, _ = lattice.aux_labels.index(
+                indexes=kept_path.values, axis=0, need_value_indexes=False
+            )
+        else:
+            assert isinstance(lattice.aux_labels, torch.Tensor)
+            aux_labels = k2.index_select(lattice.aux_labels, kept_path.values)
+            # aux_labels is a 1-D torch.Tensor. It also contains -1 and 0.
+
+        fsa = k2.linear_fsa(labels)
+        fsa.aux_labels = aux_labels
+        # Caution: fsa.scores are all 0s.
+        # `fsa` has only one extra attribute: aux_labels.
+        return Nbest(fsa=fsa, shape=utt_to_path_shape)
+
+    def intersect(self, lattice: k2.Fsa, use_double_scores=True) -> "Nbest":
+        """Intersect this Nbest object with a lattice, get 1-best
+        path from the resulting FsaVec, and return a new Nbest object.
+
+        The purpose of this function is to attach scores to an Nbest.
+
+        Args:
+          lattice:
+            An FsaVec with axes [utt][state][arc]. If it has `aux_labels`, then
+            we assume its `labels` are token IDs and `aux_labels` are word IDs.
+            If it has only `labels`, we assume its `labels` are word IDs.
+          use_double_scores:
+            True to use double precision when computing shortest path.
+            False to use single precision.
+        Returns:
+          Return a new Nbest. This new Nbest shares the same shape with `self`,
+          while its `fsa` is the 1-best path from intersecting `self.fsa` and
+          `lattice`. Also, its `fsa` has non-zero scores and inherits attributes
+          for `lattice`.
+        """
+        # Note: We view each linear FSA as a word sequence
+        # and we use the passed lattice to give each word sequence a score.
+        #
+        # We are not viewing each linear FSAs as a token sequence.
+        #
+        # So we use k2.invert() here.
+
+        # We use a word fsa to intersect with k2.invert(lattice)
+        word_fsa = k2.invert(self.fsa)
+
+        word_fsa.scores.zero_()
+        if hasattr(lattice, "aux_labels"):
+            # delete token IDs as it is not needed
+            del word_fsa.aux_labels
+            word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa)
+        else:
+            word_fsa_with_epsilon_loops = k2.linear_fst_with_self_loops(word_fsa)
+
+        path_to_utt_map = self.shape.row_ids(1)
+
+        if hasattr(lattice, "aux_labels"):
+            # lattice has token IDs as labels and word IDs as aux_labels.
+            # inv_lattice has word IDs as labels and token IDs as aux_labels
+            inv_lattice = k2.invert(lattice)
+            inv_lattice = k2.arc_sort(inv_lattice)
+        else:
+            inv_lattice = k2.arc_sort(lattice)
+
+        if inv_lattice.shape[0] == 1:
+            path_lattice = _intersect_device(
+                inv_lattice,
+                word_fsa_with_epsilon_loops,
+                b_to_a_map=torch.zeros_like(path_to_utt_map),
+                sorted_match_a=True,
+            )
+        else:
+            path_lattice = _intersect_device(
+                inv_lattice,
+                word_fsa_with_epsilon_loops,
+                b_to_a_map=path_to_utt_map,
+                sorted_match_a=True,
+            )
+
+        # path_lattice has word IDs as labels and token IDs as aux_labels
+        path_lattice = k2.top_sort(k2.connect(path_lattice))
+
+        one_best = k2.shortest_path(path_lattice, use_double_scores=use_double_scores)
+
+        one_best = k2.invert(one_best)
+        # Now one_best has token IDs as labels and word IDs as aux_labels
+
+        return Nbest(fsa=one_best, shape=self.shape)
+
+    def compute_am_scores(self) -> k2.RaggedTensor:
+        """Compute AM scores of each linear FSA (i.e., each path within
+        an utterance).
+
+        Hint:
+          `self.fsa.scores` contains two parts: acoustic scores (AM scores)
+          and n-gram language model scores (LM scores).
+
+        Caution:
+          We require that ``self.fsa`` has an attribute ``lm_scores``.
+
+        Returns:
+          Return a ragged tensor with 2 axes [utt][path_scores].
+          Its dtype is torch.float64.
+        """
+        scores_shape = self.fsa.arcs.shape().remove_axis(1)
+        # scores_shape has axes [path][arc]
+        am_scores = self.fsa.scores - self.fsa.lm_scores
+        ragged_am_scores = k2.RaggedTensor(scores_shape, am_scores.contiguous())
+        tot_scores = ragged_am_scores.sum()
+
+        return k2.RaggedTensor(self.shape, tot_scores)
+
+    def compute_lm_scores(self) -> k2.RaggedTensor:
+        """Compute LM scores of each linear FSA (i.e., each path within
+        an utterance).
+
+        Hint:
+          `self.fsa.scores` contains two parts: acoustic scores (AM scores)
+          and n-gram language model scores (LM scores).
+
+        Caution:
+          We require that ``self.fsa`` has an attribute ``lm_scores``.
+
+        Returns:
+          Return a ragged tensor with 2 axes [utt][path_scores].
+          Its dtype is torch.float64.
+        """
+        scores_shape = self.fsa.arcs.shape().remove_axis(1)
+        # scores_shape has axes [path][arc]
+
+        ragged_lm_scores = k2.RaggedTensor(
+            scores_shape, self.fsa.lm_scores.contiguous()
+        )
+
+        tot_scores = ragged_lm_scores.sum()
+
+        return k2.RaggedTensor(self.shape, tot_scores)
+
+    def tot_scores(self) -> k2.RaggedTensor:
+        """Get total scores of FSAs in this Nbest.
+
+        Note:
+          Since FSAs in Nbest are just linear FSAs, log-semiring
+          and tropical semiring produce the same total scores.
+
+        Returns:
+          Return a ragged tensor with two axes [utt][path_scores].
+          Its dtype is torch.float64.
+        """
+        scores_shape = self.fsa.arcs.shape().remove_axis(1)
+        # scores_shape has axes [path][arc]
+
+        ragged_scores = k2.RaggedTensor(scores_shape, self.fsa.scores.contiguous())
+
+        tot_scores = ragged_scores.sum()
+
+        return k2.RaggedTensor(self.shape, tot_scores)
+
+    def build_levenshtein_graphs(self) -> k2.Fsa:
+        """Return an FsaVec with axes [utt][state][arc]."""
+        word_ids = get_texts(self.fsa, return_ragged=True)
+        return k2.levenshtein_graph(word_ids)
+
+
+def one_best_decoding(
+    lattice: k2.Fsa,
+    use_double_scores: bool = True,
+    lm_scale_list: Optional[List[float]] = None,
+) -> Union[k2.Fsa, Dict[str, k2.Fsa]]:
+    """Get the best path from a lattice.
+
+    Args:
+      lattice:
+        The decoding lattice returned by :func:`get_lattice`.
+      use_double_scores:
+        True to use double precision floating point in the computation.
+        False to use single precision.
+      lm_scale_list:
+        A list of floats representing LM score scales.
+    Return:
+      An FsaVec containing linear paths.
+    """
+    if lm_scale_list is not None:
+        ans = dict()
+        saved_am_scores = lattice.scores - lattice.lm_scores
+        for lm_scale in lm_scale_list:
+            am_scores = saved_am_scores / lm_scale
+            lattice.scores = am_scores + lattice.lm_scores
+
+            best_path = k2.shortest_path(lattice, use_double_scores=use_double_scores)
+            key = f"lm_scale_{lm_scale}"
+            ans[key] = best_path
+        return ans
+
+    return k2.shortest_path(lattice, use_double_scores=use_double_scores)
+
+
+def nbest_decoding(
+    lattice: k2.Fsa,
+    num_paths: int,
+    use_double_scores: bool = True,
+    nbest_scale: float = 1.0,
+) -> k2.Fsa:
+    """It implements something like CTC prefix beam search using n-best lists.
+
+    The basic idea is to first extract `num_paths` paths from the given lattice,
+    build a word sequence from these paths, and compute the total scores
+    of the word sequence in the tropical semiring. The one with the max score
+    is used as the decoding output.
+
+    Caution:
+      Don't be confused by `best` in the name `n-best`. Paths are selected
+      **randomly**, not by ranking their scores.
+
+    Hint:
+      This decoding method is for demonstration only and it does
+      not produce a lower WER than :func:`one_best_decoding`.
+
+    Args:
+      lattice:
+        The decoding lattice, e.g., can be the return value of
+        :func:`get_lattice`. It has 3 axes [utt][state][arc].
+      num_paths:
+        It specifies the size `n` in n-best. Note: Paths are selected randomly
+        and those containing identical word sequences are removed and only one
+        of them is kept.
+      use_double_scores:
+        True to use double precision floating point in the computation.
+        False to use single precision.
+      nbest_scale:
+        It's the scale applied to the `lattice.scores`. A smaller value
+        leads to more unique paths at the risk of missing the correct path.
+    Returns:
+      An FsaVec containing **linear** FSAs. It axes are [utt][state][arc].
+    """
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # nbest.fsa.scores contains 0s
+
+    nbest = nbest.intersect(lattice)
+    # now nbest.fsa.scores gets assigned
+
+    # max_indexes contains the indexes for the path with the maximum score
+    # within an utterance.
+    max_indexes = nbest.tot_scores().argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+    return best_path
+
+
+def nbest_oracle(
+    lattice: k2.Fsa,
+    num_paths: int,
+    ref_texts: List[str],
+    word_table: k2.SymbolTable,
+    use_double_scores: bool = True,
+    nbest_scale: float = 0.5,
+    oov: str = "<UNK>",
+) -> Dict[str, List[List[int]]]:
+    """Select the best hypothesis given a lattice and a reference transcript.
+
+    The basic idea is to extract `num_paths` paths from the given lattice,
+    unique them, and select the one that has the minimum edit distance with
+    the corresponding reference transcript as the decoding output.
+
+    The decoding result returned from this function is the best result that
+    we can obtain using n-best decoding with all kinds of rescoring techniques.
+
+    This function is useful to tune the value of `nbest_scale`.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc].
+        Note: We assume its `aux_labels` contains word IDs.
+      num_paths:
+        The size of `n` in n-best.
+      ref_texts:
+        A list of reference transcript. Each entry contains space(s)
+        separated words
+      word_table:
+        It is the word symbol table.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      oov:
+        The out of vocabulary word.
+    Return:
+      Return a dict. Its key contains the information about the parameters
+      when calling this function, while its value contains the decoding output.
+      `len(ans_dict) == len(ref_texts)`
+    """
+    device = lattice.device
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    hyps = nbest.build_levenshtein_graphs()
+
+    oov_id = word_table[oov]
+    word_ids_list = []
+    for text in ref_texts:
+        word_ids = []
+        for word in text.split():
+            if word in word_table:
+                word_ids.append(word_table[word])
+            else:
+                word_ids.append(oov_id)
+        word_ids_list.append(word_ids)
+
+    refs = k2.levenshtein_graph(word_ids_list, device=device)
+
+    levenshtein_alignment = k2.levenshtein_alignment(
+        refs=refs,
+        hyps=hyps,
+        hyp_to_ref_map=nbest.shape.row_ids(1),
+        sorted_match_ref=True,
+    )
+
+    tot_scores = levenshtein_alignment.get_tot_scores(
+        use_double_scores=False, log_semiring=False
+    )
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+
+    max_indexes = ragged_tot_scores.argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+    return best_path
+
+
+def rescore_with_n_best_list(
+    lattice: k2.Fsa,
+    G: k2.Fsa,
+    num_paths: int,
+    lm_scale_list: List[float],
+    nbest_scale: float = 1.0,
+    use_double_scores: bool = True,
+) -> Dict[str, k2.Fsa]:
+    """Rescore an n-best list with an n-gram LM.
+    The path with the maximum score is used as the decoding output.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc]. It must have the following
+        attributes: ``aux_labels`` and ``lm_scores``. Its labels are
+        token IDs and ``aux_labels`` word IDs.
+      G:
+        An FsaVec containing only a single FSA. It is an n-gram LM.
+      num_paths:
+        Size of nbest list.
+      lm_scale_list:
+        A list of floats representing LM score scales.
+      nbest_scale:
+        Scale to be applied to ``lattice.score`` when sampling paths
+        using ``k2.random_paths``.
+      use_double_scores:
+        True to use double precision during computation. False to use
+        single precision.
+    Returns:
+      A dict of FsaVec, whose key is an lm_scale and the value is the
+      best decoding path for each utterance in the lattice.
+    """
+    device = lattice.device
+
+    assert len(lattice.shape) == 3
+    assert hasattr(lattice, "aux_labels")
+    assert hasattr(lattice, "lm_scores")
+
+    assert G.shape == (1, None, None)
+    assert G.device == device
+    assert hasattr(G, "aux_labels") is False
+
+    max_loop_count = 10
+    loop_count = 0
+    while loop_count <= max_loop_count:
+        try:
+            nbest = Nbest.from_lattice(
+                lattice=lattice,
+                num_paths=num_paths,
+                use_double_scores=use_double_scores,
+                nbest_scale=nbest_scale,
+            )
+            # nbest.fsa.scores are all 0s at this point
+            nbest = nbest.intersect(lattice)
+            break
+        except RuntimeError as e:
+            logging.info(f"Caught exception:\n{e}\n")
+            logging.info(f"num_paths before decreasing: {num_paths}")
+            num_paths = int(num_paths / 2)
+            if loop_count >= max_loop_count or num_paths <= 0:
+                logging.info("Return None as the resulting lattice is too large.")
+                return None
+            logging.info(
+                "This OOM is not an error. You can ignore it. "
+                "If your model does not converge well, or --max-duration "
+                "is too large, or the input sound file is difficult to "
+                "decode, you will meet this exception."
+            )
+            logging.info(f"num_paths after decreasing: {num_paths}")
+        loop_count += 1
+
+    # Now nbest.fsa has its scores set
+    assert hasattr(nbest.fsa, "lm_scores")
+
+    am_scores = nbest.compute_am_scores()
+
+    nbest = nbest.intersect(G)
+    # Now nbest contains only lm scores
+    lm_scores = nbest.tot_scores()
+
+    ans = dict()
+    for lm_scale in lm_scale_list:
+        tot_scores = am_scores.values / lm_scale + lm_scores.values
+        tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+        max_indexes = tot_scores.argmax()
+        best_path = k2.index_fsa(nbest.fsa, max_indexes)
+        key = f"lm_scale_{lm_scale}"
+        ans[key] = best_path
+    return ans
+
+
+def nbest_rescore_with_LM(
+    lattice: k2.Fsa,
+    LM: k2.Fsa,
+    num_paths: int,
+    lm_scale_list: List[float],
+    nbest_scale: float = 1.0,
+    use_double_scores: bool = True,
+) -> Dict[str, k2.Fsa]:
+    """Rescore an n-best list with an n-gram LM.
+    The path with the maximum score is used as the decoding output.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc]. It must have the following
+        attributes: ``aux_labels`` and ``lm_scores``. They are both token
+        IDs.
+      LM:
+        An FsaVec containing only a single FSA. It is one of follows:
+        - LG, L is lexicon and G is word-level n-gram LM.
+        - G, token-level n-gram LM.
+      num_paths:
+        Size of nbest list.
+      lm_scale_list:
+        A list of floats representing LM score scales.
+      nbest_scale:
+        Scale to be applied to ``lattice.score`` when sampling paths
+        using ``k2.random_paths``.
+      use_double_scores:
+        True to use double precision during computation. False to use
+        single precision.
+    Returns:
+      A dict of FsaVec, whose key is an lm_scale and the value is the
+      best decoding path for each utterance in the lattice.
+    """
+    device = lattice.device
+
+    assert len(lattice.shape) == 3
+    assert hasattr(lattice, "aux_labels")
+    assert hasattr(lattice, "lm_scores")
+
+    assert LM.shape == (1, None, None)
+    assert LM.device == device
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # nbest.fsa.scores contains 0s
+
+    nbest = nbest.intersect(lattice)
+
+    # Now nbest.fsa has its scores set
+    assert hasattr(nbest.fsa, "lm_scores")
+
+    # am scores + bi-gram scores
+    hp_scores = nbest.tot_scores()
+
+    # Now start to intersect nbest with LG or G
+    inv_fsa = k2.invert(nbest.fsa)
+    if hasattr(LM, "aux_labels"):
+        # LM is LG here
+        # delete token IDs as it is not needed
+        del inv_fsa.aux_labels
+    inv_fsa.scores.zero_()
+    inv_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(inv_fsa)
+    path_to_utt_map = nbest.shape.row_ids(1)
+
+    LM = k2.arc_sort(LM)
+    path_lattice = k2.intersect_device(
+        LM,
+        inv_fsa_with_epsilon_loops,
+        b_to_a_map=torch.zeros_like(path_to_utt_map),
+        sorted_match_a=True,
+    )
+
+    # Its labels are token IDs.
+    # If LM is G, its aux_labels are tokens IDs;
+    # If LM is LG, its aux_labels are words IDs.
+    path_lattice = k2.top_sort(k2.connect(path_lattice))
+    one_best = k2.shortest_path(path_lattice, use_double_scores=use_double_scores)
+
+    lm_scores = one_best.get_tot_scores(
+        use_double_scores=use_double_scores,
+        log_semiring=True,  # Note: we always use True
+    )
+    # If LM is LG, we might get empty paths
+    lm_scores[lm_scores == float("-inf")] = -1e9
+
+    ans = dict()
+    for lm_scale in lm_scale_list:
+        tot_scores = hp_scores.values / lm_scale + lm_scores
+        tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+        max_indexes = tot_scores.argmax()
+        best_path = k2.index_fsa(nbest.fsa, max_indexes)
+        key = f"lm_scale_{lm_scale}"
+        ans[key] = best_path
+    return ans
+
+
+def rescore_with_whole_lattice(
+    lattice: k2.Fsa,
+    G_with_epsilon_loops: k2.Fsa,
+    lm_scale_list: Optional[List[float]] = None,
+    use_double_scores: bool = True,
+) -> Union[k2.Fsa, Dict[str, k2.Fsa]]:
+    """Intersect the lattice with an n-gram LM and use shortest path
+    to decode.
+
+    The input lattice is obtained by intersecting `HLG` with
+    a DenseFsaVec, where the `G` in `HLG` is in general a 3-gram LM.
+    The input `G_with_epsilon_loops` is usually a 4-gram LM. You can consider
+    this function as a second pass decoding. In the first pass decoding, we
+    use a small G, while we use a larger G in the second pass decoding.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc]. Its `aux_lables` are word IDs.
+        It must have an attribute `lm_scores`.
+      G_with_epsilon_loops:
+        An FsaVec containing only a single FSA. It contains epsilon self-loops.
+        It is an acceptor and its labels are word IDs.
+      lm_scale_list:
+        Optional. If none, return the intersection of `lattice` and
+        `G_with_epsilon_loops`.
+        If not None, it contains a list of values to scale LM scores.
+        For each scale, there is a corresponding decoding result contained in
+        the resulting dict.
+      use_double_scores:
+        True to use double precision in the computation.
+        False to use single precision.
+    Returns:
+      If `lm_scale_list` is None, return a new lattice which is the intersection
+      result of `lattice` and `G_with_epsilon_loops`.
+      Otherwise, return a dict whose key is an entry in `lm_scale_list` and the
+      value is the decoding result (i.e., an FsaVec containing linear FSAs).
+    """
+    # Nbest is not used in this function
+    assert hasattr(lattice, "lm_scores")
+    assert G_with_epsilon_loops.shape == (1, None, None)
+
+    device = lattice.device
+    lattice.scores = lattice.scores - lattice.lm_scores
+    # We will use lm_scores from G, so remove lats.lm_scores here
+    del lattice.lm_scores
+
+    assert hasattr(G_with_epsilon_loops, "lm_scores")
+
+    # Now, lattice.scores contains only am_scores
+
+    # inv_lattice has word IDs as labels.
+    # Its `aux_labels` is token IDs
+    inv_lattice = k2.invert(lattice)
+    num_seqs = lattice.shape[0]
+
+    b_to_a_map = torch.zeros(num_seqs, device=device, dtype=torch.int32)
+
+    # NOTE: The choice of the threshold list is arbitrary here to avoid OOM.
+    # You may need to fine tune it.
+    prune_th_list = [1e-10, 1e-9, 1e-8, 1e-7, 1e-6]
+    prune_th_list += [1e-5, 1e-4, 1e-3, 1e-2, 1e-1]
+    max_loop_count = 10
+    loop_count = 0
+    while loop_count <= max_loop_count:
+        try:
+            rescoring_lattice = k2.intersect_device(
+                G_with_epsilon_loops,
+                inv_lattice,
+                b_to_a_map,
+                sorted_match_a=True,
+            )
+            rescoring_lattice = k2.top_sort(k2.connect(rescoring_lattice))
+            break
+        except RuntimeError as e:
+            logging.info(f"Caught exception:\n{e}\n")
+            if loop_count >= max_loop_count:
+                logging.info("Return None as the resulting lattice is too large.")
+                return None
+            logging.info(f"num_arcs before pruning: {inv_lattice.arcs.num_elements()}")
+            logging.info(
+                "This OOM is not an error. You can ignore it. "
+                "If your model does not converge well, or --max-duration "
+                "is too large, or the input sound file is difficult to "
+                "decode, you will meet this exception."
+            )
+            inv_lattice = k2.prune_on_arc_post(
+                inv_lattice,
+                prune_th_list[loop_count],
+                True,
+            )
+            logging.info(f"num_arcs after pruning: {inv_lattice.arcs.num_elements()}")
+        loop_count += 1
+
+    # lat has token IDs as labels
+    # and word IDs as aux_labels.
+    lat = k2.invert(rescoring_lattice)
+
+    if lm_scale_list is None:
+        return lat
+
+    ans = dict()
+    saved_am_scores = lat.scores - lat.lm_scores
+    for lm_scale in lm_scale_list:
+        am_scores = saved_am_scores / lm_scale
+        lat.scores = am_scores + lat.lm_scores
+
+        best_path = k2.shortest_path(lat, use_double_scores=use_double_scores)
+        key = f"lm_scale_{lm_scale}"
+        ans[key] = best_path
+    return ans
+
+
+def rescore_with_attention_decoder(
+    lattice: k2.Fsa,
+    num_paths: int,
+    model: torch.nn.Module,
+    memory: torch.Tensor,
+    memory_key_padding_mask: Optional[torch.Tensor],
+    sos_id: int,
+    eos_id: int,
+    nbest_scale: float = 1.0,
+    ngram_lm_scale: Optional[float] = None,
+    attention_scale: Optional[float] = None,
+    use_double_scores: bool = True,
+) -> Dict[str, k2.Fsa]:
+    """This function extracts `num_paths` paths from the given lattice and uses
+    an attention decoder to rescore them. The path with the highest score is
+    the decoding output.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc].
+      num_paths:
+        Number of paths to extract from the given lattice for rescoring.
+      model:
+        A transformer model. See the class "Transformer" in
+        conformer_ctc/transformer.py for its interface.
+      memory:
+        The encoder memory of the given model. It is the output of
+        the last torch.nn.TransformerEncoder layer in the given model.
+        Its shape is `(T, N, C)`.
+      memory_key_padding_mask:
+        The padding mask for memory with shape `(N, T)`.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      nbest_scale:
+        It's the scale applied to `lattice.scores`. A smaller value
+        leads to more unique paths at the risk of missing the correct path.
+      ngram_lm_scale:
+        Optional. It specifies the scale for n-gram LM scores.
+      attention_scale:
+        Optional. It specifies the scale for attention decoder scores.
+    Returns:
+      A dict of FsaVec, whose key contains a string
+      ngram_lm_scale_attention_scale and the value is the
+      best decoding path for each utterance in the lattice.
+    """
+    max_loop_count = 10
+    loop_count = 0
+    while loop_count <= max_loop_count:
+        try:
+            nbest = Nbest.from_lattice(
+                lattice=lattice,
+                num_paths=num_paths,
+                use_double_scores=use_double_scores,
+                nbest_scale=nbest_scale,
+            )
+            # nbest.fsa.scores are all 0s at this point
+            nbest = nbest.intersect(lattice)
+            break
+        except RuntimeError as e:
+            logging.info(f"Caught exception:\n{e}\n")
+            logging.info(f"num_paths before decreasing: {num_paths}")
+            num_paths = int(num_paths / 2)
+            if loop_count >= max_loop_count or num_paths <= 0:
+                logging.info("Return None as the resulting lattice is too large.")
+                return None
+            logging.info(
+                "This OOM is not an error. You can ignore it. "
+                "If your model does not converge well, or --max-duration "
+                "is too large, or the input sound file is difficult to "
+                "decode, you will meet this exception."
+            )
+            logging.info(f"num_paths after decreasing: {num_paths}")
+        loop_count += 1
+
+    # Now nbest.fsa has its scores set.
+    # Also, nbest.fsa inherits the attributes from `lattice`.
+    assert hasattr(nbest.fsa, "lm_scores")
+
+    am_scores = nbest.compute_am_scores()
+    ngram_lm_scores = nbest.compute_lm_scores()
+
+    # The `tokens` attribute is set inside `compile_hlg.py`
+    assert hasattr(nbest.fsa, "tokens")
+    assert isinstance(nbest.fsa.tokens, torch.Tensor)
+
+    path_to_utt_map = nbest.shape.row_ids(1).to(torch.long)
+    # the shape of memory is (T, N, C), so we use axis=1 here
+    expanded_memory = memory.index_select(1, path_to_utt_map)
+
+    if memory_key_padding_mask is not None:
+        # The shape of memory_key_padding_mask is (N, T), so we
+        # use axis=0 here.
+        expanded_memory_key_padding_mask = memory_key_padding_mask.index_select(
+            0, path_to_utt_map
+        )
+    else:
+        expanded_memory_key_padding_mask = None
+
+    # remove axis corresponding to states.
+    tokens_shape = nbest.fsa.arcs.shape().remove_axis(1)
+    tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.tokens)
+    tokens = tokens.remove_values_leq(0)
+    token_ids = tokens.tolist()
+
+    if len(token_ids) == 0:
+        print("Warning: rescore_with_attention_decoder(): empty token-ids")
+        return None
+
+    nll = model.decoder_nll(
+        memory=expanded_memory,
+        memory_key_padding_mask=expanded_memory_key_padding_mask,
+        token_ids=token_ids,
+        sos_id=sos_id,
+        eos_id=eos_id,
+    )
+    assert nll.ndim == 2
+    assert nll.shape[0] == len(token_ids)
+
+    attention_scores = -nll.sum(dim=1)
+
+    if ngram_lm_scale is None:
+        ngram_lm_scale_list = [0.01, 0.05, 0.08]
+        ngram_lm_scale_list += [0.1, 0.3, 0.5, 0.6, 0.7, 0.9, 1.0]
+        ngram_lm_scale_list += [1.1, 1.2, 1.3, 1.5, 1.7, 1.9, 2.0]
+        ngram_lm_scale_list += [2.1, 2.2, 2.3, 2.5, 3.0, 4.0, 5.0]
+    else:
+        ngram_lm_scale_list = [ngram_lm_scale]
+
+    if attention_scale is None:
+        attention_scale_list = [0.01, 0.05, 0.08]
+        attention_scale_list += [0.1, 0.3, 0.5, 0.6, 0.7, 0.9, 1.0]
+        attention_scale_list += [1.1, 1.2, 1.3, 1.5, 1.7, 1.9, 2.0]
+        attention_scale_list += [2.1, 2.2, 2.3, 2.5, 3.0, 4.0, 5.0]
+    else:
+        attention_scale_list = [attention_scale]
+
+    ans = dict()
+    for n_scale in ngram_lm_scale_list:
+        for a_scale in attention_scale_list:
+            tot_scores = (
+                am_scores.values
+                + n_scale * ngram_lm_scores.values
+                + a_scale * attention_scores
+            )
+            ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+            max_indexes = ragged_tot_scores.argmax()
+            best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+            key = f"ngram_lm_scale_{n_scale}_attention_scale_{a_scale}"
+            ans[key] = best_path
+    return ans
+
+
+def rescore_with_rnn_lm(
+    lattice: k2.Fsa,
+    num_paths: int,
+    rnn_lm_model: torch.nn.Module,
+    model: torch.nn.Module,
+    memory: torch.Tensor,
+    memory_key_padding_mask: Optional[torch.Tensor],
+    sos_id: int,
+    eos_id: int,
+    blank_id: int,
+    nbest_scale: float = 1.0,
+    ngram_lm_scale: Optional[float] = None,
+    attention_scale: Optional[float] = None,
+    rnn_lm_scale: Optional[float] = None,
+    use_double_scores: bool = True,
+) -> Dict[str, k2.Fsa]:
+    """This function extracts `num_paths` paths from the given lattice and uses
+    an attention decoder to rescore them. The path with the highest score is
+    the decoding output.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc].
+      num_paths:
+        Number of paths to extract from the given lattice for rescoring.
+      rnn_lm_model:
+        A rnn-lm model used for LM rescoring
+      model:
+        A transformer model. See the class "Transformer" in
+        conformer_ctc/transformer.py for its interface.
+      memory:
+        The encoder memory of the given model. It is the output of
+        the last torch.nn.TransformerEncoder layer in the given model.
+        Its shape is `(T, N, C)`.
+      memory_key_padding_mask:
+        The padding mask for memory with shape `(N, T)`.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      nbest_scale:
+        It's the scale applied to `lattice.scores`. A smaller value
+        leads to more unique paths at the risk of missing the correct path.
+      ngram_lm_scale:
+        Optional. It specifies the scale for n-gram LM scores.
+      attention_scale:
+        Optional. It specifies the scale for attention decoder scores.
+      rnn_lm_scale:
+        Optional. It specifies the scale for RNN LM scores.
+    Returns:
+      A dict of FsaVec, whose key contains a string
+      ngram_lm_scale_attention_scale and the value is the
+      best decoding path for each utterance in the lattice.
+    """
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # nbest.fsa.scores are all 0s at this point
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa has its scores set.
+    # Also, nbest.fsa inherits the attributes from `lattice`.
+    assert hasattr(nbest.fsa, "lm_scores")
+
+    am_scores = nbest.compute_am_scores()
+    ngram_lm_scores = nbest.compute_lm_scores()
+
+    # The `tokens` attribute is set inside `compile_hlg.py`
+    assert hasattr(nbest.fsa, "tokens")
+    assert isinstance(nbest.fsa.tokens, torch.Tensor)
+
+    path_to_utt_map = nbest.shape.row_ids(1).to(torch.long)
+    # the shape of memory is (T, N, C), so we use axis=1 here
+    expanded_memory = memory.index_select(1, path_to_utt_map)
+
+    if memory_key_padding_mask is not None:
+        # The shape of memory_key_padding_mask is (N, T), so we
+        # use axis=0 here.
+        expanded_memory_key_padding_mask = memory_key_padding_mask.index_select(
+            0, path_to_utt_map
+        )
+    else:
+        expanded_memory_key_padding_mask = None
+
+    # remove axis corresponding to states.
+    tokens_shape = nbest.fsa.arcs.shape().remove_axis(1)
+    tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.tokens)
+    tokens = tokens.remove_values_leq(0)
+    token_ids = tokens.tolist()
+
+    if len(token_ids) == 0:
+        print("Warning: rescore_with_attention_decoder(): empty token-ids")
+        return None
+
+    nll = model.decoder_nll(
+        memory=expanded_memory,
+        memory_key_padding_mask=expanded_memory_key_padding_mask,
+        token_ids=token_ids,
+        sos_id=sos_id,
+        eos_id=eos_id,
+    )
+    assert nll.ndim == 2
+    assert nll.shape[0] == len(token_ids)
+
+    attention_scores = -nll.sum(dim=1)
+
+    # Now for RNN LM
+    sos_tokens = add_sos(tokens, sos_id)
+    tokens_eos = add_eos(tokens, eos_id)
+    sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+    sentence_lengths = sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+
+    x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+    y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+    x_tokens = x_tokens.to(torch.int64)
+    y_tokens = y_tokens.to(torch.int64)
+    sentence_lengths = sentence_lengths.to(torch.int64)
+
+    rnn_lm_nll = rnn_lm_model(x=x_tokens, y=y_tokens, lengths=sentence_lengths)
+    assert rnn_lm_nll.ndim == 2
+    assert rnn_lm_nll.shape[0] == len(token_ids)
+
+    rnn_lm_scores = -1 * rnn_lm_nll.sum(dim=1)
+
+    ngram_lm_scale_list = DEFAULT_LM_SCALE
+    attention_scale_list = DEFAULT_LM_SCALE
+    rnn_lm_scale_list = DEFAULT_LM_SCALE
+
+    if ngram_lm_scale:
+        ngram_lm_scale_list = [ngram_lm_scale]
+
+    if attention_scale:
+        attention_scale_list = [attention_scale]
+
+    if rnn_lm_scale:
+        rnn_lm_scale_list = [rnn_lm_scale]
+
+    ans = dict()
+    for n_scale in ngram_lm_scale_list:
+        for a_scale in attention_scale_list:
+            for r_scale in rnn_lm_scale_list:
+                tot_scores = (
+                    am_scores.values
+                    + n_scale * ngram_lm_scores.values
+                    + a_scale * attention_scores
+                    + r_scale * rnn_lm_scores
+                )
+                ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+                max_indexes = ragged_tot_scores.argmax()
+                best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+                key = f"ngram_lm_scale_{n_scale}_attention_scale_{a_scale}_rnn_lm_scale_{r_scale}"  # noqa
+                ans[key] = best_path
+    return ans
diff --git a/icefall/diagnostics.py b/icefall/diagnostics.py
new file mode 100644
index 000000000..65b6f67b0
--- /dev/null
+++ b/icefall/diagnostics.py
@@ -0,0 +1,684 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey
+#                                                    Zengwei Yao
+#                                                    Mingshuang Luo)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+from dataclasses import dataclass
+from typing import Optional, Tuple, List
+
+import torch
+from torch import Tensor, nn
+
+
+class TensorDiagnosticOptions(object):
+    """Options object for tensor diagnostics:
+
+    Args:
+      max_eig_dim:
+        The maximum dimension for which we print out eigenvalues
+        (limited for speed reasons).
+    """
+
+    def __init__(self, max_eig_dim: int = 512):
+        self.max_eig_dim = max_eig_dim
+
+    def dim_is_summarized(self, size: int):
+        return size > 10 and size != 31
+
+
+def get_tensor_stats(
+    x: Tensor,
+    dim: int,
+    stats_type: str,
+) -> Tuple[Tensor, int]:
+    """
+    Returns the specified transformation of the Tensor (either x or x.abs()
+    or (x > 0), summed over all but the index `dim`.
+
+    Args:
+      x:
+        Tensor, tensor to be analyzed
+      dim:
+        Dimension with 0 <= dim < x.ndim
+      stats_type:
+        The stats_type includes several types:
+        "abs" -> take abs() before summing
+        "positive" -> take (x > 0) before summing
+        "rms" -> square before summing, we'll take sqrt later
+        "value"  -> just sum x itself
+        "max", "min" -> take the maximum or minimum [over all other dims but dim] instead of summing
+    Returns:
+      stats: a Tensor of shape (x.shape[dim],).
+      count: an integer saying how many items were counted in each element
+      of stats.
+    """
+
+    count = x.numel() // x.shape[dim]
+
+    if stats_type == "eigs":
+        x = x.transpose(dim, -1)
+        x = x.reshape(-1, x.shape[-1])
+        # shape of returned tensor: (s, s),
+        # where s is size of dimension `dim` of original x.
+        return torch.matmul(x.transpose(0, 1), x), count
+    elif stats_type == "abs":
+        x = x.abs()
+    elif stats_type == "rms":
+        x = x**2
+    elif stats_type == "positive":
+        x = (x > 0).to(dtype=torch.float)
+    else:
+        assert stats_type in ["value", "max", "min"]
+
+    sum_dims = [d for d in range(x.ndim) if d != dim]
+    if len(sum_dims) > 0:
+        if stats_type == "max":
+            for dim in reversed(sum_dims):
+                x = torch.max(x, dim=dim)[0]
+        elif stats_type == "min":
+            for dim in reversed(sum_dims):
+                x = torch.min(x, dim=dim)[0]
+        else:
+            x = torch.sum(x, dim=sum_dims)
+    x = x.flatten().clone()
+    return x, count
+
+
+@dataclass
+class TensorAndCount:
+    tensor: Tensor
+    count: int
+
+
+class TensorDiagnostic(object):
+    """This class is not directly used by the user, it is responsible for
+    collecting diagnostics for a module or parameter tensor of a torch.nn.Module.
+
+    Args:
+      opts:
+        Options object.
+      name:
+        The name associated with this diagnostics object, will probably be {module_name}.X
+           where X is "output" or "grad", or {parameter_name}.Y where Y is param_value or param_grad.
+    """
+
+    def __init__(self, opts: TensorDiagnosticOptions, name: str):
+        self.opts = opts
+        self.name = name
+        self.class_name = None  # will assign in accumulate()
+
+        self.stats = None  # we'll later assign a list to self.stats.
+        # It's a list of dicts, indexed by dim (i.e. by the
+        # axis of the tensor).  The dicts, in turn, are
+        # indexed by `stats-type` which are strings in
+        # ["abs", "max", "min", "positive", "value", "rms"].
+
+        # scalar_stats contains some analysis of the activations and gradients,
+        self.scalar_stats = None
+
+        # the keys into self.stats[dim] are strings, whose values can be
+        # "abs", "max", "min" ,"value", "positive", "rms", "value".
+        # The values e.g. self.stats[dim]["rms"] are lists of dataclass TensorAndCount,
+        # containing a tensor and its associated count (which is the sum of the other dims
+        # that we aggregated over, e.g. the number of frames and/or batch elements and/or
+        # channels.
+        # ... we actually accumulate the Tensors / counts any time we have the same-dim tensor,
+        # only adding a new element to the list if there was a different dim.
+        # if the string in the key is "eigs", if we detect a length mismatch we put None as the value.
+
+    def accumulate(self, x, class_name: Optional[str] = None):
+        """
+        Accumulate tensors.
+        """
+        if class_name is not None:
+            self.class_name = class_name
+        if isinstance(x, Tuple):
+            x = x[0]
+        if not isinstance(x, Tensor):
+            return
+        if x.numel() == 0:  # for empty tensor
+            return
+        x = x.detach().clone()
+        if x.ndim == 0:
+            x = x.unsqueeze(0)
+        ndim = x.ndim
+        if self.stats is None:
+            self.stats = [dict() for _ in range(ndim)]
+
+        for dim in range(ndim):
+            this_dim_stats = self.stats[dim]
+            if ndim > 1:
+                stats_types = ["abs", "max", "min", "positive", "value", "rms"]
+                if x.shape[dim] <= self.opts.max_eig_dim:
+                    stats_types.append("eigs")
+            else:
+                stats_types = ["value", "abs", "max", "min"]
+
+            for stats_type in stats_types:
+                stats, count = get_tensor_stats(x, dim, stats_type)
+                if stats_type not in this_dim_stats:
+                    this_dim_stats[stats_type] = []  # list of TensorAndCount
+
+                done = False
+                if this_dim_stats[stats_type] is None:
+                    # we can reach here if we detected for stats_type "eigs" that
+                    # where was more than one different size for this dim.  Then we
+                    # disable accumulating this stats type, as it uses too much memory.
+                    continue
+                for s in this_dim_stats[stats_type]:
+                    if s.tensor.shape == stats.shape:
+                        if stats_type == "max":
+                            s.tensor = torch.maximum(s.tensor, stats)
+
+                        elif stats_type == "min":
+                            s.tensor = torch.minimum(s.tensor, stats)
+                        else:
+                            assert stats_type != "max"
+                            s.tensor += stats
+                        s.count += count
+                        done = True
+                        break
+                if not done:
+                    if this_dim_stats[stats_type] != [] and stats_type == "eigs":
+                        # >1 size encountered on this dim, e.g. it's a batch or time dimension,
+                        # don't accumulat "eigs" stats type, it uses too much memory
+                        this_dim_stats[stats_type] = None
+                    else:
+                        this_dim_stats[stats_type].append(TensorAndCount(stats, count))
+
+    def print_diagnostics(self):
+        """Print diagnostics for each dimension of the tensor."""
+        if self.stats is None:
+            print(f"Warning: the stats of {self.name} is None.")
+            return
+        for dim, this_dim_stats in enumerate(self.stats):
+            if "rms" in this_dim_stats and "value" in this_dim_stats:
+                # produce "stddev" stats, which is centered RMS.
+                rms_stats_list = this_dim_stats["rms"]
+                value_stats_list = this_dim_stats["value"]
+                if len(rms_stats_list) == len(value_stats_list):
+                    stddev_stats_list = []
+                    for r, v in zip(rms_stats_list, value_stats_list):
+                        stddev_stats_list.append(
+                            # r.count and v.count should be the same, but we don't check this.
+                            TensorAndCount(
+                                r.tensor - v.tensor * v.tensor / (v.count + 1.0e-20),
+                                r.count,
+                            )
+                        )
+                    this_dim_stats["stddev"] = stddev_stats_list
+
+            for stats_type, stats_list in this_dim_stats.items():
+                # stats_type could be "rms", "value", "abs", "eigs", "positive", "min" or "max".
+                # "stats_list" could be a list of TensorAndCount (one list per distinct tensor
+                # shape of the stats), or None
+                if stats_list is None:
+                    assert stats_type == "eigs"
+                    continue
+
+                def get_count(count):
+                    return 1 if stats_type in ["max", "min"] else count
+
+                if len(stats_list) == 1:
+                    stats = stats_list[0].tensor / get_count(stats_list[0].count)
+                else:
+                    # a dimension that has variable size in different nnet
+                    # forwards, e.g. a time dimension in an ASR model.
+                    stats = torch.cat(
+                        [x.tensor / get_count(x.count) for x in stats_list], dim=0
+                    )
+
+                if stats_type == "eigs":
+                    try:
+                        if hasattr(torch, "linalg") and hasattr(torch.linalg, "eigh"):
+                            eigs, _ = torch.linalg.eigh(stats)
+                        else:
+                            eigs, _ = torch.symeig(stats)
+                        stats = eigs.abs().sqrt()
+                    except:  # noqa
+                        print("Error getting eigenvalues, trying another method.")
+                        if hasattr(torch, "linalg") and hasattr(torch.linalg, "eig"):
+                            eigs, _ = torch.linalg.eig(stats)
+                            eigs = eigs.abs()
+                        else:
+                            eigs, _ = torch.eig(stats)
+                            eigs = eigs.norm(dim=1)
+                        stats = eigs.sqrt()
+                        # sqrt so it reflects data magnitude, like stddev- not variance
+
+                if stats_type in ["rms", "stddev"]:
+                    # we stored the square; after aggregation we need to take sqrt.
+                    stats = stats.sqrt()
+
+                # if `summarize` we print percentiles of the stats; else,
+                # we print out individual elements.
+                summarize = (len(stats_list) > 1) or self.opts.dim_is_summarized(
+                    stats.numel()
+                )
+                if summarize:  # usually `summarize` will be true
+                    # print out percentiles.
+                    stats = stats.sort()[0]
+                    num_percentiles = 10
+                    size = stats.numel()
+                    percentiles = []
+                    for i in range(num_percentiles + 1):
+                        index = (i * (size - 1)) // num_percentiles
+                        percentiles.append(stats[index].item())
+                    percentiles = ["%.2g" % x for x in percentiles]
+                    percentiles = " ".join(percentiles)
+                    ans = f"percentiles: [{percentiles}]"
+                else:
+                    ans = stats.tolist()
+                    ans = ["%.2g" % x for x in ans]
+                    ans = "[" + " ".join(ans) + "]"
+                if stats_type in ["value", "rms", "stddev", "eigs"]:
+                    # This norm is useful because it is strictly less than the largest
+                    # sqrt(eigenvalue) of the variance, which we print out, and shows,
+                    # speaking in an approximate way, how much of that largest eigenvalue
+                    # can be attributed to the mean of the distribution.
+                    norm = (stats**2).sum().sqrt().item()
+                    ans += f", norm={norm:.2g}"
+                mean = stats.mean().item()
+                rms = (stats**2).mean().sqrt().item()
+                ans += f", mean={mean:.3g}, rms={rms:.3g}"
+
+                # OK, "ans" contains the actual stats, e.g.
+                # ans = "percentiles: [0.43 0.46 0.48 0.49 0.49 0.5 0.51 0.52 0.53 0.54 0.59], mean=0.5, rms=0.5"
+
+                sizes = [x.tensor.shape[0] for x in stats_list]
+                size_str = (
+                    f"{sizes[0]}" if len(sizes) == 1 else f"{min(sizes)}..{max(sizes)}"
+                )
+                maybe_class_name = (
+                    f" type={self.class_name}," if self.class_name is not None else ""
+                )
+                print(
+                    f"module={self.name},{maybe_class_name} dim={dim}, size={size_str}, {stats_type} {ans}"
+                )
+
+
+class ScalarDiagnostic(object):
+    """This class is not directly used by the user, it is responsible for
+    collecting diagnostics for a single module (subclass of torch.nn.Module) that
+    represents some kind of nonlinearity, e.g. ReLU, sigmoid, etc.
+    """
+
+    def __init__(self, opts: TensorDiagnosticOptions, name: str):
+        self.opts = opts
+        self.name = name
+        self.class_name = None  # will assign in accumulate()
+        self.is_forward_pass = True
+
+        self.tick_scale = None
+
+        self.saved_inputs = []
+        self.is_ok = True
+
+        self.counts = None
+        self.sum_grad = None
+        self.sum_gradsq = None
+        self.sum_abs_grad = None
+
+    def accumulate_input(self, x: Tensor, class_name: Optional[str] = None):
+        """
+        Called in forward pass.
+        """
+        if not self.is_forward_pass:
+            # in case we did a forward pass without a backward pass, for some reason.
+            self.saved_inputs = []
+            self.is_forward_pass = True
+
+        if class_name is not None:
+            self.class_name = class_name
+        if not self.is_ok:
+            return
+
+        limit = 10
+        if len(self.saved_inputs) > limit:
+            print(
+                f"ERROR: forward pass called for this module over {limit} times with no backward pass. "
+                f" Will not accumulate scalar stats."
+            )
+            self.is_ok = False
+            return
+        self.saved_inputs.append(x)
+
+    def accumulate_output_grad(self, grad: Tensor):
+        if not self.is_ok:
+            return
+        if self.is_forward_pass:
+            self.is_forward_pass = False
+
+        last_shape = (
+            "n/a" if len(self.saved_inputs) == 0 else self.saved_inputs[-1].shape
+        )
+        if len(self.saved_inputs) == 0 or grad.shape != last_shape:
+            print(
+                f"ERROR: shape mismatch or no forward activation present when backward "
+                f"pass called: grad shape ={tuple(grad.shape)}, num-saved-inputs={len(self.saved_inputs)}"
+                f", shape-of-last-saved-input={last_shape}"
+            )
+            self.is_ok = False
+            return
+
+        x = self.saved_inputs.pop()
+        self.process_input_and_grad(x, grad)
+
+    def process_input_and_grad(self, x: Tensor, grad: Tensor):
+        assert x.shape == grad.shape
+        x = x.flatten()
+        grad = grad.flatten()
+
+        num_ticks_per_side = 256
+
+        if self.tick_scale is None:
+            x_abs_sorted = x.abs().sort()[0]
+            # take the 98th percentile as the largest value we count separately.
+            index = int(x.numel() * 0.98)
+            self.tick_scale = float(x_abs_sorted[index] / num_ticks_per_side)
+
+            # integerize from tick * (-num ticks_per_side ..  num_ticks_per_side - 1]
+            self.counts = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.long, device=x.device
+            )
+            self.sum_grad = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.double, device=x.device
+            )
+            # sum_gradsq is for getting error bars.
+            self.sum_gradsq = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.double, device=x.device
+            )
+            self.sum_abs_grad = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.double, device=x.device
+            )
+
+        # this will round down.
+        x = (x / self.tick_scale).to(torch.long)
+        x = x.clamp_(min=-num_ticks_per_side, max=num_ticks_per_side - 1)
+        x = x + num_ticks_per_side
+
+        self.counts.index_add_(dim=0, index=x, source=torch.ones_like(x))
+        self.sum_grad.index_add_(dim=0, index=x, source=grad.to(torch.double))
+        self.sum_gradsq.index_add_(
+            dim=0, index=x, source=(grad * grad).to(torch.double)
+        )
+        self.sum_abs_grad.index_add_(dim=0, index=x, source=grad.abs().to(torch.double))
+
+    def print_diagnostics(self):
+        """Print diagnostics."""
+        if self.is_ok is False or self.counts is None:
+            print(f"Warning: no stats accumulated for {self.name}, is_ok={self.is_ok}")
+            return
+
+        counts = self.counts.to("cpu")
+        sum_grad = self.sum_grad.to(device="cpu", dtype=torch.float32)
+        sum_gradsq = self.sum_gradsq.to(device="cpu", dtype=torch.float32)
+        sum_abs_grad = self.sum_abs_grad.to(device="cpu", dtype=torch.float32)
+
+        counts_cumsum = counts.cumsum(dim=0)
+        counts_tot = counts_cumsum[-1]
+
+        # subdivide the distribution up into `num_bins` intervals for analysis, for greater
+        # statistical significance.  each bin corresponds to multiple of the original 'tick' intervals.
+        num_bins = 20
+
+        # integer division
+        counts_per_bin = (counts_tot // num_bins) + 1
+        bin_indexes = counts_cumsum // counts_per_bin
+        bin_indexes = bin_indexes.clamp(min=0, max=num_bins).to(torch.long)
+
+        bin_counts = torch.zeros(num_bins, dtype=torch.long)
+        bin_counts.index_add_(dim=0, index=bin_indexes, source=counts)
+        bin_grad = torch.zeros(num_bins)
+        bin_grad.index_add_(dim=0, index=bin_indexes, source=sum_grad)
+        bin_gradsq = torch.zeros(num_bins)
+        bin_gradsq.index_add_(dim=0, index=bin_indexes, source=sum_gradsq)
+        bin_abs_grad = torch.zeros(num_bins)
+        bin_abs_grad.index_add_(dim=0, index=bin_indexes, source=sum_abs_grad)
+
+        avg_grad = bin_grad / bin_counts
+        avg_grad_stddev = (bin_gradsq / bin_counts).sqrt()
+
+        bin_boundary_counts = (
+            torch.arange(num_bins + 1, dtype=torch.long) * counts_per_bin
+        )
+        bin_tick_indexes = torch.searchsorted(counts_cumsum, bin_boundary_counts)
+        # boundaries are the "x" values between the bins, e.g. corresponding to the
+        # locations of percentiles of the distribution.
+        num_ticks_per_side = counts.numel() // 2
+        bin_boundaries = (bin_tick_indexes - num_ticks_per_side) * self.tick_scale
+
+        bin_grad = bin_grad / (bin_counts + 1)
+        bin_conf_interval = bin_gradsq.sqrt() / (
+            bin_counts + 1
+        )  # consider this a standard deviation.
+        # bin_grad / bin_abs_grad will give us a sense for how important in a practical sense,
+        # the gradients are.
+        bin_abs_grad = bin_abs_grad / (bin_counts + 1)
+
+        bin_rel_grad = bin_grad / (bin_abs_grad + 1.0e-20)
+        bin_conf = bin_grad / (bin_conf_interval + 1.0e-20)
+
+        def tensor_to_str(x: Tensor):
+            x = ["%.2g" % f for f in x]
+            x = "[" + " ".join(x) + "]"
+            return x
+
+        maybe_class_name = (
+            f" type={self.class_name}," if self.class_name is not None else ""
+        )
+
+        print(
+            f"module={self.name},{maybe_class_name} bin-boundaries={tensor_to_str(bin_boundaries)}, "
+            f"rel_grad={tensor_to_str(bin_rel_grad)}, grad_conf={tensor_to_str(bin_conf)}"
+        )
+
+
+class ModelDiagnostic(object):
+    """This class stores diagnostics for all tensors in the torch.nn.Module.
+
+    Args:
+      opts:
+        Options object.
+    """
+
+    def __init__(self, opts: Optional[TensorDiagnosticOptions] = None):
+        # In this dictionary, the keys are tensors names and the values
+        # are corresponding TensorDiagnostic objects.
+        if opts is None:
+            self.opts = TensorDiagnosticOptions()
+        else:
+            self.opts = opts
+        self.diagnostics = dict()
+
+    def __getitem__(self, name: str):
+        T = ScalarDiagnostic if name[-7:] == ".scalar" else TensorDiagnostic
+        if name not in self.diagnostics:
+            self.diagnostics[name] = T(self.opts, name)
+        return self.diagnostics[name]
+
+    def print_diagnostics(self):
+        """Print diagnostics for each tensor."""
+        for k in sorted(self.diagnostics.keys()):
+            self.diagnostics[k].print_diagnostics()
+
+
+def get_class_name(module: nn.Module):
+    ans = type(module).__name__
+    # we put the below in try blocks in case anyone is using a different version of these modules that
+    # might have different member names.
+    if ans == "Balancer" or ans == "ActivationBalancer":
+        try:
+            ans += f"[{float(module.min_positive)},{float(module.max_positive)},{float(module.min_abs)},{float(module.max_abs)}]"
+        except:
+            pass
+    elif ans == "AbsValuePenalizer":
+        try:
+            ans += f"[{module.limit}]"
+        except:
+            pass
+    return ans
+
+
+def attach_diagnostics(
+    model: nn.Module, opts: Optional[TensorDiagnosticOptions] = None
+) -> ModelDiagnostic:
+    """Attach a ModelDiagnostic object to the model by
+    1) registering forward hook and backward hook on each module, to accumulate
+    its output tensors and gradient tensors, respectively;
+    2) registering backward hook on each module parameter, to accumulate its
+    values and gradients.
+
+    Args:
+      model:
+        the model to be analyzed.
+      opts:
+        Options object.
+
+    Returns:
+      The ModelDiagnostic object attached to the model.
+    """
+
+    ans = ModelDiagnostic(opts)
+    for name, module in model.named_modules():
+        if name == "":
+            name = "<top-level>"
+
+        # Setting model_diagnostic=ans and n=name below, instead of trying to
+        # capture the variables, ensures that we use the current values.
+        # (this matters for `name`, since the variable gets overwritten).
+        # These closures don't really capture by value, only by
+        # "the final value the variable got in the function" :-(
+        def forward_hook(_module, _input, _output, _model_diagnostic=ans, _name=name):
+            if isinstance(_output, tuple) and len(_output) == 1:
+                _output = _output[0]
+
+            if isinstance(_output, Tensor) and _output.dtype in (
+                torch.float32,
+                torch.float16,
+                torch.float64,
+            ):
+                _model_diagnostic[f"{_name}.output"].accumulate(
+                    _output, class_name=get_class_name(_module)
+                )
+            elif isinstance(_output, tuple):
+                for i, o in enumerate(_output):
+                    if isinstance(o, Tensor) and o.dtype in (
+                        torch.float32,
+                        torch.float16,
+                        torch.float64,
+                    ):
+                        _model_diagnostic[f"{_name}.output[{i}]"].accumulate(
+                            o, class_name=get_class_name(_module)
+                        )
+
+        def backward_hook(_module, _input, _output, _model_diagnostic=ans, _name=name):
+            if isinstance(_output, tuple) and len(_output) == 1:
+                _output = _output[0]
+            if isinstance(_output, Tensor) and _output.dtype in (
+                torch.float32,
+                torch.float16,
+                torch.float64,
+            ):
+                _model_diagnostic[f"{_name}.grad"].accumulate(
+                    _output, class_name=get_class_name(_module)
+                )
+            elif isinstance(_output, tuple):
+                for i, o in enumerate(_output):
+                    if isinstance(o, Tensor) and o.dtype in (
+                        torch.float32,
+                        torch.float16,
+                        torch.float64,
+                    ):
+                        _model_diagnostic[f"{_name}.grad[{i}]"].accumulate(
+                            o, class_name=get_class_name(_module)
+                        )
+
+        module.register_forward_hook(forward_hook)
+        module.register_backward_hook(backward_hook)
+
+        if type(module).__name__ in [
+            "Sigmoid",
+            "Tanh",
+            "ReLU",
+            "TanSwish",
+            "Swish",
+            "DoubleSwish",
+            "Swoosh",
+        ]:
+            # For these specific module types, accumulate some additional diagnostics
+            # that can help us improve the activation function.  These require a lot of memory,
+            # to save the forward activations, so limit this to some select classes.
+            # Note: this will not work correctly for all model types.
+            def scalar_forward_hook(
+                _module, _input, _output, _model_diagnostic=ans, _name=name
+            ):
+                if isinstance(_input, tuple):
+                    (_input,) = _input
+                assert isinstance(_input, Tensor)
+                _model_diagnostic[f"{_name}.scalar"].accumulate_input(
+                    _input, class_name=get_class_name(_module)
+                )
+
+            def scalar_backward_hook(
+                _module, _input, _output, _model_diagnostic=ans, _name=name
+            ):
+                if isinstance(_output, tuple):
+                    (_output,) = _output
+                assert isinstance(_output, Tensor)
+                _model_diagnostic[f"{_name}.scalar"].accumulate_output_grad(_output)
+
+            module.register_forward_hook(scalar_forward_hook)
+            module.register_backward_hook(scalar_backward_hook)
+
+    for name, parameter in model.named_parameters():
+
+        def param_backward_hook(
+            grad, _parameter=parameter, _model_diagnostic=ans, _name=name
+        ):
+            _model_diagnostic[f"{_name}.param_value"].accumulate(_parameter)
+            _model_diagnostic[f"{_name}.param_grad"].accumulate(grad)
+
+        parameter.register_hook(param_backward_hook)
+
+    return ans
+
+
+def _test_tensor_diagnostic():
+    opts = TensorDiagnosticOptions(512)
+
+    diagnostic = TensorDiagnostic(opts, "foo")
+
+    for _ in range(10):
+        diagnostic.accumulate(torch.randn(50, 100) * 10.0)
+
+    diagnostic.print_diagnostics()
+
+    model = nn.Sequential(nn.Linear(100, 50), nn.ReLU(), nn.Linear(50, 80))
+
+    diagnostic = attach_diagnostics(model, opts)
+    for _ in range(10):
+        T = random.randint(200, 300)
+        x = torch.randn(T, 100)
+        y = model(x)
+        y.sum().backward()
+
+    diagnostic.print_diagnostics()
+
+
+if __name__ == "__main__":
+    _test_tensor_diagnostic()
diff --git a/icefall/dist.py b/icefall/dist.py
new file mode 100644
index 000000000..922f31a2f
--- /dev/null
+++ b/icefall/dist.py
@@ -0,0 +1,68 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+
+import torch
+from torch import distributed as dist
+
+
+def setup_dist(
+    rank, world_size, master_port=None, use_ddp_launch=False, master_addr=None
+):
+    """
+    rank and world_size are used only if use_ddp_launch is False.
+    """
+    if "MASTER_ADDR" not in os.environ:
+        os.environ["MASTER_ADDR"] = (
+            "localhost" if master_addr is None else str(master_addr)
+        )
+
+    if "MASTER_PORT" not in os.environ:
+        os.environ["MASTER_PORT"] = "12354" if master_port is None else str(master_port)
+
+    if use_ddp_launch is False:
+        dist.init_process_group("nccl", rank=rank, world_size=world_size)
+        torch.cuda.set_device(rank)
+    else:
+        dist.init_process_group("nccl")
+
+
+def cleanup_dist():
+    dist.destroy_process_group()
+
+
+def get_world_size():
+    if "WORLD_SIZE" in os.environ:
+        return int(os.environ["WORLD_SIZE"])
+    if dist.is_available() and dist.is_initialized():
+        return dist.get_world_size()
+    else:
+        return 1
+
+
+def get_rank():
+    if "RANK" in os.environ:
+        return int(os.environ["RANK"])
+    elif dist.is_available() and dist.is_initialized():
+        return dist.get_rank()
+    else:
+        return 0
+
+
+def get_local_rank():
+    return int(os.environ.get("LOCAL_RANK", 0))
diff --git a/icefall/env.py b/icefall/env.py
new file mode 100644
index 000000000..373e9a9ff
--- /dev/null
+++ b/icefall/env.py
@@ -0,0 +1,120 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                    Wei Kang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import socket
+import subprocess
+import sys
+from pathlib import Path
+from typing import Any, Dict
+
+import k2
+import k2.version
+import lhotse
+import torch
+
+
+def get_git_sha1():
+    try:
+        git_commit = (
+            subprocess.run(
+                ["git", "rev-parse", "--short", "HEAD"],
+                check=True,
+                stdout=subprocess.PIPE,
+            )
+            .stdout.decode()
+            .rstrip("\n")
+            .strip()
+        )
+        dirty_commit = (
+            len(
+                subprocess.run(
+                    ["git", "diff", "--shortstat"],
+                    check=True,
+                    stdout=subprocess.PIPE,
+                )
+                .stdout.decode()
+                .rstrip("\n")
+                .strip()
+            )
+            > 0
+        )
+        git_commit = git_commit + "-dirty" if dirty_commit else git_commit + "-clean"
+    except:  # noqa
+        return None
+
+    return git_commit
+
+
+def get_git_date():
+    try:
+        git_date = (
+            subprocess.run(
+                ["git", "log", "-1", "--format=%ad", "--date=local"],
+                check=True,
+                stdout=subprocess.PIPE,
+            )
+            .stdout.decode()
+            .rstrip("\n")
+            .strip()
+        )
+    except:  # noqa
+        return None
+
+    return git_date
+
+
+def get_git_branch_name():
+    try:
+        git_date = (
+            subprocess.run(
+                ["git", "rev-parse", "--abbrev-ref", "HEAD"],
+                check=True,
+                stdout=subprocess.PIPE,
+            )
+            .stdout.decode()
+            .rstrip("\n")
+            .strip()
+        )
+    except:  # noqa
+        return None
+
+    return git_date
+
+
+def get_env_info() -> Dict[str, Any]:
+    """Get the environment information."""
+    return {
+        "k2-version": k2.version.__version__,
+        "k2-build-type": k2.version.__build_type__,
+        "k2-with-cuda": k2.with_cuda,
+        "k2-git-sha1": k2.version.__git_sha1__,
+        "k2-git-date": k2.version.__git_date__,
+        "lhotse-version": lhotse.__version__,
+        "torch-version": str(torch.__version__),
+        "torch-cuda-available": torch.cuda.is_available(),
+        "torch-cuda-version": torch.version.cuda,
+        "python-version": sys.version[:3],
+        "icefall-git-branch": get_git_branch_name(),
+        "icefall-git-sha1": get_git_sha1(),
+        "icefall-git-date": get_git_date(),
+        "icefall-path": str(Path(__file__).resolve().parent.parent),
+        "k2-path": str(Path(k2.__file__).resolve()),
+        "lhotse-path": str(Path(lhotse.__file__).resolve()),
+        "hostname": socket.gethostname(),
+        "IP address": socket.gethostbyname(socket.gethostname()),
+    }
diff --git a/icefall/graph_compiler.py b/icefall/graph_compiler.py
new file mode 100644
index 000000000..d26ddbbd1
--- /dev/null
+++ b/icefall/graph_compiler.py
@@ -0,0 +1,161 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+class CtcTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lexicon: Lexicon,
+        device: torch.device,
+        oov: str = "<UNK>",
+        need_repeat_flag: bool = False,
+    ):
+        """
+        Args:
+          lexicon:
+            It is built from `data/lang/lexicon.txt`.
+          device:
+            The device to use for operations compiling transcripts to FSAs.
+          oov:
+            Out of vocabulary word. When a word in the transcript
+            does not exist in the lexicon, it is replaced with `oov`.
+          need_repeat_flag:
+            If True, will add an attribute named `_is_repeat_token_` to ctc_topo
+            indicating whether this token is a repeat token in ctc graph.
+            This attribute is needed to implement delay-penalty for phone-based
+            ctc loss. See https://github.com/k2-fsa/k2/pull/1086 for more
+            details. Note: The above change MUST be included in k2 to open this
+            flag.
+        """
+        L_inv = lexicon.L_inv.to(device)
+        assert L_inv.requires_grad is False
+
+        assert oov in lexicon.word_table
+
+        self.L_inv = k2.arc_sort(L_inv)
+        self.oov_id = lexicon.word_table[oov]
+        self.word_table = lexicon.word_table
+
+        max_token_id = max(lexicon.tokens)
+        ctc_topo = k2.ctc_topo(max_token_id, modified=False)
+
+        self.ctc_topo = ctc_topo.to(device)
+
+        if need_repeat_flag:
+            self.ctc_topo._is_repeat_token_ = (
+                self.ctc_topo.labels != self.ctc_topo.aux_labels
+            )
+
+        self.device = device
+
+    def compile(self, texts: List[str]) -> k2.Fsa:
+        """Build decoding graphs by composing ctc_topo with
+        given transcripts.
+
+        Args:
+          texts:
+            A list of strings. Each string contains a sentence for an utterance.
+            A sentence consists of spaces separated words. An example `texts`
+            looks like:
+
+                ['hello icefall', 'CTC training with k2']
+
+        Returns:
+          An FsaVec, the composition result of `self.ctc_topo` and the
+          transcript FSA.
+        """
+        transcript_fsa = self.convert_transcript_to_fsa(texts)
+
+        # NOTE: k2.compose runs on CUDA only when treat_epsilons_specially
+        # is False, so we add epsilon self-loops here
+        fsa_with_self_loops = k2.remove_epsilon_and_add_self_loops(transcript_fsa)
+
+        fsa_with_self_loops = k2.arc_sort(fsa_with_self_loops)
+
+        decoding_graph = k2.compose(
+            self.ctc_topo, fsa_with_self_loops, treat_epsilons_specially=False
+        )
+
+        assert decoding_graph.requires_grad is False
+
+        return decoding_graph
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of word IDs.
+
+        Args:
+          texts:
+            It is a list of strings. Each string consists of space(s)
+            separated words. An example containing two strings is given below:
+
+                ['HELLO ICEFALL', 'HELLO k2']
+        Returns:
+          Return a list-of-list of word IDs.
+        """
+        word_ids_list = []
+        for text in texts:
+            word_ids = []
+            for word in text.split():
+                if word in self.word_table:
+                    word_ids.append(self.word_table[word])
+                else:
+                    word_ids.append(self.oov_id)
+            word_ids_list.append(word_ids)
+        return word_ids_list
+
+    def convert_transcript_to_fsa(self, texts: List[str]) -> k2.Fsa:
+        """Convert a list of transcript texts to an FsaVec.
+
+        Args:
+          texts:
+            A list of strings. Each string contains a sentence for an utterance.
+            A sentence consists of spaces separated words. An example `texts`
+            looks like:
+
+                ['hello icefall', 'CTC training with k2']
+
+        Returns:
+          Return an FsaVec, whose `shape[0]` equals to `len(texts)`.
+        """
+        word_ids_list = []
+        for text in texts:
+            word_ids = []
+            for word in text.split():
+                if word in self.word_table:
+                    word_ids.append(self.word_table[word])
+                else:
+                    word_ids.append(self.oov_id)
+            word_ids_list.append(word_ids)
+
+        word_fsa = k2.linear_fsa(word_ids_list, self.device)
+
+        word_fsa_with_self_loops = k2.add_epsilon_self_loops(word_fsa)
+
+        fsa = k2.intersect(
+            self.L_inv, word_fsa_with_self_loops, treat_epsilons_specially=False
+        )
+        # fsa has word ID as labels and token ID as aux_labels, so
+        # we need to invert it
+        ans_fsa = fsa.invert_()
+        return k2.arc_sort(ans_fsa)
diff --git a/icefall/hooks.py b/icefall/hooks.py
new file mode 100644
index 000000000..398a5f689
--- /dev/null
+++ b/icefall/hooks.py
@@ -0,0 +1,95 @@
+# Copyright  2021-2022  Xiaomi Corporation  (authors: Zengwei Yao, Daniel Povey)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import random
+
+import torch
+from torch import Tensor, nn
+
+
+def register_inf_check_hooks(model: nn.Module) -> None:
+    """Registering forward hook on each module, to check
+    whether its output tensors is not finite.
+
+    Args:
+      model:
+        the model to be analyzed.
+    """
+
+    for name, module in model.named_modules():
+        if name == "":
+            name = "<top-level>"
+
+        # default param _name is a way to capture the current value of the variable "name".
+        def forward_hook(_module, _input, _output, _name=name):
+            if isinstance(_output, Tensor):
+                if not torch.isfinite(_output.to(torch.float32).sum()):
+                    raise ValueError(
+                        f"The sum of {_name}.output is not finite: {_output}"
+                    )
+            elif isinstance(_output, tuple):
+                for i, o in enumerate(_output):
+                    if isinstance(o, tuple):
+                        o = o[0]
+                    if not isinstance(o, Tensor):
+                        continue
+                    if not torch.isfinite(o.to(torch.float32).sum()):
+                        raise ValueError(
+                            f"The sum of {_name}.output[{i}] is not finite: {_output}"
+                        )
+
+        # default param _name is a way to capture the current value of the variable "name".
+        def backward_hook(_module, _input, _output, _name=name):
+            if isinstance(_output, Tensor):
+                if not torch.isfinite(_output.to(torch.float32).sum()):
+                    logging.warning(
+                        f"The sum of {_name}.grad is not finite"  # ": {_output}"
+                    )
+            elif isinstance(_output, tuple):
+                for i, o in enumerate(_output):
+                    if isinstance(o, tuple):
+                        o = o[0]
+                    if not isinstance(o, Tensor):
+                        continue
+                    if not torch.isfinite(o.to(torch.float32).sum()):
+                        logging.warning(f"The sum of {_name}.grad[{i}] is not finite")
+
+        module.register_forward_hook(forward_hook)
+        module.register_backward_hook(backward_hook)
+
+    for name, parameter in model.named_parameters():
+
+        def param_backward_hook(grad, _name=name):
+            if not torch.isfinite(grad.to(torch.float32).sum()):
+                logging.warning(f"The sum of {_name}.param_grad is not finite")
+
+        parameter.register_hook(param_backward_hook)
+
+
+def _test_inf_check_hooks():
+    model = nn.Sequential(nn.Linear(100, 50), nn.Linear(50, 80))
+
+    register_inf_check_hooks(model)
+    for _ in range(10):
+        T = random.randint(200, 300)
+        x = torch.randn(T, 100) + float("inf") * (T % 2)
+        y = model(x)
+        y.sum().backward()
+
+
+if __name__ == "__main__":
+    _test_inf_check_hooks()
diff --git a/icefall/lexicon.py b/icefall/lexicon.py
new file mode 100644
index 000000000..22e1b78bb
--- /dev/null
+++ b/icefall/lexicon.py
@@ -0,0 +1,269 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import logging
+import re
+import sys
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+
+
+def read_lexicon(filename: str) -> List[Tuple[str, List[str]]]:
+    """Read a lexicon from `filename`.
+
+    Each line in the lexicon contains "word p1 p2 p3 ...".
+    That is, the first field is a word and the remaining
+    fields are tokens. Fields are separated by space(s).
+
+    Args:
+      filename:
+        Path to the lexicon.txt
+
+    Returns:
+      A list of tuples., e.g., [('w', ['p1', 'p2']), ('w1', ['p3, 'p4'])]
+    """
+    ans = []
+
+    with open(filename, "r", encoding="utf-8") as f:
+        whitespace = re.compile("[ \t]+")
+        for line in f:
+            a = whitespace.split(line.strip(" \t\r\n"))
+            if len(a) == 0:
+                continue
+
+            if len(a) < 2:
+                logging.info(f"Found bad line {line} in lexicon file {filename}")
+                logging.info("Every line is expected to contain at least 2 fields")
+                sys.exit(1)
+            word = a[0]
+            if word == "<eps>":
+                logging.info(f"Found bad line {line} in lexicon file {filename}")
+                logging.info("<eps> should not be a valid word")
+                sys.exit(1)
+
+            tokens = a[1:]
+            ans.append((word, tokens))
+
+    return ans
+
+
+def write_lexicon(filename: str, lexicon: List[Tuple[str, List[str]]]) -> None:
+    """Write a lexicon to a file.
+
+    Args:
+      filename:
+        Path to the lexicon file to be generated.
+      lexicon:
+        It can be the return value of :func:`read_lexicon`.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for word, tokens in lexicon:
+            f.write(f"{word} {' '.join(tokens)}\n")
+
+
+def convert_lexicon_to_ragged(
+    filename: str, word_table: k2.SymbolTable, token_table: k2.SymbolTable
+) -> k2.RaggedTensor:
+    """Read a lexicon and convert it to a ragged tensor.
+
+    The ragged tensor has two axes: [word][token].
+
+    Caution:
+      We assume that each word has a unique pronunciation.
+
+    Args:
+      filename:
+        Filename of the lexicon. It has a format that can be read
+        by :func:`read_lexicon`.
+      word_table:
+        The word symbol table.
+      token_table:
+        The token symbol table.
+    Returns:
+      A k2 ragged tensor with two axes [word][token].
+    """
+    disambig_id = word_table["#0"]
+    # We reuse the same words.txt from the phone based lexicon
+    # so that we can share the same G.fst. Here, we have to
+    # exclude some words present only in the phone based lexicon.
+    excluded_words = ["<eps>", "!SIL", "<SPOKEN_NOISE>"]
+
+    # epsilon is not a word, but it occupies a position
+    #
+    row_splits = [0]
+    token_ids_list = []
+
+    lexicon_tmp = read_lexicon(filename)
+    lexicon = dict(lexicon_tmp)
+    if len(lexicon_tmp) != len(lexicon):
+        raise RuntimeError("It's assumed that each word has a unique pronunciation")
+
+    for i in range(disambig_id):
+        w = word_table[i]
+        if w in excluded_words:
+            row_splits.append(row_splits[-1])
+            continue
+        tokens = lexicon[w]
+        token_ids = [token_table[k] for k in tokens]
+
+        row_splits.append(row_splits[-1] + len(token_ids))
+        token_ids_list.extend(token_ids)
+
+    cached_tot_size = row_splits[-1]
+    row_splits = torch.tensor(row_splits, dtype=torch.int32)
+
+    shape = k2.ragged.create_ragged_shape2(
+        row_splits,
+        None,
+        cached_tot_size,
+    )
+    values = torch.tensor(token_ids_list, dtype=torch.int32)
+
+    return k2.RaggedTensor(shape, values)
+
+
+class Lexicon(object):
+    """Phone based lexicon."""
+
+    def __init__(
+        self,
+        lang_dir: Path,
+        disambig_pattern: str = re.compile(r"^#\d+$"),
+    ):
+        """
+        Args:
+          lang_dir:
+            Path to the lang directory. It is expected to contain the following
+            files:
+                - tokens.txt
+                - words.txt
+                - L.pt
+            The above files are produced by the script `prepare.sh`. You
+            should have run that before running the training code.
+          disambig_pattern:
+            It contains the pattern for disambiguation symbols.
+        """
+        lang_dir = Path(lang_dir)
+        self.token_table = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        self.word_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        if (lang_dir / "Linv.pt").exists():
+            logging.info(f"Loading pre-compiled {lang_dir}/Linv.pt")
+            L_inv = k2.Fsa.from_dict(torch.load(lang_dir / "Linv.pt"))
+        else:
+            logging.info("Converting L.pt to Linv.pt")
+            L = k2.Fsa.from_dict(torch.load(lang_dir / "L.pt"))
+            L_inv = k2.arc_sort(L.invert())
+            torch.save(L_inv.as_dict(), lang_dir / "Linv.pt")
+
+        # We save L_inv instead of L because it will be used to intersect with
+        # transcript FSAs, both of whose labels are word IDs.
+        self.L_inv = L_inv
+        self.disambig_pattern = disambig_pattern
+
+    @property
+    def tokens(self) -> List[int]:
+        """Return a list of token IDs excluding those from
+        disambiguation symbols.
+
+        Caution:
+          0 is not a token ID so it is excluded from the return value.
+        """
+        symbols = self.token_table.symbols
+        ans = []
+        for s in symbols:
+            if not self.disambig_pattern.match(s):
+                ans.append(self.token_table[s])
+        if 0 in ans:
+            ans.remove(0)
+        ans.sort()
+        return ans
+
+
+class UniqLexicon(Lexicon):
+    def __init__(
+        self,
+        lang_dir: Path,
+        uniq_filename: str = "uniq_lexicon.txt",
+        disambig_pattern: str = re.compile(r"^#\d+$"),
+    ):
+        """
+        Refer to the help information in Lexicon.__init__.
+
+        uniq_filename: It is assumed to be inside the given `lang_dir`.
+
+        Each word in the lexicon is assumed to have a unique pronunciation.
+        """
+        lang_dir = Path(lang_dir)
+        super().__init__(lang_dir=lang_dir, disambig_pattern=disambig_pattern)
+
+        self.ragged_lexicon = convert_lexicon_to_ragged(
+            filename=lang_dir / uniq_filename,
+            word_table=self.word_table,
+            token_table=self.token_table,
+        )
+        # TODO: should we move it to a certain device ?
+
+    def texts_to_token_ids(
+        self, texts: List[str], oov: str = "<UNK>"
+    ) -> k2.RaggedTensor:
+        """
+        Args:
+          texts:
+            A list of transcripts. Each transcript contains space(s)
+            separated words. An example texts is::
+
+                ['HELLO k2', 'HELLO icefall']
+          oov:
+            The OOV word. If a word in `texts` is not in the lexicon, it is
+            replaced with `oov`.
+        Returns:
+          Return a ragged int tensor with 2 axes [utterance][token_id]
+        """
+        oov_id = self.word_table[oov]
+
+        word_ids_list = []
+        for text in texts:
+            word_ids = []
+            for word in text.split():
+                if word in self.word_table:
+                    word_ids.append(self.word_table[word])
+                else:
+                    word_ids.append(oov_id)
+            word_ids_list.append(word_ids)
+        ragged_indexes = k2.RaggedTensor(word_ids_list, dtype=torch.int32)
+        ans = self.ragged_lexicon.index(ragged_indexes)
+        ans = ans.remove_axis(ans.num_axes - 2)
+        return ans
+
+    def words_to_token_ids(self, words: List[str]) -> k2.RaggedTensor:
+        """Convert a list of words to a ragged tensor containing token IDs.
+
+        We assume there are no OOVs in "words".
+        """
+        word_ids = [self.word_table[w] for w in words]
+        word_ids = torch.tensor(word_ids, dtype=torch.int32)
+
+        ragged, _ = self.ragged_lexicon.index(
+            indexes=word_ids,
+            axis=0,
+            need_value_indexes=False,
+        )
+        return ragged
diff --git a/icefall/lm_wrapper.py b/icefall/lm_wrapper.py
new file mode 100644
index 000000000..5e2783a47
--- /dev/null
+++ b/icefall/lm_wrapper.py
@@ -0,0 +1,254 @@
+# Copyright (c)  2022  Xiaomi Corporation (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+
+import torch
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.transformer_lm.model import TransformerLM
+from icefall.utils import AttributeDict, str2bool
+
+
+class LmScorer(torch.nn.Module):
+    """This is a wrapper for NN LMs
+    The language models supported include:
+        RNN,
+        Transformer
+    """
+
+    def __init__(
+        self,
+        lm_type: str,
+        params: AttributeDict,
+        device,
+        lm_scale: float = 0.3,
+    ):
+        super(LmScorer, self).__init__()
+        assert lm_type in ["rnn", "transformer"], f"{lm_type} is not supported"
+        self.lm_type = lm_type
+        self.lm = self.get_lm(lm_type, device, params)
+        self.lm_scale = lm_scale
+        self.params = params
+
+    @classmethod
+    def add_arguments(cls, parser):
+        # LM general arguments
+        parser.add_argument(
+            "--lm-vocab-size",
+            type=int,
+            default=500,
+        )
+
+        parser.add_argument(
+            "--lm-epoch",
+            type=int,
+            default=7,
+            help="""Which epoch to be used
+            """,
+        )
+
+        parser.add_argument(
+            "--lm-avg",
+            type=int,
+            default=1,
+            help="""Number of checkpoints to be averaged
+            """,
+        )
+
+        parser.add_argument("--lm-exp-dir", type=str, help="Path to LM experiments")
+
+        # Now RNNLM related arguments
+        parser.add_argument(
+            "--rnn-lm-embedding-dim",
+            type=int,
+            default=2048,
+            help="Embedding dim of the model",
+        )
+
+        parser.add_argument(
+            "--rnn-lm-hidden-dim",
+            type=int,
+            default=2048,
+            help="Hidden dim of the model",
+        )
+
+        parser.add_argument(
+            "--rnn-lm-num-layers",
+            type=int,
+            default=3,
+            help="Number of RNN layers the model",
+        )
+
+        parser.add_argument(
+            "--rnn-lm-tie-weights",
+            type=str2bool,
+            default=True,
+            help="""True to share the weights between the input embedding layer and the
+            last output linear layer
+            """,
+        )
+
+        # Now transformers
+        parser.add_argument(
+            "--transformer-lm-exp-dir", type=str, help="Directory of transformer LM exp"
+        )
+
+        parser.add_argument(
+            "--transformer-lm-dim-feedforward",
+            type=int,
+            default=2048,
+            help="Dimension of FFW module in transformer",
+        )
+
+        parser.add_argument(
+            "--transformer-lm-encoder-dim",
+            type=int,
+            default=768,
+            help="Encoder dimension of transformer",
+        )
+
+        parser.add_argument(
+            "--transformer-lm-embedding-dim",
+            type=int,
+            default=768,
+            help="Input embedding dimension of transformer",
+        )
+
+        parser.add_argument(
+            "--transformer-lm-nhead",
+            type=int,
+            default=8,
+            help="Number of attention heads in transformer",
+        )
+
+        parser.add_argument(
+            "--transformer-lm-num-layers",
+            type=int,
+            default=16,
+            help="Number of encoder layers in transformer",
+        )
+
+        parser.add_argument(
+            "--transformer-lm-tie-weights",
+            type=str2bool,
+            default=True,
+            help="If tie weights in transformer LM",
+        )
+
+    def get_lm(self, lm_type: str, device, params: AttributeDict) -> torch.nn.Module:
+        """Return the neural network LM
+
+        Args:
+            lm_type (str): Type name of NN LM
+        """
+        if lm_type == "rnn":
+            model = RnnLmModel(
+                vocab_size=params.vocab_size,
+                embedding_dim=params.rnn_lm_embedding_dim,
+                hidden_dim=params.rnn_lm_hidden_dim,
+                num_layers=params.rnn_lm_num_layers,
+                tie_weights=params.rnn_lm_tie_weights,
+            )
+
+            if params.lm_avg == 1:
+                load_checkpoint(
+                    f"{params.lm_exp_dir}/epoch-{params.lm_epoch}.pt", model
+                )
+                model.to(device)
+            else:
+                start = params.lm_epoch - params.lm_avg + 1
+                filenames = []
+                for i in range(start, params.lm_epoch + 1):
+                    if start >= 0:
+                        filenames.append(f"{params.lm_exp_dir}/epoch-{i}.pt")
+                logging.info(f"averaging {filenames}")
+                model.to(device)
+                model.load_state_dict(average_checkpoints(filenames, device=device))
+
+        elif lm_type == "transformer":
+            model = TransformerLM(
+                vocab_size=params.vocab_size,
+                d_model=params.transformer_lm_encoder_dim,
+                embedding_dim=params.transformer_lm_embedding_dim,
+                dim_feedforward=params.transformer_lm_dim_feedforward,
+                nhead=params.transformer_lm_nhead,
+                num_layers=params.transformer_lm_num_layers,
+                tie_weights=params.transformer_lm_tie_weights,
+                params=params,
+            )
+
+            if params.lm_avg == 1:
+                load_checkpoint(
+                    f"{params.lm_exp_dir}/epoch-{params.lm_epoch}.pt", model
+                )
+                model.to(device)
+            else:
+                start = params.lm_epoch - params.lm_avg + 1
+                filenames = []
+                for i in range(start, params.lm_epoch + 1):
+                    if start >= 0:
+                        filenames.append(f"{params.lm_exp_dir}/epoch-{i}.pt")
+                logging.info(f"averaging {filenames}")
+                model.to(device)
+                model.load_state_dict(average_checkpoints(filenames, device=device))
+        else:
+            raise NotImplementedError()
+
+        return model
+
+    def score_token(self, x: torch.Tensor, x_lens: torch.Tensor, state=None):
+        """Score the input and return the prediction
+        This requires the lm to have the method `score_token`
+        Args:
+            x (torch.Tensor): Input tokens
+            x_lens (torch.Tensor): Length of the input tokens
+            state (optional): LM states
+
+        """
+        return self.lm.score_token(x, x_lens, state)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = AttributeDict()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    Scorer = LmScorer(params=params, device=device)
+    Scorer.eval()
+
+    x = (
+        torch.tensor([[1, 4, 19, 256, 77], [1, 4, 19, 256, 77]])
+        .to(device)
+        .to(torch.int64)
+    )
+    x_lens = torch.tensor([5, 5]).to(device)
+
+    state = None
+
+    score, state = Scorer.score(x, x_lens)
+    print(score.shape)
+    print(score[0])
+    print(score[1])
diff --git a/icefall/mmi.py b/icefall/mmi.py
new file mode 100644
index 000000000..b7777b434
--- /dev/null
+++ b/icefall/mmi.py
@@ -0,0 +1,221 @@
+from typing import List
+
+import k2
+import torch
+from torch import nn
+
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+
+
+def _compute_mmi_loss_exact_optimized(
+    dense_fsa_vec: k2.DenseFsaVec,
+    texts: List[str],
+    graph_compiler: MmiTrainingGraphCompiler,
+    den_scale: float = 1.0,
+    beam_size: float = 8.0,
+) -> torch.Tensor:
+    """
+    The function name contains `exact`, which means it uses a version of
+    intersection without pruning.
+
+    `optimized` in the function name means this function is optimized
+    in that it calls k2.intersect_dense only once
+
+    Note:
+      It is faster at the cost of using more memory.
+
+    Args:
+      dense_fsa_vec:
+        It contains the neural network output.
+      texts:
+        The transcript. Each element consists of space(s) separated words.
+      graph_compiler:
+        Used to build num_graphs and den_graphs
+      den_scale:
+        The scale applied to the denominator tot_scores.
+    Returns:
+      Return a scalar loss. It is the sum over utterances in a batch,
+      without normalization.
+    """
+    num_graphs, den_graphs = graph_compiler.compile(texts, replicate_den=False)
+
+    device = num_graphs.device
+
+    num_fsas = num_graphs.shape[0]
+    assert dense_fsa_vec.dim0() == num_fsas
+
+    assert den_graphs.shape[0] == 1
+
+    # The motivation to concatenate num_graphs and den_graphs
+    # is to reduce the number of calls to k2.intersect_dense.
+    num_den_graphs = k2.cat([num_graphs, den_graphs])
+
+    # NOTE: The a_to_b_map in k2.intersect_dense must be sorted
+    # so the following reorders num_den_graphs.
+    #
+    # The following code computes a_to_b_map
+
+    # [0, 1, 2, ... ]
+    num_graphs_indexes = torch.arange(num_fsas, dtype=torch.int32)
+
+    # [num_fsas, num_fsas, num_fsas, ... ]
+    den_graphs_indexes = torch.tensor([num_fsas] * num_fsas, dtype=torch.int32)
+
+    # [0, num_fsas, 1, num_fsas, 2, num_fsas, ... ]
+    num_den_graphs_indexes = (
+        torch.stack([num_graphs_indexes, den_graphs_indexes]).t().reshape(-1).to(device)
+    )
+
+    num_den_reordered_graphs = k2.index(num_den_graphs, num_den_graphs_indexes)
+
+    # [[0, 1, 2, ...]]
+    a_to_b_map = torch.arange(num_fsas, dtype=torch.int32).reshape(1, -1)
+
+    # [[0, 1, 2, ...]] -> [0, 0, 1, 1, 2, 2, ... ]
+    a_to_b_map = a_to_b_map.repeat(2, 1).t().reshape(-1).to(device)
+
+    num_den_lats = k2.intersect_dense(
+        num_den_reordered_graphs,
+        dense_fsa_vec,
+        output_beam=beam_size,
+        a_to_b_map=a_to_b_map,
+    )
+
+    num_den_tot_scores = num_den_lats.get_tot_scores(
+        log_semiring=True, use_double_scores=True
+    )
+
+    num_tot_scores = num_den_tot_scores[::2]
+    den_tot_scores = num_den_tot_scores[1::2]
+
+    tot_scores = num_tot_scores - den_scale * den_tot_scores
+    loss = -1 * tot_scores.sum()
+    return loss
+
+
+def _compute_mmi_loss_exact_non_optimized(
+    dense_fsa_vec: k2.DenseFsaVec,
+    texts: List[str],
+    graph_compiler: MmiTrainingGraphCompiler,
+    den_scale: float = 1.0,
+    beam_size: float = 8.0,
+) -> torch.Tensor:
+    """
+    See :func:`_compute_mmi_loss_exact_optimized` for the meaning
+    of the arguments.
+
+    It's more readable, though it invokes k2.intersect_dense twice.
+
+    Note:
+      It uses less memory at the cost of speed. It is slower.
+    """
+    num_graphs, den_graphs = graph_compiler.compile(texts, replicate_den=True)
+
+    # TODO: pass output_beam as function argument
+    num_lats = k2.intersect_dense(
+        num_graphs, dense_fsa_vec, output_beam=beam_size, max_arcs=2147483600
+    )
+    den_lats = k2.intersect_dense(
+        den_graphs, dense_fsa_vec, output_beam=beam_size, max_arcs=2147483600
+    )
+
+    num_tot_scores = num_lats.get_tot_scores(log_semiring=True, use_double_scores=True)
+
+    den_tot_scores = den_lats.get_tot_scores(log_semiring=True, use_double_scores=True)
+
+    tot_scores = num_tot_scores - den_scale * den_tot_scores
+
+    loss = -1 * tot_scores.sum()
+    return loss
+
+
+def _compute_mmi_loss_pruned(
+    dense_fsa_vec: k2.DenseFsaVec,
+    texts: List[str],
+    graph_compiler: MmiTrainingGraphCompiler,
+    den_scale: float = 1.0,
+    beam_size: float = 8.0,
+) -> torch.Tensor:
+    """
+    See :func:`_compute_mmi_loss_exact_optimized` for the meaning
+    of the arguments.
+
+    `pruned` means it uses k2.intersect_dense_pruned
+
+    Note:
+      It uses the least amount of memory, but the loss is not exact due
+      to pruning.
+    """
+    num_graphs, den_graphs = graph_compiler.compile(texts, replicate_den=False)
+
+    num_lats = k2.intersect_dense(num_graphs, dense_fsa_vec, output_beam=8.0)
+
+    # the values for search_beam/output_beam/min_active_states/max_active_states
+    # are not tuned. You may want to tune them.
+    den_lats = k2.intersect_dense_pruned(
+        den_graphs,
+        dense_fsa_vec,
+        search_beam=20.0,
+        output_beam=beam_size,
+        min_active_states=30,
+        max_active_states=10000,
+    )
+
+    num_tot_scores = num_lats.get_tot_scores(log_semiring=True, use_double_scores=True)
+
+    den_tot_scores = den_lats.get_tot_scores(log_semiring=True, use_double_scores=True)
+
+    tot_scores = num_tot_scores - den_scale * den_tot_scores
+
+    loss = -1 * tot_scores.sum()
+    return loss
+
+
+class LFMMILoss(nn.Module):
+    """
+    Computes Lattice-Free Maximum Mutual Information (LFMMI) loss.
+
+    TODO: more detailed description
+    """
+
+    def __init__(
+        self,
+        graph_compiler: MmiTrainingGraphCompiler,
+        use_pruned_intersect: bool = False,
+        den_scale: float = 1.0,
+        beam_size: float = 8.0,
+    ):
+        super().__init__()
+        self.graph_compiler = graph_compiler
+        self.den_scale = den_scale
+        self.use_pruned_intersect = use_pruned_intersect
+        self.beam_size = beam_size
+
+    def forward(
+        self,
+        dense_fsa_vec: k2.DenseFsaVec,
+        texts: List[str],
+    ) -> torch.Tensor:
+        """
+        Args:
+          dense_fsa_vec:
+            It contains the neural network output.
+          texts:
+            A list of strings. Each string contains space(s) separated words.
+        Returns:
+          Return a scalar loss. It is the sum over utterances in a batch,
+          without normalization.
+        """
+        if self.use_pruned_intersect:
+            func = _compute_mmi_loss_pruned
+        else:
+            func = _compute_mmi_loss_exact_non_optimized
+            #  func = _compute_mmi_loss_exact_optimized
+
+        return func(
+            dense_fsa_vec=dense_fsa_vec,
+            texts=texts,
+            graph_compiler=self.graph_compiler,
+            den_scale=self.den_scale,
+            beam_size=self.beam_size,
+        )
diff --git a/icefall/mmi_graph_compiler.py b/icefall/mmi_graph_compiler.py
new file mode 100644
index 000000000..600f09f2b
--- /dev/null
+++ b/icefall/mmi_graph_compiler.py
@@ -0,0 +1,219 @@
+import logging
+from pathlib import Path
+from typing import Iterable, List, Tuple, Union
+
+import k2
+import torch
+
+from icefall.lexicon import UniqLexicon
+
+
+class MmiTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lang_dir: Path,
+        uniq_filename: str = "uniq_lexicon.txt",
+        device: Union[str, torch.device] = "cpu",
+        oov: str = "<UNK>",
+        sos_id: int = 1,
+        eos_id: int = 1,
+    ):
+        """
+        Args:
+          lang_dir:
+            Path to the lang directory. It is expected to contain the
+            following files::
+
+                - tokens.txt
+                - words.txt
+                - P.fst.txt
+
+            The above files are generated by the script `prepare.sh`. You
+            should have run it before running the training code.
+          uniq_filename:
+            File name to the lexicon in which every word has exactly one
+            pronunciation. We assume this file is inside the given `lang_dir`.
+
+          device:
+            It indicates CPU or CUDA.
+          oov:
+            Out of vocabulary word. When a word in the transcript
+            does not exist in the lexicon, it is replaced with `oov`.
+        """
+        self.lang_dir = Path(lang_dir)
+        self.lexicon = UniqLexicon(lang_dir, uniq_filename=uniq_filename)
+        self.device = torch.device(device)
+
+        self.L_inv = self.lexicon.L_inv.to(self.device)
+
+        self.oov_id = self.lexicon.word_table[oov]
+        self.sos_id = sos_id
+        self.eos_id = eos_id
+
+        self.build_ctc_topo_P()
+
+    def build_ctc_topo_P(self):
+        """Built ctc_topo_P, the composition result of
+        ctc_topo and P, where P is a pre-trained bigram
+        word piece LM.
+        """
+        # Note: there is no need to save a pre-compiled P and ctc_topo
+        # as it is very fast to generate them.
+        logging.info(f"Loading P from {self.lang_dir/'P.fst.txt'}")
+        with open(self.lang_dir / "P.fst.txt") as f:
+            # P is not an acceptor because there is
+            # a back-off state, whose incoming arcs
+            # have label #0 and aux_label 0 (i.e., <eps>).
+            P = k2.Fsa.from_openfst(f.read(), acceptor=False)
+
+        first_token_disambig_id = self.lexicon.token_table["#0"]
+
+        # P.aux_labels is not needed in later computations, so
+        # remove it here.
+        del P.aux_labels
+        # CAUTION: The following line is crucial.
+        # Arcs entering the back-off state have label equal to #0.
+        # We have to change it to 0 here.
+        labels = P.labels.clone()
+        labels[labels >= first_token_disambig_id] = 0
+        P.labels = labels
+
+        P = k2.remove_epsilon(P)
+        P = k2.arc_sort(P)
+        P = P.to(self.device)
+        # Add epsilon self-loops to P because we want the
+        # following operation "k2.intersect" to run on GPU.
+        P_with_self_loops = k2.add_epsilon_self_loops(P)
+
+        max_token_id = max(self.lexicon.tokens)
+        logging.info(
+            f"Building ctc_topo (modified=False). max_token_id: {max_token_id}"
+        )
+        ctc_topo = k2.ctc_topo(max_token_id, modified=False, device=self.device)
+
+        ctc_topo_inv = k2.arc_sort(ctc_topo.invert_())
+
+        logging.info("Building ctc_topo_P")
+        ctc_topo_P = k2.intersect(
+            ctc_topo_inv, P_with_self_loops, treat_epsilons_specially=False
+        ).invert()
+
+        self.ctc_topo_P = k2.arc_sort(ctc_topo_P)
+        logging.info(f"ctc_topo_P num_arcs: {self.ctc_topo_P.num_arcs}")
+
+    def compile(
+        self, texts: Iterable[str], replicate_den: bool = True
+    ) -> Tuple[k2.Fsa, k2.Fsa]:
+        """Create numerator and denominator graphs from transcripts
+        and the bigram phone LM.
+
+        Args:
+          texts:
+            A list of transcripts. Within a transcript, words are
+            separated by spaces. An example `texts` is given below::
+
+                ["Hello icefall", "LF-MMI training with icefall using k2"]
+
+          replicate_den:
+            If True, the returned den_graph is replicated to match the number
+            of FSAs in the returned num_graph; if False, the returned den_graph
+            contains only a single FSA
+        Returns:
+          A tuple (num_graph, den_graph), where
+
+            - `num_graph` is the numerator graph. It is an FsaVec with
+              shape `(len(texts), None, None)`.
+
+            - `den_graph` is the denominator graph. It is an FsaVec
+              with the same shape of the `num_graph` if replicate_den is
+              True; otherwise, it is an FsaVec containing only a single FSA.
+        """
+        transcript_fsa = self.build_transcript_fsa(texts)
+
+        # remove word IDs from transcript_fsa since it is not needed
+        del transcript_fsa.aux_labels
+        # NOTE: You can comment out the above statement
+        # if you want to run test/test_mmi_graph_compiler.py
+
+        transcript_fsa_with_self_loops = k2.remove_epsilon_and_add_self_loops(
+            transcript_fsa
+        )
+
+        transcript_fsa_with_self_loops = k2.arc_sort(transcript_fsa_with_self_loops)
+
+        num = k2.compose(
+            self.ctc_topo_P,
+            transcript_fsa_with_self_loops,
+            treat_epsilons_specially=False,
+        )
+
+        # CAUTION: Due to the presence of P,
+        # the resulting `num` may not be connected
+        num = k2.connect(num)
+
+        num = k2.arc_sort(num)
+
+        ctc_topo_P_vec = k2.create_fsa_vec([self.ctc_topo_P])
+        if replicate_den:
+            indexes = torch.zeros(len(texts), dtype=torch.int32, device=self.device)
+            den = k2.index_fsa(ctc_topo_P_vec, indexes)
+        else:
+            den = ctc_topo_P_vec
+
+        return num, den
+
+    def build_transcript_fsa(self, texts: List[str]) -> k2.Fsa:
+        """Convert transcripts to an FsaVec with the help of a lexicon
+        and word symbol table.
+
+        Args:
+          texts:
+            Each element is a transcript containing words separated by space(s).
+            For instance, it may be 'HELLO icefall', which contains
+            two words.
+
+        Returns:
+          Return an FST (FsaVec) corresponding to the transcript.
+          Its `labels` is token IDs and `aux_labels` is word IDs.
+        """
+        word_ids_list = []
+        for text in texts:
+            word_ids = []
+            for word in text.split():
+                if word in self.lexicon.word_table:
+                    word_ids.append(self.lexicon.word_table[word])
+                else:
+                    word_ids.append(self.oov_id)
+            word_ids_list.append(word_ids)
+
+        fsa = k2.linear_fsa(word_ids_list, self.device)
+        fsa = k2.add_epsilon_self_loops(fsa)
+
+        # The reason to use `invert_()` at the end is as follows:
+        #
+        # (1) The `labels` of L_inv is word IDs and `aux_labels` is token IDs
+        # (2) `fsa.labels` is word IDs
+        # (3) after intersection, the `labels` is still word IDs
+        # (4) after `invert_()`, the `labels` is token IDs
+        #     and `aux_labels` is word IDs
+        transcript_fsa = k2.intersect(
+            self.L_inv, fsa, treat_epsilons_specially=False
+        ).invert_()
+        transcript_fsa = k2.arc_sort(transcript_fsa)
+        return transcript_fsa
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of piece IDs.
+
+        Args:
+          texts:
+            It is a list of strings. Each string consists of space(s)
+            separated words. An example containing two strings is given below:
+
+                ['HELLO ICEFALL', 'HELLO k2']
+            We assume it contains no OOVs. Otherwise, it will raise an
+            exception.
+        Returns:
+          Return a list-of-list of token IDs.
+        """
+        return self.lexicon.texts_to_token_ids(texts).tolist()
diff --git a/icefall/ngram_lm.py b/icefall/ngram_lm.py
new file mode 100644
index 000000000..63885a9d0
--- /dev/null
+++ b/icefall/ngram_lm.py
@@ -0,0 +1,171 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections import defaultdict
+from typing import List, Optional, Tuple
+
+from icefall.utils import is_module_available
+
+
+class NgramLm:
+    def __init__(
+        self,
+        fst_filename: str,
+        backoff_id: int,
+        is_binary: bool = False,
+    ):
+        """
+        Args:
+          fst_filename:
+            Path to the FST.
+          backoff_id:
+            ID of the backoff symbol.
+          is_binary:
+            True if the given file is a binary FST.
+        """
+        if not is_module_available("kaldifst"):
+            raise ValueError("Please 'pip install kaldifst' first.")
+
+        import kaldifst
+
+        if is_binary:
+            lm = kaldifst.StdVectorFst.read(fst_filename)
+        else:
+            with open(fst_filename, "r") as f:
+                lm = kaldifst.compile(f.read(), acceptor=False)
+
+        if not lm.is_ilabel_sorted:
+            kaldifst.arcsort(lm, sort_type="ilabel")
+
+        self.lm = lm
+        self.backoff_id = backoff_id
+
+    def _process_backoff_arcs(
+        self,
+        state: int,
+        cost: float,
+    ) -> List[Tuple[int, float]]:
+        """Similar to ProcessNonemitting() from Kaldi, this function
+        returns the list of states reachable from the given state via
+        backoff arcs.
+
+        Args:
+          state:
+            The input state.
+          cost:
+            The cost of reaching the given state from the start state.
+        Returns:
+          Return a list, where each element contains a tuple with two entries:
+            - next_state
+            - cost of next_state
+          If there is no backoff arc leaving the input state, then return
+          an empty list.
+        """
+        ans = []
+
+        next_state, next_cost = self._get_next_state_and_cost_without_backoff(
+            state=state,
+            label=self.backoff_id,
+        )
+        if next_state is None:
+            return ans
+        ans.append((next_state, next_cost + cost))
+        ans += self._process_backoff_arcs(next_state, next_cost + cost)
+        return ans
+
+    def _get_next_state_and_cost_without_backoff(
+        self, state: int, label: int
+    ) -> Tuple[int, float]:
+        """TODO: Add doc."""
+        import kaldifst
+
+        arc_iter = kaldifst.ArcIterator(self.lm, state)
+        num_arcs = self.lm.num_arcs(state)
+
+        # The LM is arc sorted by ilabel, so we use binary search below.
+        left = 0
+        right = num_arcs - 1
+        while left <= right:
+            mid = (left + right) // 2
+            arc_iter.seek(mid)
+            arc = arc_iter.value
+            if arc.ilabel < label:
+                left = mid + 1
+            elif arc.ilabel > label:
+                right = mid - 1
+            else:
+                return arc.nextstate, arc.weight.value
+
+        return None, None
+
+    def get_next_state_and_cost(
+        self,
+        state: int,
+        label: int,
+    ) -> Tuple[List[int], List[float]]:
+        states = [state]
+        costs = [0]
+
+        extra_states_costs = self._process_backoff_arcs(
+            state=state,
+            cost=0,
+        )
+
+        for s, c in extra_states_costs:
+            states.append(s)
+            costs.append(c)
+
+        next_states = []
+        next_costs = []
+        for s, c in zip(states, costs):
+            ns, nc = self._get_next_state_and_cost_without_backoff(s, label)
+            if ns:
+                next_states.append(ns)
+                next_costs.append(c + nc)
+
+        return next_states, next_costs
+
+
+class NgramLmStateCost:
+    def __init__(self, ngram_lm: NgramLm, state_cost: Optional[dict] = None):
+        assert ngram_lm.lm.start == 0, ngram_lm.lm.start
+        self.ngram_lm = ngram_lm
+        if state_cost is not None:
+            self.state_cost = state_cost
+        else:
+            self.state_cost = defaultdict(lambda: float("inf"))
+
+            # At the very beginning, we are at the start state with cost 0
+            self.state_cost[0] = 0.0
+
+    def forward_one_step(self, label: int) -> "NgramLmStateCost":
+        state_cost = defaultdict(lambda: float("inf"))
+        for s, c in self.state_cost.items():
+            next_states, next_costs = self.ngram_lm.get_next_state_and_cost(
+                s,
+                label,
+            )
+            for ns, nc in zip(next_states, next_costs):
+                state_cost[ns] = min(state_cost[ns], c + nc)
+
+        return NgramLmStateCost(ngram_lm=self.ngram_lm, state_cost=state_cost)
+
+    @property
+    def lm_score(self) -> float:
+        if len(self.state_cost) == 0:
+            return float("-inf")
+
+        return -1 * min(self.state_cost.values())
diff --git a/icefall/otc_graph_compiler.py b/icefall/otc_graph_compiler.py
new file mode 100644
index 000000000..bfd679452
--- /dev/null
+++ b/icefall/otc_graph_compiler.py
@@ -0,0 +1,246 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#                2023  Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from pathlib import Path
+from typing import List, Union
+
+import k2
+import sentencepiece as spm
+import torch
+
+from icefall.utils import str2bool
+
+
+class OtcTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lang_dir: Path,
+        otc_token: str,
+        device: Union[str, torch.device] = "cpu",
+        sos_token: str = "<sos/eos>",
+        eos_token: str = "<sos/eos>",
+        initial_bypass_weight: float = 0.0,
+        initial_self_loop_weight: float = 0.0,
+        bypass_weight_decay: float = 0.0,
+        self_loop_weight_decay: float = 0.0,
+    ) -> None:
+        """
+        Args:
+          lang_dir:
+            This directory is expected to contain the following files:
+
+                - bpe.model
+                - words.txt
+          otc_token:
+            The special token in OTC that represent all non-blank tokens
+          device:
+            It indicates CPU or CUDA.
+          sos_token:
+            The word piece that represents sos.
+          eos_token:
+            The word piece that represents eos.
+        """
+        lang_dir = Path(lang_dir)
+        bpe_model_file = lang_dir / "bpe.model"
+        sp = spm.SentencePieceProcessor()
+        sp.load(str(bpe_model_file))
+        self.sp = sp
+        self.token_table = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+
+        self.otc_token = otc_token
+        assert self.otc_token in self.token_table
+
+        self.device = device
+
+        self.sos_id = self.sp.piece_to_id(sos_token)
+        self.eos_id = self.sp.piece_to_id(eos_token)
+
+        assert self.sos_id != self.sp.unk_id()
+        assert self.eos_id != self.sp.unk_id()
+
+        max_token_id = self.get_max_token_id()
+        ctc_topo = k2.ctc_topo(max_token_id, modified=False)
+        self.ctc_topo = ctc_topo.to(self.device)
+
+        self.initial_bypass_weight = initial_bypass_weight
+        self.initial_self_loop_weight = initial_self_loop_weight
+        self.bypass_weight_decay = bypass_weight_decay
+        self.self_loop_weight_decay = self_loop_weight_decay
+
+    def get_max_token_id(self):
+        max_token_id = 0
+        for symbol in self.token_table.symbols:
+            if not symbol.startswith("#"):
+                max_token_id = max(self.token_table[symbol], max_token_id)
+        assert max_token_id > 0
+
+        return max_token_id
+
+    def make_arc(
+        self,
+        from_state: int,
+        to_state: int,
+        symbol: Union[str, int],
+        weight: float,
+    ):
+        return f"{from_state} {to_state} {symbol} {weight}"
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of piece IDs.
+
+        Args:
+          texts:
+            It is a list of strings. Each string consists of space(s)
+            separated words. An example containing two strings is given below:
+
+                ['HELLO ICEFALL', 'HELLO k2']
+        Returns:
+          Return a list-of-list of piece IDs.
+        """
+        return self.sp.encode(texts, out_type=int)
+
+    def compile(
+        self,
+        texts: List[str],
+        allow_bypass_arc: str2bool = True,
+        allow_self_loop_arc: str2bool = True,
+        bypass_weight: float = 0.0,
+        self_loop_weight: float = 0.0,
+    ) -> k2.Fsa:
+        """Build a OTC graph from a texts (list of words).
+
+        Args:
+          texts:
+            A list of strings. Each string contains a sentence for an utterance.
+            A sentence consists of spaces separated words. An example `texts`
+            looks like:
+              ['hello icefall', 'CTC training with k2']
+          allow_bypass_arc:
+            Whether to add bypass arc to training graph for substitution
+            and insertion errors (wrong or extra words in the transcript).
+          allow_self_loop_arc:
+            Whether to add self-loop arc to training graph for deletion
+            errors (missing words in the transcript).
+          bypass_weight:
+            Weight associated with bypass arc.
+          self_loop_weight:
+            Weight associated with self-loop arc.
+
+        Return:
+          Return an FsaVec, which is the result of composing a
+          CTC topology with OTC FSAs constructed from the given texts.
+        """
+
+        transcript_fsa = self.convert_transcript_to_fsa(
+            texts,
+            self.otc_token,
+            allow_bypass_arc,
+            allow_self_loop_arc,
+            bypass_weight,
+            self_loop_weight,
+        )
+        transcript_fsa = transcript_fsa.to(self.device)
+        fsa_with_self_loop = k2.remove_epsilon_and_add_self_loops(transcript_fsa)
+        fsa_with_self_loop = k2.arc_sort(fsa_with_self_loop)
+
+        graph = k2.compose(
+            self.ctc_topo,
+            fsa_with_self_loop,
+            treat_epsilons_specially=False,
+        )
+        assert graph.requires_grad is False
+
+        return graph
+
+    def convert_transcript_to_fsa(
+        self,
+        texts: List[str],
+        otc_token: str,
+        allow_bypass_arc: str2bool = True,
+        allow_self_loop_arc: str2bool = True,
+        bypass_weight: float = 0.0,
+        self_loop_weight: float = 0.0,
+    ):
+        otc_token_id = self.token_table[otc_token]
+
+        transcript_fsa_list = []
+        for text in texts:
+            text_piece_ids = []
+
+            for word in text.split():
+                piece_ids = self.sp.encode(word, out_type=int)
+                text_piece_ids.append(piece_ids)
+
+            arcs = []
+            start_state = 0
+            cur_state = start_state
+            next_state = 1
+
+            for piece_ids in text_piece_ids:
+                bypass_cur_state = cur_state
+
+                if allow_self_loop_arc:
+                    self_loop_arc = self.make_arc(
+                        cur_state,
+                        cur_state,
+                        otc_token_id,
+                        self_loop_weight,
+                    )
+                    arcs.append(self_loop_arc)
+
+                for piece_id in piece_ids:
+                    arc = self.make_arc(cur_state, next_state, piece_id, 0.0)
+                    arcs.append(arc)
+
+                    cur_state = next_state
+                    next_state += 1
+
+                bypass_next_state = cur_state
+                if allow_bypass_arc:
+                    bypass_arc = self.make_arc(
+                        bypass_cur_state,
+                        bypass_next_state,
+                        otc_token_id,
+                        bypass_weight,
+                    )
+                    arcs.append(bypass_arc)
+                bypass_cur_state = cur_state
+
+            if allow_self_loop_arc:
+                self_loop_arc = self.make_arc(
+                    cur_state,
+                    cur_state,
+                    otc_token_id,
+                    self_loop_weight,
+                )
+                arcs.append(self_loop_arc)
+
+            # Deal with final state
+            final_state = next_state
+            final_arc = self.make_arc(cur_state, final_state, -1, 0.0)
+            arcs.append(final_arc)
+            arcs.append(f"{final_state}")
+            sorted_arcs = sorted(arcs, key=lambda a: int(a.split()[0]))
+
+            transcript_fsa = k2.Fsa.from_str("\n".join(sorted_arcs))
+            transcript_fsa = k2.arc_sort(transcript_fsa)
+            transcript_fsa_list.append(transcript_fsa)
+
+        transcript_fsa_vec = k2.create_fsa_vec(transcript_fsa_list)
+
+        return transcript_fsa_vec
diff --git a/icefall/profiler.py b/icefall/profiler.py
new file mode 100644
index 000000000..49e138579
--- /dev/null
+++ b/icefall/profiler.py
@@ -0,0 +1,910 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This is modified from https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/profiling/flops_profiler/profiler.py
+
+import k2
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+from typing import List, Optional
+from collections import OrderedDict
+import numpy as np
+
+Tensor = torch.Tensor
+
+module_flop_count = []
+old_functions = {}
+
+
+class FlopsProfiler(object):
+    """Measures the latency, number of estimated floating-point operations and parameters of each module in a PyTorch model.
+
+    The flops-profiler profiles the forward pass of a PyTorch model and prints the model graph with the measured profile attached to each module. It shows how latency, flops and parameters are spent in the model and which modules or layers could be the bottleneck. It also outputs the names of the top k modules in terms of aggregated latency, flops, and parameters at depth l with k and l specified by the user. The output profile is computed for each batch of input.
+
+    To profile a trained model in inference, use the `get_model_profile` API.
+
+    Args:
+        object (torch.nn.Module): The PyTorch model to profile.
+    """
+
+    def __init__(self, model, module_hoop_mapping=None):
+        self.model = model
+        self.started = False
+        self.func_patched = False
+        self.module_hoop_mapping = (
+            module_hoop_mapping
+            if module_hoop_mapping is not None
+            else MODULE_HOOK_MAPPING
+        )
+
+    def start_profile(self, ignore_list=None):
+        """Starts profiling.
+
+        Extra attributes are added recursively to all the modules and the profiled torch.nn.functionals are monkey patched.
+
+        Args:
+            ignore_list (list, optional): the list of modules to ignore while profiling. Defaults to None.
+        """
+        self.reset_profile()
+        _patch_functionals()
+        _patch_tensor_methods()
+
+        def register_module_hooks(module, ignore_list):
+            if ignore_list and type(module) in ignore_list:
+                return
+
+            # if computing the flops of a module directly
+            if type(module) in self.module_hoop_mapping:
+                if not hasattr(module, "__flops_handle__"):
+                    module.__flops_handle__ = module.register_forward_hook(
+                        self.module_hoop_mapping[type(module)]
+                    )
+                return
+
+            # if computing the flops of the functionals in a module
+            def pre_hook(module, input):
+                module_flop_count.append([])
+
+            if not hasattr(module, "__pre_hook_handle__"):
+                module.__pre_hook_handle__ = module.register_forward_pre_hook(pre_hook)
+
+            def post_hook(module, input, output):
+                if module_flop_count:
+                    module.__flops__ += sum([elem[1] for elem in module_flop_count[-1]])
+                    module_flop_count.pop()
+
+            if not hasattr(module, "__post_hook_handle__"):
+                module.__post_hook_handle__ = module.register_forward_hook(post_hook)
+
+        self.model.apply(partial(register_module_hooks, ignore_list=ignore_list))
+        self.started = True
+        self.func_patched = True
+
+    def stop_profile(self):
+        """Stop profiling.
+
+        All torch.nn.functionals are restored to their originals.
+        """
+        if self.started and self.func_patched:
+            _reload_functionals()
+            _reload_tensor_methods()
+            self.func_patched = False
+
+        def remove_profile_attrs(module):
+            if hasattr(module, "__pre_hook_handle__"):
+                module.__pre_hook_handle__.remove()
+                del module.__pre_hook_handle__
+            if hasattr(module, "__post_hook_handle__"):
+                module.__post_hook_handle__.remove()
+                del module.__post_hook_handle__
+            if hasattr(module, "__flops_handle__"):
+                module.__flops_handle__.remove()
+                del module.__flops_handle__
+
+        self.model.apply(remove_profile_attrs)
+
+    def reset_profile(self):
+        """Resets the profiling.
+
+        Adds or resets the extra attributes.
+        """
+
+        def add_or_reset_attrs(module):
+            module.__flops__ = 0
+            module.__params__ = sum(p.numel() for p in module.parameters())
+
+        self.model.apply(add_or_reset_attrs)
+
+    def end_profile(self):
+        """Ends profiling.
+
+        The added attributes and handles are removed recursively on all the modules.
+        """
+        if not self.started:
+            return
+        self.stop_profile()
+        self.started = False
+
+        def remove_profile_attrs(module):
+            if hasattr(module, "__flops__"):
+                del module.__flops__
+            if hasattr(module, "__params__"):
+                del module.__params__
+
+        self.model.apply(remove_profile_attrs)
+
+    def get_total_flops(self, as_string=False):
+        """Returns the total flops of the model.
+
+        Args:
+            as_string (bool, optional): whether to output the flops as string. Defaults to False.
+
+        Returns:
+            The number of multiply-accumulate operations of the model forward pass.
+        """
+        total_flops = get_module_flops(self.model)
+        return num_to_string(total_flops) if as_string else total_flops
+
+    def get_total_params(self, as_string=False):
+        """Returns the total parameters of the model.
+
+        Args:
+            as_string (bool, optional): whether to output the parameters as string. Defaults to False.
+
+        Returns:
+            The number of parameters in the model.
+        """
+        return (
+            params_to_string(self.model.__params__)
+            if as_string
+            else self.model.__params__
+        )
+
+
+def _prod(dims):
+    p = 1
+    for v in dims:
+        p *= v
+    return p
+
+
+def _linear_flops_compute(input, weight, bias=None):
+    out_features = weight.shape[0]
+    macs = input.numel() * out_features
+    return 2 * macs
+
+
+def _relu_flops_compute(input, inplace=False):
+    return input.numel()
+
+
+def _prelu_flops_compute(input: Tensor, weight: Tensor):
+    return input.numel()
+
+
+def _elu_flops_compute(input: Tensor, alpha: float = 1.0, inplace: bool = False):
+    return input.numel()
+
+
+def _leaky_relu_flops_compute(
+    input: Tensor, negative_slope: float = 0.01, inplace: bool = False
+):
+    return input.numel()
+
+
+def _relu6_flops_compute(input: Tensor, inplace: bool = False):
+    return input.numel()
+
+
+def _silu_flops_compute(input: Tensor, inplace: bool = False):
+    return input.numel()
+
+
+def _gelu_flops_compute(input, **kwargs):
+    return input.numel()
+
+
+def _pool_flops_compute(
+    input,
+    kernel_size,
+    stride=None,
+    padding=0,
+    dilation=None,
+    ceil_mode=False,
+    count_include_pad=True,
+    divisor_override=None,
+    return_indices=None,
+):
+    return input.numel()
+
+
+def _conv_flops_compute(
+    input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1
+):
+    assert weight.shape[1] * groups == input.shape[1]
+
+    batch_size = input.shape[0]
+    in_channels = input.shape[1]
+    out_channels = weight.shape[0]
+    kernel_dims = list(weight.shape[2:])
+    input_dims = list(input.shape[2:])
+
+    length = len(input_dims)
+
+    paddings = padding if type(padding) is tuple else (padding,) * length
+    strides = stride if type(stride) is tuple else (stride,) * length
+    dilations = dilation if type(dilation) is tuple else (dilation,) * length
+
+    output_dims = []
+    for idx, input_dim in enumerate(input_dims):
+        output_dim = (
+            input_dim
+            + 2 * paddings[idx]
+            - (dilations[idx] * (kernel_dims[idx] - 1) + 1)
+        ) // strides[idx] + 1
+        output_dims.append(output_dim)
+
+    filters_per_channel = out_channels // groups
+    conv_per_position_macs = int(_prod(kernel_dims)) * in_channels * filters_per_channel
+    active_elements_count = batch_size * int(_prod(output_dims))
+    overall_conv_macs = conv_per_position_macs * active_elements_count
+    overall_conv_flops = 2 * overall_conv_macs
+
+    bias_flops = 0
+    if bias is not None:
+        bias_flops = out_channels * active_elements_count
+
+    return int(overall_conv_flops + bias_flops)
+
+
+def _conv_trans_flops_compute(
+    input,
+    weight,
+    bias=None,
+    stride=1,
+    padding=0,
+    output_padding=0,
+    groups=1,
+    dilation=1,
+):
+    batch_size = input.shape[0]
+    in_channels = input.shape[1]
+    out_channels = weight.shape[0]
+    kernel_dims = list(weight.shape[2:])
+    input_dims = list(input.shape[2:])
+
+    length = len(input_dims)
+
+    paddings = padding if type(padding) is tuple else (padding,) * length
+    strides = stride if type(stride) is tuple else (stride,) * length
+    dilations = dilation if type(dilation) is tuple else (dilation,) * length
+
+    output_dims = []
+    for idx, input_dim in enumerate(input_dims):
+        output_dim = (
+            input_dim
+            + 2 * paddings[idx]
+            - (dilations[idx] * (kernel_dims[idx] - 1) + 1)
+        ) // strides[idx] + 1
+        output_dims.append(output_dim)
+
+    paddings = padding if type(padding) is tuple else (padding, padding)
+    strides = stride if type(stride) is tuple else (stride, stride)
+    dilations = dilation if type(dilation) is tuple else (dilation, dilation)
+
+    filters_per_channel = out_channels // groups
+    conv_per_position_macs = int(_prod(kernel_dims)) * in_channels * filters_per_channel
+    active_elements_count = batch_size * int(_prod(input_dims))
+    overall_conv_macs = conv_per_position_macs * active_elements_count
+    overall_conv_flops = 2 * overall_conv_macs
+
+    bias_flops = 0
+    if bias is not None:
+        bias_flops = out_channels * batch_size * int(_prod(output_dims))
+
+    return int(overall_conv_flops + bias_flops)
+
+
+def _batch_norm_flops_compute(
+    input,
+    running_mean,
+    running_var,
+    weight=None,
+    bias=None,
+    training=False,
+    momentum=0.1,
+    eps=1e-05,
+):
+    has_affine = weight is not None
+    if training:
+        # estimation
+        return input.numel() * (5 if has_affine else 4), 0
+    flops = input.numel() * (2 if has_affine else 1)
+    return flops
+
+
+def _layer_norm_flops_compute(
+    input: Tensor,
+    normalized_shape: List[int],
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    eps: float = 1e-5,
+):
+    has_affine = weight is not None
+    # estimation
+    return input.numel() * (5 if has_affine else 4)
+
+
+def _group_norm_flops_compute(
+    input: Tensor,
+    num_groups: int,
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    eps: float = 1e-5,
+):
+    has_affine = weight is not None
+    # estimation
+    return input.numel() * (5 if has_affine else 4)
+
+
+def _instance_norm_flops_compute(
+    input: Tensor,
+    running_mean: Optional[Tensor] = None,
+    running_var: Optional[Tensor] = None,
+    weight: Optional[Tensor] = None,
+    bias: Optional[Tensor] = None,
+    use_input_stats: bool = True,
+    momentum: float = 0.1,
+    eps: float = 1e-5,
+):
+    has_affine = weight is not None
+    # estimation
+    return input.numel() * (5 if has_affine else 4)
+
+
+def _upsample_flops_compute(input, **kwargs):
+    size = kwargs.get("size", None)
+    if size is not None:
+        if isinstance(size, tuple) or isinstance(size, list):
+            return int(_prod(size)), 0
+        else:
+            return int(size), 0
+    scale_factor = kwargs.get("scale_factor", None)
+    assert scale_factor is not None, "either size or scale_factor should be defined"
+    flops = input.numel()
+    if isinstance(scale_factor, tuple) and len(scale_factor) == len(input):
+        flops * int(_prod(scale_factor))
+    else:
+        flops * scale_factor ** len(input)
+    return flops
+
+
+def _softmax_flops_compute(input, dim=None, _stacklevel=3, dtype=None):
+    return input.numel()
+
+
+def _sigmoid_flops_compute(input):
+    return input.numel()
+
+
+def _embedding_flops_compute(
+    input,
+    weight,
+    padding_idx=None,
+    max_norm=None,
+    norm_type=2.0,
+    scale_grad_by_freq=False,
+    sparse=False,
+):
+    return 0
+
+
+def _dropout_flops_compute(input, p=0.5, training=True, inplace=False):
+    return 0
+
+
+def _matmul_flops_compute(input, other, *, out=None):
+    """
+    Count flops for the matmul operation.
+    """
+    macs = _prod(input.shape) * other.shape[-1]
+    return 2 * macs
+
+
+def _addmm_flops_compute(input, mat1, mat2, *, beta=1, alpha=1, out=None):
+    """
+    Count flops for the addmm operation.
+    """
+    macs = _prod(mat1.shape) * mat2.shape[-1]
+    return 2 * macs + _prod(input.shape)
+
+
+def _einsum_flops_compute(equation, *operands):
+    """
+    Count flops for the einsum operation.
+    """
+    equation = equation.replace(" ", "")
+    input_shapes = [o.shape for o in operands]
+
+    # Re-map equation so that same equation with different alphabet
+    # representations will look the same.
+    letter_order = OrderedDict((k, 0) for k in equation if k.isalpha()).keys()
+    mapping = {ord(x): 97 + i for i, x in enumerate(letter_order)}
+    equation = equation.translate(mapping)
+
+    np_arrs = [np.zeros(s) for s in input_shapes]
+    optim = np.einsum_path(equation, *np_arrs, optimize="optimal")[1]
+    for line in optim.split("\n"):
+        if "optimized flop" in line.lower():
+            flop = int(float(line.split(":")[-1]))
+            return flop
+    raise NotImplementedError("Unsupported einsum operation.")
+
+
+def _tensor_addmm_flops_compute(self, mat1, mat2, *, beta=1, alpha=1, out=None):
+    """
+    Count flops for the tensor addmm operation.
+    """
+    macs = _prod(mat1.shape) * mat2.shape[-1]
+    return 2 * macs + _prod(self.shape)
+
+
+def _mul_flops_compute(input, other, *, out=None):
+    print("mul")
+    return _elementwise_flops_compute(input, other)
+
+
+def _add_flops_compute(input, other, *, alpha=1, out=None):
+    print("add")
+    return _elementwise_flops_compute(input, other)
+
+
+def _sum_flops_compute(input, dim, keepdim=False):
+    return input.numel()
+
+
+def _elementwise_flops_compute(input, other):
+    if not torch.is_tensor(input):
+        if torch.is_tensor(other):
+            return _prod(other.shape)
+        else:
+            return 1
+    elif not torch.is_tensor(other):
+        return _prod(input.shape)
+    else:
+        dim_input = len(input.shape)
+        dim_other = len(other.shape)
+        max_dim = max(dim_input, dim_other)
+
+        final_shape = []
+        for i in range(max_dim):
+            in_i = input.shape[i] if i < dim_input else 1
+            ot_i = other.shape[i] if i < dim_other else 1
+            if in_i > ot_i:
+                final_shape.append(in_i)
+            else:
+                final_shape.append(ot_i)
+        flops = _prod(final_shape)
+        return flops
+
+
+def _tanh_flops_compute(input):
+    return input.numel()
+
+
+def _k2_swoosh_flops_compute(input):
+    # For SwooshLForward and SwooshRForward
+    # estimate as swish/silu
+    return input.numel()
+
+
+def wrapFunc(func, funcFlopCompute):
+    oldFunc = func
+    name = func.__str__
+    old_functions[name] = oldFunc
+
+    def newFunc(*args, **kwds):
+        flops = funcFlopCompute(*args, **kwds)
+        if module_flop_count:
+            module_flop_count[-1].append((name, flops))
+        return oldFunc(*args, **kwds)
+
+    newFunc.__str__ = func.__str__
+
+    return newFunc
+
+
+def _patch_functionals():
+    # FC
+    F.linear = wrapFunc(F.linear, _linear_flops_compute)
+
+    # convolutions
+    F.conv1d = wrapFunc(F.conv1d, _conv_flops_compute)
+    F.conv2d = wrapFunc(F.conv2d, _conv_flops_compute)
+    F.conv3d = wrapFunc(F.conv3d, _conv_flops_compute)
+
+    # conv transposed
+    F.conv_transpose1d = wrapFunc(F.conv_transpose1d, _conv_trans_flops_compute)
+    F.conv_transpose2d = wrapFunc(F.conv_transpose2d, _conv_trans_flops_compute)
+    F.conv_transpose3d = wrapFunc(F.conv_transpose3d, _conv_trans_flops_compute)
+
+    # activations
+    F.relu = wrapFunc(F.relu, _relu_flops_compute)
+    F.prelu = wrapFunc(F.prelu, _prelu_flops_compute)
+    F.elu = wrapFunc(F.elu, _elu_flops_compute)
+    F.leaky_relu = wrapFunc(F.leaky_relu, _leaky_relu_flops_compute)
+    F.relu6 = wrapFunc(F.relu6, _relu6_flops_compute)
+    if hasattr(F, "silu"):
+        F.silu = wrapFunc(F.silu, _silu_flops_compute)
+    F.gelu = wrapFunc(F.gelu, _gelu_flops_compute)
+
+    # Normalizations
+    F.batch_norm = wrapFunc(F.batch_norm, _batch_norm_flops_compute)
+    F.layer_norm = wrapFunc(F.layer_norm, _layer_norm_flops_compute)
+    F.instance_norm = wrapFunc(F.instance_norm, _instance_norm_flops_compute)
+    F.group_norm = wrapFunc(F.group_norm, _group_norm_flops_compute)
+
+    # poolings
+    F.avg_pool1d = wrapFunc(F.avg_pool1d, _pool_flops_compute)
+    F.avg_pool2d = wrapFunc(F.avg_pool2d, _pool_flops_compute)
+    F.avg_pool3d = wrapFunc(F.avg_pool3d, _pool_flops_compute)
+    F.max_pool1d = wrapFunc(F.max_pool1d, _pool_flops_compute)
+    F.max_pool2d = wrapFunc(F.max_pool2d, _pool_flops_compute)
+    F.max_pool3d = wrapFunc(F.max_pool3d, _pool_flops_compute)
+    F.adaptive_avg_pool1d = wrapFunc(F.adaptive_avg_pool1d, _pool_flops_compute)
+    F.adaptive_avg_pool2d = wrapFunc(F.adaptive_avg_pool2d, _pool_flops_compute)
+    F.adaptive_avg_pool3d = wrapFunc(F.adaptive_avg_pool3d, _pool_flops_compute)
+    F.adaptive_max_pool1d = wrapFunc(F.adaptive_max_pool1d, _pool_flops_compute)
+    F.adaptive_max_pool2d = wrapFunc(F.adaptive_max_pool2d, _pool_flops_compute)
+    F.adaptive_max_pool3d = wrapFunc(F.adaptive_max_pool3d, _pool_flops_compute)
+
+    # upsample
+    F.upsample = wrapFunc(F.upsample, _upsample_flops_compute)
+    F.interpolate = wrapFunc(F.interpolate, _upsample_flops_compute)
+
+    # softmax
+    F.softmax = wrapFunc(F.softmax, _softmax_flops_compute)
+
+    # sigmoid
+    F.sigmoid = wrapFunc(F.sigmoid, _sigmoid_flops_compute)
+
+    # embedding
+    F.embedding = wrapFunc(F.embedding, _embedding_flops_compute)
+
+    # swoosh functions in k2
+    k2.swoosh_l_forward = wrapFunc(k2.swoosh_l_forward, _k2_swoosh_flops_compute)
+    k2.swoosh_r_forward = wrapFunc(k2.swoosh_r_forward, _k2_swoosh_flops_compute)
+    k2.swoosh_l = wrapFunc(k2.swoosh_l, _k2_swoosh_flops_compute)
+    k2.swoosh_r = wrapFunc(k2.swoosh_r, _k2_swoosh_flops_compute)
+
+
+def _patch_tensor_methods():
+    torch.matmul = wrapFunc(torch.matmul, _matmul_flops_compute)
+    torch.Tensor.matmul = wrapFunc(torch.Tensor.matmul, _matmul_flops_compute)
+    torch.mm = wrapFunc(torch.mm, _matmul_flops_compute)
+    torch.Tensor.mm = wrapFunc(torch.Tensor.mm, _matmul_flops_compute)
+    torch.bmm = wrapFunc(torch.bmm, _matmul_flops_compute)
+    torch.Tensor.bmm = wrapFunc(torch.Tensor.bmm, _matmul_flops_compute)
+
+    torch.addmm = wrapFunc(torch.addmm, _addmm_flops_compute)
+    torch.Tensor.addmm = wrapFunc(torch.Tensor.addmm, _tensor_addmm_flops_compute)
+
+    torch.mul = wrapFunc(torch.mul, _mul_flops_compute)
+    torch.Tensor.mul = wrapFunc(torch.Tensor.mul, _mul_flops_compute)
+
+    torch.add = wrapFunc(torch.add, _add_flops_compute)
+    torch.Tensor.add = wrapFunc(torch.Tensor.add, _add_flops_compute)
+
+    torch.sum = wrapFunc(torch.sum, _sum_flops_compute)
+    torch.Tensor.sum = wrapFunc(torch.Tensor.sum, _sum_flops_compute)
+
+    torch.einsum = wrapFunc(torch.einsum, _einsum_flops_compute)
+
+    torch.baddbmm = wrapFunc(torch.baddbmm, _tensor_addmm_flops_compute)
+
+    torch.tanh = wrapFunc(torch.tanh, _tanh_flops_compute)
+
+    torch.Tensor.softmax = wrapFunc(torch.Tensor.softmax, _softmax_flops_compute)
+
+    torch.sigmoid = wrapFunc(torch.sigmoid, _sigmoid_flops_compute)
+    torch.Tensor.sigmoid = wrapFunc(torch.Tensor.sigmoid, _sigmoid_flops_compute)
+
+
+def _reload_functionals():
+    # torch.nn.functional does not support importlib.reload()
+    F.linear = old_functions[F.linear.__str__]
+    F.conv1d = old_functions[F.conv1d.__str__]
+    F.conv2d = old_functions[F.conv2d.__str__]
+    F.conv3d = old_functions[F.conv3d.__str__]
+    F.conv_transpose1d = old_functions[F.conv_transpose1d.__str__]
+    F.conv_transpose2d = old_functions[F.conv_transpose2d.__str__]
+    F.conv_transpose3d = old_functions[F.conv_transpose3d.__str__]
+    F.relu = old_functions[F.relu.__str__]
+    F.prelu = old_functions[F.prelu.__str__]
+    F.elu = old_functions[F.elu.__str__]
+    F.leaky_relu = old_functions[F.leaky_relu.__str__]
+    F.relu6 = old_functions[F.relu6.__str__]
+    if hasattr(F, "silu"):
+        F.silu = old_functions[F.silu.__str__]
+    F.gelu = old_functions[F.gelu.__str__]
+    F.batch_norm = old_functions[F.batch_norm.__str__]
+    F.layer_norm = old_functions[F.layer_norm.__str__]
+    F.instance_norm = old_functions[F.instance_norm.__str__]
+    F.group_norm = old_functions[F.group_norm.__str__]
+    F.avg_pool1d = old_functions[F.avg_pool1d.__str__]
+    F.avg_pool2d = old_functions[F.avg_pool2d.__str__]
+    F.avg_pool3d = old_functions[F.avg_pool3d.__str__]
+    F.max_pool1d = old_functions[F.max_pool1d.__str__]
+    F.max_pool2d = old_functions[F.max_pool2d.__str__]
+    F.max_pool3d = old_functions[F.max_pool3d.__str__]
+    F.adaptive_avg_pool1d = old_functions[F.adaptive_avg_pool1d.__str__]
+    F.adaptive_avg_pool2d = old_functions[F.adaptive_avg_pool2d.__str__]
+    F.adaptive_avg_pool3d = old_functions[F.adaptive_avg_pool3d.__str__]
+    F.adaptive_max_pool1d = old_functions[F.adaptive_max_pool1d.__str__]
+    F.adaptive_max_pool2d = old_functions[F.adaptive_max_pool2d.__str__]
+    F.adaptive_max_pool3d = old_functions[F.adaptive_max_pool3d.__str__]
+    F.upsample = old_functions[F.upsample.__str__]
+    F.interpolate = old_functions[F.interpolate.__str__]
+    F.softmax = old_functions[F.softmax.__str__]
+    F.sigmoid = old_functions[F.sigmoid.__str__]
+    F.embedding = old_functions[F.embedding.__str__]
+    # swoosh functions in k2
+    k2.swoosh_l = old_functions[k2.swoosh_l.__str__]
+    k2.swoosh_r = old_functions[k2.swoosh_r.__str__]
+    k2.swoosh_l_forward = old_functions[k2.swoosh_l_forward.__str__]
+    k2.swoosh_r_forward = old_functions[k2.swoosh_r_forward.__str__]
+
+
+def _reload_tensor_methods():
+    torch.matmul = old_functions[torch.matmul.__str__]
+    torch.Tensor.matmul = old_functions[torch.Tensor.matmul.__str__]
+    torch.mm = old_functions[torch.mm.__str__]
+    torch.Tensor.mm = old_functions[torch.Tensor.mm.__str__]
+    torch.bmm = old_functions[torch.matmul.__str__]
+    torch.Tensor.bmm = old_functions[torch.Tensor.bmm.__str__]
+    torch.addmm = old_functions[torch.addmm.__str__]
+    torch.Tensor.addmm = old_functions[torch.Tensor.addmm.__str__]
+    torch.mul = old_functions[torch.mul.__str__]
+    torch.Tensor.mul = old_functions[torch.Tensor.mul.__str__]
+    torch.add = old_functions[torch.add.__str__]
+    torch.Tensor.add = old_functions[torch.Tensor.add.__str__]
+    torch.sum = old_functions[torch.sum.__str__]
+    torch.Tensor.sum = old_functions[torch.Tensor.sum.__str__]
+
+    torch.einsum = old_functions[torch.einsum.__str__]
+
+    torch.baddbmm = old_functions[torch.baddbmm.__str__]
+
+    torch.Tensor.softmax = old_functions[torch.Tensor.softmax.__str__]
+
+    torch.sigmoid = old_functions[torch.sigmoid.__str__]
+    torch.Tensor.sigmoid = old_functions[torch.Tensor.sigmoid.__str__]
+
+
+def _rnn_flops(flops, rnn_module, w_ih, w_hh, input_size):
+    # matrix matrix mult ih state and internal state
+    flops += w_ih.shape[0] * w_ih.shape[1]
+    # matrix matrix mult hh state and internal state
+    flops += w_hh.shape[0] * w_hh.shape[1]
+    if isinstance(rnn_module, (nn.RNN, nn.RNNCell)):
+        # add both operations
+        flops += rnn_module.hidden_size
+    elif isinstance(rnn_module, (nn.GRU, nn.GRUCell)):
+        # hadamard of r
+        flops += rnn_module.hidden_size
+        # adding operations from both states
+        flops += rnn_module.hidden_size * 3
+        # last two hadamard _product and add
+        flops += rnn_module.hidden_size * 3
+    elif isinstance(rnn_module, (nn.LSTM, nn.LSTMCell)):
+        # adding operations from both states
+        flops += rnn_module.hidden_size * 4
+        # two hadamard _product and add for C state
+        flops += (
+            rnn_module.hidden_size + rnn_module.hidden_size + rnn_module.hidden_size
+        )
+        # final hadamard
+        flops += (
+            rnn_module.hidden_size + rnn_module.hidden_size + rnn_module.hidden_size
+        )
+    return flops
+
+
+def _rnn_forward_hook(rnn_module, input, output):
+    flops = 0
+    # input is a tuple containing a sequence to process and (optionally) hidden state
+    inp = input[0]
+    batch_size = inp.shape[0]
+    seq_length = inp.shape[1]
+    num_layers = rnn_module.num_layers
+
+    for i in range(num_layers):
+        w_ih = rnn_module.__getattr__("weight_ih_l" + str(i))
+        w_hh = rnn_module.__getattr__("weight_hh_l" + str(i))
+        if i == 0:
+            input_size = rnn_module.input_size
+        else:
+            input_size = rnn_module.hidden_size
+        flops = _rnn_flops(flops, rnn_module, w_ih, w_hh, input_size)
+        if rnn_module.bias:
+            b_ih = rnn_module.__getattr__("bias_ih_l" + str(i))
+            b_hh = rnn_module.__getattr__("bias_hh_l" + str(i))
+            flops += b_ih.shape[0] + b_hh.shape[0]
+
+    flops *= batch_size
+    flops *= seq_length
+    if rnn_module.bidirectional:
+        flops *= 2
+    rnn_module.__flops__ += int(flops)
+
+
+def _rnn_cell_forward_hook(rnn_cell_module, input, output):
+    flops = 0
+    inp = input[0]
+    batch_size = inp.shape[0]
+    w_ih = rnn_cell_module.__getattr__("weight_ih")
+    w_hh = rnn_cell_module.__getattr__("weight_hh")
+    input_size = inp.shape[1]
+    flops = _rnn_flops(flops, rnn_cell_module, w_ih, w_hh, input_size)
+    if rnn_cell_module.bias:
+        b_ih = rnn_cell_module.__getattr__("bias_ih")
+        b_hh = rnn_cell_module.__getattr__("bias_hh")
+        flops += b_ih.shape[0] + b_hh.shape[0]
+
+    flops *= batch_size
+    rnn_cell_module.__flops__ += int(flops)
+
+
+MODULE_HOOK_MAPPING = {
+    # RNN
+    nn.RNN: _rnn_forward_hook,
+    nn.GRU: _rnn_forward_hook,
+    nn.LSTM: _rnn_forward_hook,
+    nn.RNNCell: _rnn_cell_forward_hook,
+    nn.LSTMCell: _rnn_cell_forward_hook,
+    nn.GRUCell: _rnn_cell_forward_hook,
+}
+
+
+def num_to_string(num, precision=2):
+    if num // 10**9 > 0:
+        return str(round(num / 10.0**9, precision)) + " G"
+    elif num // 10**6 > 0:
+        return str(round(num / 10.0**6, precision)) + " M"
+    elif num // 10**3 > 0:
+        return str(round(num / 10.0**3, precision)) + " K"
+    else:
+        return str(num)
+
+
+def number_to_string(num, units=None, precision=2):
+    if units is None:
+        if num // 10**9 > 0:
+            return str(round(num / 10.0**9, precision)) + " G"
+        elif num // 10**6 > 0:
+            return str(round(num / 10.0**6, precision)) + " M"
+        elif num // 10**3 > 0:
+            return str(round(num / 10.0**3, precision)) + " K"
+        else:
+            return str(num) + " "
+    else:
+        if units == "G":
+            return str(round(num / 10.0**9, precision)) + " " + units
+        elif units == "M":
+            return str(round(num / 10.0**6, precision)) + " " + units
+        elif units == "K":
+            return str(round(num / 10.0**3, precision)) + " " + units
+        else:
+            return str(num) + " "
+
+
+def flops_to_string(flops, units=None, precision=2):
+    if units is None:
+        if flops // 10**12 > 0:
+            return str(round(flops / 10.0**12, precision)) + " TFLOPS"
+        if flops // 10**9 > 0:
+            return str(round(flops / 10.0**9, precision)) + " GFLOPS"
+        elif flops // 10**6 > 0:
+            return str(round(flops / 10.0**6, precision)) + " MFLOPS"
+        elif flops // 10**3 > 0:
+            return str(round(flops / 10.0**3, precision)) + " KFLOPS"
+        else:
+            return str(flops) + " FLOPS"
+    else:
+        if units == "TFLOPS":
+            return str(round(flops / 10.0**12, precision)) + " " + units
+        if units == "GFLOPS":
+            return str(round(flops / 10.0**9, precision)) + " " + units
+        elif units == "MFLOPS":
+            return str(round(flops / 10.0**6, precision)) + " " + units
+        elif units == "KFLOPS":
+            return str(round(flops / 10.0**3, precision)) + " " + units
+        else:
+            return str(flops) + " FLOPS"
+
+
+def params_to_string(params_num, units=None, precision=2):
+    if units is None:
+        if params_num // 10**6 > 0:
+            return str(round(params_num / 10**6, 2)) + " M"
+        elif params_num // 10**3:
+            return str(round(params_num / 10**3, 2)) + " k"
+        else:
+            return str(params_num)
+    else:
+        if units == "M":
+            return str(round(params_num / 10.0**6, precision)) + " " + units
+        elif units == "K":
+            return str(round(params_num / 10.0**3, precision)) + " " + units
+        else:
+            return str(params_num)
+
+
+def get_module_flops(module):
+    sum = module.__flops__
+    # iterate over immediate children modules
+    for child in module.children():
+        sum += get_module_flops(child)
+    return sum
+
+
+def get_module_duration(module):
+    duration = module.__duration__
+    if duration == 0:  # e.g. ModuleList
+        for m in module.children():
+            duration += m.__duration__
+    return duration
+
+
+def get_model_profile(
+    model,
+    args=[],
+    as_string=True,
+    ignore_modules=None,
+    module_hoop_mapping=None,
+):
+    """Returns the total floating-point operations, MACs, and parameters of a model.
+
+    Example:
+
+    .. code-block:: python
+
+        model = torchvision.models.alexnet()
+        batch_size = 256
+        flops, params = get_model_profile(model=model, args=(feature, feature_lens))
+
+    Args:
+        model ([torch.nn.Module]): the PyTorch model to be profiled.
+        args (list): list of positional arguments to the model.
+        top_modules (int, optional): the number of top modules to print in the aggregated profile. Defaults to 3.
+        as_string (bool, optional): whether to print the output as string. Defaults to True.
+        ignore_modules ([type], optional): the list of modules to ignore during profiling. Defaults to None.
+
+    Returns:
+        The number of floating-point operations, multiply-accumulate operations (MACs), and parameters in the model.
+    """
+    assert isinstance(model, nn.Module), "model must be a PyTorch module"
+    prof = FlopsProfiler(model, module_hoop_mapping=module_hoop_mapping)
+    model.eval()
+
+    assert len(args) > 0, "input args must be specified"
+
+    prof.start_profile(ignore_list=ignore_modules)
+
+    _ = model(*args)
+
+    flops = prof.get_total_flops()
+    params = prof.get_total_params()
+
+    prof.end_profile()
+    if as_string:
+        return (
+            number_to_string(flops),
+            params_to_string(params),
+        )
+
+    return flops, params
diff --git a/icefall/rnn_lm/.gitignore b/icefall/rnn_lm/.gitignore
new file mode 100644
index 000000000..877fb1e18
--- /dev/null
+++ b/icefall/rnn_lm/.gitignore
@@ -0,0 +1 @@
+icefall-librispeech-rnn-lm
diff --git a/icefall/rnn_lm/__init__.py b/icefall/rnn_lm/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/icefall/rnn_lm/check-onnx-streaming.py b/icefall/rnn_lm/check-onnx-streaming.py
new file mode 100755
index 000000000..28b908f82
--- /dev/null
+++ b/icefall/rnn_lm/check-onnx-streaming.py
@@ -0,0 +1,128 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation
+
+"""
+Usage:
+
+./check-onnx-streaming.py \
+  --jit ./icefall-librispeech-rnn-lm/exp/cpu_jit.pt \
+  --onnx ./icefall-librispeech-rnn-lm/exp/with-state-epoch-99-avg-1.onnx
+
+Note: You can download pre-trained models from
+https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+
+"""
+
+import argparse
+import logging
+from typing import Tuple
+
+import onnxruntime as ort
+import torch
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit",
+        required=True,
+        type=str,
+        help="Path to the torchscript model",
+    )
+
+    parser.add_argument(
+        "--onnx",
+        required=True,
+        type=str,
+        help="Path to the onnx model",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(self, filename: str):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.model = ort.InferenceSession(
+            filename,
+            sess_options=session_opts,
+        )
+
+        meta_data = self.model.get_modelmeta().custom_metadata_map
+        self.sos_id = int(meta_data["sos_id"])
+        self.eos_id = int(meta_data["eos_id"])
+        self.vocab_size = int(meta_data["vocab_size"])
+        self.num_layers = int(meta_data["num_layers"])
+        self.hidden_size = int(meta_data["hidden_size"])
+        print(meta_data)
+
+    def __call__(
+        self, x: torch.Tensor, y: torch.Tensor, h0: torch.Tensor, c0: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+                self.model.get_outputs()[2].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: y.numpy(),
+                self.model.get_inputs()[2].name: h0.numpy(),
+                self.model.get_inputs()[3].name: c0.numpy(),
+            },
+        )
+        return (
+            torch.from_numpy(out[0]),
+            torch.from_numpy(out[1]),
+            torch.from_numpy(out[2]),
+        )
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_model = torch.jit.load(args.jit).cpu()
+    onnx_model = OnnxModel(args.onnx)
+    N = torch.arange(1, 5).tolist()
+
+    num_layers = onnx_model.num_layers
+    hidden_size = onnx_model.hidden_size
+
+    for n in N:
+        L = torch.randint(low=1, high=100, size=(1,)).item()
+        x = torch.randint(
+            low=1, high=onnx_model.vocab_size, size=(n, L), dtype=torch.int64
+        )
+        h0 = torch.rand(num_layers, n, hidden_size)
+        c0 = torch.rand(num_layers, n, hidden_size)
+
+        torch_log_prob, torch_h0, torch_c0 = torch_model.score_token_onnx(x, h0, c0)
+        onnx_log_prob, onnx_h0, onnx_c0 = onnx_model(x, h0, c0)
+
+        for torch_v, onnx_v in zip(
+            (torch_log_prob, torch_h0, torch_c0), (onnx_log_prob, onnx_h0, onnx_c0)
+        ):
+            assert torch.allclose(torch_v, onnx_v, atol=1e-5), (
+                torch_v.shape,
+                onnx_v.shape,
+                (torch_v - onnx_v).abs().max(),
+            )
+            print(n, L, torch_v.sum(), onnx_v.sum())
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20230423)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/rnn_lm/check-onnx.py b/icefall/rnn_lm/check-onnx.py
new file mode 100755
index 000000000..24c5395f8
--- /dev/null
+++ b/icefall/rnn_lm/check-onnx.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation
+
+"""
+Usage:
+
+./check-onnx.py \
+  --jit ./icefall-librispeech-rnn-lm/exp/cpu_jit.pt \
+  --onnx ./icefall-librispeech-rnn-lm/exp/no-state-epoch-99-avg-1.onnx
+
+Note: You can download pre-trained models from
+https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+
+"""
+
+import argparse
+import logging
+
+import onnxruntime as ort
+import torch
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--jit",
+        required=True,
+        type=str,
+        help="Path to the torchscript model",
+    )
+
+    parser.add_argument(
+        "--onnx",
+        required=True,
+        type=str,
+        help="Path to the onnx model",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(self, filename: str):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.model = ort.InferenceSession(
+            filename,
+            sess_options=session_opts,
+        )
+
+        meta_data = self.model.get_modelmeta().custom_metadata_map
+        self.sos_id = int(meta_data["sos_id"])
+        self.eos_id = int(meta_data["eos_id"])
+        self.vocab_size = int(meta_data["vocab_size"])
+        print(meta_data)
+
+    def __call__(self, x: torch.Tensor, x_lens: torch.Tensor) -> torch.Tensor:
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0])
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    logging.info(vars(args))
+
+    torch_model = torch.jit.load(args.jit).cpu()
+    onnx_model = OnnxModel(args.onnx)
+    N = torch.arange(1, 5).tolist()
+
+    for n in N:
+        L = torch.randint(low=1, high=100, size=(1,)).item()
+        x = torch.randint(
+            low=1, high=onnx_model.vocab_size, size=(n, L), dtype=torch.int64
+        )
+        x_lens = torch.full((n,), fill_value=L, dtype=torch.int64)
+        if n > 1:
+            x_lens[0] = L // 2 + 1
+
+        sos = torch.full((1,), fill_value=onnx_model.sos_id).expand(n, 1)
+        sos_x = torch.cat([sos, x], dim=1)
+
+        pad_col = torch.zeros((1,), dtype=x.dtype).expand(n, 1)
+        x_eos = torch.cat([x, pad_col], dim=1)
+
+        row_index = torch.arange(0, n, dtype=x.dtype)
+        x_eos[row_index, x_lens] = onnx_model.eos_id
+
+        torch_nll = torch_model(sos_x, x_eos, x_lens + 1).sum(dim=-1)
+        onnx_nll = onnx_model(x, x_lens)
+        # Note: For int8 models, the differences may be quite large,
+        # e.g., within 0.9
+        assert torch.allclose(torch_nll, onnx_nll), (
+            torch_nll,
+            onnx_nll,
+        )
+        print(n, L, torch_nll, onnx_nll)
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20230420)
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/rnn_lm/compute_perplexity.py b/icefall/rnn_lm/compute_perplexity.py
new file mode 100755
index 000000000..cc566bd92
--- /dev/null
+++ b/icefall/rnn_lm/compute_perplexity.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  ./rnn_lm/compute_perplexity.py \
+    --epoch 4 \
+    --avg 2 \
+    --lm-data ./data/lm_training_bpe_500/sorted_lm_data-test.pt
+
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+
+import torch
+from dataset import get_dataloader
+from model import RnnLmModel
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=49,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        help="Path to the LM test data for computing perplexity",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=50,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--max-sent-len",
+        type=int,
+        default=100,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--sos-id",
+        type=int,
+        default=1,
+        help="SOS ID",
+    )
+
+    parser.add_argument(
+        "--eos-id",
+        type=int,
+        default=1,
+        help="EOS ID",
+    )
+
+    parser.add_argument(
+        "--blank-id",
+        type=int,
+        default=0,
+        help="Blank ID",
+    )
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lm_data = Path(args.lm_data)
+
+    params = AttributeDict(vars(args))
+
+    if params.iter > 0:
+        setup_logger(
+            f"{params.exp_dir}/log-ppl/log-ppl-iter-{params.iter}-avg-{params.avg}"
+        )
+    else:
+        setup_logger(
+            f"{params.exp_dir}/log-ppl/log-ppl-epoch-{params.epoch}-avg-{params.avg}"
+        )
+    logging.info("Computing perplexity started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    logging.info("About to create model")
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if i >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    num_param_requires_grad = sum(
+        [p.numel() for p in model.parameters() if p.requires_grad]
+    )
+
+    logging.info(f"Number of model parameters: {num_param}")
+    logging.info(
+        f"Number of model parameters (requires_grad): "
+        f"{num_param_requires_grad} "
+        f"({num_param_requires_grad/num_param_requires_grad*100}%)"
+    )
+
+    logging.info(f"Loading LM test data from {params.lm_data}")
+    test_dl = get_dataloader(
+        filename=params.lm_data,
+        is_distributed=False,
+        params=params,
+    )
+
+    tot_loss = 0.0
+    num_tokens = 0
+    num_sentences = 0
+    for batch_idx, batch in enumerate(test_dl):
+        x, y, sentence_lengths = batch
+        x = x.to(device)
+        y = y.to(device)
+        sentence_lengths = sentence_lengths.to(device)
+
+        nll = model(x, y, sentence_lengths)
+        loss = nll.sum().cpu().item()
+
+        tot_loss += loss
+        num_tokens += sentence_lengths.sum().cpu().item()
+        num_sentences += x.size(0)
+
+    ppl = math.exp(tot_loss / num_tokens)
+    logging.info(
+        f"total nll: {tot_loss}, num tokens: {num_tokens}, "
+        f"num sentences: {num_sentences}, ppl: {ppl:.3f}"
+    )
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/rnn_lm/dataset.py b/icefall/rnn_lm/dataset.py
new file mode 100644
index 000000000..53be53f64
--- /dev/null
+++ b/icefall/rnn_lm/dataset.py
@@ -0,0 +1,214 @@
+# Copyright (c)  2021  Xiaomi Corporation (authors: Daniel Povey, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Tuple
+
+import k2
+import torch
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
+
+from icefall.utils import AttributeDict, add_eos, add_sos
+
+
+class LmDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        sentences: k2.RaggedTensor,
+        words: k2.RaggedTensor,
+        sentence_lengths: torch.Tensor,
+        max_sent_len: int,
+        batch_size: int,
+    ):
+        """
+        Args:
+          sentences:
+            A ragged tensor of dtype torch.int32 with 2 axes [sentence][word].
+          words:
+            A ragged tensor of dtype torch.int32 with 2 axes [word][token].
+          sentence_lengths:
+            A 1-D tensor of dtype torch.int32 containing number of tokens
+            of each sentence.
+          max_sent_len:
+            Maximum sentence length. It is used to change the batch size
+            dynamically. In general, we try to keep the product of
+            "max_sent_len in a batch" and "num_of_sent in a batch" being
+            a constant.
+          batch_size:
+            The expected batch size. It is changed dynamically according
+            to the "max_sent_len".
+
+        See `../local/prepare_lm_training_data.py` for how `sentences` and
+        `words` are generated. We assume that `sentences` are sorted by length.
+        See `../local/sort_lm_training_data.py`.
+        """
+        super().__init__()
+        self.sentences = sentences
+        self.words = words
+
+        sentence_lengths = sentence_lengths.tolist()
+
+        assert batch_size > 0, batch_size
+        assert max_sent_len > 1, max_sent_len
+        batch_indexes = []
+        num_sentences = sentences.dim0
+        cur = 0
+        while cur < num_sentences:
+            sz = sentence_lengths[cur] // max_sent_len + 1
+            # Assume the current sentence has 3 * max_sent_len tokens,
+            # in the worst case, the subsequent sentences also have
+            # this number of tokens, we should reduce the batch size
+            # so that this batch will not contain too many tokens
+            actual_batch_size = batch_size // sz + 1
+            actual_batch_size = min(actual_batch_size, batch_size)
+            end = cur + actual_batch_size
+            end = min(end, num_sentences)
+            this_batch_indexes = torch.arange(cur, end).tolist()
+            batch_indexes.append(this_batch_indexes)
+            cur = end
+        assert batch_indexes[-1][-1] == num_sentences - 1
+
+        self.batch_indexes = k2.RaggedTensor(batch_indexes)
+
+    def __len__(self) -> int:
+        """Return number of batches in this dataset"""
+        return self.batch_indexes.dim0
+
+    def __getitem__(self, i: int) -> k2.RaggedTensor:
+        """Get the i'th batch in this dataset
+        Return a ragged tensor with 2 axes [sentence][token].
+        """
+        assert 0 <= i < len(self), i
+
+        # indexes is a 1-D tensor containing sentence indexes
+        indexes = self.batch_indexes[i]
+
+        # sentence_words is a ragged tensor with 2 axes
+        # [sentence][word]
+        sentence_words = self.sentences[indexes]
+
+        # in case indexes contains only 1 entry, the returned
+        # sentence_words is a 1-D tensor, we have to convert
+        # it to a ragged tensor
+        if isinstance(sentence_words, torch.Tensor):
+            sentence_words = k2.RaggedTensor(sentence_words.unsqueeze(0))
+
+        # sentence_word_tokens is a ragged tensor with 3 axes
+        # [sentence][word][token]
+        sentence_word_tokens = self.words.index(sentence_words)
+        assert sentence_word_tokens.num_axes == 3
+
+        sentence_tokens = sentence_word_tokens.remove_axis(1)
+        return sentence_tokens
+
+
+class LmDatasetCollate:
+    def __init__(self, sos_id: int, eos_id: int, blank_id: int):
+        """
+        Args:
+          sos_id:
+            Token ID of the SOS symbol.
+          eos_id:
+            Token ID of the EOS symbol.
+          blank_id:
+            Token ID of the blank symbol.
+        """
+        self.sos_id = sos_id
+        self.eos_id = eos_id
+        self.blank_id = blank_id
+
+    def __call__(
+        self, batch: List[k2.RaggedTensor]
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Return a tuple containing 3 tensors:
+
+        - x, a 2-D tensor of dtype torch.int32; each row contains tokens
+             for a sentence starting with `self.sos_id`. It is padded to
+             the max sentence length with `self.blank_id`.
+
+        - y, a 2-D tensor of dtype torch.int32; each row contains tokens
+             for a sentence ending with `self.eos_id` before padding.
+             Then it is padded to the max sentence length with
+             `self.blank_id`.
+
+        - lengths, a 2-D tensor of dtype torch.int32, containing the number of
+                   tokens of each sentence before padding.
+        """
+        # The batching stuff has already been done in LmDataset
+        assert len(batch) == 1
+        sentence_tokens = batch[0]
+        row_splits = sentence_tokens.shape.row_splits(1)
+        sentence_token_lengths = row_splits[1:] - row_splits[:-1]
+        sentence_tokens_with_sos = add_sos(sentence_tokens, self.sos_id)
+        sentence_tokens_with_eos = add_eos(sentence_tokens, self.eos_id)
+
+        x = sentence_tokens_with_sos.pad(mode="constant", padding_value=self.blank_id)
+        y = sentence_tokens_with_eos.pad(mode="constant", padding_value=self.blank_id)
+        sentence_token_lengths += 1  # plus 1 since we added a SOS
+
+        return x.to(torch.int64), y.to(torch.int64), sentence_token_lengths
+
+
+def get_dataloader(
+    filename: str,
+    is_distributed: bool,
+    params: AttributeDict,
+) -> torch.utils.data.DataLoader:
+    """Get dataloader for LM training.
+
+    Args:
+      filename:
+        Path to the file containing LM data. The file is assumed to
+        be generated by `../local/sort_lm_training_data.py`.
+      is_distributed:
+        True if using DDP training. False otherwise.
+      params:
+        Set `get_params()` from `rnn_lm/train.py`
+    Returns:
+      Return a dataloader containing the LM data.
+    """
+    lm_data = torch.load(filename)
+
+    words = lm_data["words"]
+    sentences = lm_data["sentences"]
+    sentence_lengths = lm_data["sentence_lengths"]
+
+    dataset = LmDataset(
+        sentences=sentences,
+        words=words,
+        sentence_lengths=sentence_lengths,
+        max_sent_len=params.max_sent_len,
+        batch_size=params.batch_size,
+    )
+    if is_distributed:
+        sampler = DistributedSampler(dataset, shuffle=True, drop_last=True)
+    else:
+        sampler = None
+
+    collate_fn = LmDatasetCollate(
+        sos_id=params.sos_id,
+        eos_id=params.eos_id,
+        blank_id=params.blank_id,
+    )
+
+    dataloader = DataLoader(
+        dataset,
+        batch_size=1,
+        collate_fn=collate_fn,
+        sampler=sampler,
+        shuffle=sampler is None,
+    )
+    return dataloader
diff --git a/icefall/rnn_lm/export-onnx.py b/icefall/rnn_lm/export-onnx.py
new file mode 100755
index 000000000..dfede708b
--- /dev/null
+++ b/icefall/rnn_lm/export-onnx.py
@@ -0,0 +1,395 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation
+
+import argparse
+import logging
+from pathlib import Path
+
+import onnx
+import torch
+from model import RnnLmModel
+from onnxruntime.quantization import QuantType, quantize_dynamic
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import AttributeDict, str2bool
+from typing import Dict
+from train import get_params
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+# A wrapper for RnnLm model to simpily the C++ calling code
+# when exporting the model to ONNX.
+#
+# TODO(fangjun): The current wrapper works only for non-streaming ASR
+# since we don't expose the LM state and it is used to score
+# a complete sentence at once.
+class RnnLmModelWrapper(torch.nn.Module):
+    def __init__(self, model: RnnLmModel, sos_id: int, eos_id: int):
+        super().__init__()
+        self.model = model
+        self.sos_id = sos_id
+        self.eos_id = eos_id
+
+    def forward(self, x: torch.Tensor, x_lens: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 2-D tensor of shape (N, L) with dtype torch.int64.
+            It does not contain SOS or EOS. We will add SOS and EOS inside
+            this function.
+          x_lens:
+            A 1-D tensor of shape (N,) with dtype torch.int64. It contains
+            number of valid tokens in ``x`` before padding.
+        Returns:
+          Return a 1-D tensor of shape (N,) containing negative loglikelihood.
+          Its dtype is torch.float32
+        """
+        N = x.size(0)
+
+        sos_tensor = torch.full((1,), fill_value=self.sos_id, dtype=x.dtype).expand(
+            N, 1
+        )
+        sos_x = torch.cat([sos_tensor, x], dim=1)
+
+        pad_col = torch.zeros((1,), dtype=x.dtype).expand(N, 1)
+        x_eos = torch.cat([x, pad_col], dim=1)
+
+        row_index = torch.arange(0, N, dtype=x.dtype)
+        x_eos[row_index, x_lens] = self.eos_id
+
+        # use x_lens + 1 here since we prepended x with sos
+        return (
+            self.model(x=sos_x, y=x_eos, lengths=x_lens + 1)
+            .to(torch.float32)
+            .sum(dim=1)
+        )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=True,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    return parser
+
+
+def export_without_state(
+    model: RnnLmModel,
+    filename: str,
+    params: AttributeDict,
+    opset_version: int,
+):
+    model_wrapper = RnnLmModelWrapper(
+        model,
+        sos_id=params.sos_id,
+        eos_id=params.eos_id,
+    )
+
+    N = 1
+    L = 20
+    x = torch.randint(low=1, high=params.vocab_size, size=(N, L), dtype=torch.int64)
+    x_lens = torch.full((N,), fill_value=L, dtype=torch.int64)
+
+    # Note(fangjun): The following warnings can be ignored.
+    # We can use ./check-onnx.py to validate the exported model with batch_size > 1
+    """
+    torch/onnx/symbolic_opset9.py:2119: UserWarning: Exporting a model to ONNX
+    with a batch_size other than 1, with a variable length with LSTM can cause
+    an error when running the ONNX model with a different batch size. Make sure
+    to save the model with a batch size of 1, or define the initial states
+    (h0/c0) as inputs of the model. warnings.warn("Exporting a model to ONNX
+    with a batch_size other than 1, " +
+    """
+
+    torch.onnx.export(
+        model_wrapper,
+        (x, x_lens),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["nll"],
+        dynamic_axes={
+            "x": {0: "N", 1: "L"},
+            "x_lens": {0: "N"},
+            "nll": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "rnnlm",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "rnnlm without state",
+        "sos_id": str(params.sos_id),
+        "eos_id": str(params.eos_id),
+        "vocab_size": str(params.vocab_size),
+        "url": "https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=filename, meta_data=meta_data)
+
+
+def export_with_state(
+    model: RnnLmModel,
+    filename: str,
+    params: AttributeDict,
+    opset_version: int,
+):
+    N = 1
+    L = 20
+    num_layers = model.rnn.num_layers
+    hidden_size = model.rnn.hidden_size
+    embedding_dim = model.embedding_dim
+
+    x = torch.randint(low=1, high=params.vocab_size, size=(N, L), dtype=torch.int64)
+    h0 = torch.zeros(num_layers, N, hidden_size)
+    c0 = torch.zeros(num_layers, N, hidden_size)
+
+    # Note(fangjun): The following warnings can be ignored.
+    # We can use ./check-onnx.py to validate the exported model with batch_size > 1
+    """
+    torch/onnx/symbolic_opset9.py:2119: UserWarning: Exporting a model to ONNX
+    with a batch_size other than 1, with a variable length with LSTM can cause
+    an error when running the ONNX model with a different batch size. Make sure
+    to save the model with a batch size of 1, or define the initial states
+    (h0/c0) as inputs of the model. warnings.warn("Exporting a model to ONNX
+    with a batch_size other than 1, " +
+    """
+
+    torch.onnx.export(
+        model,
+        (x, h0, c0),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "h0", "c0"],
+        output_names=["log_softmax", "next_h0", "next_c0"],
+        dynamic_axes={
+            "x": {0: "N", 1: "L"},
+            "h0": {1: "N"},
+            "c0": {1: "N"},
+            "log_softmax": {0: "N"},
+            "next_h0": {1: "N"},
+            "next_c0": {1: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "rnnlm",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "rnnlm state",
+        "sos_id": str(params.sos_id),
+        "eos_id": str(params.eos_id),
+        "vocab_size": str(params.vocab_size),
+        "num_layers": str(num_layers),
+        "hidden_size": str(hidden_size),
+        "embedding_dim": str(embedding_dim),
+        "url": "https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    device = torch.device("cpu")
+    logging.info(f"device: {device}")
+
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if i >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting model without state")
+    filename = params.exp_dir / f"no-state-{suffix}.onnx"
+    export_without_state(
+        model=model,
+        filename=filename,
+        params=params,
+        opset_version=opset_version,
+    )
+
+    filename_int8 = params.exp_dir / f"no-state-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=filename,
+        model_output=filename_int8,
+        weight_type=QuantType.QInt8,
+    )
+
+    # now for streaming export
+    saved_forward = model.__class__.forward
+    model.__class__.forward = model.__class__.score_token_onnx
+    streaming_filename = params.exp_dir / f"with-state-{suffix}.onnx"
+    export_with_state(
+        model=model,
+        filename=streaming_filename,
+        params=params,
+        opset_version=opset_version,
+    )
+    model.__class__.forward = saved_forward
+
+    streaming_filename_int8 = params.exp_dir / f"with-state-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=streaming_filename,
+        model_output=streaming_filename_int8,
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/rnn_lm/export-onnx.sh b/icefall/rnn_lm/export-onnx.sh
new file mode 100755
index 000000000..6e3262b5e
--- /dev/null
+++ b/icefall/rnn_lm/export-onnx.sh
@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+
+# We use the model from
+# https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main
+# as an example
+
+export CUDA_VISIBLE_DEVICES=
+
+if [ ! -f ./icefall-librispeech-rnn-lm/exp/pretrained.pt ]; then
+  GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+  pushd icefall-librispeech-rnn-lm/exp
+  git lfs pull --include "pretrained.pt"
+  ln -s pretrained.pt epoch-99.pt
+  popd
+fi
+
+python3 ./export-onnx.py \
+  --exp-dir ./icefall-librispeech-rnn-lm/exp \
+  --epoch 99 \
+  --avg 1 \
+  --vocab-size 500 \
+  --embedding-dim 2048 \
+  --hidden-dim 2048 \
+  --num-layers 3 \
+  --tie-weights 1
+
diff --git a/icefall/rnn_lm/export.py b/icefall/rnn_lm/export.py
new file mode 100644
index 000000000..dadf23009
--- /dev/null
+++ b/icefall/rnn_lm/export.py
@@ -0,0 +1,205 @@
+#!/usr/bin/env python3
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from model import RnnLmModel
+
+from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=True,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = AttributeDict({})
+    params.update(vars(args))
+
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    model.to(device)
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+    elif params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if i >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints(filenames, device=device), strict=False
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+
+        model.__class__.score_token_onnx = torch.jit.export(
+            model.__class__.score_token_onnx
+        )
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/rnn_lm/export.sh b/icefall/rnn_lm/export.sh
new file mode 100755
index 000000000..678bc294e
--- /dev/null
+++ b/icefall/rnn_lm/export.sh
@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+
+# We use the model from
+# https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main
+# as an example
+
+export CUDA_VISIBLE_DEVICES=
+
+if [ ! -f ./icefall-librispeech-rnn-lm/exp/pretrained.pt ]; then
+  GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+  pushd icefall-librispeech-rnn-lm/exp
+  git lfs pull --include "pretrained.pt"
+  ln -s pretrained.pt epoch-99.pt
+  popd
+fi
+
+python3 ./export.py \
+  --exp-dir ./icefall-librispeech-rnn-lm/exp \
+  --epoch 99 \
+  --avg 1 \
+  --vocab-size 500 \
+  --embedding-dim 2048 \
+  --hidden-dim 2048 \
+  --num-layers 3 \
+  --tie-weights 1 \
+  --jit 1
+
diff --git a/icefall/rnn_lm/model.py b/icefall/rnn_lm/model.py
new file mode 100644
index 000000000..5eacf5d40
--- /dev/null
+++ b/icefall/rnn_lm/model.py
@@ -0,0 +1,295 @@
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from typing import Tuple
+
+import torch
+import torch.nn.functional as F
+
+from icefall.utils import add_eos, add_sos, make_pad_mask
+
+
+class RnnLmModel(torch.nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        hidden_dim: int,
+        num_layers: int,
+        tie_weights: bool = False,
+    ):
+        """
+        Args:
+          vocab_size:
+            Vocabulary size of BPE model.
+          embedding_dim:
+            Input embedding dimension.
+          hidden_dim:
+            Hidden dimension of RNN layers.
+          num_layers:
+            Number of RNN layers.
+          tie_weights:
+            True to share the weights between the input embedding layer and the
+            last output linear layer. See https://arxiv.org/abs/1608.05859
+            and https://arxiv.org/abs/1611.01462
+        """
+        super().__init__()
+        self.vocab_size = vocab_size
+        self.embedding_dim = embedding_dim
+        self.hidden_dim = hidden_dim
+        self.num_layers = num_layers
+        self.tie_weights = tie_weights
+
+        self.input_embedding = torch.nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+        )
+
+        self.rnn = torch.nn.LSTM(
+            input_size=embedding_dim,
+            hidden_size=hidden_dim,
+            num_layers=num_layers,
+            batch_first=True,
+        )
+
+        self.output_linear = torch.nn.Linear(
+            in_features=hidden_dim, out_features=vocab_size
+        )
+
+        self.vocab_size = vocab_size
+        if tie_weights:
+            logging.info("Tying weights")
+            assert embedding_dim == hidden_dim, (embedding_dim, hidden_dim)
+            self.output_linear.weight = self.input_embedding.weight
+        else:
+            logging.info("Not tying weights")
+
+        self.cache = {}
+
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        h0: torch.Tensor,
+        c0: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 2-D tensor of shape (N, L). We won't prepend it with SOS.
+          y:
+            A 2-D tensor of shape (N, L). We won't append it with EOS.
+          h0:
+            A 3-D tensor of shape (num_layers, N, hidden_size).
+            (If proj_size > 0, then it is (num_layers, N, proj_size))
+          c0:
+            A 3-D tensor of shape (num_layers, N, hidden_size).
+        Returns:
+          Return a tuple containing 3 tensors:
+            - negative loglike (nll), a 1-D tensor of shape (N,)
+            - next_h0, a 3-D tensor with the same shape as h0
+            - next_c0, a 3-D tensor with the same shape as c0
+        """
+        assert x.ndim == y.ndim == 2, (x.ndim, y.ndim)
+        assert x.shape == y.shape, (x.shape, y.shape)
+
+        # embedding is of shape (N, L, embedding_dim)
+        embedding = self.input_embedding(x)
+        # Note: We use batch_first==True
+        rnn_out, (next_h0, next_c0) = self.rnn(embedding, (h0, c0))
+        logits = self.output_linear(rnn_out)
+        nll_loss = F.cross_entropy(
+            logits.reshape(-1, self.vocab_size), y.reshape(-1), reduction="none"
+        )
+
+        batch_size = x.size(0)
+        nll_loss = nll_loss.reshape(batch_size, -1).sum(dim=1)
+        return nll_loss, next_h0, next_c0
+
+    def forward(
+        self, x: torch.Tensor, y: torch.Tensor, lengths: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 2-D tensor with shape (N, L). Each row
+            contains token IDs for a sentence and starts with the SOS token.
+          y:
+            A shifted version of `x` and with EOS appended.
+          lengths:
+            A 1-D tensor of shape (N,). It contains the sentence lengths
+            before padding.
+        Returns:
+          Return a 2-D tensor of shape (N, L) containing negative log-likelihood
+          loss values. Note: Loss values for padding positions are set to 0.
+        """
+        assert x.ndim == y.ndim == 2, (x.ndim, y.ndim)
+        assert lengths.ndim == 1, lengths.ndim
+        assert x.shape == y.shape, (x.shape, y.shape)
+
+        batch_size = x.size(0)
+        assert lengths.size(0) == batch_size, (lengths.size(0), batch_size)
+
+        # embedding is of shape (N, L, embedding_dim)
+        embedding = self.input_embedding(x)
+
+        # Note: We use batch_first==True
+        rnn_out, _ = self.rnn(embedding)
+        logits = self.output_linear(rnn_out)
+
+        # Note: No need to use `log_softmax()` here
+        # since F.cross_entropy() expects unnormalized probabilities
+
+        # nll_loss is of shape (N*L,)
+        # nll -> negative log-likelihood
+        nll_loss = F.cross_entropy(
+            logits.reshape(-1, self.vocab_size), y.reshape(-1), reduction="none"
+        )
+        # Set loss values for padding positions to 0
+        mask = make_pad_mask(lengths).reshape(-1)
+        nll_loss.masked_fill_(mask, 0)
+
+        nll_loss = nll_loss.reshape(batch_size, -1)
+
+        return nll_loss
+
+    def predict_batch(self, tokens, token_lens, sos_id, eos_id, blank_id):
+        device = next(self.parameters()).device
+        batch_size = len(token_lens)
+
+        sos_tokens = add_sos(tokens, sos_id)
+        tokens_eos = add_eos(tokens, eos_id)
+        sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+
+        sentence_lengths = sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+
+        x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+        y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+        x_tokens = x_tokens.to(torch.int64).to(device)
+        y_tokens = y_tokens.to(torch.int64).to(device)
+        sentence_lengths = sentence_lengths.to(torch.int64).to(device)
+
+        embedding = self.input_embedding(x_tokens)
+
+        # Note: We use batch_first==True
+        rnn_out, states = self.rnn(embedding)
+        logits = self.output_linear(rnn_out)
+        mask = torch.zeros(logits.shape).bool().to(device)
+        for i in range(batch_size):
+            mask[i, token_lens[i], :] = True
+        logits = logits[mask].reshape(batch_size, -1)
+
+        return logits[:, :].log_softmax(-1), states
+
+    def clean_cache(self):
+        self.cache = {}
+
+    def score_token(self, x: torch.Tensor, x_lens: torch.Tensor, state=None):
+        """Score a batch of tokens, i.e each sample in the batch should be a
+        single token. For example, x = torch.tensor([[5],[10],[20]])
+
+
+        Args:
+            x (torch.Tensor):
+                A batch of tokens
+            x_lens (torch.Tensor):
+                The length of tokens in the batch before padding
+            state (optional):
+                Either None or a tuple of two torch.Tensor. Each tensor has
+                the shape of (num_layers, bs, hidden_dim)
+
+        Returns:
+            _type_: _description_
+        """
+        device = next(self.parameters()).device
+        batch_size = x.size(0)
+        if state:
+            h, c = state
+        else:
+            h = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.hidden_size).to(
+                device
+            )
+            c = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.hidden_size).to(
+                device
+            )
+
+        embedding = self.input_embedding(x)
+        rnn_out, states = self.rnn(embedding, (h, c))
+        logits = self.output_linear(rnn_out)
+
+        return logits[:, 0].log_softmax(-1), states
+
+    def score_token_onnx(
+        self,
+        x: torch.Tensor,
+        state_h: torch.Tensor,
+        state_c: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Score a batch of tokens, i.e each sample in the batch should be a
+        single token. For example, x = torch.tensor([[5],[10],[20]])
+
+
+        Args:
+            x (torch.Tensor):
+                A batch of tokens
+            state_h:
+                state h of RNN has the shape of (num_layers, bs, hidden_dim)
+            state_c:
+                state c of RNN has the shape of (num_layers, bs, hidden_dim)
+
+        Returns:
+            _type_: _description_
+        """
+        embedding = self.input_embedding(x)
+        rnn_out, (next_h0, next_c0) = self.rnn(embedding, (state_h, state_c))
+        logits = self.output_linear(rnn_out)
+
+        return logits[:, 0].log_softmax(-1), next_h0, next_c0
+
+    def forward_with_state(
+        self, tokens, token_lens, sos_id, eos_id, blank_id, state=None
+    ):
+        batch_size = len(token_lens)
+        if state:
+            h, c = state
+        else:
+            h = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.hidden_size)
+            c = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.hidden_size)
+
+        device = next(self.parameters()).device
+
+        sos_tokens = add_sos(tokens, sos_id)
+        tokens_eos = add_eos(tokens, eos_id)
+        sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+
+        sentence_lengths = sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+
+        x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+        y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+        x_tokens = x_tokens.to(torch.int64).to(device)
+        y_tokens = y_tokens.to(torch.int64).to(device)
+        sentence_lengths = sentence_lengths.to(torch.int64).to(device)
+
+        embedding = self.input_embedding(x_tokens)
+
+        # Note: We use batch_first==True
+        rnn_out, states = self.rnn(embedding, (h, c))
+        logits = self.output_linear(rnn_out)
+
+        return logits, states
diff --git a/icefall/rnn_lm/test_dataset.py b/icefall/rnn_lm/test_dataset.py
new file mode 100755
index 000000000..bf961f54b
--- /dev/null
+++ b/icefall/rnn_lm/test_dataset.py
@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import k2
+import torch
+from rnn_lm.dataset import LmDataset, LmDatasetCollate
+
+
+def main():
+    sentences = k2.RaggedTensor(
+        [[0, 1, 2], [1, 0, 1], [0, 1], [1, 3, 0, 2, 0], [3], [0, 2, 1]]
+    )
+    words = k2.RaggedTensor([[3, 6], [2, 8, 9, 3], [5], [5, 6, 7, 8, 9]])
+
+    num_sentences = sentences.dim0
+
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    sentence_lengths = torch.tensor(sentence_lengths, dtype=torch.int32)
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+    sentences = sentences[indices.to(torch.int32)]
+    sentence_lengths = sentence_lengths[indices]
+
+    dataset = LmDataset(
+        sentences=sentences,
+        words=words,
+        sentence_lengths=sentence_lengths,
+        max_sent_len=3,
+        batch_size=4,
+    )
+
+    collate_fn = LmDatasetCollate(sos_id=1, eos_id=-1, blank_id=0)
+    dataloader = torch.utils.data.DataLoader(
+        dataset, batch_size=1, collate_fn=collate_fn
+    )
+
+    for i in dataloader:
+        print(i)
+    # I've checked the output manually; the output is as expected.
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/rnn_lm/test_dataset_ddp.py b/icefall/rnn_lm/test_dataset_ddp.py
new file mode 100755
index 000000000..48fbb19cb
--- /dev/null
+++ b/icefall/rnn_lm/test_dataset_ddp.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import k2
+import torch
+import torch.multiprocessing as mp
+from rnn_lm.dataset import LmDataset, LmDatasetCollate
+from torch import distributed as dist
+
+
+def generate_data():
+    sentences = k2.RaggedTensor(
+        [[0, 1, 2], [1, 0, 1], [0, 1], [1, 3, 0, 2, 0], [3], [0, 2, 1]]
+    )
+    words = k2.RaggedTensor([[3, 6], [2, 8, 9, 3], [5], [5, 6, 7, 8, 9]])
+
+    num_sentences = sentences.dim0
+
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    sentence_lengths = torch.tensor(sentence_lengths, dtype=torch.int32)
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+    sentences = sentences[indices.to(torch.int32)]
+    sentence_lengths = sentence_lengths[indices]
+
+    return sentences, words, sentence_lengths
+
+
+def run(rank, world_size):
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = "12352"
+
+    dist.init_process_group("nccl", rank=rank, world_size=world_size)
+    torch.cuda.set_device(rank)
+
+    sentences, words, sentence_lengths = generate_data()
+
+    dataset = LmDataset(
+        sentences=sentences,
+        words=words,
+        sentence_lengths=sentence_lengths,
+        max_sent_len=3,
+        batch_size=4,
+    )
+    sampler = torch.utils.data.distributed.DistributedSampler(
+        dataset, shuffle=True, drop_last=False
+    )
+
+    collate_fn = LmDatasetCollate(sos_id=1, eos_id=-1, blank_id=0)
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=1,
+        collate_fn=collate_fn,
+        sampler=sampler,
+        shuffle=False,
+    )
+
+    for i in dataloader:
+        print(f"rank: {rank}", i)
+
+    dist.destroy_process_group()
+
+
+def main():
+    world_size = 2
+    mp.spawn(run, args=(world_size,), nprocs=world_size, join=True)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/rnn_lm/test_model.py b/icefall/rnn_lm/test_model.py
new file mode 100755
index 000000000..5a216a3fb
--- /dev/null
+++ b/icefall/rnn_lm/test_model.py
@@ -0,0 +1,69 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from rnn_lm.model import RnnLmModel
+
+
+def test_rnn_lm_model():
+    vocab_size = 4
+    model = RnnLmModel(
+        vocab_size=vocab_size, embedding_dim=10, hidden_dim=10, num_layers=2
+    )
+    x = torch.tensor(
+        [
+            [1, 3, 2, 2],
+            [1, 2, 2, 0],
+            [1, 2, 0, 0],
+        ]
+    )
+    y = torch.tensor(
+        [
+            [3, 2, 2, 1],
+            [2, 2, 1, 0],
+            [2, 1, 0, 0],
+        ]
+    )
+    lengths = torch.tensor([4, 3, 2])
+    nll_loss = model(x, y, lengths)
+    print(nll_loss)
+    """
+    tensor([[1.1180, 1.3059, 1.2426, 1.7773],
+            [1.4231, 1.2783, 1.7321, 0.0000],
+            [1.4231, 1.6752, 0.0000, 0.0000]], grad_fn=<ViewBackward>)
+    """
+
+
+def test_rnn_lm_model_tie_weights():
+    model = RnnLmModel(
+        vocab_size=10,
+        embedding_dim=10,
+        hidden_dim=10,
+        num_layers=2,
+        tie_weights=True,
+    )
+    assert model.input_embedding.weight is model.output_linear.weight
+
+
+def main():
+    test_rnn_lm_model()
+    test_rnn_lm_model_tie_weights()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20211122)
+    main()
diff --git a/icefall/rnn_lm/train.py b/icefall/rnn_lm/train.py
new file mode 100755
index 000000000..0178b80bf
--- /dev/null
+++ b/icefall/rnn_lm/train.py
@@ -0,0 +1,689 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+    ./rnn_lm/train.py \
+        --start-epoch 0 \
+        --world-size 2 \
+        --num-epochs 1 \
+        --use-fp16 0 \
+        --tie-weights 0 \
+        --embedding-dim 800 \
+        --hidden-dim 200 \
+        --num-layers 2 \
+        --batch-size 400
+
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from dataset import get_dataloader
+from lhotse.utils import fix_random_seed
+from model import RnnLmModel
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import save_checkpoint_with_global_batch_idx
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        exp_dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, logs, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=True,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=400,
+    )
+
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        default="data/lm_training_bpe_500/sorted_lm_data.pt",
+        help="LM training data",
+    )
+
+    parser.add_argument(
+        "--lm-data-valid",
+        type=str,
+        default="data/lm_training_bpe_500/sorted_lm_data-valid.pt",
+        help="LM validation data",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=True,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--lr",
+        type=float,
+        default=1e-3,
+    )
+
+    parser.add_argument(
+        "--max-sent-len",
+        type=int,
+        default=200,
+        help="""Maximum number of tokens in a sentence. This is used
+        to adjust batch-size dynamically""",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters."""
+
+    params = AttributeDict(
+        {
+            "max_sent_len": 200,
+            "sos_id": 1,
+            "eos_id": 1,
+            "blank_id": 0,
+            "weight_decay": 1e-6,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 100,
+            "reset_interval": 2000,
+            "valid_interval": 200,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    logging.info(f"Loading checkpoint: {filename}")
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    model: nn.Module,
+    x: torch.Tensor,
+    y: torch.Tensor,
+    sentence_lengths: torch.Tensor,
+    is_training: bool,
+) -> Tuple[torch.Tensor, MetricsTracker]:
+    """Compute the negative log-likelihood loss given a model and its input.
+    Args:
+      model:
+        The NN model, e.g., RnnLmModel.
+      x:
+        A 2-D tensor. Each row contains BPE token IDs for a sentence. Also,
+        each row starts with SOS ID.
+      y:
+        A 2-D tensor. Each row is a shifted version of the corresponding row
+        in `x` but ends with an EOS ID (before padding).
+     sentence_lengths:
+       A 1-D tensor containing number of tokens of each sentence
+       before padding.
+     is_training:
+       True for training. False for validation.
+    """
+    with torch.set_grad_enabled(is_training):
+        device = model.device
+        x = x.to(device)
+        y = y.to(device)
+        sentence_lengths = sentence_lengths.to(device)
+
+        nll = model(x, y, sentence_lengths)
+        loss = nll.sum()
+
+        num_tokens = sentence_lengths.sum().item()
+
+        loss_info = MetricsTracker()
+        # Note: Due to how MetricsTracker() is designed,
+        # we use "frames" instead of "num_tokens" as a key here
+        loss_info["frames"] = num_tokens
+        loss_info["loss"] = loss.detach().item()
+    return loss, loss_info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        x, y, sentence_lengths = batch
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                model=model,
+                x=x,
+                y=y,
+                sentence_lengths=sentence_lengths,
+                is_training=False,
+            )
+
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all sentences is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        x, y, sentence_lengths = batch
+        batch_size = x.size(0)
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                model=model,
+                x=x,
+                y=y,
+                sentence_lengths=sentence_lengths,
+                is_training=True,
+            )
+
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                params=params,
+                optimizer=optimizer,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+
+        if batch_idx % params.log_interval == 0:
+            # Note: "frames" here means "num_tokens"
+            this_batch_ppl = math.exp(loss_info["loss"] / loss_info["frames"])
+            tot_ppl = math.exp(tot_loss["loss"] / tot_loss["frames"])
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}, ppl: {this_batch_ppl}] "
+                f"tot_loss[{tot_loss}, ppl: {tot_ppl}], "
+                f"batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+                tb_writer.add_scalar(
+                    "train/current_ppl", this_batch_ppl, params.batch_idx_train
+                )
+
+                tb_writer.add_scalar("train/tot_ppl", tot_ppl, params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+
+            valid_ppl = math.exp(valid_info["loss"] / valid_info["frames"])
+            logging.info(
+                f"Epoch {params.cur_epoch}, validation: {valid_info}, "
+                f"ppl: {valid_ppl}"
+            )
+
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+                tb_writer.add_scalar(
+                    "train/valid_ppl", valid_ppl, params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    is_distributed = world_size > 1
+
+    fix_random_seed(params.seed)
+    if is_distributed:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    logging.info(f"Device: {device}")
+
+    logging.info("About to create model")
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if is_distributed:
+        model = DDP(model, device_ids=[rank])
+
+    model.device = device
+
+    optimizer = optim.Adam(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    if checkpoints:
+        logging.info("Load optimizer state_dict from checkpoint")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    logging.info(f"Loading LM training data from {params.lm_data}")
+    train_dl = get_dataloader(
+        filename=params.lm_data,
+        is_distributed=is_distributed,
+        params=params,
+    )
+
+    logging.info(f"Loading LM validation data from {params.lm_data_valid}")
+    valid_dl = get_dataloader(
+        filename=params.lm_data_valid,
+        is_distributed=is_distributed,
+        params=params,
+    )
+
+    # Note: No learning rate scheduler is used here
+    for epoch in range(params.start_epoch, params.num_epochs):
+        if is_distributed:
+            train_dl.sampler.set_epoch(epoch)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if is_distributed:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/shared/convert-k2-to-openfst.py b/icefall/shared/convert-k2-to-openfst.py
new file mode 100755
index 000000000..29a2cd7f7
--- /dev/null
+++ b/icefall/shared/convert-k2-to-openfst.py
@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script takes as input an FST in k2 format and convert it
+to an FST in OpenFST format.
+
+The generated FST is saved into a binary file and its type is
+StdVectorFst.
+
+Usage examples:
+(1) Convert an acceptor
+
+  ./convert-k2-to-openfst.py in.pt binary.fst
+
+(2) Convert a transducer
+
+  ./convert-k2-to-openfst.py --olabels aux_labels in.pt binary.fst
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import kaldifst.utils
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--olabels",
+        type=str,
+        default=None,
+        help="""If not empty, the input FST is assumed to be a transducer
+        and we use its attribute specified by "olabels" as the output labels.
+        """,
+    )
+    parser.add_argument(
+        "input_filename",
+        type=str,
+        help="Path to the input FST in k2 format",
+    )
+
+    parser.add_argument(
+        "output_filename",
+        type=str,
+        help="Path to the output FST in OpenFst format",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    logging.info(f"{vars(args)}")
+
+    input_filename = args.input_filename
+    output_filename = args.output_filename
+    olabels = args.olabels
+
+    if Path(output_filename).is_file():
+        logging.info(f"{output_filename} already exists - skipping")
+        return
+
+    assert Path(input_filename).is_file(), f"{input_filename} does not exist"
+    logging.info(f"Loading {input_filename}")
+    k2_fst = k2.Fsa.from_dict(torch.load(input_filename))
+    if olabels:
+        assert hasattr(k2_fst, olabels), f"No such attribute: {olabels}"
+
+    p = Path(output_filename).parent
+    if not p.is_dir():
+        logging.info(f"Creating {p}")
+        p.mkdir(parents=True)
+
+    logging.info("Converting (May take some time if the input FST is large)")
+    fst = kaldifst.utils.k2_to_openfst(k2_fst, olabels=olabels)
+    logging.info(f"Saving to {output_filename}")
+    fst.write(output_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/shared/make_kn_lm.py b/icefall/shared/make_kn_lm.py
new file mode 100755
index 000000000..231aca7f1
--- /dev/null
+++ b/icefall/shared/make_kn_lm.py
@@ -0,0 +1,442 @@
+#!/usr/bin/env python3
+
+# Copyright 2016  Johns Hopkins University (Author: Daniel Povey)
+#           2018  Ruizhe Huang
+# Apache 2.0.
+
+# This is an implementation of computing Kneser-Ney smoothed language model
+# in the same way as srilm. This is a back-off, unmodified version of
+# Kneser-Ney smoothing, which produces the same results as the following
+# command (as an example) of srilm:
+#
+# $ ngram-count -order 4 -kn-modify-counts-at-end -ukndiscount -gt1min 0 -gt2min 0 -gt3min 0 -gt4min 0 \
+# -text corpus.txt -lm lm.arpa
+#
+# The data structure is based on: kaldi/egs/wsj/s5/utils/lang/make_phone_lm.py
+# The smoothing algorithm is based on: http://www.speech.sri.com/projects/srilm/manpages/ngram-discount.7.html
+
+import argparse
+import io
+import math
+import os
+import re
+import sys
+from collections import Counter, defaultdict
+
+parser = argparse.ArgumentParser(
+    description="""
+    Generate kneser-ney language model as arpa format. By default,
+    it will read the corpus from standard input, and output to standard output.
+    """
+)
+parser.add_argument(
+    "-ngram-order",
+    type=int,
+    default=4,
+    choices=[2, 3, 4, 5, 6, 7],
+    help="Order of n-gram",
+)
+parser.add_argument("-text", type=str, default=None, help="Path to the corpus file")
+parser.add_argument(
+    "-lm", type=str, default=None, help="Path to output arpa file for language models"
+)
+parser.add_argument(
+    "-verbose", type=int, default=0, choices=[0, 1, 2, 3, 4, 5], help="Verbose level"
+)
+args = parser.parse_args()
+
+# For encoding-agnostic scripts, we assume byte stream as input.
+# Need to be very careful about the use of strip() and split()
+# in this case, because there is a latin-1 whitespace character
+# (nbsp) which is part of the unicode encoding range.
+# Ref: kaldi/egs/wsj/s5/utils/lang/bpe/prepend_words.py @ 69cd717
+default_encoding = "latin-1"
+
+strip_chars = " \t\r\n"
+whitespace = re.compile("[ \t]+")
+
+
+class CountsForHistory:
+    # This class (which is more like a struct) stores the counts seen in a
+    # particular history-state.  It is used inside class NgramCounts.
+    # It really does the job of a dict from int to float, but it also
+    # keeps track of the total count.
+    def __init__(self):
+        # The 'lambda: defaultdict(float)' is an anonymous function taking no
+        # arguments that returns a new defaultdict(float).
+        self.word_to_count = defaultdict(int)
+        # using a set to count the number of unique contexts
+        self.word_to_context = defaultdict(set)
+        self.word_to_f = dict()  # discounted probability
+        self.word_to_bow = dict()  # back-off weight
+        self.total_count = 0
+
+    def words(self):
+        return self.word_to_count.keys()
+
+    def __str__(self):
+        # e.g. returns ' total=12: 3->4, 4->6, -1->2'
+        return " total={0}: {1}".format(
+            str(self.total_count),
+            ", ".join(
+                [
+                    "{0} -> {1}".format(word, count)
+                    for word, count in self.word_to_count.items()
+                ]
+            ),
+        )
+
+    def add_count(self, predicted_word, context_word, count):
+        assert count >= 0
+
+        self.total_count += count
+        self.word_to_count[predicted_word] += count
+        if context_word is not None:
+            self.word_to_context[predicted_word].add(context_word)
+
+
+class NgramCounts:
+    # A note on data-structure.  Firstly, all words are represented as
+    # integers.  We store n-gram counts as an array, indexed by (history-length
+    # == n-gram order minus one) (note: python calls arrays "lists") of dicts
+    # from histories to counts, where histories are arrays of integers and
+    # "counts" are dicts from integer to float.  For instance, when
+    # accumulating the 4-gram count for the '8' in the sequence '5 6 7 8', we'd
+    # do as follows: self.counts[3][[5,6,7]][8] += 1.0 where the [3] indexes an
+    # array, the [[5,6,7]] indexes a dict, and the [8] indexes a dict.
+    def __init__(self, ngram_order, bos_symbol="<s>", eos_symbol="</s>"):
+        assert ngram_order >= 2
+
+        self.ngram_order = ngram_order
+        self.bos_symbol = bos_symbol
+        self.eos_symbol = eos_symbol
+
+        self.counts = []
+        for n in range(ngram_order):
+            self.counts.append(defaultdict(lambda: CountsForHistory()))
+
+        self.d = []  # list of discounting factor for each order of ngram
+
+    # adds a raw count (called while processing input data).
+    # Suppose we see the sequence '6 7 8 9' and ngram_order=4, 'history'
+    # would be (6,7,8) and 'predicted_word' would be 9; 'count' would be
+    # 1.
+    def add_count(self, history, predicted_word, context_word, count):
+        self.counts[len(history)][history].add_count(
+            predicted_word, context_word, count
+        )
+
+    # 'line' is a string containing a sequence of integer word-ids.
+    # This function adds the un-smoothed counts from this line of text.
+    def add_raw_counts_from_line(self, line):
+        if line == "":
+            words = [self.bos_symbol, self.eos_symbol]
+        else:
+            words = [self.bos_symbol] + whitespace.split(line) + [self.eos_symbol]
+
+        for i in range(len(words)):
+            for n in range(1, self.ngram_order + 1):
+                if i + n > len(words):
+                    break
+                ngram = words[i : i + n]
+                predicted_word = ngram[-1]
+                history = tuple(ngram[:-1])
+                if i == 0 or n == self.ngram_order:
+                    context_word = None
+                else:
+                    context_word = words[i - 1]
+
+                self.add_count(history, predicted_word, context_word, 1)
+
+    def add_raw_counts_from_standard_input(self):
+        lines_processed = 0
+        # byte stream as input
+        infile = io.TextIOWrapper(sys.stdin.buffer, encoding=default_encoding)
+        for line in infile:
+            line = line.strip(strip_chars)
+            self.add_raw_counts_from_line(line)
+            lines_processed += 1
+        if lines_processed == 0 or args.verbose > 0:
+            print(
+                "make_phone_lm.py: processed {0} lines of input".format(
+                    lines_processed
+                ),
+                file=sys.stderr,
+            )
+
+    def add_raw_counts_from_file(self, filename):
+        lines_processed = 0
+        with open(filename, encoding=default_encoding) as fp:
+            for line in fp:
+                line = line.strip(strip_chars)
+                self.add_raw_counts_from_line(line)
+                lines_processed += 1
+        if lines_processed == 0 or args.verbose > 0:
+            print(
+                "make_phone_lm.py: processed {0} lines of input".format(
+                    lines_processed
+                ),
+                file=sys.stderr,
+            )
+
+    def cal_discounting_constants(self):
+        # For each order N of N-grams, we calculate discounting constant D_N = n1_N / (n1_N + 2 * n2_N),
+        # where n1_N is the number of unique N-grams with count = 1 (counts-of-counts).
+        # This constant is used similarly to absolute discounting.
+        # Return value: d is a list of floats, where d[N+1] = D_N
+
+        # for the lowest order, i.e., 1-gram, we do not need to discount, thus the constant is 0
+        # This is a special case: as we currently assumed having seen all vocabularies in the dictionary,
+        # but perhaps this is not the case for some other scenarios.
+        self.d = [0]
+        for n in range(1, self.ngram_order):
+            this_order_counts = self.counts[n]
+            n1 = 0
+            n2 = 0
+            for hist, counts_for_hist in this_order_counts.items():
+                stat = Counter(counts_for_hist.word_to_count.values())
+                n1 += stat[1]
+                n2 += stat[2]
+            assert n1 + 2 * n2 > 0
+
+            # We are doing this max(0.001, xxx) to avoid zero discounting constant D due to n1=0,
+            # which could happen if the number of symbols is small.
+            # Otherwise, zero discounting constant can cause division by zero in computing BOW.
+            self.d.append(max(0.1, n1 * 1.0) / (n1 + 2 * n2))
+
+    def cal_f(self):
+        # f(a_z) is a probability distribution of word sequence a_z.
+        # Typically f(a_z) is discounted to be less than the ML estimate so we have
+        # some leftover probability for the z words unseen in the context (a_).
+        #
+        # f(a_z) = (c(a_z) - D0) / c(a_)    ;; for highest order N-grams
+        # f(_z)  = (n(*_z) - D1) / n(*_*)	;; for lower order N-grams
+
+        # highest order N-grams
+        n = self.ngram_order - 1
+        this_order_counts = self.counts[n]
+        for hist, counts_for_hist in this_order_counts.items():
+            for w, c in counts_for_hist.word_to_count.items():
+                counts_for_hist.word_to_f[w] = (
+                    max((c - self.d[n]), 0) * 1.0 / counts_for_hist.total_count
+                )
+
+        # lower order N-grams
+        for n in range(0, self.ngram_order - 1):
+            this_order_counts = self.counts[n]
+            for hist, counts_for_hist in this_order_counts.items():
+                n_star_star = 0
+                for w in counts_for_hist.word_to_count.keys():
+                    n_star_star += len(counts_for_hist.word_to_context[w])
+
+                if n_star_star != 0:
+                    for w in counts_for_hist.word_to_count.keys():
+                        n_star_z = len(counts_for_hist.word_to_context[w])
+                        counts_for_hist.word_to_f[w] = (
+                            max((n_star_z - self.d[n]), 0) * 1.0 / n_star_star
+                        )
+                else:  # patterns begin with <s>, they do not have "modified count", so use raw count instead
+                    for w in counts_for_hist.word_to_count.keys():
+                        n_star_z = counts_for_hist.word_to_count[w]
+                        counts_for_hist.word_to_f[w] = (
+                            max((n_star_z - self.d[n]), 0)
+                            * 1.0
+                            / counts_for_hist.total_count
+                        )
+
+    def cal_bow(self):
+        # Backoff weights are only necessary for ngrams which form a prefix of a longer ngram.
+        # Thus, two sorts of ngrams do not have a bow:
+        # 1) highest order ngram
+        # 2) ngrams ending in </s>
+        #
+        # bow(a_) = (1 - Sum_Z1 f(a_z)) / (1 - Sum_Z1 f(_z))
+        # Note that Z1 is the set of all words with c(a_z) > 0
+
+        # highest order N-grams
+        n = self.ngram_order - 1
+        this_order_counts = self.counts[n]
+        for hist, counts_for_hist in this_order_counts.items():
+            for w in counts_for_hist.word_to_count.keys():
+                counts_for_hist.word_to_bow[w] = None
+
+        # lower order N-grams
+        for n in range(0, self.ngram_order - 1):
+            this_order_counts = self.counts[n]
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    if w == self.eos_symbol:
+                        counts_for_hist.word_to_bow[w] = None
+                    else:
+                        a_ = hist + (w,)
+
+                        assert len(a_) < self.ngram_order
+                        assert a_ in self.counts[len(a_)].keys()
+
+                        a_counts_for_hist = self.counts[len(a_)][a_]
+
+                        sum_z1_f_a_z = 0
+                        for u in a_counts_for_hist.word_to_count.keys():
+                            sum_z1_f_a_z += a_counts_for_hist.word_to_f[u]
+
+                        sum_z1_f_z = 0
+                        _ = a_[1:]
+                        _counts_for_hist = self.counts[len(_)][_]
+                        # Should be careful here: what is Z1
+                        for u in a_counts_for_hist.word_to_count.keys():
+                            sum_z1_f_z += _counts_for_hist.word_to_f[u]
+
+                        if sum_z1_f_z < 1:
+                            # assert sum_z1_f_a_z < 1
+                            counts_for_hist.word_to_bow[w] = (1.0 - sum_z1_f_a_z) / (
+                                1.0 - sum_z1_f_z
+                            )
+                        else:
+                            counts_for_hist.word_to_bow[w] = None
+
+    def print_raw_counts(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    res.append(
+                        "{0}\t{1}".format(ngram, counts_for_hist.word_to_count[w])
+                    )
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_modified_counts(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    modified_count = len(counts_for_hist.word_to_context[w])
+                    raw_count = counts_for_hist.word_to_count[w]
+
+                    if modified_count == 0:
+                        res.append("{0}\t{1}".format(ngram, raw_count))
+                    else:
+                        res.append("{0}\t{1}".format(ngram, modified_count))
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_f(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    f = counts_for_hist.word_to_f[w]
+                    if f == 0:  # f(<s>) is always 0
+                        f = 1e-99
+
+                    res.append("{0}\t{1}".format(ngram, math.log(f, 10)))
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_f_and_bow(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    f = counts_for_hist.word_to_f[w]
+                    if f == 0:  # f(<s>) is always 0
+                        f = 1e-99
+
+                    bow = counts_for_hist.word_to_bow[w]
+                    if bow is None:
+                        res.append("{1}\t{0}".format(ngram, math.log(f, 10)))
+                    else:
+                        res.append(
+                            "{1}\t{0}\t{2}".format(
+                                ngram, math.log(f, 10), math.log(bow, 10)
+                            )
+                        )
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_as_arpa(
+        self, fout=io.TextIOWrapper(sys.stdout.buffer, encoding="latin-1")
+    ):
+        # print as ARPA format.
+
+        print("\\data\\", file=fout)
+        for hist_len in range(self.ngram_order):
+            # print the number of n-grams.
+            print(
+                "ngram {0}={1}".format(
+                    hist_len + 1,
+                    sum(
+                        [
+                            len(counts_for_hist.word_to_f)
+                            for counts_for_hist in self.counts[hist_len].values()
+                        ]
+                    ),
+                ),
+                file=fout,
+            )
+
+        print("", file=fout)
+
+        for hist_len in range(self.ngram_order):
+            print("\\{0}-grams:".format(hist_len + 1), file=fout)
+
+            this_order_counts = self.counts[hist_len]
+            for hist, counts_for_hist in this_order_counts.items():
+                for word in counts_for_hist.word_to_count.keys():
+                    ngram = hist + (word,)
+                    prob = counts_for_hist.word_to_f[word]
+                    bow = counts_for_hist.word_to_bow[word]
+
+                    if prob == 0:  # f(<s>) is always 0
+                        prob = 1e-99
+
+                    line = "{0}\t{1}".format("%.7f" % math.log10(prob), " ".join(ngram))
+                    if bow is not None:
+                        line += "\t{0}".format("%.7f" % math.log10(bow))
+                    print(line, file=fout)
+            print("", file=fout)
+        print("\\end\\", file=fout)
+
+
+if __name__ == "__main__":
+    ngram_counts = NgramCounts(args.ngram_order)
+
+    if args.text is None:
+        ngram_counts.add_raw_counts_from_standard_input()
+    else:
+        assert os.path.isfile(args.text)
+        ngram_counts.add_raw_counts_from_file(args.text)
+
+    ngram_counts.cal_discounting_constants()
+    ngram_counts.cal_f()
+    ngram_counts.cal_bow()
+
+    if args.lm is None:
+        ngram_counts.print_as_arpa()
+    else:
+        with open(args.lm, "w", encoding=default_encoding) as f:
+            ngram_counts.print_as_arpa(fout=f)
diff --git a/icefall/shared/ngram_entropy_pruning.py b/icefall/shared/ngram_entropy_pruning.py
new file mode 100755
index 000000000..b1ebee9ea
--- /dev/null
+++ b/icefall/shared/ngram_entropy_pruning.py
@@ -0,0 +1,630 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+#
+# Copyright 2021  Johns Hopkins University (Author: Ruizhe Huang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+./ngram_entropy_pruning.py \
+    -threshold 1e-8 \
+    -lm download/lm/4gram.arpa \
+    -write-lm download/lm/4gram_pruned_1e8.arpa
+
+This file is from Kaldi `egs/wsj/s5/utils/lang/ngram_entropy_pruning.py`.
+This is an implementation of ``Entropy-based Pruning of Backoff Language Models''
+in the same way as SRILM.
+"""
+
+
+import argparse
+import gzip
+import logging
+import math
+import re
+from collections import OrderedDict, defaultdict
+from enum import Enum, unique
+from io import StringIO
+
+parser = argparse.ArgumentParser(
+    description="""
+    Prune an n-gram language model based on the relative entropy 
+    between the original and the pruned model, based on Andreas Stolcke's paper.
+    An n-gram entry is removed, if the removal causes (training set) perplexity 
+    of the model to increase by less than threshold relative.
+    
+    The command takes an arpa file and a pruning threshold as input, 
+    and outputs a pruned arpa file.
+    """
+)
+parser.add_argument("-threshold", type=float, default=1e-6, help="Order of n-gram")
+parser.add_argument("-lm", type=str, default=None, help="Path to the input arpa file")
+parser.add_argument(
+    "-write-lm", type=str, default=None, help="Path to output arpa file after pruning"
+)
+parser.add_argument(
+    "-minorder",
+    type=int,
+    default=1,
+    help="The minorder parameter limits pruning to ngrams of that length and above.",
+)
+parser.add_argument(
+    "-encoding", type=str, default="utf-8", help="Encoding of the arpa file"
+)
+parser.add_argument(
+    "-verbose",
+    type=int,
+    default=2,
+    choices=[0, 1, 2, 3, 4, 5],
+    help="Verbose level, where 0 is most noisy; 5 is most silent",
+)
+args = parser.parse_args()
+
+default_encoding = args.encoding
+logging.basicConfig(
+    format="%(asctime)s — %(levelname)s — %(funcName)s:%(lineno)d — %(message)s",
+    level=args.verbose * 10,
+)
+
+
+class Context(dict):
+    """
+    This class stores data for a context h.
+    It behaves like a python dict object, except that it has several
+    additional attributes.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.log_bo = None
+
+
+class Arpa:
+    """
+    This is a class that implement the data structure of an APRA LM.
+    It (as well as some other classes) is modified based on the library
+    by Stefan Fischer:
+    https://github.com/sfischer13/python-arpa
+    """
+
+    UNK = "<unk>"
+    SOS = "<s>"
+    EOS = "</s>"
+    FLOAT_NDIGITS = 7
+    base = 10
+
+    @staticmethod
+    def _check_input(my_input):
+        if not my_input:
+            raise ValueError
+        elif isinstance(my_input, tuple):
+            return my_input
+        elif isinstance(my_input, list):
+            return tuple(my_input)
+        elif isinstance(my_input, str):
+            return tuple(my_input.strip().split(" "))
+        else:
+            raise ValueError
+
+    @staticmethod
+    def _check_word(input_word):
+        if not isinstance(input_word, str):
+            raise ValueError
+        if " " in input_word:
+            raise ValueError
+
+    def _replace_unks(self, words):
+        return tuple((w if w in self else self._unk) for w in words)
+
+    def __init__(self, path=None, encoding=None, unk=None):
+        self._counts = OrderedDict()
+        self._ngrams = (
+            OrderedDict()
+        )  # Use self._ngrams[len(h)][h][w] for saving the entry of (h,w)
+        self._vocabulary = set()
+        if unk is None:
+            self._unk = self.UNK
+
+        if path is not None:
+            self.loadf(path, encoding)
+
+    def __contains__(self, ngram):
+        h = ngram[:-1]  # h is a tuple
+        w = ngram[-1]  # w is a string/word
+        return h in self._ngrams[len(h)] and w in self._ngrams[len(h)][h]
+
+    def contains_word(self, word):
+        self._check_word(word)
+        return word in self._vocabulary
+
+    def add_count(self, order, count):
+        self._counts[order] = count
+        self._ngrams[order - 1] = defaultdict(Context)
+
+    def update_counts(self):
+        for order in range(1, self.order() + 1):
+            count = sum([len(wlist) for _, wlist in self._ngrams[order - 1].items()])
+            if count > 0:
+                self._counts[order] = count
+
+    def add_entry(self, ngram, p, bo=None, order=None):
+        # Note: ngram is a tuple of strings, e.g. ("w1", "w2", "w3")
+        h = ngram[:-1]  # h is a tuple
+        w = ngram[-1]  # w is a string/word
+
+        # Note that p and bo here are in fact in the log domain (self.base = 10)
+        h_context = self._ngrams[len(h)][h]
+        h_context[w] = p
+        if bo is not None:
+            self._ngrams[len(ngram)][ngram].log_bo = bo
+
+        for word in ngram:
+            self._vocabulary.add(word)
+
+    def counts(self):
+        return sorted(self._counts.items())
+
+    def order(self):
+        return max(self._counts.keys(), default=None)
+
+    def vocabulary(self, sort=True):
+        if sort:
+            return sorted(self._vocabulary)
+        else:
+            return self._vocabulary
+
+    def _entries(self, order):
+        return (
+            self._entry(h, w)
+            for h, wlist in self._ngrams[order - 1].items()
+            for w in wlist
+        )
+
+    def _entry(self, h, w):
+        # return the entry for the ngram (h, w)
+        ngram = h + (w,)
+        log_p = self._ngrams[len(h)][h][w]
+        log_bo = self._log_bo(ngram)
+        if log_bo is not None:
+            return (
+                round(log_p, self.FLOAT_NDIGITS),
+                ngram,
+                round(log_bo, self.FLOAT_NDIGITS),
+            )
+        else:
+            return round(log_p, self.FLOAT_NDIGITS), ngram
+
+    def _log_bo(self, ngram):
+        if len(ngram) in self._ngrams and ngram in self._ngrams[len(ngram)]:
+            return self._ngrams[len(ngram)][ngram].log_bo
+        else:
+            return None
+
+    def _log_p(self, ngram):
+        h = ngram[:-1]  # h is a tuple
+        w = ngram[-1]  # w is a string/word
+        if h in self._ngrams[len(h)] and w in self._ngrams[len(h)][h]:
+            return self._ngrams[len(h)][h][w]
+        else:
+            return None
+
+    def log_p_raw(self, ngram):
+        log_p = self._log_p(ngram)
+        if log_p is not None:
+            return log_p
+        else:
+            if len(ngram) == 1:
+                raise KeyError
+            else:
+                log_bo = self._log_bo(ngram[:-1])
+                if log_bo is None:
+                    log_bo = 0
+                return log_bo + self.log_p_raw(ngram[1:])
+
+    def log_joint_prob(self, sequence):
+        # Compute the joint prob of the sequence based on the chain rule
+        # Note that sequence should be a tuple of strings
+        #
+        # Reference:
+        # https://github.com/BitSpeech/SRILM/blob/d571a4424fb0cf08b29fbfccfddd092ea969eae3/lm/src/LM.cc#L527
+
+        log_joint_p = 0
+        seq = sequence
+        while len(seq) > 0:
+            log_joint_p += self.log_p_raw(seq)
+            seq = seq[:-1]
+
+            # If we're computing the marginal probability of the unigram
+            # <s> context we have to look up </s> instead since the former
+            # has prob = 0.
+            if len(seq) == 1 and seq[0] == self.SOS:
+                seq = (self.EOS,)
+
+        return log_joint_p
+
+    def set_new_context(self, h):
+        old_context = self._ngrams[len(h)][h]
+        self._ngrams[len(h)][h] = Context()
+        return old_context
+
+    def log_p(self, ngram):
+        words = self._check_input(ngram)
+        if self._unk:
+            words = self._replace_unks(words)
+        return self.log_p_raw(words)
+
+    def log_s(self, sentence, sos=SOS, eos=EOS):
+        words = self._check_input(sentence)
+        if self._unk:
+            words = self._replace_unks(words)
+        if sos:
+            words = (sos,) + words
+        if eos:
+            words = words + (eos,)
+        result = sum(self.log_p_raw(words[:i]) for i in range(1, len(words) + 1))
+        if sos:
+            result = result - self.log_p_raw(words[:1])
+        return result
+
+    def p(self, ngram):
+        return self.base ** self.log_p(ngram)
+
+    def s(self, sentence):
+        return self.base ** self.log_s(sentence)
+
+    def write(self, fp):
+        fp.write("\n\\data\\\n")
+        for order, count in self.counts():
+            fp.write("ngram {}={}\n".format(order, count))
+        fp.write("\n")
+        for order, _ in self.counts():
+            fp.write("\\{}-grams:\n".format(order))
+            for e in self._entries(order):
+                prob = e[0]
+                ngram = " ".join(e[1])
+                if len(e) == 2:
+                    fp.write("{}\t{}\n".format(prob, ngram))
+                elif len(e) == 3:
+                    backoff = e[2]
+                    fp.write("{}\t{}\t{}\n".format(prob, ngram, backoff))
+                else:
+                    raise ValueError
+            fp.write("\n")
+        fp.write("\\end\\\n")
+
+
+class ArpaParser:
+    """
+    This is a class that implement a parser of an arpa file
+    """
+
+    @unique
+    class State(Enum):
+        DATA = 1
+        COUNT = 2
+        HEADER = 3
+        ENTRY = 4
+
+    re_count = re.compile(r"^ngram (\d+)=(\d+)$")
+    re_header = re.compile(r"^\\(\d+)-grams:$")
+    re_entry = re.compile(
+        "^(-?\\d+(\\.\\d+)?([eE]-?\\d+)?)"
+        "\t"
+        "(\\S+( \\S+)*)"
+        "(\t((-?\\d+(\\.\\d+)?)([eE]-?\\d+)?))?$"
+    )
+
+    def _parse(self, fp):
+        self._result = []
+        self._state = self.State.DATA
+        self._tmp_model = None
+        self._tmp_order = None
+        for line in fp:
+            line = line.strip()
+            if self._state == self.State.DATA:
+                self._data(line)
+            elif self._state == self.State.COUNT:
+                self._count(line)
+            elif self._state == self.State.HEADER:
+                self._header(line)
+            elif self._state == self.State.ENTRY:
+                self._entry(line)
+        if self._state != self.State.DATA:
+            raise Exception(line)
+        return self._result
+
+    def _data(self, line):
+        if line == "\\data\\":
+            self._state = self.State.COUNT
+            self._tmp_model = Arpa()
+        else:
+            pass  # skip comment line
+
+    def _count(self, line):
+        match = self.re_count.match(line)
+        if match:
+            order = match.group(1)
+            count = match.group(2)
+            self._tmp_model.add_count(int(order), int(count))
+        elif not line:
+            self._state = self.State.HEADER  # there are no counts
+        else:
+            raise Exception(line)
+
+    def _header(self, line):
+        match = self.re_header.match(line)
+        if match:
+            self._state = self.State.ENTRY
+            self._tmp_order = int(match.group(1))
+        elif line == "\\end\\":
+            self._result.append(self._tmp_model)
+            self._state = self.State.DATA
+            self._tmp_model = None
+            self._tmp_order = None
+        elif not line:
+            pass  # skip empty line
+        else:
+            raise Exception(line)
+
+    def _entry(self, line):
+        match = self.re_entry.match(line)
+        if match:
+            p = self._float_or_int(match.group(1))
+            ngram = tuple(match.group(4).split(" "))
+            bo_match = match.group(7)
+            bo = self._float_or_int(bo_match) if bo_match else None
+            self._tmp_model.add_entry(ngram, p, bo, self._tmp_order)
+        elif not line:
+            self._state = self.State.HEADER  # last entry
+        else:
+            raise Exception(line)
+
+    @staticmethod
+    def _float_or_int(s):
+        f = float(s)
+        i = int(f)
+        if str(i) == s:  # don't drop trailing ".0"
+            return i
+        else:
+            return f
+
+    def load(self, fp):
+        """Deserialize fp (a file-like object) to a Python object."""
+        return self._parse(fp)
+
+    def loadf(self, path, encoding=None):
+        """Deserialize path (.arpa, .gz) to a Python object."""
+        path = str(path)
+        if path.endswith(".gz"):
+            with gzip.open(path, mode="rt", encoding=encoding) as f:
+                return self.load(f)
+        else:
+            with open(path, mode="rt", encoding=encoding) as f:
+                return self.load(f)
+
+    def loads(self, s):
+        """Deserialize s (a str) to a Python object."""
+        with StringIO(s) as f:
+            return self.load(f)
+
+    def dump(self, obj, fp):
+        """Serialize obj to fp (a file-like object) in ARPA format."""
+        obj.write(fp)
+
+    def dumpf(self, obj, path, encoding=None):
+        """Serialize obj to path in ARPA format (.arpa, .gz)."""
+        path = str(path)
+        if path.endswith(".gz"):
+            with gzip.open(path, mode="wt", encoding=encoding) as f:
+                return self.dump(obj, f)
+        else:
+            with open(path, mode="wt", encoding=encoding) as f:
+                self.dump(obj, f)
+
+    def dumps(self, obj):
+        """Serialize obj to an ARPA formatted str."""
+        with StringIO() as f:
+            self.dump(obj, f)
+            return f.getvalue()
+
+
+def add_log_p(prev_log_sum, log_p, base):
+    return math.log(base**log_p + base**prev_log_sum, base)
+
+
+def compute_numerator_denominator(lm, h):
+    log_sum_seen_h = -math.inf
+    log_sum_seen_h_lower = -math.inf
+    base = lm.base
+    for w, log_p in lm._ngrams[len(h)][h].items():
+        log_sum_seen_h = add_log_p(log_sum_seen_h, log_p, base)
+
+        ngram = h + (w,)
+        log_p_lower = lm.log_p_raw(ngram[1:])
+        log_sum_seen_h_lower = add_log_p(log_sum_seen_h_lower, log_p_lower, base)
+
+    numerator = 1.0 - base**log_sum_seen_h
+    denominator = 1.0 - base**log_sum_seen_h_lower
+    return numerator, denominator
+
+
+def prune(lm, threshold, minorder):
+    # Reference:
+    # https://github.com/BitSpeech/SRILM/blob/d571a4424fb0cf08b29fbfccfddd092ea969eae3/lm/src/NgramLM.cc#L2330
+
+    for i in range(
+        lm.order(), max(minorder - 1, 1), -1
+    ):  # i is the order of the ngram (h, w)
+        logging.info("processing %d-grams ..." % i)
+        count_pruned_ngrams = 0
+
+        h_dict = lm._ngrams[i - 1]
+        for h in list(h_dict.keys()):
+            # old backoff weight, BOW(h)
+            log_bow = lm._log_bo(h)
+            if log_bow is None:
+                log_bow = 0
+
+            # Compute numerator and denominator of the backoff weight,
+            # so that we can quickly compute the BOW adjustment due to
+            # leaving out one prob.
+            numerator, denominator = compute_numerator_denominator(lm, h)
+
+            # assert abs(math.log(numerator, lm.base) - math.log(denominator, lm.base) - h_dict[h].log_bo) < 1e-5
+
+            # Compute the marginal probability of the context, P(h)
+            h_log_p = lm.log_joint_prob(h)
+
+            all_pruned = True
+            pruned_w_set = set()
+
+            for w, log_p in h_dict[h].items():
+                ngram = h + (w,)
+
+                # lower-order estimate for ngramProb, P(w|h')
+                backoff_prob = lm.log_p_raw(ngram[1:])
+
+                # Compute BOW after removing ngram, BOW'(h)
+                new_log_bow = math.log(
+                    numerator + lm.base**log_p, lm.base
+                ) - math.log(denominator + lm.base**backoff_prob, lm.base)
+
+                # Compute change in entropy due to removal of ngram
+                delta_prob = backoff_prob + new_log_bow - log_p
+                delta_entropy = -(lm.base**h_log_p) * (
+                    (lm.base**log_p) * delta_prob
+                    + numerator * (new_log_bow - log_bow)
+                )
+
+                # compute relative change in model (training set) perplexity
+                perp_change = lm.base**delta_entropy - 1.0
+
+                pruned = threshold > 0 and perp_change < threshold
+
+                # Make sure we don't prune ngrams whose backoff nodes are needed
+                if (
+                    pruned
+                    and len(ngram) in lm._ngrams
+                    and len(lm._ngrams[len(ngram)][ngram]) > 0
+                ):
+                    pruned = False
+
+                logging.debug(
+                    "CONTEXT "
+                    + str(h)
+                    + " WORD "
+                    + w
+                    + " CONTEXTPROB %f " % h_log_p
+                    + " OLDPROB %f " % log_p
+                    + " NEWPROB %f " % (backoff_prob + new_log_bow)
+                    + " DELTA-H %f " % delta_entropy
+                    + " DELTA-LOGP %f " % delta_prob
+                    + " PPL-CHANGE %f " % perp_change
+                    + " PRUNED "
+                    + str(pruned)
+                )
+
+                if pruned:
+                    pruned_w_set.add(w)
+                    count_pruned_ngrams += 1
+                else:
+                    all_pruned = False
+
+            # If we removed all ngrams for this context we can
+            # remove the context itself, but only if the present
+            # context is not a prefix to a longer one.
+            if all_pruned and len(pruned_w_set) == len(h_dict[h]):
+                del h_dict[
+                    h
+                ]  # this context h is no longer needed, as its ngram prob is stored at its own context h'
+            elif len(pruned_w_set) > 0:
+                # The pruning for this context h is actually done here
+                old_context = lm.set_new_context(h)
+
+                for w, p_w in old_context.items():
+                    if w not in pruned_w_set:
+                        lm.add_entry(
+                            h + (w,), p_w
+                        )  # the entry hw is stored at the context h
+
+                # We need to recompute the back-off weight, but
+                # this can only be done after completing the pruning
+                # of the lower-order ngrams.
+                # Reference:
+                # https://github.com/BitSpeech/SRILM/blob/d571a4424fb0cf08b29fbfccfddd092ea969eae3/flm/src/FNgramLM.cc#L2124
+
+        logging.info("pruned %d %d-grams" % (count_pruned_ngrams, i))
+
+    # recompute backoff weights
+    for i in range(
+        max(minorder - 1, 1) + 1, lm.order() + 1
+    ):  # be careful of this order: from low- to high-order
+        for h in lm._ngrams[i - 1]:
+            numerator, denominator = compute_numerator_denominator(lm, h)
+            new_log_bow = math.log(numerator, lm.base) - math.log(denominator, lm.base)
+            lm._ngrams[len(h)][h].log_bo = new_log_bow
+
+    # update counts
+    lm.update_counts()
+
+    return
+
+
+def check_h_is_valid(lm, h):
+    sum_under_h = sum(
+        [lm.base ** lm.log_p_raw(h + (w,)) for w in lm.vocabulary(sort=False)]
+    )
+    if abs(sum_under_h - 1.0) > 1e-6:
+        logging.info("warning: %s %f" % (str(h), sum_under_h))
+        return False
+    else:
+        return True
+
+
+def validate_lm(lm):
+    # sanity check if the conditional probability sums to one under each context h
+    for i in range(lm.order(), 0, -1):  # i is the order of the ngram (h, w)
+        logging.info("validating %d-grams ..." % i)
+        h_dict = lm._ngrams[i - 1]
+        for h in h_dict.keys():
+            check_h_is_valid(lm, h)
+
+
+def compare_two_apras(path1, path2):
+    pass
+
+
+if __name__ == "__main__":
+    # load an arpa file
+    logging.info("Loading the arpa file from %s" % args.lm)
+    parser = ArpaParser()
+    models = parser.loadf(args.lm, encoding=default_encoding)
+    lm = models[0]  # ARPA files may contain several models.
+    logging.info("Stats before pruning:")
+    for i, cnt in lm.counts():
+        logging.info("ngram %d=%d" % (i, cnt))
+
+    # prune it, the language model will be modified in-place
+    logging.info("Start pruning the model with threshold=%.3E..." % args.threshold)
+    prune(lm, args.threshold, args.minorder)
+
+    # validate_lm(lm)
+
+    # write the arpa language model to a file
+    logging.info("Stats after pruning:")
+    for i, cnt in lm.counts():
+        logging.info("ngram %d=%d" % (i, cnt))
+    logging.info("Saving the pruned arpa file to %s" % args.write_lm)
+    parser.dumpf(lm, args.write_lm, encoding=default_encoding)
+    logging.info("Done.")
diff --git a/icefall/shared/parse_options.sh b/icefall/shared/parse_options.sh
new file mode 100755
index 000000000..71fb9e5ea
--- /dev/null
+++ b/icefall/shared/parse_options.sh
@@ -0,0 +1,97 @@
+#!/usr/bin/env bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### Now we process the command line options
+###
+while true; do
+  [ -z "${1:-}" ] && break;  # break if there are no arguments
+  case "$1" in
+    # If the enclosing script is called with --help option, print the help
+    # message and exit.  Scripts should put help messages in $help_message
+    --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
+      else printf "$help_message\n" 1>&2 ; fi;
+      exit 0 ;;
+    --*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
+      exit 1 ;;
+    # If the first command-line argument begins with "--" (e.g. --foo-bar),
+    # then work out the variable name as $name, which will equal "foo_bar".
+    --*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
+      # Next we test whether the variable in question is undefned-- if so it's
+      # an invalid option and we die.  Note: $0 evaluates to the name of the
+      # enclosing script.
+      # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
+      # is undefined.  We then have to wrap this test inside "eval" because
+      # foo_bar is itself inside a variable ($name).
+      eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+
+      oldval="`eval echo \\$$name`";
+      # Work out whether we seem to be expecting a Boolean argument.
+      if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      eval $name=\"$2\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
+        exit 1;
+      fi
+      shift 2;
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.
diff --git a/icefall/transformer_lm/__init__.py b/icefall/transformer_lm/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/icefall/transformer_lm/attention.py b/icefall/transformer_lm/attention.py
new file mode 100644
index 000000000..5ce83b15e
--- /dev/null
+++ b/icefall/transformer_lm/attention.py
@@ -0,0 +1,510 @@
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+
+from icefall.transformer_lm.scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from icefall.utils import is_jit_tracing
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = False,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            left_context=left_context,
+        )
+
+    def rel_shift(self, x: Tensor, left_context: int = 0) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1+left_context).
+                time1 means the length of query vector.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+
+        time2 = time1 + left_context
+        if not is_jit_tracing():
+            assert (
+                n == left_context + 2 * time1 - 1
+            ), f"{n} == {left_context} + 2 * {time1} - 1"
+
+        if is_jit_tracing():
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(time2)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, time1, time2)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time1_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, time1, time2),
+                (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+                storage_offset=n_stride * (time1 - 1),
+            )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = False,
+        attn_mask: Optional[Tensor] = None,
+        left_context: int = 0,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        if not is_jit_tracing():
+            assert embed_dim == embed_dim_to_check
+            assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        if not is_jit_tracing():
+            assert (
+                head_dim * num_heads == embed_dim
+            ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None and not is_jit_tracing():
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        if not is_jit_tracing():
+            assert pos_emb_bsz in (1, bsz)  # actually it is 1
+
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1)
+        p = p.permute(0, 2, 3, 1)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(q_with_bias_v, p)  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd, left_context)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        if not is_jit_tracing():
+            assert list(attn_output_weights.size()) == [
+                bsz * num_heads,
+                tgt_len,
+                src_len,
+            ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+
+        # If we are using dynamic_chunk_training and setting a limited
+        # num_left_chunks, the attention may only see the padding values which
+        # will also be masked out by `key_padding_mask`, at this circumstances,
+        # the whole column of `attn_output_weights` will be `-inf`
+        # (i.e. be `nan` after softmax), so, we fill `0.0` at the masking
+        # positions to avoid invalid loss value below.
+        if (
+            attn_mask is not None
+            and attn_mask.dtype == torch.bool
+            and key_padding_mask is not None
+        ):
+            if attn_mask.size(0) != 1:
+                attn_mask = attn_mask.view(bsz, num_heads, tgt_len, src_len)
+                combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2)
+            else:
+                # attn_mask.shape == (1, tgt_len, src_len)
+                combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze(
+                    1
+                ).unsqueeze(2)
+
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0)
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+
+        if not is_jit_tracing():
+            assert list(attn_output.size()) == [
+                bsz * num_heads,
+                tgt_len,
+                head_dim,
+            ]
+
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
diff --git a/icefall/transformer_lm/compute_perplexity.py b/icefall/transformer_lm/compute_perplexity.py
new file mode 100644
index 000000000..72d7c477b
--- /dev/null
+++ b/icefall/transformer_lm/compute_perplexity.py
@@ -0,0 +1,195 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import math
+from pathlib import Path
+
+import torch
+from dataset import get_dataloader
+from train import get_params
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.transformer_lm.model import TransformerLM
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=7,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transformer_lm/exp_full_libri_16layer_maxlen200_8gpu",
+    )
+
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        help="Path to the LM test data for computing perplexity",
+        default="transformer_lm/libri_lm_training_bpe500/sorted_lm_data-test.pt",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=16,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=50,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--max-sent-len",
+        type=int,
+        default=100,
+        help="Number of RNN layers the model",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lm_data = Path(args.lm_data)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-ppl/")
+    logging.info("Computing perplexity started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    logging.info("About to create model")
+    model = TransformerLM(
+        vocab_size=params.vocab_size,
+        d_model=params.encoder_dim,
+        embedding_dim=params.embedding_dim,
+        dim_feedforward=params.dim_feedforward,
+        nhead=params.nhead,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+        params=params,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        model.to(device)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    num_param_requires_grad = sum(
+        [p.numel() for p in model.parameters() if p.requires_grad]
+    )
+
+    logging.info(f"Number of model parameters: {num_param}")
+    logging.info(
+        f"Number of model parameters (requires_grad): "
+        f"{num_param_requires_grad} "
+        f"({num_param_requires_grad/num_param_requires_grad*100}%)"
+    )
+
+    logging.info(f"Loading LM test data from {params.lm_data}")
+    test_dl = get_dataloader(
+        filename=params.lm_data,
+        is_distributed=False,
+        params=params,
+    )
+
+    tot_loss = 0.0
+    num_tokens = 0
+    num_sentences = 0
+    for batch_idx, batch in enumerate(test_dl):
+        x, y, sentence_lengths = batch
+        x = x.to(device)
+        y = y.to(device)
+        sentence_lengths = sentence_lengths.to(device)
+
+        nll = model(x, y, sentence_lengths)
+        loss = nll.sum().cpu().item()
+
+        tot_loss += loss
+        num_tokens += sentence_lengths.sum().cpu().item()
+        num_sentences += x.size(0)
+
+    ppl = math.exp(tot_loss / num_tokens)
+    logging.info(
+        f"total nll: {tot_loss}, num tokens: {num_tokens}, "
+        f"num sentences: {num_sentences}, ppl: {ppl:.3f}"
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/transformer_lm/dataset.py b/icefall/transformer_lm/dataset.py
new file mode 120000
index 000000000..5792a6cf0
--- /dev/null
+++ b/icefall/transformer_lm/dataset.py
@@ -0,0 +1 @@
+../rnn_lm/dataset.py
\ No newline at end of file
diff --git a/icefall/transformer_lm/encoder.py b/icefall/transformer_lm/encoder.py
new file mode 100644
index 000000000..4357b83d7
--- /dev/null
+++ b/icefall/transformer_lm/encoder.py
@@ -0,0 +1,329 @@
+# Copyright (c)  2021  Xiaomi Corporation (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+from icefall.transformer_lm.attention import RelPositionMultiheadAttention
+from icefall.transformer_lm.scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from icefall.utils import is_jit_tracing, make_pad_mask
+
+
+class Transformer(torch.nn.Module):
+    """_summary_
+
+    Args:
+        input_dim (int): Input feature dimension
+        d_mode (int): The dimension of the transformer
+        dim_feedforward (int ): The dimension of the ffw module
+        nhead (int): The number of attention heads
+        dropout_rate (float): dropout rate
+        att_dropout (float): dropout rate in attention module
+    """
+
+    def __init__(
+        self,
+        input_dim: int,
+        d_model: int,
+        dim_feedforward: int,
+        nhead: int = 4,
+        num_layers: int = 6,
+        dropout_rate: float = 0.1,
+        att_dropout: float = 0.0,
+    ):
+        super().__init__()
+
+        self.encoder_layers = num_layers
+        self.d_model = d_model
+
+        self.embed = ScaledLinear(input_dim, d_model)
+        self.norm_before = BasicNorm(d_model, learn_eps=False)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout_rate)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            dim_feedforward=dim_feedforward,
+            nhead=nhead,
+            dropout_rate=dropout_rate,
+        )
+
+        self.encoder = TransformerEncoder(encoder_layer, num_layers)
+
+    def _create_attention_mask(self, x_lens: torch.Tensor):
+        # create a 2D attention mask to mask out
+        # the upper right half of the attention matrix
+        max_len = max(x_lens)
+        ones = torch.ones(max_len, max_len, device=x_lens.device, dtype=torch.bool)
+        return torch.triu(ones, diagonal=1)
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Transformer forward
+
+        Args:
+            x (torch.Tensor): Input tensor (B,T,input_dim)
+            x_lens (torch.Tensor): The length of input tensors before padding (B,)
+
+        Returns:
+            Return a tuple of 2 tensors:
+            - x: output feature of the transformer (B,T,d_model)
+            - x_lens: output feature lens of the transformer
+        """
+
+        attention_mask = self._create_attention_mask(x_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+
+        x = self.norm_before(self.embed(x))
+
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)
+
+        x = self.encoder(
+            x,
+            pos_emb,
+            mask=attention_mask,  # pass the attention mast
+            src_key_padding_mask=src_key_padding_mask,
+        )  # (T, N, C)
+
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        return x, x_lens
+
+
+class TransformerEncoder(torch.nn.Module):
+    def __init__(self, encoder_layer: torch.nn.Module, num_layers: int) -> None:
+        """TransformerEncoder is a stack of N encoder layers
+
+        Args:
+            encoder_layer (torch.nn.Module): an instance of the TransformerEncoderLayer()
+            num_layers (int): Number of layers to be stacked
+        """
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        pos_emb: torch.Tensor,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """_summary_
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            output: transformer encoded features
+        """
+        output = src
+
+        for layer_index, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_key_padding_mask=src_key_padding_mask,
+                src_mask=mask,
+            )
+
+        return output
+
+
+class TransformerEncoderLayer(torch.nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        dim_feedforward: int,
+        nhead: int,
+        dropout_rate: float,
+    ):
+        """TransformerEncoderLayer is made up of self-attn and feedforward module
+
+        Args:
+            d_model (int): The model size
+            dim_feedforward (int): Dimension of ffw module
+            nhead (int): Number of heads
+            dropout_rate (float): Dropout rate
+        """
+        super().__init__()
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout_rate),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout_rate)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        pos_emb: torch.Tensor,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        src_mask: Optional[torch.Tensor] = None,
+        cache=None,
+    ):
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            src_mask: the mask for the src sequence (optional).
+        """
+        src_orig = src
+
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+
+        src = src + self.dropout(src_att)
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        return src
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        if is_jit_tracing():
+            # 10k frames correspond to ~100k ms, e.g., 100 seconds, i.e.,
+            # It assumes that the maximum input won't have more than
+            # 10k frames.
+            #
+            # TODO(fangjun): Use torch.jit.script() for this module
+            max_len = 10000
+
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: torch.Tensor, left_context: int = 0) -> None:
+        """Reset the positional encodings."""
+        x_size_1 = x.size(1) + left_context
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x_size_1 * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vector and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x_size_1, self.d_model)
+        pe_negative = torch.zeros(x_size_1, self.d_model)
+        position = torch.arange(0, x_size_1, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        left_context: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+            left_context (int): left context (in frames) used during streaming decoding.
+                this is used only in real streaming decoding, in other circumstances,
+                it MUST be 0.
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x, left_context)
+        x_size_1 = x.size(1) + left_context
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x_size_1
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
diff --git a/icefall/transformer_lm/export.py b/icefall/transformer_lm/export.py
new file mode 100644
index 000000000..c08982e37
--- /dev/null
+++ b/icefall/transformer_lm/export.py
@@ -0,0 +1,186 @@
+#!/usr/bin/env python3
+# Copyright (c)  2022  Xiaomi Corporation (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from model import TransformerLM
+
+from icefall.checkpoint import load_checkpoint
+from icefall.utils import AttributeDict, load_averaged_model, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=11,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=768,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=768,
+        help="Encoder dim of the model",
+    )
+
+    parser.add_argument(
+        "--dim_feedforward",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=16,
+        help="Number of Transformer layers",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=True,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = AttributeDict({})
+    params.update(vars(args))
+
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("About to create model")
+    model = TransformerLM(
+        vocab_size=params.vocab_size,
+        d_model=params.encoder_dim,
+        embedding_dim=params.embedding_dim,
+        dim_feedforward=params.dim_feedforward,
+        nhead=params.nhead,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+        params=params,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        model = load_averaged_model(
+            params.exp_dir, model, params.epoch, params.avg, device
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/transformer_lm/model.py b/icefall/transformer_lm/model.py
new file mode 100644
index 000000000..c78cf1821
--- /dev/null
+++ b/icefall/transformer_lm/model.py
@@ -0,0 +1,114 @@
+# Copyright (c)  2022  Xiaomi Corporation (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+
+from icefall.transformer_lm.encoder import Transformer
+from icefall.utils import AttributeDict, add_eos, add_sos, make_pad_mask
+
+
+class TransformerLM(torch.nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        d_model: int,
+        dim_feedforward: int,
+        nhead: int = 8,
+        num_layers: int = 16,
+        tie_weights: bool = True,
+        dropout: float = 0.1,
+        emb_dropout_rate: float = 0.0,
+        params: AttributeDict = None,
+    ):
+        super().__init__()
+
+        self.vocab_size = vocab_size
+        self.params = params
+
+        self.input_embedding = torch.nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+        )
+
+        self.encoder = Transformer(
+            input_dim=embedding_dim,
+            d_model=d_model,
+            dim_feedforward=dim_feedforward,
+            nhead=nhead,
+            num_layers=num_layers,
+            dropout_rate=dropout,
+        )
+
+        self.output_linear = torch.nn.Linear(
+            in_features=d_model, out_features=vocab_size
+        )
+        if tie_weights:
+            logging.info("Tying weights")
+            assert d_model == embedding_dim, (d_model, embedding_dim)
+            self.output_linear.weight = self.input_embedding.weight
+        else:
+            logging.info("Not tying weights")
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        x_lens: torch.Tensor,
+        return_logits: bool = False,
+    ):
+        """Forward transformer language model
+
+        Args:
+            x (torch.Tensor): Input tokens (B,L)
+            y (torch.Tensor): Output tokens (with EOS appended) (B,L)
+            x_lens (torch.Tensor): Length of input tokens before padding (B,)
+            return_logits (bool, optional): Return logits instead of NLL
+
+        """
+
+        x = self.input_embedding(x)
+
+        x, x_lens = self.encoder(x, x_lens)
+
+        logits = self.output_linear(x)
+
+        if return_logits:
+            return logits
+
+        nll_loss = F.cross_entropy(
+            logits.reshape(-1, self.vocab_size), y.reshape(-1), reduction="none"
+        )
+
+        mask = make_pad_mask(x_lens).reshape(-1)
+        nll_loss.masked_fill_(mask, 0)
+
+        return nll_loss
+
+    def score_token(self, x: torch.Tensor, x_lens: torch.Tensor, state=None):
+        bs = x.size(0)
+
+        state = None
+        logits = self.forward(x, x, x_lens, return_logits=True)
+        index = torch.arange(bs)
+
+        last_logits = logits[index, x_lens - 1, :]
+
+        return last_logits.log_softmax(-1), state
diff --git a/icefall/transformer_lm/scaling.py b/icefall/transformer_lm/scaling.py
new file mode 120000
index 000000000..0876c0704
--- /dev/null
+++ b/icefall/transformer_lm/scaling.py
@@ -0,0 +1 @@
+../../egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/icefall/transformer_lm/train.py b/icefall/transformer_lm/train.py
new file mode 100644
index 000000000..c36abfcdf
--- /dev/null
+++ b/icefall/transformer_lm/train.py
@@ -0,0 +1,609 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+Usage:
+    ./transformer_lm/train.py \
+        --start-epoch 0 \
+        --world-size 2 \
+        --num-epochs 1 \
+        --use-fp16 0 \
+        --num-layers 12 \
+        --batch-size 400
+
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from dataset import get_dataloader
+from lhotse.utils import fix_random_seed
+from model import TransformerLM
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        exp_dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transformer_lm/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, logs, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=True,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=400,
+    )
+
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        default="data/lm_training_bpe_500/sorted_lm_data.pt",
+        help="LM training data",
+    )
+
+    parser.add_argument(
+        "--lm-data-valid",
+        type=str,
+        default="data/lm_training_bpe_500/sorted_lm_data-valid.pt",
+        help="LM validation data",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=12,
+        help="Number of Transformer layers in the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=True,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters."""
+
+    params = AttributeDict(
+        {
+            "max_sent_len": 200,
+            "sos_id": 1,
+            "eos_id": 1,
+            "blank_id": 0,
+            "lr": 1e-3,
+            "weight_decay": 1e-6,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 200,
+            "reset_interval": 2000,
+            "valid_interval": 1000,
+            "nhead": 8,
+            "embedding_dim": 768,
+            "encoder_dim": 768,
+            "dim_feedforward": 2048,
+            "dropout": 0.1,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    logging.info(f"Loading checkpoint: {filename}")
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    model: nn.Module,
+    x: torch.Tensor,
+    y: torch.Tensor,
+    sentence_lengths: torch.Tensor,
+    is_training: bool,
+) -> Tuple[torch.Tensor, MetricsTracker]:
+    """Compute the negative log-likelihood loss given a model and its input.
+    Args:
+      model:
+        The NN model,
+      x:
+        A 2-D tensor. Each row contains BPE token IDs for a sentence. Also,
+        each row starts with SOS ID.
+      y:
+        A 2-D tensor. Each row is a shifted version of the corresponding row
+        in `x` but ends with an EOS ID (before padding).
+     sentence_lengths:
+       A 1-D tensor containing number of tokens of each sentence
+       before padding.
+     is_training:
+       True for training. False for validation.
+    """
+    with torch.set_grad_enabled(is_training):
+        device = model.device
+        x = x.to(device)
+        y = y.to(device)
+        sentence_lengths = sentence_lengths.to(device)
+
+        nll = model(x, y, sentence_lengths)
+        loss = nll.sum()
+
+        num_tokens = sentence_lengths.sum().item()
+
+        loss_info = MetricsTracker()
+        # Note: Due to how MetricsTracker() is designed,
+        # we use "frames" instead of "num_tokens" as a key here
+        loss_info["frames"] = num_tokens
+        loss_info["loss"] = loss.detach().item()
+    return loss, loss_info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        x, y, sentence_lengths = batch
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                model=model,
+                x=x,
+                y=y,
+                sentence_lengths=sentence_lengths,
+                is_training=False,
+            )
+
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all sentences is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        x, y, sentence_lengths = batch
+        batch_size = x.size(0)
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                model=model,
+                x=x,
+                y=y,
+                sentence_lengths=sentence_lengths,
+                is_training=True,
+            )
+
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            # Note: "frames" here means "num_tokens"
+            this_batch_ppl = math.exp(loss_info["loss"] / loss_info["frames"])
+            tot_ppl = math.exp(tot_loss["loss"] / tot_loss["frames"])
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}, ppl: {this_batch_ppl}] "
+                f"tot_loss[{tot_loss}, ppl: {tot_ppl}], "
+                f"batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+                tb_writer.add_scalar(
+                    "train/current_ppl", this_batch_ppl, params.batch_idx_train
+                )
+
+                tb_writer.add_scalar("train/tot_ppl", tot_ppl, params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+
+            valid_ppl = math.exp(valid_info["loss"] / valid_info["frames"])
+            logging.info(
+                f"Epoch {params.cur_epoch}, validation: {valid_info}, "
+                f"ppl: {valid_ppl}"
+            )
+
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+                tb_writer.add_scalar(
+                    "train/valid_ppl", valid_ppl, params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    is_distributed = world_size > 1
+
+    fix_random_seed(params.seed)
+    if is_distributed:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    logging.info(f"Device: {device}")
+
+    logging.info("About to create model")
+    model = TransformerLM(
+        vocab_size=params.vocab_size,
+        d_model=params.encoder_dim,
+        embedding_dim=params.embedding_dim,
+        dim_feedforward=params.dim_feedforward,
+        nhead=params.nhead,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+        params=params,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if is_distributed:
+        model = DDP(model, device_ids=[rank])
+
+    model.device = device
+
+    optimizer = optim.Adam(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    if checkpoints:
+        logging.info("Load optimizer state_dict from checkpoint")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    logging.info(f"Loading LM training data from {params.lm_data}")
+    train_dl = get_dataloader(
+        filename=params.lm_data,
+        is_distributed=is_distributed,
+        params=params,
+    )
+
+    logging.info(f"Loading LM validation data from {params.lm_data_valid}")
+    valid_dl = get_dataloader(
+        filename=params.lm_data_valid,
+        is_distributed=is_distributed,
+        params=params,
+    )
+
+    # Note: No learning rate scheduler is used here
+    for epoch in range(params.start_epoch, params.num_epochs):
+        if is_distributed:
+            train_dl.sampler.set_epoch(epoch)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if is_distributed:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/utils.py b/icefall/utils.py
new file mode 100644
index 000000000..a9e8a81b9
--- /dev/null
+++ b/icefall/utils.py
@@ -0,0 +1,2173 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Mingshuang Luo,
+#                                                    Zengwei Yao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import collections
+import logging
+import os
+import re
+import subprocess
+from collections import defaultdict
+from contextlib import contextmanager
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from shutil import copyfile
+from typing import Dict, Iterable, List, Optional, TextIO, Tuple, Union
+
+import k2
+import k2.version
+import kaldialign
+import sentencepiece as spm
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import average_checkpoints
+
+Pathlike = Union[str, Path]
+
+
+# Pytorch issue: https://github.com/pytorch/pytorch/issues/47379
+# Fixed: https://github.com/pytorch/pytorch/pull/49853
+# The fix was included in v1.9.0
+# https://github.com/pytorch/pytorch/releases/tag/v1.9.0
+def is_jit_tracing():
+    if torch.jit.is_scripting():
+        return False
+    elif torch.jit.is_tracing():
+        return True
+    return False
+
+
+@contextmanager
+def get_executor():
+    # We'll either return a process pool or a distributed worker pool.
+    # Note that this has to be a context manager because we might use multiple
+    # context manager ("with" clauses) inside, and this way everything will
+    # free up the resources at the right time.
+    try:
+        # If this is executed on the CLSP grid, we will try to use the
+        # Grid Engine to distribute the tasks.
+        # Other clusters can also benefit from that, provided a
+        # cluster-specific wrapper.
+        # (see https://github.com/pzelasko/plz for reference)
+        #
+        # The following must be installed:
+        # $ pip install dask distributed
+        # $ pip install git+https://github.com/pzelasko/plz
+        name = subprocess.check_output("hostname -f", shell=True, text=True)
+        if name.strip().endswith(".clsp.jhu.edu"):
+            import plz
+            from distributed import Client
+
+            with plz.setup_cluster() as cluster:
+                cluster.scale(80)
+                yield Client(cluster)
+            return
+    except Exception:
+        pass
+    # No need to return anything - compute_and_store_features
+    # will just instantiate the pool itself.
+    yield None
+
+
+def str2bool(v):
+    """Used in argparse.ArgumentParser.add_argument to indicate
+    that a type is a bool type and user can enter
+
+        - yes, true, t, y, 1, to represent True
+        - no, false, f, n, 0, to represent False
+
+    See https://stackoverflow.com/questions/15008758/parsing-boolean-values-with-argparse  # noqa
+    """
+    if isinstance(v, bool):
+        return v
+    if v.lower() in ("yes", "true", "t", "y", "1"):
+        return True
+    elif v.lower() in ("no", "false", "f", "n", "0"):
+        return False
+    else:
+        raise argparse.ArgumentTypeError("Boolean value expected.")
+
+
+def setup_logger(
+    log_filename: Pathlike,
+    log_level: str = "info",
+    use_console: bool = True,
+) -> None:
+    """Setup log level.
+
+    Args:
+      log_filename:
+        The filename to save the log.
+      log_level:
+        The log level to use, e.g., "debug", "info", "warning", "error",
+        "critical"
+      use_console:
+        True to also print logs to console.
+    """
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+    if dist.is_available() and dist.is_initialized():
+        world_size = dist.get_world_size()
+        rank = dist.get_rank()
+        formatter = f"%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] ({rank}/{world_size}) %(message)s"  # noqa
+        log_filename = f"{log_filename}-{date_time}-{rank}"
+    else:
+        formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+        log_filename = f"{log_filename}-{date_time}"
+
+    os.makedirs(os.path.dirname(log_filename), exist_ok=True)
+
+    level = logging.ERROR
+    if log_level == "debug":
+        level = logging.DEBUG
+    elif log_level == "info":
+        level = logging.INFO
+    elif log_level == "warning":
+        level = logging.WARNING
+    elif log_level == "critical":
+        level = logging.CRITICAL
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=level,
+        filemode="w",
+    )
+    if use_console:
+        console = logging.StreamHandler()
+        console.setLevel(level)
+        console.setFormatter(logging.Formatter(formatter))
+        logging.getLogger("").addHandler(console)
+
+
+class AttributeDict(dict):
+    def __getattr__(self, key):
+        if key in self:
+            return self[key]
+        raise AttributeError(f"No such attribute '{key}'")
+
+    def __setattr__(self, key, value):
+        self[key] = value
+
+    def __delattr__(self, key):
+        if key in self:
+            del self[key]
+            return
+        raise AttributeError(f"No such attribute '{key}'")
+
+
+def encode_supervisions(
+    supervisions: dict,
+    subsampling_factor: int,
+    token_ids: Optional[List[List[int]]] = None,
+) -> Tuple[torch.Tensor, Union[List[str], List[List[int]]]]:
+    """
+    Encodes Lhotse's ``batch["supervisions"]`` dict into
+    a pair of torch Tensor, and a list of transcription strings or token indexes
+
+    The supervision tensor has shape ``(batch_size, 3)``.
+    Its second dimension contains information about sequence index [0],
+    start frames [1] and num frames [2].
+
+    The batch items might become re-ordered during this operation -- the
+    returned tensor and list of strings are guaranteed to be consistent with
+    each other.
+    """
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    indices = torch.argsort(supervision_segments[:, 2], descending=True)
+    supervision_segments = supervision_segments[indices]
+
+    if token_ids is None:
+        texts = supervisions["text"]
+        res = [texts[idx] for idx in indices]
+    else:
+        res = [token_ids[idx] for idx in indices]
+
+    return supervision_segments, res
+
+
+def get_texts(
+    best_paths: k2.Fsa, return_ragged: bool = False
+) -> Union[List[List[int]], k2.RaggedTensor]:
+    """Extract the texts (as word IDs) from the best-path FSAs.
+    Args:
+      best_paths:
+        A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
+        containing multiple FSAs, which is expected to be the result
+        of k2.shortest_path (otherwise the returned values won't
+        be meaningful).
+      return_ragged:
+        True to return a ragged tensor with two axes [utt][word_id].
+        False to return a list-of-list word IDs.
+    Returns:
+      Returns a list of lists of int, containing the label sequences we
+      decoded.
+    """
+    if isinstance(best_paths.aux_labels, k2.RaggedTensor):
+        # remove 0's and -1's.
+        aux_labels = best_paths.aux_labels.remove_values_leq(0)
+        # TODO: change arcs.shape() to arcs.shape
+        aux_shape = best_paths.arcs.shape().compose(aux_labels.shape)
+
+        # remove the states and arcs axes.
+        aux_shape = aux_shape.remove_axis(1)
+        aux_shape = aux_shape.remove_axis(1)
+        aux_labels = k2.RaggedTensor(aux_shape, aux_labels.values)
+    else:
+        # remove axis corresponding to states.
+        aux_shape = best_paths.arcs.shape().remove_axis(1)
+        aux_labels = k2.RaggedTensor(aux_shape, best_paths.aux_labels)
+        # remove 0's and -1's.
+        aux_labels = aux_labels.remove_values_leq(0)
+
+    assert aux_labels.num_axes == 2
+    if return_ragged:
+        return aux_labels
+    else:
+        return aux_labels.tolist()
+
+
+def encode_supervisions_otc(
+    supervisions: dict,
+    subsampling_factor: int,
+    token_ids: Optional[List[List[int]]] = None,
+) -> Tuple[torch.Tensor, Union[List[str], List[List[int]]]]:
+    """
+    Encodes Lhotse's ``batch["supervisions"]`` dict into
+    a pair of torch Tensor, and a list of transcription strings or token indexes
+
+    The supervision tensor has shape ``(batch_size, 3)``.
+    Its second dimension contains information about sequence index [0],
+    start frames [1] and num frames [2].
+
+    The batch items might become re-ordered during this operation -- the
+    returned tensor and list of strings are guaranteed to be consistent with
+    each other.
+    """
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    indices = torch.argsort(supervision_segments[:, 2], descending=True)
+    supervision_segments = supervision_segments[indices]
+
+    ids = []
+    verbatim_texts = []
+    sorted_ids = []
+    sorted_verbatim_texts = []
+
+    for cut in supervisions["cut"]:
+        id = cut.id
+        if hasattr(cut.supervisions[0], "verbatim_text"):
+            verbatim_text = cut.supervisions[0].verbatim_text
+        else:
+            verbatim_text = ""
+        ids.append(id)
+        verbatim_texts.append(verbatim_text)
+
+    for index in indices.tolist():
+        sorted_ids.append(ids[index])
+        sorted_verbatim_texts.append(verbatim_texts[index])
+
+    if token_ids is None:
+        texts = supervisions["text"]
+        res = [texts[idx] for idx in indices]
+    else:
+        res = [token_ids[idx] for idx in indices]
+
+    return supervision_segments, res, sorted_ids, sorted_verbatim_texts
+
+
+@dataclass
+class DecodingResults:
+    # timestamps[i][k] contains the frame number on which tokens[i][k]
+    # is decoded
+    timestamps: List[List[int]]
+
+    # hyps[i] is the recognition results, i.e., word IDs or token IDs
+    # for the i-th utterance with fast_beam_search_nbest_LG.
+    hyps: Union[List[List[int]], k2.RaggedTensor]
+
+    # scores[i][k] contains the log-prob of tokens[i][k]
+    scores: Optional[List[List[float]]] = None
+
+
+def get_texts_with_timestamp(
+    best_paths: k2.Fsa, return_ragged: bool = False
+) -> DecodingResults:
+    """Extract the texts (as word IDs) and timestamps (as frame indexes)
+    from the best-path FSAs.
+    Args:
+      best_paths:
+        A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
+        containing multiple FSAs, which is expected to be the result
+        of k2.shortest_path (otherwise the returned values won't
+        be meaningful).
+      return_ragged:
+        True to return a ragged tensor with two axes [utt][word_id].
+        False to return a list-of-list word IDs.
+    Returns:
+      Returns a list of lists of int, containing the label sequences we
+      decoded.
+    """
+    if isinstance(best_paths.aux_labels, k2.RaggedTensor):
+        all_aux_shape = (
+            best_paths.arcs.shape().remove_axis(1).compose(best_paths.aux_labels.shape)
+        )
+        all_aux_labels = k2.RaggedTensor(all_aux_shape, best_paths.aux_labels.values)
+        # remove 0's and -1's.
+        aux_labels = best_paths.aux_labels.remove_values_leq(0)
+        # TODO: change arcs.shape() to arcs.shape
+        aux_shape = best_paths.arcs.shape().compose(aux_labels.shape)
+        # remove the states and arcs axes.
+        aux_shape = aux_shape.remove_axis(1)
+        aux_shape = aux_shape.remove_axis(1)
+        aux_labels = k2.RaggedTensor(aux_shape, aux_labels.values)
+    else:
+        # remove axis corresponding to states.
+        aux_shape = best_paths.arcs.shape().remove_axis(1)
+        all_aux_labels = k2.RaggedTensor(aux_shape, best_paths.aux_labels)
+        # remove 0's and -1's.
+        aux_labels = all_aux_labels.remove_values_leq(0)
+
+    assert aux_labels.num_axes == 2
+
+    timestamps = []
+    if isinstance(best_paths.aux_labels, k2.RaggedTensor):
+        for p in range(all_aux_labels.dim0):
+            time = []
+            for i, arc in enumerate(all_aux_labels[p].tolist()):
+                if len(arc) == 1 and arc[0] > 0:
+                    time.append(i)
+            timestamps.append(time)
+    else:
+        for labels in all_aux_labels.tolist():
+            time = [i for i, v in enumerate(labels) if v > 0]
+            timestamps.append(time)
+
+    return DecodingResults(
+        timestamps=timestamps,
+        hyps=aux_labels if return_ragged else aux_labels.tolist(),
+    )
+
+
+def get_alignments(best_paths: k2.Fsa, kind: str) -> List[List[int]]:
+    """Extract labels or aux_labels from the best-path FSAs.
+
+    Args:
+      best_paths:
+        A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
+        containing multiple FSAs, which is expected to be the result
+        of k2.shortest_path (otherwise the returned values won't
+        be meaningful).
+      kind:
+        Possible values are: "labels" and "aux_labels". Caution: When it is
+        "labels", the resulting alignments contain repeats.
+    Returns:
+      Returns a list of lists of int, containing the token sequences we
+      decoded. For `ans[i]`, its length equals to the number of frames
+      after subsampling of the i-th utterance in the batch.
+
+    Example:
+      When `kind` is `labels`, one possible alignment example is (with
+      repeats)::
+
+        c c c blk a a blk blk t t t blk blk
+
+     If `kind` is `aux_labels`, the above example changes to::
+
+        c blk blk blk a blk blk blk t blk blk blk blk
+
+    """
+    assert kind in ("labels", "aux_labels")
+    # arc.shape() has axes [fsa][state][arc], we remove "state"-axis here
+    token_shape = best_paths.arcs.shape().remove_axis(1)
+    # token_shape has axes [fsa][arc]
+    tokens = k2.RaggedTensor(token_shape, getattr(best_paths, kind).contiguous())
+    tokens = tokens.remove_values_eq(-1)
+    return tokens.tolist()
+
+
+def save_alignments(
+    alignments: Dict[str, List[int]],
+    subsampling_factor: int,
+    filename: str,
+) -> None:
+    """Save alignments to a file.
+
+    Args:
+      alignments:
+        A dict containing alignments. Keys of the dict are utterances and
+        values are the corresponding framewise alignments after subsampling.
+      subsampling_factor:
+        The subsampling factor of the model.
+      filename:
+        Path to save the alignments.
+    Returns:
+      Return None.
+    """
+    ali_dict = {
+        "subsampling_factor": subsampling_factor,
+        "alignments": alignments,
+    }
+    torch.save(ali_dict, filename)
+
+
+def load_alignments(filename: str) -> Tuple[int, Dict[str, List[int]]]:
+    """Load alignments from a file.
+
+    Args:
+      filename:
+        Path to the file containing alignment information.
+        The file should be saved by :func:`save_alignments`.
+    Returns:
+      Return a tuple containing:
+        - subsampling_factor: The subsampling_factor used to compute
+          the alignments.
+        - alignments: A dict containing utterances and their corresponding
+          framewise alignment, after subsampling.
+    """
+    ali_dict = torch.load(filename)
+    subsampling_factor = ali_dict["subsampling_factor"]
+    alignments = ali_dict["alignments"]
+    return subsampling_factor, alignments
+
+
+def store_transcripts(
+    filename: Pathlike, texts: Iterable[Tuple[str, str, str]], char_level: bool = False
+) -> None:
+    """Save predicted results and reference transcripts to a file.
+
+    Args:
+      filename:
+        File to save the results to.
+      texts:
+        An iterable of tuples. The first element is the cur_id, the second is
+        the reference transcript and the third element is the predicted result.
+        If it is a multi-talker ASR system, the ref and hyp may also be lists of
+        strings.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf8") as f:
+        for cut_id, ref, hyp in texts:
+            if char_level:
+                ref = list("".join(ref))
+                hyp = list("".join(hyp))
+            print(f"{cut_id}:\tref={ref}", file=f)
+            print(f"{cut_id}:\thyp={hyp}", file=f)
+
+
+def store_transcripts_and_timestamps(
+    filename: Pathlike,
+    texts: Iterable[Tuple[str, List[str], List[str], List[float], List[float]]],
+) -> None:
+    """Save predicted results and reference transcripts as well as their timestamps
+    to a file.
+
+    Args:
+      filename:
+        File to save the results to.
+      texts:
+        An iterable of tuples. The first element is the cur_id, the second is
+        the reference transcript and the third element is the predicted result.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf8") as f:
+        for cut_id, ref, hyp, time_ref, time_hyp in texts:
+            print(f"{cut_id}:\tref={ref}", file=f)
+            print(f"{cut_id}:\thyp={hyp}", file=f)
+
+            if len(time_ref) > 0:
+                if isinstance(time_ref[0], tuple):
+                    # each element is <start, end> pair
+                    s = (
+                        "["
+                        + ", ".join(["(%0.3f, %.03f)" % (i, j) for (i, j) in time_ref])
+                        + "]"
+                    )
+                else:
+                    # each element is a float number
+                    s = "[" + ", ".join(["%0.3f" % i for i in time_ref]) + "]"
+                print(f"{cut_id}:\ttimestamp_ref={s}", file=f)
+
+            if len(time_hyp) > 0:
+                if isinstance(time_hyp[0], tuple):
+                    # each element is <start, end> pair
+                    s = (
+                        "["
+                        + ", ".join(["(%0.3f, %.03f)" % (i, j) for (i, j) in time_hyp])
+                        + "]"
+                    )
+                else:
+                    # each element is a float number
+                    s = "[" + ", ".join(["%0.3f" % i for i in time_hyp]) + "]"
+                print(f"{cut_id}:\ttimestamp_hyp={s}", file=f)
+
+
+def write_error_stats(
+    f: TextIO,
+    test_set_name: str,
+    results: List[Tuple[str, str]],
+    enable_log: bool = True,
+    compute_CER: bool = False,
+    sclite_mode: bool = False,
+) -> float:
+    """Write statistics based on predicted results and reference transcripts.
+
+    It will write the following to the given file:
+
+        - WER
+        - number of insertions, deletions, substitutions, corrects and total
+          reference words. For example::
+
+              Errors: 23 insertions, 57 deletions, 212 substitutions, over 2606
+              reference words (2337 correct)
+
+        - The difference between the reference transcript and predicted result.
+          An instance is given below::
+
+            THE ASSOCIATION OF (EDISON->ADDISON) ILLUMINATING COMPANIES
+
+          The above example shows that the reference word is `EDISON`,
+          but it is predicted to `ADDISON` (a substitution error).
+
+          Another example is::
+
+            FOR THE FIRST DAY (SIR->*) I THINK
+
+          The reference word `SIR` is missing in the predicted
+          results (a deletion error).
+      results:
+        An iterable of tuples. The first element is the cut_id, the second is
+        the reference transcript and the third element is the predicted result.
+      enable_log:
+        If True, also print detailed WER to the console.
+        Otherwise, it is written only to the given file.
+    Returns:
+      Return None.
+    """
+    subs: Dict[Tuple[str, str], int] = defaultdict(int)
+    ins: Dict[str, int] = defaultdict(int)
+    dels: Dict[str, int] = defaultdict(int)
+
+    # `words` stores counts per word, as follows:
+    #   corr, ref_sub, hyp_sub, ins, dels
+    words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
+    num_corr = 0
+    ERR = "*"
+
+    if compute_CER:
+        for i, res in enumerate(results):
+            cut_id, ref, hyp = res
+            ref = list("".join(ref))
+            hyp = list("".join(hyp))
+            results[i] = (cut_id, ref, hyp)
+
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR, sclite_mode=sclite_mode)
+        for ref_word, hyp_word in ali:
+            if ref_word == ERR:
+                ins[hyp_word] += 1
+                words[hyp_word][3] += 1
+            elif hyp_word == ERR:
+                dels[ref_word] += 1
+                words[ref_word][4] += 1
+            elif hyp_word != ref_word:
+                subs[(ref_word, hyp_word)] += 1
+                words[ref_word][1] += 1
+                words[hyp_word][2] += 1
+            else:
+                words[ref_word][0] += 1
+                num_corr += 1
+    ref_len = sum([len(r) for _, r, _ in results])
+    sub_errs = sum(subs.values())
+    ins_errs = sum(ins.values())
+    del_errs = sum(dels.values())
+    tot_errs = sub_errs + ins_errs + del_errs
+    tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len)
+
+    if enable_log:
+        logging.info(
+            f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} "
+            f"[{tot_errs} / {ref_len}, {ins_errs} ins, "
+            f"{del_errs} del, {sub_errs} sub ]"
+        )
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    print(
+        f"Errors: {ins_errs} insertions, {del_errs} deletions, "
+        f"{sub_errs} substitutions, over {ref_len} reference "
+        f"words ({num_corr} correct)",
+        file=f,
+    )
+    print(
+        "Search below for sections starting with PER-UTT DETAILS:, "
+        "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:",
+        file=f,
+    )
+
+    print("", file=f)
+    print("PER-UTT DETAILS: corr or (ref->hyp)  ", file=f)
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        combine_successive_errors = True
+        if combine_successive_errors:
+            ali = [[[x], [y]] for x, y in ali]
+            for i in range(len(ali) - 1):
+                if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]:
+                    ali[i + 1][0] = ali[i][0] + ali[i + 1][0]
+                    ali[i + 1][1] = ali[i][1] + ali[i + 1][1]
+                    ali[i] = [[], []]
+            ali = [
+                [
+                    list(filter(lambda a: a != ERR, x)),
+                    list(filter(lambda a: a != ERR, y)),
+                ]
+                for x, y in ali
+            ]
+            ali = list(filter(lambda x: x != [[], []], ali))
+            ali = [
+                [
+                    ERR if x == [] else " ".join(x),
+                    ERR if y == [] else " ".join(y),
+                ]
+                for x, y in ali
+            ]
+
+        print(
+            f"{cut_id}:\t"
+            + " ".join(
+                (
+                    ref_word if ref_word == hyp_word else f"({ref_word}->{hyp_word})"
+                    for ref_word, hyp_word in ali
+                )
+            ),
+            file=f,
+        )
+
+    print("", file=f)
+    print("SUBSTITUTIONS: count ref -> hyp", file=f)
+
+    for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True):
+        print(f"{count}   {ref} -> {hyp}", file=f)
+
+    print("", file=f)
+    print("DELETIONS: count ref", file=f)
+    for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True):
+        print(f"{count}   {ref}", file=f)
+
+    print("", file=f)
+    print("INSERTIONS: count hyp", file=f)
+    for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True):
+        print(f"{count}   {hyp}", file=f)
+
+    print("", file=f)
+    print("PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp", file=f)
+    for _, word, counts in sorted(
+        [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True
+    ):
+        (corr, ref_sub, hyp_sub, ins, dels) = counts
+        tot_errs = ref_sub + hyp_sub + ins + dels
+        ref_count = corr + ref_sub + dels
+        hyp_count = corr + hyp_sub + ins
+
+        print(f"{word}   {corr} {tot_errs} {ref_count} {hyp_count}", file=f)
+    return float(tot_err_rate)
+
+
+def write_error_stats_with_timestamps(
+    f: TextIO,
+    test_set_name: str,
+    results: List[
+        Tuple[
+            str,
+            List[str],
+            List[str],
+            List[Union[float, Tuple[float, float]]],
+            List[Union[float, Tuple[float, float]]],
+        ]
+    ],
+    enable_log: bool = True,
+    with_end_time: bool = False,
+) -> Tuple[float, Union[float, Tuple[float, float]], Union[float, Tuple[float, float]]]:
+    """Write statistics based on predicted results and reference transcripts
+    as well as their timestamps.
+
+    It will write the following to the given file:
+
+        - WER
+        - number of insertions, deletions, substitutions, corrects and total
+          reference words. For example::
+
+              Errors: 23 insertions, 57 deletions, 212 substitutions, over 2606
+              reference words (2337 correct)
+
+        - The difference between the reference transcript and predicted result.
+          An instance is given below::
+
+            THE ASSOCIATION OF (EDISON->ADDISON) ILLUMINATING COMPANIES
+
+          The above example shows that the reference word is `EDISON`,
+          but it is predicted to `ADDISON` (a substitution error).
+
+          Another example is::
+
+            FOR THE FIRST DAY (SIR->*) I THINK
+
+          The reference word `SIR` is missing in the predicted
+          results (a deletion error).
+      results:
+        An iterable of tuples. The first element is the cur_id, the second is
+        the reference transcript and the third element is the predicted result.
+      enable_log:
+        If True, also print detailed WER to the console.
+        Otherwise, it is written only to the given file.
+      with_end_time:
+        Whether use end timestamps.
+
+    Returns:
+      Return total word error rate and mean delay.
+    """
+    subs: Dict[Tuple[str, str], int] = defaultdict(int)
+    ins: Dict[str, int] = defaultdict(int)
+    dels: Dict[str, int] = defaultdict(int)
+
+    # `words` stores counts per word, as follows:
+    #   corr, ref_sub, hyp_sub, ins, dels
+    words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
+    num_corr = 0
+    ERR = "*"
+    # Compute mean alignment delay on the correct words
+    all_delay = []
+    for cut_id, ref, hyp, time_ref, time_hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        has_time = len(time_ref) > 0 and len(time_hyp) > 0
+        if has_time:
+            # pointer to timestamp_hyp
+            p_hyp = 0
+            # pointer to timestamp_ref
+            p_ref = 0
+        for ref_word, hyp_word in ali:
+            if ref_word == ERR:
+                ins[hyp_word] += 1
+                words[hyp_word][3] += 1
+                if has_time:
+                    p_hyp += 1
+            elif hyp_word == ERR:
+                dels[ref_word] += 1
+                words[ref_word][4] += 1
+                if has_time:
+                    p_ref += 1
+            elif hyp_word != ref_word:
+                subs[(ref_word, hyp_word)] += 1
+                words[ref_word][1] += 1
+                words[hyp_word][2] += 1
+                if has_time:
+                    p_hyp += 1
+                    p_ref += 1
+            else:
+                words[ref_word][0] += 1
+                num_corr += 1
+                if has_time:
+                    if with_end_time:
+                        all_delay.append(
+                            (
+                                time_hyp[p_hyp][0] - time_ref[p_ref][0],
+                                time_hyp[p_hyp][1] - time_ref[p_ref][1],
+                            )
+                        )
+                    else:
+                        all_delay.append(time_hyp[p_hyp] - time_ref[p_ref])
+                    p_hyp += 1
+                    p_ref += 1
+        if has_time:
+            assert p_hyp == len(hyp), (p_hyp, len(hyp))
+            assert p_ref == len(ref), (p_ref, len(ref))
+
+    ref_len = sum([len(r) for _, r, _, _, _ in results])
+    sub_errs = sum(subs.values())
+    ins_errs = sum(ins.values())
+    del_errs = sum(dels.values())
+    tot_errs = sub_errs + ins_errs + del_errs
+    tot_err_rate = float("%.2f" % (100.0 * tot_errs / ref_len))
+
+    if with_end_time:
+        mean_delay = (float("inf"), float("inf"))
+        var_delay = (float("inf"), float("inf"))
+    else:
+        mean_delay = float("inf")
+        var_delay = float("inf")
+    num_delay = len(all_delay)
+    if num_delay > 0:
+        if with_end_time:
+            all_delay_start = [i[0] for i in all_delay]
+            mean_delay_start = sum(all_delay_start) / num_delay
+            var_delay_start = (
+                sum([(i - mean_delay_start) ** 2 for i in all_delay_start]) / num_delay
+            )
+
+            all_delay_end = [i[1] for i in all_delay]
+            mean_delay_end = sum(all_delay_end) / num_delay
+            var_delay_end = (
+                sum([(i - mean_delay_end) ** 2 for i in all_delay_end]) / num_delay
+            )
+
+            mean_delay = (
+                float("%.3f" % mean_delay_start),
+                float("%.3f" % mean_delay_end),
+            )
+            var_delay = (float("%.3f" % var_delay_start), float("%.3f" % var_delay_end))
+        else:
+            mean_delay = sum(all_delay) / num_delay
+            var_delay = sum([(i - mean_delay) ** 2 for i in all_delay]) / num_delay
+            mean_delay = float("%.3f" % mean_delay)
+            var_delay = float("%.3f" % var_delay)
+
+    if enable_log:
+        logging.info(
+            f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} "
+            f"[{tot_errs} / {ref_len}, {ins_errs} ins, "
+            f"{del_errs} del, {sub_errs} sub ]"
+        )
+        logging.info(
+            f"[{test_set_name}] %symbol-delay mean (s): "
+            f"{mean_delay}, variance: {var_delay} "  # noqa
+            f"computed on {num_delay} correct words"
+        )
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    print(
+        f"Errors: {ins_errs} insertions, {del_errs} deletions, "
+        f"{sub_errs} substitutions, over {ref_len} reference "
+        f"words ({num_corr} correct)",
+        file=f,
+    )
+    print(
+        "Search below for sections starting with PER-UTT DETAILS:, "
+        "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:",
+        file=f,
+    )
+
+    print("", file=f)
+    print("PER-UTT DETAILS: corr or (ref->hyp)  ", file=f)
+    for cut_id, ref, hyp, _, _ in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        combine_successive_errors = True
+        if combine_successive_errors:
+            ali = [[[x], [y]] for x, y in ali]
+            for i in range(len(ali) - 1):
+                if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]:
+                    ali[i + 1][0] = ali[i][0] + ali[i + 1][0]
+                    ali[i + 1][1] = ali[i][1] + ali[i + 1][1]
+                    ali[i] = [[], []]
+            ali = [
+                [
+                    list(filter(lambda a: a != ERR, x)),
+                    list(filter(lambda a: a != ERR, y)),
+                ]
+                for x, y in ali
+            ]
+            ali = list(filter(lambda x: x != [[], []], ali))
+            ali = [
+                [
+                    ERR if x == [] else " ".join(x),
+                    ERR if y == [] else " ".join(y),
+                ]
+                for x, y in ali
+            ]
+
+        print(
+            f"{cut_id}:\t"
+            + " ".join(
+                (
+                    ref_word if ref_word == hyp_word else f"({ref_word}->{hyp_word})"
+                    for ref_word, hyp_word in ali
+                )
+            ),
+            file=f,
+        )
+
+    print("", file=f)
+    print("SUBSTITUTIONS: count ref -> hyp", file=f)
+
+    for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True):
+        print(f"{count}   {ref} -> {hyp}", file=f)
+
+    print("", file=f)
+    print("DELETIONS: count ref", file=f)
+    for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True):
+        print(f"{count}   {ref}", file=f)
+
+    print("", file=f)
+    print("INSERTIONS: count hyp", file=f)
+    for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True):
+        print(f"{count}   {hyp}", file=f)
+
+    print("", file=f)
+    print("PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp", file=f)
+    for _, word, counts in sorted(
+        [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True
+    ):
+        (corr, ref_sub, hyp_sub, ins, dels) = counts
+        tot_errs = ref_sub + hyp_sub + ins + dels
+        ref_count = corr + ref_sub + dels
+        hyp_count = corr + hyp_sub + ins
+
+        print(f"{word}   {corr} {tot_errs} {ref_count} {hyp_count}", file=f)
+    return float(tot_err_rate), float(mean_delay), float(var_delay)
+
+
+def write_surt_error_stats(
+    f: TextIO,
+    test_set_name: str,
+    results: List[Tuple[str, str]],
+    enable_log: bool = True,
+    num_channels: int = 2,
+) -> float:
+    """Write statistics based on predicted results and reference transcripts for SURT
+    multi-talker ASR systems. The difference between this and the `write_error_stats`
+    is that this function finds the optimal speaker-agnostic WER using the ``meeteval``
+    toolkit.
+
+    Args:
+        f: File to write the statistics to.
+        test_set_name: Name of the test set.
+        results: List of tuples containing the utterance ID and the predicted
+            transcript.
+        enable_log: Whether to enable logging.
+        num_channels: Number of output channels/branches. Defaults to 2.
+    Returns:
+      Return None.
+    """
+    from meeteval.wer import wer
+
+    subs: Dict[Tuple[str, str], int] = defaultdict(int)
+    ins: Dict[str, int] = defaultdict(int)
+    dels: Dict[str, int] = defaultdict(int)
+    ref_lens: List[int] = []
+
+    print(
+        "Search below for sections starting with PER-UTT DETAILS:, "
+        "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:",
+        file=f,
+    )
+
+    print("", file=f)
+    print("PER-UTT DETAILS: corr or (ref->hyp)  ", file=f)
+
+    # `words` stores counts per word, as follows:
+    #   corr, ref_sub, hyp_sub, ins, dels
+    words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
+    num_corr = 0
+    ERR = "*"
+    for cut_id, ref, hyp in results:
+        # First compute the optimal assignment of references to output channels
+        orc_wer = wer.orc_word_error_rate(ref, hyp)
+        assignment = orc_wer.assignment
+        refs = [[] for _ in range(num_channels)]
+        # Assign references to channels
+        for i, ref_text in zip(assignment, ref):
+            refs[i] += ref_text.split()
+        hyps = [hyp_text.split() for hyp_text in hyp]
+        # Now compute the WER for each channel
+        for ref_c, hyp_c in zip(refs, hyps):
+            ref_lens.append(len(ref_c))
+            ali = kaldialign.align(ref_c, hyp_c, ERR)
+            for ref_word, hyp_word in ali:
+                if ref_word == ERR:
+                    ins[hyp_word] += 1
+                    words[hyp_word][3] += 1
+                elif hyp_word == ERR:
+                    dels[ref_word] += 1
+                    words[ref_word][4] += 1
+                elif hyp_word != ref_word:
+                    subs[(ref_word, hyp_word)] += 1
+                    words[ref_word][1] += 1
+                    words[hyp_word][2] += 1
+                else:
+                    words[ref_word][0] += 1
+                    num_corr += 1
+            combine_successive_errors = True
+            if combine_successive_errors:
+                ali = [[[x], [y]] for x, y in ali]
+                for i in range(len(ali) - 1):
+                    if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]:
+                        ali[i + 1][0] = ali[i][0] + ali[i + 1][0]
+                        ali[i + 1][1] = ali[i][1] + ali[i + 1][1]
+                        ali[i] = [[], []]
+                ali = [
+                    [
+                        list(filter(lambda a: a != ERR, x)),
+                        list(filter(lambda a: a != ERR, y)),
+                    ]
+                    for x, y in ali
+                ]
+                ali = list(filter(lambda x: x != [[], []], ali))
+                ali = [
+                    [
+                        ERR if x == [] else " ".join(x),
+                        ERR if y == [] else " ".join(y),
+                    ]
+                    for x, y in ali
+                ]
+
+            print(
+                f"{cut_id}:\t"
+                + " ".join(
+                    (
+                        ref_word
+                        if ref_word == hyp_word
+                        else f"({ref_word}->{hyp_word})"
+                        for ref_word, hyp_word in ali
+                    )
+                ),
+                file=f,
+            )
+    ref_len = sum(ref_lens)
+    sub_errs = sum(subs.values())
+    ins_errs = sum(ins.values())
+    del_errs = sum(dels.values())
+    tot_errs = sub_errs + ins_errs + del_errs
+    tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len)
+
+    if enable_log:
+        logging.info(
+            f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} "
+            f"[{tot_errs} / {ref_len}, {ins_errs} ins, "
+            f"{del_errs} del, {sub_errs} sub ]"
+        )
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    print(
+        f"Errors: {ins_errs} insertions, {del_errs} deletions, "
+        f"{sub_errs} substitutions, over {ref_len} reference "
+        f"words ({num_corr} correct)",
+        file=f,
+    )
+
+    print("", file=f)
+    print("SUBSTITUTIONS: count ref -> hyp", file=f)
+
+    for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True):
+        print(f"{count}   {ref} -> {hyp}", file=f)
+
+    print("", file=f)
+    print("DELETIONS: count ref", file=f)
+    for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True):
+        print(f"{count}   {ref}", file=f)
+
+    print("", file=f)
+    print("INSERTIONS: count hyp", file=f)
+    for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True):
+        print(f"{count}   {hyp}", file=f)
+
+    print("", file=f)
+    print("PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp", file=f)
+    for _, word, counts in sorted(
+        [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True
+    ):
+        (corr, ref_sub, hyp_sub, ins, dels) = counts
+        tot_errs = ref_sub + hyp_sub + ins + dels
+        ref_count = corr + ref_sub + dels
+        hyp_count = corr + hyp_sub + ins
+
+        print(f"{word}   {corr} {tot_errs} {ref_count} {hyp_count}", file=f)
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    return float(tot_err_rate)
+
+
+class MetricsTracker(collections.defaultdict):
+    def __init__(self):
+        # Passing the type 'int' to the base-class constructor
+        # makes undefined items default to int() which is zero.
+        # This class will play a role as metrics tracker.
+        # It can record many metrics, including but not limited to loss.
+        super(MetricsTracker, self).__init__(int)
+
+    def __add__(self, other: "MetricsTracker") -> "MetricsTracker":
+        ans = MetricsTracker()
+        for k, v in self.items():
+            ans[k] = v
+        for k, v in other.items():
+            ans[k] = ans[k] + v
+        return ans
+
+    def __mul__(self, alpha: float) -> "MetricsTracker":
+        ans = MetricsTracker()
+        for k, v in self.items():
+            ans[k] = v * alpha
+        return ans
+
+    def __str__(self) -> str:
+        ans_frames = ""
+        ans_utterances = ""
+        for k, v in self.norm_items():
+            norm_value = "%.4g" % v
+            if "utt_" not in k:
+                ans_frames += str(k) + "=" + str(norm_value) + ", "
+            else:
+                ans_utterances += str(k) + "=" + str(norm_value)
+                if k == "utt_duration":
+                    ans_utterances += " frames, "
+                elif k == "utt_pad_proportion":
+                    ans_utterances += ", "
+                else:
+                    raise ValueError(f"Unexpected key: {k}")
+        frames = "%.2f" % self["frames"]
+        ans_frames += "over " + str(frames) + " frames. "
+        if ans_utterances != "":
+            utterances = "%.2f" % self["utterances"]
+            ans_utterances += "over " + str(utterances) + " utterances."
+
+        return ans_frames + ans_utterances
+
+    def norm_items(self) -> List[Tuple[str, float]]:
+        """
+        Returns a list of pairs, like:
+          [('ctc_loss', 0.1), ('att_loss', 0.07)]
+        """
+        num_frames = self["frames"] if "frames" in self else 1
+        num_utterances = self["utterances"] if "utterances" in self else 1
+        ans = []
+        for k, v in self.items():
+            if k == "frames" or k == "utterances":
+                continue
+            norm_value = (
+                float(v) / num_frames if "utt_" not in k else float(v) / num_utterances
+            )
+            ans.append((k, norm_value))
+        return ans
+
+    def reduce(self, device):
+        """
+        Reduce using torch.distributed, which I believe ensures that
+        all processes get the total.
+        """
+        keys = sorted(self.keys())
+        s = torch.tensor([float(self[k]) for k in keys], device=device)
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        for k, v in zip(keys, s.cpu().tolist()):
+            self[k] = v
+
+    def write_summary(
+        self,
+        tb_writer: SummaryWriter,
+        prefix: str,
+        batch_idx: int,
+    ) -> None:
+        """Add logging information to a TensorBoard writer.
+
+        Args:
+            tb_writer: a TensorBoard writer
+            prefix: a prefix for the name of the loss, e.g. "train/valid_",
+                or "train/current_"
+            batch_idx: The current batch index, used as the x-axis of the plot.
+        """
+        for k, v in self.norm_items():
+            tb_writer.add_scalar(prefix + k, v, batch_idx)
+
+
+def concat(ragged: k2.RaggedTensor, value: int, direction: str) -> k2.RaggedTensor:
+    """Prepend a value to the beginning of each sublist or append a value.
+    to the end of each sublist.
+
+    Args:
+      ragged:
+        A ragged tensor with two axes.
+      value:
+        The value to prepend or append.
+      direction:
+        It can be either "left" or "right". If it is "left", we
+        prepend the value to the beginning of each sublist;
+        if it is "right", we append the value to the end of each
+        sublist.
+
+    Returns:
+      Return a new ragged tensor, whose sublists either start with
+      or end with the given value.
+
+    >>> a = k2.RaggedTensor([[1, 3], [5]])
+    >>> a
+    [ [ 1 3 ] [ 5 ] ]
+    >>> concat(a, value=0, direction="left")
+    [ [ 0 1 3 ] [ 0 5 ] ]
+    >>> concat(a, value=0, direction="right")
+    [ [ 1 3 0 ] [ 5 0 ] ]
+
+    """
+    dtype = ragged.dtype
+    device = ragged.device
+
+    assert ragged.num_axes == 2, f"num_axes: {ragged.num_axes}"
+    pad_values = torch.full(
+        size=(ragged.tot_size(0), 1),
+        fill_value=value,
+        device=device,
+        dtype=dtype,
+    )
+    pad = k2.RaggedTensor(pad_values)
+
+    if direction == "left":
+        ans = k2.ragged.cat([pad, ragged], axis=1)
+    elif direction == "right":
+        ans = k2.ragged.cat([ragged, pad], axis=1)
+    else:
+        raise ValueError(
+            f'Unsupported direction: {direction}. " \
+            "Expect either "left" or "right"'
+        )
+    return ans
+
+
+def add_sos(ragged: k2.RaggedTensor, sos_id: int) -> k2.RaggedTensor:
+    """Add SOS to each sublist.
+
+    Args:
+      ragged:
+        A ragged tensor with two axes.
+      sos_id:
+        The ID of the SOS symbol.
+
+    Returns:
+      Return a new ragged tensor, where each sublist starts with SOS.
+
+    >>> a = k2.RaggedTensor([[1, 3], [5]])
+    >>> a
+    [ [ 1 3 ] [ 5 ] ]
+    >>> add_sos(a, sos_id=0)
+    [ [ 0 1 3 ] [ 0 5 ] ]
+
+    """
+    return concat(ragged, sos_id, direction="left")
+
+
+def add_eos(ragged: k2.RaggedTensor, eos_id: int) -> k2.RaggedTensor:
+    """Add EOS to each sublist.
+
+    Args:
+      ragged:
+        A ragged tensor with two axes.
+      eos_id:
+        The ID of the EOS symbol.
+
+    Returns:
+      Return a new ragged tensor, where each sublist ends with EOS.
+
+    >>> a = k2.RaggedTensor([[1, 3], [5]])
+    >>> a
+    [ [ 1 3 ] [ 5 ] ]
+    >>> add_eos(a, eos_id=0)
+    [ [ 1 3 0 ] [ 5 0 ] ]
+
+    """
+    return concat(ragged, eos_id, direction="right")
+
+
+def make_pad_mask(lengths: torch.Tensor, max_len: int = 0) -> torch.Tensor:
+    """
+    Args:
+      lengths:
+        A 1-D tensor containing sentence lengths.
+      max_len:
+        The length of masks.
+    Returns:
+      Return a 2-D bool tensor, where masked positions
+      are filled with `True` and non-masked positions are
+      filled with `False`.
+
+    >>> lengths = torch.tensor([1, 3, 2, 5])
+    >>> make_pad_mask(lengths)
+    tensor([[False,  True,  True,  True,  True],
+            [False, False, False,  True,  True],
+            [False, False,  True,  True,  True],
+            [False, False, False, False, False]])
+    """
+    assert lengths.ndim == 1, lengths.ndim
+    max_len = max(max_len, lengths.max())
+    n = lengths.size(0)
+    seq_range = torch.arange(0, max_len, device=lengths.device)
+    expaned_lengths = seq_range.unsqueeze(0).expand(n, max_len)
+
+    return expaned_lengths >= lengths.unsqueeze(-1)
+
+
+# Copied and modified from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/mask.py
+def subsequent_chunk_mask(
+    size: int,
+    chunk_size: int,
+    num_left_chunks: int = -1,
+    device: torch.device = torch.device("cpu"),
+) -> torch.Tensor:
+    """Create mask for subsequent steps (size, size) with chunk size,
+       this is for streaming encoder
+    Args:
+        size (int): size of mask
+        chunk_size (int): size of chunk
+        num_left_chunks (int): number of left chunks
+            <0: use full chunk
+            >=0: use num_left_chunks
+        device (torch.device): "cpu" or "cuda" or torch.Tensor.device
+    Returns:
+        torch.Tensor: mask
+    Examples:
+        >>> subsequent_chunk_mask(4, 2)
+        [[1, 1, 0, 0],
+         [1, 1, 0, 0],
+         [1, 1, 1, 1],
+         [1, 1, 1, 1]]
+    """
+    ret = torch.zeros(size, size, device=device, dtype=torch.bool)
+    for i in range(size):
+        if num_left_chunks < 0:
+            start = 0
+        else:
+            start = max((i // chunk_size - num_left_chunks) * chunk_size, 0)
+        ending = min((i // chunk_size + 1) * chunk_size, size)
+        ret[i, start:ending] = True
+    return ret
+
+
+def l1_norm(x):
+    return torch.sum(torch.abs(x))
+
+
+def l2_norm(x):
+    return torch.sum(torch.pow(x, 2))
+
+
+def linf_norm(x):
+    return torch.max(torch.abs(x))
+
+
+def measure_weight_norms(model: nn.Module, norm: str = "l2") -> Dict[str, float]:
+    """
+    Compute the norms of the model's parameters.
+
+    :param model: a torch.nn.Module instance
+    :param norm: how to compute the norm. Available values: 'l1', 'l2', 'linf'
+    :return: a dict mapping from parameter's name to its norm.
+    """
+    with torch.no_grad():
+        norms = {}
+        for name, param in model.named_parameters():
+            if norm == "l1":
+                val = l1_norm(param)
+            elif norm == "l2":
+                val = l2_norm(param)
+            elif norm == "linf":
+                val = linf_norm(param)
+            else:
+                raise ValueError(f"Unknown norm type: {norm}")
+            norms[name] = val.item()
+        return norms
+
+
+def measure_gradient_norms(model: nn.Module, norm: str = "l1") -> Dict[str, float]:
+    """
+    Compute the norms of the gradients for each of model's parameters.
+
+    :param model: a torch.nn.Module instance
+    :param norm: how to compute the norm. Available values: 'l1', 'l2', 'linf'
+    :return: a dict mapping from parameter's name to its gradient's norm.
+    """
+    with torch.no_grad():
+        norms = {}
+        for name, param in model.named_parameters():
+            if norm == "l1":
+                val = l1_norm(param.grad)
+            elif norm == "l2":
+                val = l2_norm(param.grad)
+            elif norm == "linf":
+                val = linf_norm(param.grad)
+            else:
+                raise ValueError(f"Unknown norm type: {norm}")
+            norms[name] = val.item()
+        return norms
+
+
+def get_parameter_groups_with_lrs(
+    model: nn.Module,
+    lr: float,
+    include_names: bool = False,
+    freeze_modules: List[str] = [],
+) -> List[dict]:
+    """
+    This is for use with the ScaledAdam optimizers (more recent versions that accept lists of
+    named-parameters; we can, if needed, create a version without the names).
+
+    It provides a way to specify learning-rate scales inside the module, so that if
+    any nn.Module in the hierarchy has a floating-point parameter 'lr_scale', it will
+    scale the LR of any parameters inside that module or its submodules.  Note: you
+    can set module parameters outside the __init__ function, e.g.:
+      >>> a = nn.Linear(10, 10)
+      >>> a.lr_scale = 0.5
+
+    Returns: a list of dicts, of the following form:
+      if include_names == False:
+        [  { 'params': [ tensor1, tensor2, ... ], 'lr': 0.01 },
+           { 'params': [ tensor3, tensor4, ... ], 'lr': 0.005 },
+         ...   ]
+      if include_names == true:
+        [  { 'named_params': [ (name1, tensor1, (name2, tensor2), ... ], 'lr': 0.01 },
+           { 'named_params': [ (name3, tensor3), (name4, tensor4), ... ], 'lr': 0.005 },
+         ...   ]
+
+    """
+    named_modules = list(model.named_modules())
+
+    # flat_lr_scale just contains the lr_scale explicitly specified
+    # for each prefix of the name, e.g. 'encoder.layers.3', these need
+    # to be multiplied for all prefix of the name of any given parameter.
+    flat_lr_scale = defaultdict(lambda: 1.0)
+    names = []
+    for name, m in model.named_modules():
+        names.append(name)
+        if hasattr(m, "lr_scale"):
+            flat_lr_scale[name] = m.lr_scale
+
+    # lr_to_parames is a dict from learning rate (floating point) to: if
+    # include_names == true, a list of (name, parameter) for that learning rate;
+    # otherwise a list of parameters for that learning rate.
+    lr_to_params = defaultdict(list)
+
+    for name, parameter in model.named_parameters():
+        split_name = name.split(".")
+        # caution: as a special case, if the name is '', split_name will be [ '' ].
+        prefix = split_name[0]
+        if prefix == "module":  # DDP
+            module_name = split_name[1]
+            if module_name in freeze_modules:
+                logging.info(f"Remove {name} from parameters")
+                continue
+        else:
+            if prefix in freeze_modules:
+                logging.info(f"Remove {name} from parameters")
+                continue
+        cur_lr = lr * flat_lr_scale[prefix]
+        if prefix != "":
+            cur_lr *= flat_lr_scale[""]
+        for part in split_name[1:]:
+            prefix = ".".join([prefix, part])
+            cur_lr *= flat_lr_scale[prefix]
+        lr_to_params[cur_lr].append((name, parameter) if include_names else parameter)
+
+    if include_names:
+        return [{"named_params": pairs, "lr": lr} for lr, pairs in lr_to_params.items()]
+    else:
+        return [{"params": params, "lr": lr} for lr, params in lr_to_params.items()]
+
+
+def optim_step_and_measure_param_change(
+    model: nn.Module,
+    old_parameters: Dict[str, nn.parameter.Parameter],
+) -> Dict[str, float]:
+    """
+    Measure the "relative change in parameters per minibatch."
+    It is understood as a ratio between the L2 norm of the difference between original and updates parameters,
+    and the L2 norm of the original parameter. It is given by the formula:
+
+        .. math::
+            \begin{aligned}
+                \delta = \frac{\Vert\theta - \theta_{new}\Vert^2}{\Vert\theta\Vert^2}
+            \end{aligned}
+
+    This function is supposed to be used as follows:
+
+      .. code-block:: python
+
+        old_parameters = {
+            n: p.detach().clone() for n, p in model.named_parameters()
+        }
+
+        optimizer.step()
+
+        deltas = optim_step_and_measure_param_change(old_parameters)
+
+    Args:
+      model: A torch.nn.Module instance.
+      old_parameters:
+        A Dict of named_parameters before optimizer.step().
+
+    Return:
+      A Dict containing the relative change for each parameter.
+    """
+    relative_change = {}
+    with torch.no_grad():
+        for n, p_new in model.named_parameters():
+            p_orig = old_parameters[n]
+            delta = l2_norm(p_orig - p_new) / l2_norm(p_orig)
+            relative_change[n] = delta.item()
+    return relative_change
+
+
+def load_averaged_model(
+    model_dir: str,
+    model: torch.nn.Module,
+    epoch: int,
+    avg: int,
+    device: torch.device,
+):
+    """
+    Load a model which is the average of all checkpoints
+
+    :param model_dir: a str of the experiment directory
+    :param model: a torch.nn.Module instance
+
+    :param epoch: the last epoch to load from
+    :param avg: how many models to average from
+    :param device: move model to this device
+
+    :return: A model averaged
+    """
+
+    # start cannot be negative
+    start = max(epoch - avg + 1, 0)
+    filenames = [f"{model_dir}/epoch-{i}.pt" for i in range(start, epoch + 1)]
+
+    logging.info(f"averaging {filenames}")
+    model.to(device)
+    model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    return model
+
+
+def tokenize_by_bpe_model(
+    sp: spm.SentencePieceProcessor,
+    txt: str,
+) -> str:
+    """
+    Tokenize text with bpe model. This function is from
+    https://github1s.com/wenet-e2e/wenet/blob/main/wenet/dataset/processor.py#L322-L342.
+    Args:
+      sp: spm.SentencePieceProcessor.
+      txt: str
+
+    Return:
+      A new string which includes chars and bpes.
+    """
+    tokens = []
+    # CJK(China Japan Korea) unicode range is [U+4E00, U+9FFF], ref:
+    # https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+    pattern = re.compile(r"([\u4e00-\u9fff])")
+    # Example:
+    #   txt   = "你好 ITS'S OKAY 的"
+    #   chars = ["你", "好", " ITS'S OKAY ", "的"]
+    chars = pattern.split(txt.upper())
+    mix_chars = [w for w in chars if len(w.strip()) > 0]
+    for ch_or_w in mix_chars:
+        # ch_or_w is a single CJK character(i.e., "你"), do nothing.
+        if pattern.fullmatch(ch_or_w) is not None:
+            tokens.append(ch_or_w)
+        # ch_or_w contains non-CJK characters(i.e., " IT'S OKAY "),
+        # encode ch_or_w using bpe_model.
+        else:
+            for p in sp.encode_as_pieces(ch_or_w):
+                tokens.append(p)
+    txt_with_bpe = "/".join(tokens)
+
+    return txt_with_bpe
+
+
+def tokenize_by_CJK_char(line: str) -> str:
+    """
+    Tokenize a line of text with CJK char.
+
+    Note: All return characters will be upper case.
+
+    Example:
+      input = "你好世界是 hello world 的中文"
+      output = "你 好 世 界 是 HELLO WORLD 的 中 文"
+
+    Args:
+      line:
+        The input text.
+
+    Return:
+      A new string tokenize by CJK char.
+    """
+    # The CJK ranges is from https://github.com/alvations/nltk/blob/79eed6ddea0d0a2c212c1060b477fc268fec4d4b/nltk/tokenize/util.py
+    pattern = re.compile(
+        r"([\u1100-\u11ff\u2e80-\ua4cf\ua840-\uD7AF\uF900-\uFAFF\uFE30-\uFE4F\uFF65-\uFFDC\U00020000-\U0002FFFF])"
+    )
+    chars = pattern.split(line.strip().upper())
+    return " ".join([w.strip() for w in chars if w.strip()])
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def convert_timestamp(
+    frames: List[int],
+    subsampling_factor: int,
+    frame_shift_ms: float = 10,
+) -> List[float]:
+    """Convert frame numbers to time (in seconds) given subsampling factor
+    and frame shift (in milliseconds).
+
+    Args:
+      frames:
+        A list of frame numbers after subsampling.
+      subsampling_factor:
+        The subsampling factor of the model.
+      frame_shift_ms:
+        Frame shift in milliseconds between two contiguous frames.
+    Return:
+      Return the time in seconds corresponding to each given frame.
+    """
+    frame_shift = frame_shift_ms / 1000.0
+    time = []
+    for f in frames:
+        time.append(round(f * subsampling_factor * frame_shift, ndigits=3))
+
+    return time
+
+
+def parse_timestamp(tokens: List[str], timestamp: List[float]) -> List[float]:
+    """
+    Parse timestamp of each word.
+
+    Args:
+      tokens:
+        List of tokens.
+      timestamp:
+        List of timestamp of each token.
+
+    Returns:
+      List of timestamp of each word.
+    """
+    start_token = b"\xe2\x96\x81".decode()  # '_'
+    assert len(tokens) == len(timestamp), (len(tokens), len(timestamp))
+    ans = []
+    for i in range(len(tokens)):
+        flag = False
+        if i == 0 or tokens[i].startswith(start_token):
+            flag = True
+            if len(tokens[i]) == 1 and tokens[i].startswith(start_token):
+                # tokens[i] == start_token
+                if i == len(tokens) - 1:
+                    # it is the last token
+                    flag = False
+                elif tokens[i + 1].startswith(start_token):
+                    # the next token also starts with start_token
+                    flag = False
+        if flag:
+            ans.append(timestamp[i])
+    return ans
+
+
+def parse_hyp_and_timestamp(
+    res: DecodingResults,
+    subsampling_factor: int,
+    frame_shift_ms: float = 10,
+    sp: Optional[spm.SentencePieceProcessor] = None,
+    word_table: Optional[k2.SymbolTable] = None,
+) -> Tuple[List[List[str]], List[List[float]]]:
+    """Parse hypothesis and timestamp.
+
+    Args:
+      res:
+        A DecodingResults object.
+      subsampling_factor:
+        The integer subsampling factor.
+      frame_shift_ms:
+        The float frame shift used for feature extraction.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+
+    Returns:
+       Return a list of hypothesis and timestamp.
+    """
+    hyps = []
+    timestamps = []
+
+    N = len(res.hyps)
+    assert len(res.timestamps) == N, (len(res.timestamps), N)
+    use_word_table = False
+    if word_table is not None:
+        assert sp is None
+        use_word_table = True
+    else:
+        assert sp is not None and word_table is None
+
+    for i in range(N):
+        time = convert_timestamp(res.timestamps[i], subsampling_factor, frame_shift_ms)
+        if use_word_table:
+            words = [word_table[i] for i in res.hyps[i]]
+        else:
+            tokens = sp.id_to_piece(res.hyps[i])
+            words = sp.decode_pieces(tokens).split()
+            time = parse_timestamp(tokens, time)
+        assert len(time) == len(words), (len(time), len(words))
+
+        hyps.append(words)
+        timestamps.append(time)
+
+    return hyps, timestamps
+
+
+# `is_module_available` is copied from
+# https://github.com/pytorch/audio/blob/6bad3a66a7a1c7cc05755e9ee5931b7391d2b94c/torchaudio/_internal/module_utils.py#L9
+def is_module_available(*modules: str) -> bool:
+    r"""Returns if a top-level module with :attr:`name` exists *without**
+    importing it. This is generally safer than try-catch block around a
+    `import X`.
+
+    Note: "borrowed" from torchaudio:
+    """
+    import importlib
+
+    return all(importlib.util.find_spec(m) is not None for m in modules)
+
+
+def filter_uneven_sized_batch(batch: dict, allowed_max_frames: int):
+    """For the uneven-sized batch, the total duration after padding would possibly
+    cause OOM. Hence, for each batch, which is sorted in descending order by length,
+    we simply drop the last few shortest samples, so that the retained total frames
+    (after padding) would not exceed the given allow_max_frames.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      allowed_max_frames:
+        The allowed max number of frames in batch.
+    """
+    features = batch["inputs"]
+    supervisions = batch["supervisions"]
+
+    N, T, _ = features.size()
+    assert T == supervisions["num_frames"].max(), (T, supervisions["num_frames"].max())
+    kept_num_utt = allowed_max_frames // T
+
+    if kept_num_utt >= N or kept_num_utt == 0:
+        return batch
+
+    # Note: we assume the samples in batch is sorted descendingly by length
+    logging.info(
+        f"Filtering uneven-sized batch, original batch size is {N}, "
+        f"retained batch size is {kept_num_utt}."
+    )
+    batch["inputs"] = features[:kept_num_utt]
+    for k, v in supervisions.items():
+        assert len(v) == N, (len(v), N)
+        batch["supervisions"][k] = v[:kept_num_utt]
+
+    return batch
+
+
+def parse_bpe_start_end_pairs(
+    tokens: List[str], is_first_token: List[bool]
+) -> List[Tuple[int, int]]:
+    """Parse pairs of start and end frame indexes for each word.
+
+    Args:
+      tokens:
+        List of BPE tokens.
+      is_first_token:
+        List of bool values, which indicates whether it is the first token,
+        i.e., not repeat or blank.
+
+    Returns:
+      List of (start-frame-index, end-frame-index) pairs for each word.
+    """
+    assert len(tokens) == len(is_first_token), (len(tokens), len(is_first_token))
+
+    start_token = b"\xe2\x96\x81".decode()  # '_'
+    blank_token = "<blk>"
+
+    non_blank_idx = [i for i in range(len(tokens)) if tokens[i] != blank_token]
+    num_non_blank = len(non_blank_idx)
+
+    pairs = []
+    start = -1
+    end = -1
+    for j in range(num_non_blank):
+        # The index in all frames
+        i = non_blank_idx[j]
+
+        found_start = False
+        if is_first_token[i] and (j == 0 or tokens[i].startswith(start_token)):
+            found_start = True
+            if tokens[i] == start_token:
+                if j == num_non_blank - 1:
+                    # It is the last non-blank token
+                    found_start = False
+                elif is_first_token[non_blank_idx[j + 1]] and tokens[
+                    non_blank_idx[j + 1]
+                ].startswith(start_token):
+                    # The next not-blank token is a first-token and also starts with start_token
+                    found_start = False
+        if found_start:
+            start = i
+
+        if start != -1:
+            found_end = False
+            if j == num_non_blank - 1:
+                # It is the last non-blank token
+                found_end = True
+            elif is_first_token[non_blank_idx[j + 1]] and tokens[
+                non_blank_idx[j + 1]
+            ].startswith(start_token):
+                # The next not-blank token is a first-token and also starts with start_token
+                found_end = True
+            if found_end:
+                end = i
+
+        if start != -1 and end != -1:
+            if not all([tokens[t] == start_token for t in range(start, end + 1)]):
+                # except the case of all start_token
+                pairs.append((start, end))
+            # Reset start and end
+            start = -1
+            end = -1
+
+    return pairs
+
+
+def parse_bpe_timestamps_and_texts(
+    best_paths: k2.Fsa, sp: spm.SentencePieceProcessor
+) -> Tuple[List[Tuple[int, int]], List[List[str]]]:
+    """Parse timestamps (frame indexes) and texts.
+
+    Args:
+      best_paths:
+        A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
+        containing multiple FSAs, which is expected to be the result
+        of k2.shortest_path (otherwise the returned values won't
+        be meaningful). Its attributes `labels` and `aux_labels`
+        are both BPE tokens.
+      sp:
+        The BPE model.
+
+    Returns:
+      utt_index_pairs:
+        A list of pair list. utt_index_pairs[i] is a list of
+        (start-frame-index, end-frame-index) pairs for each word in
+        utterance-i.
+      utt_words:
+        A list of str list. utt_words[i] is a word list of utterence-i.
+    """
+    shape = best_paths.arcs.shape().remove_axis(1)
+
+    # labels: [utt][arcs]
+    labels = k2.RaggedTensor(shape, best_paths.labels.contiguous())
+    # remove -1's.
+    labels = labels.remove_values_eq(-1)
+    labels = labels.tolist()
+
+    # aux_labels: [utt][arcs]
+    aux_labels = k2.RaggedTensor(shape, best_paths.aux_labels.contiguous())
+
+    # remove -1's.
+    all_aux_labels = aux_labels.remove_values_eq(-1)
+    # len(all_aux_labels[i]) is equal to the number of frames
+    all_aux_labels = all_aux_labels.tolist()
+
+    # remove 0's and -1's.
+    out_aux_labels = aux_labels.remove_values_leq(0)
+    # len(out_aux_labels[i]) is equal to the number of output BPE tokens
+    out_aux_labels = out_aux_labels.tolist()
+
+    utt_index_pairs = []
+    utt_words = []
+    for i in range(len(labels)):
+        tokens = sp.id_to_piece(labels[i])
+        words = sp.decode(out_aux_labels[i]).split()
+
+        # Indicates whether it is the first token, i.e., not-repeat and not-blank.
+        is_first_token = [a != 0 for a in all_aux_labels[i]]
+        index_pairs = parse_bpe_start_end_pairs(tokens, is_first_token)
+        assert len(index_pairs) == len(words), (len(index_pairs), len(words), tokens)
+        utt_index_pairs.append(index_pairs)
+        utt_words.append(words)
+
+    return utt_index_pairs, utt_words
+
+
+def parse_timestamps_and_texts(
+    best_paths: k2.Fsa, word_table: k2.SymbolTable
+) -> Tuple[List[Tuple[int, int]], List[List[str]]]:
+    """Parse timestamps (frame indexes) and texts.
+
+    Args:
+      best_paths:
+        A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
+        containing multiple FSAs, which is expected to be the result
+        of k2.shortest_path (otherwise the returned values won't
+        be meaningful). Attribute `labels` is the prediction unit,
+        e.g., phone or BPE tokens. Attribute `aux_labels` is the word index.
+      word_table:
+        The word symbol table.
+
+    Returns:
+      utt_index_pairs:
+        A list of pair list. utt_index_pairs[i] is a list of
+        (start-frame-index, end-frame-index) pairs for each word in
+        utterance-i.
+      utt_words:
+        A list of str list. utt_words[i] is a word list of utterence-i.
+    """
+    # [utt][words]
+    word_ids = get_texts(best_paths)
+
+    shape = best_paths.arcs.shape().remove_axis(1)
+
+    # labels: [utt][arcs]
+    labels = k2.RaggedTensor(shape, best_paths.labels.contiguous())
+    # remove -1's.
+    labels = labels.remove_values_eq(-1)
+    labels = labels.tolist()
+
+    # aux_labels: [utt][arcs]
+    aux_shape = shape.compose(best_paths.aux_labels.shape)
+    aux_labels = k2.RaggedTensor(aux_shape, best_paths.aux_labels.values.contiguous())
+    aux_labels = aux_labels.tolist()
+
+    utt_index_pairs = []
+    utt_words = []
+    for i, (label, aux_label) in enumerate(zip(labels, aux_labels)):
+        num_arcs = len(label)
+        # The last arc of aux_label is the arc entering the final state
+        assert num_arcs == len(aux_label) - 1, (num_arcs, len(aux_label))
+
+        index_pairs = []
+        start = -1
+        end = -1
+        for arc in range(num_arcs):
+            # len(aux_label[arc]) is 0 or 1
+            if label[arc] != 0 and len(aux_label[arc]) != 0:
+                if start != -1 and end != -1:
+                    index_pairs.append((start, end))
+                start = arc
+            if label[arc] != 0:
+                end = arc
+        if start != -1 and end != -1:
+            index_pairs.append((start, end))
+
+        words = [word_table[w] for w in word_ids[i]]
+        assert len(index_pairs) == len(words), (len(index_pairs), len(words))
+
+        utt_index_pairs.append(index_pairs)
+        utt_words.append(words)
+
+    return utt_index_pairs, utt_words
+
+
+def parse_fsa_timestamps_and_texts(
+    best_paths: k2.Fsa,
+    sp: Optional[spm.SentencePieceProcessor] = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    subsampling_factor: int = 4,
+    frame_shift_ms: float = 10,
+) -> Tuple[List[Tuple[float, float]], List[List[str]]]:
+    """Parse timestamps (in seconds) and texts for given decoded fsa paths.
+    Currently it supports two cases:
+    (1) ctc-decoding, the attributes `labels` and `aux_labels`
+        are both BPE tokens. In this case, sp should be provided.
+    (2) HLG-based 1best, the attribtute `labels` is the prediction unit,
+        e.g., phone or BPE tokens; attribute `aux_labels` is the word index.
+        In this case, word_table should be provided.
+
+    Args:
+      best_paths:
+        A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
+        containing multiple FSAs, which is expected to be the result
+        of k2.shortest_path (otherwise the returned values won't
+        be meaningful).
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      subsampling_factor:
+        The subsampling factor of the model.
+      frame_shift_ms:
+        Frame shift in milliseconds between two contiguous frames.
+
+    Returns:
+      utt_time_pairs:
+        A list of pair list. utt_time_pairs[i] is a list of
+        (start-time, end-time) pairs for each word in
+        utterance-i.
+      utt_words:
+        A list of str list. utt_words[i] is a word list of utterence-i.
+    """
+    if sp is not None:
+        assert word_table is None, "word_table is not needed if sp is provided."
+        utt_index_pairs, utt_words = parse_bpe_timestamps_and_texts(
+            best_paths=best_paths, sp=sp
+        )
+    elif word_table is not None:
+        assert sp is None, "sp is not needed if word_table is provided."
+        utt_index_pairs, utt_words = parse_timestamps_and_texts(
+            best_paths=best_paths, word_table=word_table
+        )
+    else:
+        raise ValueError("Either sp or word_table should be provided.")
+
+    utt_time_pairs = []
+    for utt in utt_index_pairs:
+        start = convert_timestamp(
+            frames=[i[0] for i in utt],
+            subsampling_factor=subsampling_factor,
+            frame_shift_ms=frame_shift_ms,
+        )
+        end = convert_timestamp(
+            # The duration in frames is (end_frame_index - start_frame_index + 1)
+            frames=[i[1] + 1 for i in utt],
+            subsampling_factor=subsampling_factor,
+            frame_shift_ms=frame_shift_ms,
+        )
+        utt_time_pairs.append(list(zip(start, end)))
+
+    return utt_time_pairs, utt_words
+
+
+# Copied from https://github.com/alvations/nltk/blob/79eed6ddea0d0a2c212c1060b477fc268fec4d4b/nltk/tokenize/util.py
+def is_cjk(character):
+    """
+    Python port of Moses' code to check for CJK character.
+
+    >>> is_cjk(u'\u33fe')
+    True
+    >>> is_cjk(u'\uFE5F')
+    False
+
+    :param character: The character that needs to be checked.
+    :type character: char
+    :return: bool
+    """
+    return any(
+        [
+            start <= ord(character) <= end
+            for start, end in [
+                (4352, 4607),
+                (11904, 42191),
+                (43072, 43135),
+                (44032, 55215),
+                (63744, 64255),
+                (65072, 65103),
+                (65381, 65500),
+                (131072, 196607),
+            ]
+        ]
+    )
+
+
+def symlink_or_copy(exp_dir: Path, src: str, dst: str):
+    """
+    In the experiment directory, create a symlink pointing to src named dst.
+    If symlink creation fails (Windows?), fall back to copyfile."""
+
+    dir_fd = os.open(exp_dir, os.O_RDONLY)
+    try:
+        os.remove(dst, dir_fd=dir_fd)
+    except FileNotFoundError:
+        pass
+    try:
+        os.symlink(src=src, dst=dst, dir_fd=dir_fd)
+    except OSError:
+        copyfile(src=exp_dir / src, dst=exp_dir / dst)
+    os.close(dir_fd)
+
+
+def num_tokens(
+    token_table: k2.SymbolTable, disambig_pattern: str = re.compile(r"^#\d+$")
+) -> int:
+    """Return the number of tokens excluding those from
+    disambiguation symbols.
+
+    Caution:
+      0 is not a token ID so it is excluded from the return value.
+    """
+    symbols = token_table.symbols
+    ans = []
+    for s in symbols:
+        if not disambig_pattern.match(s):
+            ans.append(token_table[s])
+    num_tokens = len(ans)
+    if 0 in ans:
+        num_tokens -= 1
+    return num_tokens
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 000000000..435256416
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,18 @@
+[tool.isort]
+profile = "black"
+skip = ["icefall/__init__.py"]
+
+[tool.black]
+line-length = 88
+exclude = '''
+/(
+    \.git
+  | \.github
+)/
+  | make_kn_lm.py
+  | icefall\/__init__\.py
+  | icefall\/diagnostics\.py
+  | icefall\/profiler\.py
+  | egs\/librispeech\/ASR\/zipformer
+  | egs\/ljspeech\/TTS\/vits
+'''
diff --git a/setup.py b/setup.py
new file mode 100644
index 000000000..ccd2503ff
--- /dev/null
+++ b/setup.py
@@ -0,0 +1,32 @@
+#!/usr/bin/env python3
+
+from pathlib import Path
+
+from setuptools import find_packages, setup
+
+icefall_dir = Path(__file__).parent
+install_requires = (icefall_dir / "requirements.txt").read_text().splitlines()
+
+setup(
+    name="icefall",
+    version="1.0",
+    python_requires=">=3.6.0",
+    description="Speech processing recipes using k2 and Lhotse.",
+    author="The k2 and Lhotse Development Team",
+    license="Apache-2.0 License",
+    packages=find_packages(),
+    install_requires=install_requires,
+    classifiers=[
+        "Development Status :: 3 - Alpha",
+        "Programming Language :: Python :: 3.6",
+        "Programming Language :: Python :: 3.7",
+        "Programming Language :: Python :: 3.8",
+        "Intended Audience :: Science/Research",
+        "Operating System :: POSIX :: Linux",
+        "License :: OSI Approved :: Apache Software License",
+        "Topic :: Multimedia :: Sound/Audio :: Speech",
+        "Topic :: Scientific/Engineering :: Artificial Intelligence",
+        "Topic :: Software Development :: Libraries :: Python Modules",
+        "Typing :: Typed",
+    ],
+)
diff --git a/test/test_ali.py b/test/test_ali.py
new file mode 100755
index 000000000..d607e40aa
--- /dev/null
+++ b/test/test_ali.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Runt his file using one of the following two ways:
+#  (1) python3 ./test/test_ali.py
+#  (2) pytest ./test/test_ali.py
+
+# The purpose of this file is to show that if we build a mask
+# from alignments and add it to a randomly generated nnet_output,
+# we can decode the correct transcript.
+
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest
+from lhotse.dataset import K2SpeechRecognitionDataset, SimpleCutSampler
+from lhotse.dataset.collation import collate_custom_field
+from torch.utils.data import DataLoader
+
+ICEFALL_DIR = Path(__file__).resolve().parent.parent
+egs_dir = ICEFALL_DIR / "egs/librispeech/ASR"
+lang_dir = egs_dir / "data/lang_bpe_500"
+cuts_json = egs_dir / "data/ali/cuts_dev-clean.json.gz"
+
+
+def data_exists():
+    return cuts_json.exists() and lang_dir.exists()
+
+
+def get_dataloader():
+    cuts = load_manifest(cuts_json)
+    print(cuts[0])
+    cuts = cuts.with_features_path_prefix(egs_dir)
+    sampler = SimpleCutSampler(
+        cuts,
+        max_duration=10,
+        shuffle=False,
+    )
+
+    dataset = K2SpeechRecognitionDataset(return_cuts=True)
+
+    dl = DataLoader(
+        dataset,
+        sampler=sampler,
+        batch_size=None,
+        num_workers=1,
+        persistent_workers=False,
+    )
+    return dl
+
+
+def test():
+    if not data_exists():
+        return
+    dl = get_dataloader()
+    for batch in dl:
+        supervisions = batch["supervisions"]
+        cuts = supervisions["cut"]
+        labels_alignment, labels_alignment_length = collate_custom_field(
+            CutSet.from_cuts(cuts), "labels_alignment"
+        )
+
+        (
+            aux_labels_alignment,
+            aux_labels_alignment_length,
+        ) = collate_custom_field(CutSet.from_cuts(cuts), "aux_labels_alignment")
+
+        print(labels_alignment)
+        print(aux_labels_alignment)
+        print(labels_alignment_length)
+        print(aux_labels_alignment_length)
+        break
+
+
+if __name__ == "__main__":
+    test()
diff --git a/test/test_bpe_graph_compiler.py b/test/test_bpe_graph_compiler.py
new file mode 100755
index 000000000..6c9073c4c
--- /dev/null
+++ b/test/test_bpe_graph_compiler.py
@@ -0,0 +1,41 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from pathlib import Path
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.lexicon import UniqLexicon
+
+ICEFALL_DIR = Path(__file__).resolve().parent.parent
+
+
+def test():
+    lang_dir = ICEFALL_DIR / "egs/librispeech/ASR/data/lang_bpe"
+    if not lang_dir.is_dir():
+        return
+
+    compiler = BpeCtcTrainingGraphCompiler(lang_dir)
+    ids = compiler.texts_to_ids(["HELLO", "WORLD ZZZ"])
+    compiler.compile(ids)
+
+    lexicon = UniqLexicon(lang_dir, uniq_filename="lexicon.txt")
+    ids0 = lexicon.words_to_piece_ids(["HELLO"])
+    assert ids[0] == ids0.values().tolist()
+
+    ids1 = lexicon.words_to_piece_ids(["WORLD", "ZZZ"])
+    assert ids[1] == ids1.values().tolist()
diff --git a/test/test_checkpoint.py b/test/test_checkpoint.py
new file mode 100644
index 000000000..34e829642
--- /dev/null
+++ b/test/test_checkpoint.py
@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import pytest
+import torch
+import torch.nn as nn
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint, save_checkpoint
+
+
+@pytest.fixture
+def checkpoints1(tmp_path):
+    f = tmp_path / "f.pt"
+    m = nn.Module()
+    m.p1 = nn.Parameter(torch.tensor([10.0, 20.0]), requires_grad=False)
+    m.register_buffer("p2", torch.tensor([10, 100]))
+
+    params = {"a": 10, "b": 20}
+    save_checkpoint(f, m, params=params)
+    return f
+
+
+@pytest.fixture
+def checkpoints2(tmp_path):
+    f = tmp_path / "f2.pt"
+    m = nn.Module()
+    m.p1 = nn.Parameter(torch.Tensor([50, 30.0]))
+    m.register_buffer("p2", torch.tensor([1, 3]))
+    params = {"a": 100, "b": 200}
+
+    save_checkpoint(f, m, params=params)
+    return f
+
+
+def test_load_checkpoints(checkpoints1):
+    m = nn.Module()
+    m.p1 = nn.Parameter(torch.Tensor([0, 0.0]))
+    m.p2 = nn.Parameter(torch.Tensor([0, 0]))
+    params = load_checkpoint(checkpoints1, m)
+    assert torch.allclose(m.p1, torch.Tensor([10.0, 20]))
+    assert params["a"] == 10
+    assert params["b"] == 20
+
+
+def test_average_checkpoints(checkpoints1, checkpoints2):
+    state_dict = average_checkpoints([checkpoints1, checkpoints2])
+    assert torch.allclose(state_dict["p1"], torch.Tensor([30, 25.0]))
+    assert torch.allclose(state_dict["p2"], torch.tensor([5, 51]))
diff --git a/test/test_decode.py b/test/test_decode.py
new file mode 100644
index 000000000..4c2e192a7
--- /dev/null
+++ b/test/test_decode.py
@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+You can run this file in one of the two ways:
+
+    (1) cd icefall; pytest test/test_decode.py
+    (2) cd icefall; ./test/test_decode.py
+"""
+
+import k2
+
+from icefall.decode import Nbest
+
+
+def test_nbest_from_lattice():
+    s = """
+        0 1 1 10 0.1
+        0 1 5 10 0.11
+        0 1 2 20 0.2
+        1 2 3 30 0.3
+        1 2 4 40 0.4
+        2 3 -1 -1 0.5
+        3
+    """
+    lattice = k2.Fsa.from_str(s, acceptor=False)
+    lattice = k2.Fsa.from_fsas([lattice, lattice])
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=10,
+        use_double_scores=True,
+        nbest_scale=0.5,
+    )
+    # each lattice has only 4 distinct paths that have different word sequences:
+    # 10->30
+    # 10->40
+    # 20->30
+    # 20->40
+    #
+    # So there should be only 4 paths for each lattice in the Nbest object
+    assert nbest.fsa.shape[0] == 4 * 2
+    assert nbest.shape.row_splits(1).tolist() == [0, 4, 8]
+
+    nbest2 = nbest.intersect(lattice)
+    tot_scores = nbest2.tot_scores()
+    argmax = tot_scores.argmax()
+    best_path = k2.index_fsa(nbest2.fsa, argmax)
+    print(best_path[0])
diff --git a/test/test_graph_compiler.py b/test/test_graph_compiler.py
new file mode 100644
index 000000000..10443cf22
--- /dev/null
+++ b/test/test_graph_compiler.py
@@ -0,0 +1,172 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import re
+
+import k2
+import pytest
+import torch
+
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import get_texts
+
+
+@pytest.fixture
+def lexicon():
+    """
+    We use the following test data:
+
+    lexicon.txt
+
+        foo f o o
+        bar b a r
+        baz b a z
+        <UNK> SPN
+
+    phones.txt
+
+        <eps> 0
+        a 1
+        b 2
+        f 3
+        o 4
+        r 5
+        z 6
+        SPN 7
+
+    words.txt:
+
+        <eps> 0
+        foo 1
+        bar 2
+        baz 3
+        <UNK> 4
+    """
+    L = k2.Fsa.from_str(
+        """
+        0 0 7 4 0
+        0 7 -1 -1 0
+        0 1 3 1 0
+        0 3 2 2 0
+        0 5 2 3 0
+        1 2 4 0 0
+        2 0 4 0 0
+        3 4 1 0 0
+        4 0 5 0 0
+        5 6 1 0 0
+        6 0 6 0 0
+        7
+    """,
+        num_aux_labels=1,
+    )
+    L.labels_sym = k2.SymbolTable.from_str(
+        """
+        a 1
+        b 2
+        f 3
+        o 4
+        r 5
+        z 6
+        SPN 7
+    """
+    )
+    L.aux_labels_sym = k2.SymbolTable.from_str(
+        """
+        foo 1
+        bar 2
+        baz 3
+        <UNK> 4
+    """
+    )
+    ans = Lexicon.__new__(Lexicon)
+    ans.token_table = L.labels_sym
+    ans.word_table = L.aux_labels_sym
+    ans.L_inv = k2.arc_sort(L.invert_())
+    ans.disambig_pattern = re.compile(r"^#\d+$")
+
+    return ans
+
+
+@pytest.fixture
+def compiler(lexicon):
+    return CtcTrainingGraphCompiler(lexicon, device=torch.device("cpu"))
+
+
+class TestCtcTrainingGraphCompiler(object):
+    @staticmethod
+    def test_convert_transcript_to_fsa(compiler, lexicon):
+        texts = ["bar foo", "baz ok"]
+        fsa = compiler.convert_transcript_to_fsa(texts)
+        labels0 = fsa[0].labels[:-1].tolist()
+        aux_labels0 = fsa[0].aux_labels[:-1]
+        aux_labels0 = aux_labels0[aux_labels0 != 0].tolist()
+
+        labels1 = fsa[1].labels[:-1].tolist()
+        aux_labels1 = fsa[1].aux_labels[:-1]
+        aux_labels1 = aux_labels1[aux_labels1 != 0].tolist()
+
+        labels0 = [lexicon.token_table[i] for i in labels0]
+        labels1 = [lexicon.token_table[i] for i in labels1]
+
+        aux_labels0 = [lexicon.word_table[i] for i in aux_labels0]
+        aux_labels1 = [lexicon.word_table[i] for i in aux_labels1]
+
+        assert labels0 == ["b", "a", "r", "f", "o", "o"]
+        assert aux_labels0 == ["bar", "foo"]
+
+        assert labels1 == ["b", "a", "z", "SPN"]
+        assert aux_labels1 == ["baz", "<UNK>"]
+
+    @staticmethod
+    def test_compile(compiler, lexicon):
+        texts = ["bar foo", "baz ok"]
+        decoding_graph = compiler.compile(texts)
+        input1 = ["b", "b", "<blk>", "<blk>", "a", "a", "r", "<blk>", "<blk>"]
+        input1 += ["f", "f", "<blk>", "<blk>", "o", "o", "<blk>", "o", "o"]
+
+        input2 = ["b", "b", "a", "a", "a", "<blk>", "<blk>", "z", "z"]
+        input2 += ["<blk>", "<blk>", "SPN", "SPN", "<blk>", "<blk>"]
+
+        lexicon.token_table._id2sym[0] == "<blk>"
+        lexicon.token_table._sym2id["<blk>"] = 0
+
+        input1 = [lexicon.token_table[i] for i in input1]
+        input2 = [lexicon.token_table[i] for i in input2]
+
+        fsa1 = k2.linear_fsa(input1)
+        fsa2 = k2.linear_fsa(input2)
+        fsas = k2.Fsa.from_fsas([fsa1, fsa2])
+
+        decoding_graph = k2.arc_sort(decoding_graph)
+        lattice = k2.intersect(decoding_graph, fsas, treat_epsilons_specially=False)
+        lattice = k2.connect(lattice)
+
+        aux_labels0 = lattice[0].aux_labels[:-1]
+        aux_labels0 = aux_labels0[aux_labels0 != 0].tolist()
+        aux_labels0 = [lexicon.word_table[i] for i in aux_labels0]
+        assert aux_labels0 == ["bar", "foo"]
+
+        aux_labels1 = lattice[1].aux_labels[:-1]
+        aux_labels1 = aux_labels1[aux_labels1 != 0].tolist()
+        aux_labels1 = [lexicon.word_table[i] for i in aux_labels1]
+        assert aux_labels1 == ["baz", "<UNK>"]
+
+        texts = get_texts(lattice)
+        texts = [[lexicon.word_table[i] for i in words] for words in texts]
+        assert texts == [["bar", "foo"], ["baz", "<UNK>"]]
diff --git a/test/test_lexicon.py b/test/test_lexicon.py
new file mode 100755
index 000000000..b1beab3f6
--- /dev/null
+++ b/test/test_lexicon.py
@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+You can run this file in one of the two ways:
+
+    (1) cd icefall; pytest test/test_lexicon.py
+    (2) cd icefall; ./test/test_lexicon.py
+"""
+
+
+import os
+import shutil
+import sys
+from pathlib import Path
+from typing import List
+
+import sentencepiece as spm
+
+from icefall.lexicon import UniqLexicon
+
+TMP_DIR = "/tmp/icefall-test-lexicon"
+USING_PYTEST = "pytest" in sys.modules
+ICEFALL_DIR = Path(__file__).resolve().parent.parent
+
+
+def generate_test_data():
+    Path(TMP_DIR).mkdir(exist_ok=True)
+    sentences = """
+cat tac cat cat
+at
+tac at ta at at
+at cat ct ct ta
+cat cat cat cat
+at at at at at at at
+    """
+
+    transcript = Path(TMP_DIR) / "transcript_words.txt"
+    with open(transcript, "w") as f:
+        for line in sentences.strip().split("\n"):
+            f.write(f"{line}\n")
+
+    words = """
+<eps> 0
+<UNK> 1
+at 2
+cat 3
+ct 4
+ta 5
+tac 6
+#0 7
+<s> 8
+</s> 9
+"""
+    word_txt = Path(TMP_DIR) / "words.txt"
+    with open(word_txt, "w") as f:
+        for line in words.strip().split("\n"):
+            f.write(f"{line}\n")
+
+    vocab_size = 8
+
+    os.system(
+        f"""
+cd {ICEFALL_DIR}/egs/librispeech/ASR
+
+./local/train_bpe_model.py \
+  --lang-dir {TMP_DIR} \
+  --vocab-size {vocab_size} \
+  --transcript {transcript}
+
+./local/prepare_lang_bpe.py --lang-dir {TMP_DIR} --debug 1
+"""
+    )
+
+
+def delete_test_data():
+    shutil.rmtree(TMP_DIR)
+
+
+def uniq_lexicon_test():
+    lexicon = UniqLexicon(lang_dir=TMP_DIR, uniq_filename="lexicon.txt")
+
+    # case 1: No OOV
+    texts = ["cat cat", "at ct", "at tac cat"]
+    token_ids = lexicon.texts_to_token_ids(texts)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(f"{TMP_DIR}/bpe.model")
+
+    expected_token_ids: List[List[int]] = sp.encode(texts, out_type=int)
+    assert token_ids.tolist() == expected_token_ids
+
+    # case 2: With OOV
+    texts = ["ca"]
+    token_ids = lexicon.texts_to_token_ids(texts)
+    expected_token_ids = sp.encode(texts, out_type=int)
+    assert token_ids.tolist() != expected_token_ids
+    # Note: sentencepiece breaks "ca" into "_ c a"
+    # But there is no word "ca" in the lexicon, so our
+    # implementation returns the id of "<UNK>"
+    print(token_ids, expected_token_ids)
+    assert token_ids.tolist() == [[sp.piece_to_id("▁"), sp.unk_id()]]
+
+    # case 3: With OOV
+    texts = ["foo"]
+    token_ids = lexicon.texts_to_token_ids(texts)
+    expected_token_ids = sp.encode(texts, out_type=int)
+    print(token_ids)
+    print(expected_token_ids)
+
+    # test ragged lexicon
+    ragged_lexicon = lexicon.ragged_lexicon.tolist()
+    word_disambig_id = lexicon.word_table["#0"]
+    for i in range(2, word_disambig_id):
+        piece_id = ragged_lexicon[i]
+        word = lexicon.word_table[i]
+        assert word == sp.decode(piece_id)
+        assert piece_id == sp.encode(word)
+
+
+def test_main():
+    generate_test_data()
+
+    uniq_lexicon_test()
+
+    if USING_PYTEST:
+        delete_test_data()
+
+
+def main():
+    test_main()
+
+
+if __name__ == "__main__" and not USING_PYTEST:
+    main()
diff --git a/test/test_mmi_graph_compiler.py b/test/test_mmi_graph_compiler.py
new file mode 100755
index 000000000..653c57b59
--- /dev/null
+++ b/test/test_mmi_graph_compiler.py
@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+You can run this file in one of the two ways:
+
+    (1) cd icefall; pytest test/test_mmi_graph_compiler.py
+    (2) cd icefall; ./test/test_mmi_graph_compiler.py
+"""
+
+import copy
+import os
+import shutil
+import sys
+from pathlib import Path
+
+import k2
+import sentencepiece as spm
+
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+
+TMP_DIR = "/tmp/icefall-test-mmi-graph-compiler"
+USING_PYTEST = "pytest" in sys.modules
+ICEFALL_DIR = Path(__file__).resolve().parent.parent
+
+
+def generate_test_data():
+    Path(TMP_DIR).mkdir(exist_ok=True)
+    sentences = """
+cat tac cat cat
+at at cat at cat cat
+tac at ta at at
+at cat ct ct ta ct ct cat tac
+cat cat cat cat
+at at at at at at at
+    """
+
+    transcript = Path(TMP_DIR) / "transcript_words.txt"
+    with open(transcript, "w") as f:
+        for line in sentences.strip().split("\n"):
+            f.write(f"{line}\n")
+
+    words = """
+<eps> 0
+<UNK> 1
+at 2
+cat 3
+ct 4
+ta 5
+tac 6
+#0 7
+<s> 8
+</s> 9
+"""
+    word_txt = Path(TMP_DIR) / "words.txt"
+    with open(word_txt, "w") as f:
+        for line in words.strip().split("\n"):
+            f.write(f"{line}\n")
+
+    vocab_size = 8
+
+    os.system(
+        f"""
+cd {ICEFALL_DIR}/egs/librispeech/ASR
+
+./local/train_bpe_model.py \
+  --lang-dir {TMP_DIR} \
+  --vocab-size {vocab_size} \
+  --transcript {transcript}
+
+./local/prepare_lang_bpe.py --lang-dir {TMP_DIR} --debug 0
+
+./local/convert_transcript_words_to_tokens.py \
+--lexicon {TMP_DIR}/lexicon.txt \
+--transcript {transcript} \
+--oov "<UNK>" \
+> {TMP_DIR}/transcript_tokens.txt
+
+./shared/make_kn_lm.py \
+-ngram-order 2 \
+-text {TMP_DIR}/transcript_tokens.txt \
+-lm {TMP_DIR}/P.arpa
+
+python3 -m kaldilm \
+--read-symbol-table="{TMP_DIR}/tokens.txt" \
+--disambig-symbol='#0' \
+--max-order=2 \
+{TMP_DIR}/P.arpa > {TMP_DIR}/P.fst.txt
+"""
+    )
+
+
+def delete_test_data():
+    shutil.rmtree(TMP_DIR)
+
+
+def mmi_graph_compiler_test():
+    # Caution:
+    # You have to uncomment
+    #  del transcript_fsa.aux_labels
+    # in mmi_graph_compiler.py
+    # to see the correct aux_labels in *.svg
+    graph_compiler = MmiTrainingGraphCompiler(
+        lang_dir=TMP_DIR, uniq_filename="lexicon.txt"
+    )
+    print(graph_compiler.device)
+    L_inv = graph_compiler.L_inv
+    L = k2.invert(L_inv)
+
+    L.labels_sym = graph_compiler.lexicon.token_table
+    L.aux_labels_sym = graph_compiler.lexicon.word_table
+    L.draw(f"{TMP_DIR}/L.svg", title="L")
+
+    L_inv.labels_sym = graph_compiler.lexicon.word_table
+    L_inv.aux_labels_sym = graph_compiler.lexicon.token_table
+    L_inv.draw(f"{TMP_DIR}/L_inv.svg", title="L")
+
+    ctc_topo_P = graph_compiler.ctc_topo_P
+    ctc_topo_P.labels_sym = copy.deepcopy(graph_compiler.lexicon.token_table)
+    ctc_topo_P.labels_sym._id2sym[0] = "<blk>"
+    ctc_topo_P.labels_sym._sym2id["<blk>"] = 0
+    ctc_topo_P.aux_labels_sym = graph_compiler.lexicon.token_table
+    ctc_topo_P.draw(f"{TMP_DIR}/ctc_topo_P.svg", title="ctc_topo_P")
+
+    print(ctc_topo_P.num_arcs)
+    print(k2.connect(ctc_topo_P).num_arcs)
+
+    with open(str(TMP_DIR) + "/P.fst.txt") as f:
+        # P is not an acceptor because there is
+        # a back-off state, whose incoming arcs
+        # have label #0 and aux_label 0 (i.e., <eps>).
+        P = k2.Fsa.from_openfst(f.read(), acceptor=False)
+    P.labels_sym = graph_compiler.lexicon.token_table
+    P.aux_labels_sym = graph_compiler.lexicon.token_table
+    P.draw(f"{TMP_DIR}/P.svg", title="P")
+
+    ctc_topo = k2.ctc_topo(max(graph_compiler.lexicon.tokens), False)
+    ctc_topo.labels_sym = ctc_topo_P.labels_sym
+    ctc_topo.aux_labels_sym = graph_compiler.lexicon.token_table
+    ctc_topo.draw(f"{TMP_DIR}/ctc_topo.svg", title="ctc_topo")
+    print("p num arcs", P.num_arcs)
+    print("ctc_topo num arcs", ctc_topo.num_arcs)
+    print("ctc_topo_P num arcs", ctc_topo_P.num_arcs)
+
+    texts = ["cat at ct", "at ta", "cat tac"]
+    transcript_fsa = graph_compiler.build_transcript_fsa(texts)
+    transcript_fsa[0].draw(f"{TMP_DIR}/cat_at_ct.svg", title="cat_at_ct")
+    transcript_fsa[1].draw(f"{TMP_DIR}/at_ta.svg", title="at_ta")
+    transcript_fsa[2].draw(f"{TMP_DIR}/cat_tac.svg", title="cat_tac")
+
+    num_graphs, den_graphs = graph_compiler.compile(texts, replicate_den=True)
+    num_graphs[0].draw(f"{TMP_DIR}/num_cat_at_ct.svg", title="num_cat_at_ct")
+    num_graphs[1].draw(f"{TMP_DIR}/num_at_ta.svg", title="num_at_ta")
+    num_graphs[2].draw(f"{TMP_DIR}/num_cat_tac.svg", title="num_cat_tac")
+
+    den_graphs[0].draw(f"{TMP_DIR}/den_cat_at_ct.svg", title="den_cat_at_ct")
+    den_graphs[2].draw(f"{TMP_DIR}/den_cat_tac.svg", title="den_cat_tac")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(f"{TMP_DIR}/bpe.model")
+
+    texts = ["cat at cat", "at tac"]
+    token_ids = graph_compiler.texts_to_ids(texts)
+    expected_token_ids = sp.encode(texts)
+    assert token_ids == expected_token_ids
+
+
+def test_main():
+    generate_test_data()
+
+    mmi_graph_compiler_test()
+
+    if USING_PYTEST:
+        delete_test_data()
+
+
+def main():
+    test_main()
+
+
+if __name__ == "__main__" and not USING_PYTEST:
+    main()
diff --git a/test/test_ngram_lm.py b/test/test_ngram_lm.py
new file mode 100755
index 000000000..838c792d2
--- /dev/null
+++ b/test/test_ngram_lm.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import graphviz
+
+from icefall import is_module_available
+
+if not is_module_available("kaldifst"):
+    raise ValueError("Please 'pip install kaldifst' first.")
+
+import kaldifst
+
+from icefall import NgramLm, NgramLmStateCost
+
+
+def generate_fst(filename: str):
+    s = """
+3	5	1	1	3.00464
+3	0	3	0	5.75646
+0	1	1	1	12.0533
+0	2	2	2	7.95954
+0	9.97787
+1	4	2	2	3.35436
+1	0	3	0	7.59853
+2	0	3	0
+4	2	3	0	7.43735
+4	0.551239
+5	4	2	2	0.804938
+5	1	3	0	9.67086
+"""
+    fst = kaldifst.compile(s=s, acceptor=False)
+    fst.write(filename)
+    fst_dot = kaldifst.draw(fst, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile=f"{filename}.svg")
+
+
+def main():
+    filename = "test.fst"
+    generate_fst(filename)
+    ngram_lm = NgramLm(filename, backoff_id=3, is_binary=True)
+    for label in [1, 2, 3, 4, 5]:
+        print("---label---", label)
+        p = ngram_lm.get_next_state_and_cost(state=5, label=label)
+        print(p)
+        print("---")
+
+    state_cost = NgramLmStateCost(ngram_lm)
+    s0 = state_cost.forward_one_step(1)
+    print(s0.state_cost)
+
+    s1 = s0.forward_one_step(2)
+    print(s1.state_cost)
+
+    s2 = s1.forward_one_step(2)
+    print(s2.state_cost)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/test/test_parse_timestamp.py b/test/test_parse_timestamp.py
new file mode 100755
index 000000000..92bfb49c6
--- /dev/null
+++ b/test/test_parse_timestamp.py
@@ -0,0 +1,154 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from pathlib import Path
+
+import k2
+import sentencepiece as spm
+import torch
+
+from icefall.lexicon import Lexicon
+from icefall.utils import parse_bpe_timestamps_and_texts, parse_timestamps_and_texts
+
+ICEFALL_DIR = Path(__file__).resolve().parent.parent
+
+
+def test_parse_bpe_timestamps_and_texts():
+    lang_dir = ICEFALL_DIR / "egs/librispeech/ASR/data/lang_bpe_500"
+    if not lang_dir.is_dir():
+        print(f"{lang_dir} does not exist.")
+        return
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(lang_dir / "bpe.model"))
+
+    text_1 = "HELLO WORLD"
+    token_ids_1 = sp.encode(text_1, out_type=int)
+    # out_type=str: ['_HE', 'LL', 'O', '_WORLD']
+    # out_type=int: [22, 58, 24, 425]
+
+    # [22, 22, 58, 24, 0, 0, 425, 425, 425, 0, 0]
+    labels_1 = (
+        token_ids_1[0:1] * 2
+        + token_ids_1[1:3]
+        + [0] * 2
+        + token_ids_1[3:4] * 3
+        + [0] * 2
+    )
+    # [22, 0, 58, 24, 0, 0, 425, 0, 0, 0, 0, -1]
+    aux_labels_1 = (
+        token_ids_1[0:1]
+        + [0]
+        + token_ids_1[1:3]
+        + [0] * 2
+        + token_ids_1[3:4]
+        + [0] * 4
+        + [-1]
+    )
+    fsa_1 = k2.linear_fsa(labels_1)
+    fsa_1.aux_labels = torch.tensor(aux_labels_1).to(torch.int32)
+
+    text_2 = "SAY GOODBYE"
+    token_ids_2 = sp.encode(text_2, out_type=int)
+    # out_type=str: ['_SAY', '_GOOD', 'B', 'Y', 'E']
+    # out_type=int: [289, 286, 41, 16, 11]
+
+    # [289, 0, 0, 286, 286, 41, 16, 11, 0, 0]
+    labels_2 = (
+        token_ids_2[0:1] + [0] * 2 + token_ids_2[1:2] * 2 + token_ids_2[2:5] + [0] * 2
+    )
+    # [289, 0, 0, 286, 0, 41, 16, 11, 0, 0, -1]
+    aux_labels_2 = (
+        token_ids_2[0:1]
+        + [0] * 2
+        + token_ids_2[1:2]
+        + [0]
+        + token_ids_2[2:5]
+        + [0] * 2
+        + [-1]
+    )
+    fsa_2 = k2.linear_fsa(labels_2)
+    fsa_2.aux_labels = torch.tensor(aux_labels_2).to(torch.int32)
+
+    fsa_vec = k2.create_fsa_vec([fsa_1, fsa_2])
+
+    utt_index_pairs, utt_words = parse_bpe_timestamps_and_texts(fsa_vec, sp)
+    assert utt_index_pairs[0] == [(0, 3), (6, 8)], utt_index_pairs[0]
+    assert utt_words[0] == ["HELLO", "WORLD"], utt_words[0]
+    assert utt_index_pairs[1] == [(0, 0), (3, 7)], utt_index_pairs[1]
+    assert utt_words[1] == ["SAY", "GOODBYE"], utt_words[1]
+
+
+def test_parse_timestamps_and_texts():
+    lang_dir = ICEFALL_DIR / "egs/librispeech/ASR/data/lang_bpe_500"
+    if not lang_dir.is_dir():
+        print(f"{lang_dir} does not exist.")
+        return
+
+    lexicon = Lexicon(lang_dir)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(lang_dir / "bpe.model"))
+    word_table = lexicon.word_table
+
+    text_1 = "HELLO WORLD"
+    token_ids_1 = sp.encode(text_1, out_type=int)
+    # out_type=str: ['_HE', 'LL', 'O', '_WORLD']
+    # out_type=int: [22, 58, 24, 425]
+    word_ids_1 = [word_table[s] for s in text_1.split()]  # [79677, 196937]
+    # [22, 22, 58, 24, 0, 0, 425, 425, 425, 0, 0]
+    labels_1 = (
+        token_ids_1[0:1] * 2
+        + token_ids_1[1:3]
+        + [0] * 2
+        + token_ids_1[3:4] * 3
+        + [0] * 2
+    )
+    # [[79677], [], [], [], [], [], [196937], [], [], [], [], []]
+    aux_labels_1 = [word_ids_1[0:1]] + [[]] * 5 + [word_ids_1[1:2]] + [[]] * 5
+
+    fsa_1 = k2.linear_fsa(labels_1)
+    fsa_1.aux_labels = k2.RaggedTensor(aux_labels_1)
+
+    text_2 = "SAY GOODBYE"
+    token_ids_2 = sp.encode(text_2, out_type=int)
+    # out_type=str: ['_SAY', '_GOOD', 'B', 'Y', 'E']
+    # out_type=int: [289, 286, 41, 16, 11]
+    word_ids_2 = [word_table[s] for s in text_2.split()]  # [154967, 72079]
+    # [289, 0, 0, 286, 286, 41, 16, 11, 0, 0]
+    labels_2 = (
+        token_ids_2[0:1] + [0] * 2 + token_ids_2[1:2] * 2 + token_ids_2[2:5] + [0] * 2
+    )
+    # [[154967], [], [], [72079], [], [], [], [], [], [], []]
+    aux_labels_2 = [word_ids_2[0:1]] + [[]] * 2 + [word_ids_2[1:2]] + [[]] * 7
+
+    fsa_2 = k2.linear_fsa(labels_2)
+    fsa_2.aux_labels = k2.RaggedTensor(aux_labels_2)
+
+    fsa_vec = k2.create_fsa_vec([fsa_1, fsa_2])
+
+    utt_index_pairs, utt_words = parse_timestamps_and_texts(fsa_vec, word_table)
+    assert utt_index_pairs[0] == [(0, 3), (6, 8)], utt_index_pairs[0]
+    assert utt_words[0] == ["HELLO", "WORLD"], utt_words[0]
+    assert utt_index_pairs[1] == [(0, 0), (3, 7)], utt_index_pairs[1]
+    assert utt_words[1] == ["SAY", "GOODBYE"], utt_words[1]
+
+
+if __name__ == "__main__":
+    test_parse_bpe_timestamps_and_texts()
+    test_parse_timestamps_and_texts()
diff --git a/test/test_utils.py b/test/test_utils.py
new file mode 100644
index 000000000..31f06bd51
--- /dev/null
+++ b/test/test_utils.py
@@ -0,0 +1,165 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import k2
+import pytest
+import torch
+
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    add_eos,
+    add_sos,
+    encode_supervisions,
+    get_texts,
+    make_pad_mask,
+)
+
+
+@pytest.fixture
+def sup():
+    sequence_idx = torch.tensor([0, 1, 2])
+    start_frame = torch.tensor([1, 3, 9])
+    num_frames = torch.tensor([20, 30, 10])
+    text = ["one", "two", "three"]
+    return {
+        "sequence_idx": sequence_idx,
+        "start_frame": start_frame,
+        "num_frames": num_frames,
+        "text": text,
+    }
+
+
+def test_encode_supervisions(sup):
+    supervision_segments, texts = encode_supervisions(sup, subsampling_factor=4)
+    assert torch.all(
+        torch.eq(
+            supervision_segments,
+            torch.tensor([[1, 0, 30 // 4], [0, 0, 20 // 4], [2, 9 // 4, 10 // 4]]),
+        )
+    )
+    assert texts == ["two", "one", "three"]
+
+
+def test_get_texts_ragged():
+    fsa1 = k2.Fsa.from_str(
+        """
+        0 1 1 10
+        1 2 2 20
+        2 3 3 30
+        3 4 -1 0
+        4
+    """
+    )
+    fsa1.aux_labels = k2.RaggedTensor("[ [1 3 0 2] [] [4 0 1] [-1]]")
+
+    fsa2 = k2.Fsa.from_str(
+        """
+        0 1 1 1
+        1 2 2 2
+        2 3 -1 0
+        3
+    """
+    )
+    fsa2.aux_labels = k2.RaggedTensor("[[3 0 5 0 8] [0 9 7 0] [-1]]")
+    fsas = k2.Fsa.from_fsas([fsa1, fsa2])
+    texts = get_texts(fsas)
+    assert texts == [[1, 3, 2, 4, 1], [3, 5, 8, 9, 7]]
+
+
+def test_get_texts_regular():
+    fsa1 = k2.Fsa.from_str(
+        """
+        0 1 1 3 10
+        1 2 2 0 20
+        2 3 3 2 30
+        3 4 -1 -1 0
+        4
+    """,
+        num_aux_labels=1,
+    )
+
+    fsa2 = k2.Fsa.from_str(
+        """
+        0 1 1 10 1
+        1 2 2 5 2
+        2 3 -1 -1 0
+        3
+    """,
+        num_aux_labels=1,
+    )
+    fsas = k2.Fsa.from_fsas([fsa1, fsa2])
+    texts = get_texts(fsas)
+    assert texts == [[3, 2], [10, 5]]
+
+
+def test_attribute_dict():
+    s = AttributeDict({"a": 10, "b": 20})
+    assert s.a == 10
+    assert s["b"] == 20
+    s.c = 100
+    assert s["c"] == 100
+
+    assert hasattr(s, "a")
+    assert hasattr(s, "b")
+    assert getattr(s, "a") == 10
+    del s.a
+    assert hasattr(s, "a") is False
+    setattr(s, "c", 100)
+    s.c = 100
+    try:
+        del s.a
+    except AttributeError as ex:
+        print(f"Caught exception: {ex}")
+
+
+def test_get_env_info():
+    s = get_env_info()
+    print(s)
+
+
+def test_makd_pad_mask():
+    lengths = torch.tensor([1, 3, 2])
+    mask = make_pad_mask(lengths)
+    expected = torch.tensor(
+        [
+            [False, True, True],
+            [False, False, False],
+            [False, False, True],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected))
+    assert (~expected).sum() == lengths.sum()
+
+
+def test_add_sos():
+    sos_id = 100
+    ragged = k2.RaggedTensor([[1, 2], [3], [0]])
+    sos_ragged = add_sos(ragged, sos_id)
+    expected = k2.RaggedTensor([[sos_id, 1, 2], [sos_id, 3], [sos_id, 0]])
+    assert str(sos_ragged) == str(expected)
+
+
+def test_add_eos():
+    eos_id = 30
+    ragged = k2.RaggedTensor([[1, 2], [3], [], [5, 8, 9]])
+    ragged_eos = add_eos(ragged, eos_id)
+    expected = k2.RaggedTensor(
+        [[1, 2, eos_id], [3, eos_id], [eos_id], [5, 8, 9, eos_id]]
+    )
+    assert str(ragged_eos) == str(expected)